<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[Stories by Alex Pavlov on Medium]]></title>
        <description><![CDATA[Stories by Alex Pavlov on Medium]]></description>
        <link>https://medium.com/@alex-on-ai?source=rss-93cf9798cdda------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*30oE8V6ejvfpwvo9kqm63A.jpeg</url>
            <title>Stories by Alex Pavlov on Medium</title>
            <link>https://medium.com/@alex-on-ai?source=rss-93cf9798cdda------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Wed, 18 Nov 2026 20:38:15 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@alex-on-ai/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[What Is Sales Automation When the Lead Is Also a Patient]]></title>
            <link>https://alex-on-ai.medium.com/what-is-sales-automation-when-the-lead-is-also-a-patient-7c077013e6e7?source=rss-93cf9798cdda------2</link>
            <guid isPermaLink="false">https://medium.com/p/7c077013e6e7</guid>
            <category><![CDATA[telemedicine]]></category>
            <category><![CDATA[hipaa]]></category>
            <category><![CDATA[software-development]]></category>
            <category><![CDATA[automation]]></category>
            <category><![CDATA[healthcare]]></category>
            <dc:creator><![CDATA[Alex Pavlov]]></dc:creator>
            <pubDate>Tue, 04 Aug 2026 11:23:24 GMT</pubDate>
            <atom:updated>2026-08-04T11:23:24.113Z</atom:updated>
            <content:encoded><![CDATA[<p>You are deciding whether the sales side of your healthcare product gets bought or built, and the vendor demos all look the same. So it is worth being precise about what is sales automation in a setting where the lead is also a patient. It is the software that carries a person from first contact to booked, verified, paid, and documented, with nobody retyping a phone number into a second system.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*h-GIDxfoWY1HJROB.jpg" /></figure><p>Outside healthcare, that definition is uncontroversial. The standard <a href="https://highcraft.io/blog/what-is-sales-automation/?utm_source=medium&amp;utm_medium=article&amp;utm_campaign=blog-what-is-sales-automation">breakdown of sales automation and where it actually saves time</a> holds up: capture, route, sequence, score, hand off. Inside healthcare, every one of those verbs touches data with legal weight. Consent, retention, and audit logging arrive with it, uninvited, on day one.</p><p>So the decision sits at a boundary. Which fields live in the marketing system, and which live in the record of care. Almost everything else about the build follows from where you draw that line.</p><h3>Sales automation is a workflow engine with a CRM bolted on</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*8edVePyStmGlSYWA.jpg" /></figure><p>Strip the category down and you get a state machine. A trigger fires, an entity moves states, a queue picks up work, something retries when the third-party API returns a 500. HubSpot workflows are that, with a friendly canvas on top. Zapier is that, minus memory and minus a real retry story. n8n is that with queues, self-hosting, and an ops surface you now own.</p><p>The sales-specific parts are thin by comparison. Sequences are timers. Lead scoring is a weighted sum somebody will tune twice and then never look at again. Round-robin routing is a modulo operator with a nice name.</p><p>That matters for your decision, because you are not buying “sales features.” You are buying somebody else’s opinion about state, retries, and where data rests. In a regulated product, those three opinions are the whole cost.</p><h3>PHI shows up earlier than anybody plans for</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*B8hhgvcVEc1nag8p.jpg" /></figure><p>Every healthcare intake form I have seen has a free-text field. Someone labels it “reason for visit” or “anything we should know,” and within a week it contains a diagnosis, a medication list, and an unprompted paragraph about a spouse. That field is now protected health information sitting in your lead pipeline.</p><p>The same thing happens through side doors. Chat widget transcripts. Call recordings attached to a contact. A screenshot pasted into a deal note by a coordinator trying to be helpful. A drip campaign carrying a diagnosis is not a drip campaign. It is a leak.</p><p>This is where a lot of marketing stacks quietly fail the buy decision. Some vendors sign a BAA, some sign one only on specific tiers, and some will sign one that excludes exactly the module your team wanted to use. Then there is the tracking layer. Analytics and ad pixels on a page that renders appointment details will happily ship identifiers to third parties who never signed anything. (The BAA covers the vendor. It does not cover your ops lead’s Slack habit.)</p><p>I ask two questions before anything else in a healthcare stack review. Where does this data rest, and who can read it without a ticket. If a tool cannot answer both in one sentence, it lives outside the PHI boundary or it does not live in the stack.</p><h3>The boundary of the record decides buy versus build</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*woNNaQgdH08GZbno.jpg" /></figure><p>Here is the split that has worked on the telehealth and clinic-facing platforms we have shipped.</p><p>Top of funnel gets rented. Anonymous traffic, content, ad attribution, generic newsletters. No identifiers tied to conditions, no exceptions carved out because a campaign was in a hurry. Off-the-shelf tooling is genuinely good at this and building it yourself is a waste of a quarter.</p><p>The moment a person identifies themselves and states a reason for contact, they cross into your system. Intake, eligibility, scheduling, reminders, consent capture, payment. That path gets built on infrastructure you control, with your own audit trail and your own retention rules.</p><p>Between those two sits the piece most teams underestimate: an orchestration service that owns the state machine and speaks to both sides. It holds the pipeline states, it writes to the record system, and it sends the marketing platform only what it needs. Usually that is an opaque contact ID and a stage name. The CRM knows a person reached stage four. It does not know why.</p><p>That middle layer is the part I argue for hardest, because it is what makes the rest replaceable. Change email vendors and you change one adapter. Without it, your pipeline logic is scattered across forty automations in a UI, with no version history worth the name and no way to test a change before it emails two thousand patients at 6am.</p><h3>No-code tools stop being cheaper at a predictable point</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*PzyfHIwHsMkErSV1.jpg" /></figure><p>Zapier is excellent for the first six weeks and honest about what it is. The problems come in order: task history storing payloads you now have to account for, no clean idempotency when a webhook fires twice, and error handling that means “an email to whoever set it up.” For a low-stakes lead handoff, fine. For an appointment confirmation that must go out exactly once, that is not a foundation.</p><p>n8n changes the calculus, because you can self-host it inside your own network with logs you control. We have run it that way behind Azure private networking as the connective tissue for internal ops workflows. Worth being honest about the cost: self-hosted n8n is a service, and services need upgrades, backups, secret rotation, and someone who notices when the worker queue stalls. Budget a slice of an engineer’s ongoing time, not a one-off setup week.</p><p>The rule of thumb I use: no-code for internal, low-consequence, human-in-the-loop steps. Real code for anything that writes to the record, moves money, or sends a message a patient will act on medically.</p><h3>Model the process before you shop for a tool</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*MaskDcZ8rwPGKsHa.jpg" /></figure><p>Before any vendor call, draw the process with swimlanes and named states. <a href="https://www.bpmn.org/">BPMN</a> is worth the two hours it takes to learn the eight symbols you will actually use. Not because anyone will admire the diagram, but because it forces the questions that vendor demos skip.</p><p>What happens when insurance verification comes back unresolved. Who owns that state, for how long, and what fires when the timer expires. What happens when a patient books, cancels, and rebooks in four minutes. Which of your automations is now holding a stale appointment ID.</p><p>The failure I keep meeting in production is not a missing feature. It is a state with no owner and no timeout, so records pile up in it silently until someone runs a report in month five and finds two hundred people who were never called back. A dead-end state is the automation equivalent of a voicemail nobody set up.</p><p>Once the diagram exists, the buy-versus-build conversation gets short. You can point at the boxes a platform covers and the boxes it does not.</p><h3>Integration work is where the estimate breaks</h3><p>If your automation has to talk to a hospital or a large group practice, that integration is the project. Everything else is scaffolding around it.</p><p>FHIR R4 is the pleasant case. Well-specified resources, decent sandboxes, and an actual conversation about scopes. HL7v2 is the common case, which means you are parsing pipe-delimited segments and discovering that this particular site uses a Z-segment that nobody documented. Both are workable. Neither is a two-week ticket.</p><p>The asymmetry that reshapes plans: reading is engineering, writing is politics. Pulling an ADT feed to know a patient showed up is tractable. Writing an appointment back into Epic means a vendor program, a sandbox, security review, and a go-live window set by someone else’s calendar. On one platform, the read side went live in three weeks. The write side took two quarters.</p><p>So I sequence around it. Ship the automation with a human confirming the booking in the source system, measure how much time that costs, then invest in write-back once the volume justifies the fight. That order also gives you real data for the eligibility edge cases, which no amount of diagramming will surface. A plan that assumes write access on day one usually finds out in week nine, which is a poor week to find out.</p><h3>What I rent, and what I insist on owning</h3><p>Rented: email and SMS delivery, calendar availability, e-signature, payments through Stripe, ad attribution above the identification line. These are commodities with better uptime than anything I would build for the same money.</p><p>Owned: the state machine, the field mapping, the consent record, and the audit log. Especially the audit log. Every automated write in ours carries the rule version that produced it, alongside the actor and the timestamp. When someone asks in month nine why a patient got a second reminder at 6:04am, that column answers it in one query instead of three days of archaeology across two vendors’ export files.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=7c077013e6e7" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Choosing Sales Automation Software When Your Pipeline Doesn’t Fit the CRM’s Data Model]]></title>
            <link>https://alex-on-ai.medium.com/choosing-sales-automation-software-when-your-pipeline-doesnt-fit-the-crm-s-data-model-286ad6878fba?source=rss-93cf9798cdda------2</link>
            <guid isPermaLink="false">https://medium.com/p/286ad6878fba</guid>
            <category><![CDATA[software-development]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[product-management]]></category>
            <category><![CDATA[healthcare]]></category>
            <category><![CDATA[automation]]></category>
            <dc:creator><![CDATA[Alex Pavlov]]></dc:creator>
            <pubDate>Fri, 31 Jul 2026 19:26:52 GMT</pubDate>
            <atom:updated>2026-07-31T19:26:52.019Z</atom:updated>
            <content:encoded><![CDATA[<p>Your pipeline works, mostly. A rep pulls a lead out of an enrichment tool, pastes it into the CRM, writes a first touch, sets a reminder, and forgets to log the call. So you start shopping for sales automation software and find forty products promising the same six things. The choice feels like a tool comparison. It is a data question.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*eytFXTa1JtoT6BC8.jpg" /></figure><p>What decides whether an automated motion survives a real quarter is where the state of a deal lives. State is the answer to one sentence: which step is this account on, since when, and what happens if nobody replies. Every tool has an opinion about that. Most buyers never read it, and then pay for the difference with a quarter of somebody’s time.</p><p>I write this from the build side. When we scope <a href="https://highcraft.io/sales-automation/?utm_source=medium&amp;utm_medium=article&amp;utm_campaign=sales-automation">a sales automation build for a revenue team</a>, the first hour goes to that question rather than to the vendor shortlist. Below is what I check, and where the named tools actually bend under load.</p><h3>Each tool picks a different place to keep the state</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*Lbq_ezNx7A3uuArC.jpg" /></figure><p>Zapier holds nothing between runs. A Zap is one pass through the steps, triggered and forgotten. To remember that a lead is on touch three and has been silent since Tuesday, you need Zapier Tables, an Airtable base, a Google Sheet, or a CRM field you have quietly promoted to a database column. That store has no constraints and no migrations, and nobody owns it. When a step fails repeatedly, Zapier switches the Zap off and emails you. (You will read that email on Thursday.)</p><p>n8n gives you real branching, wait nodes, error branches, and a code node for when a vendor API is odd. Self-host it and the execution history lives in your own Postgres, where you can query it like data instead of scrolling a run log. You also own the runtime: version upgrades, execution-data pruning, credential rotation, and queue mode with Redis once one long workflow starts blocking the rest.</p><p>HubSpot workflows put the state on the record itself, as properties plus enrollment criteria, with re-enrollment rules you set per workflow. That is clean while the state is per-contact and moves in one direction. It gets strange the moment two workflows write the same property from different branches.</p><h3>Your CRM’s object model is an argument about how you sell</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*pVjKxhS9RU7dTnt0.jpg" /></figure><p>HubSpot deduplicates contacts on email and companies on domain. Push five thousand rows out of Apollo or Clay and you meet every variant of that rule: aliases, plus-addressing, personal addresses on a free provider, two people sharing an ops@ inbox, a company whose domain differs from its email domain after an acquisition. Do the normalisation before the API call, in code you control.</p><p>A HubSpot deal carries one pipeline, one stage, one owner. If an account can be in a new-business motion and an expansion motion at the same time, you either create two deals and lose the account-level roll-up, or you buy custom objects on the Enterprise tier and build the roll-up yourself. Neither is wrong. Both are a decision somebody makes in week two, usually by accident.</p><p>Salesforce splits the same idea into Opportunity and Opportunity Line Items, then layers record types, validation rules, sharing rules, and an order of execution you will eventually read line by line. A Flow that writes a field can trigger another Flow, and in some configurations that happens twice per save. That is the debugging you inherit the day your automation starts writing into Salesforce rather than just reading from it.</p><h3>The integration bill arrives at the edges</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*RVwqcVCl7YFa3_XL.jpg" /></figure><p>The first sync takes two days. The next three weeks are the edges.</p><p>Webhooks are delivered at least once, and not in order. Update a contact twice inside 200 milliseconds and you get two callbacks that can land either way round. Without an idempotency key and a last-write-wins rule on a real timestamp, your enrichment overwrites the rep’s manual correction, and the rep stops trusting the system that week. HubSpot rate limits are per app per portal, with daily caps, so a backfill and your live sync compete for the same budget. Pagination cursors expire mid-crawl on more APIs than you would expect.</p><p>Reply detection is its own project. Gmail threading is reliable, out-of-office autoresponders are not. An automation that treats any inbound message as a reply will pull a prospect out of the sequence because their laptop is closed until the 14th.</p><h3>Sequencers own the reply, and you should draw that boundary yourself</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*I2v6EPXqOqGJqFqG.jpg" /></figure><p>Outreach, Salesloft, and Apollo own sending and reply detection well. Their CRM sync is a field mapping, and field mappings drift the moment someone adds a picklist value. The sharper problem is overlap: the same prospect can sit in two sequences started by two reps, and neither system will tell the other. Pick one system of record for “are we currently touching this person” and make every other tool read from it. The CRM is usually the right choice, because reporting already lives there.</p><h3>Draw the process before you shortlist</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*or9Ta36rGcxV_lTE.jpg" /></figure><p>Before anyone opens a pricing page, draw the motion. <a href="https://www.bpmn.org/">BPMN</a> is more notation than you need, and that is fine, use the boxes and the diamonds and ignore the rest. Count the states, the human decisions, and the points where an outside system can refuse.</p><p>My rule of thumb: under eight states with one human decision point, a visual builder is comfortable. Past fifteen states, with retries and parallel branches and a compensating path for a deal that goes cold and comes back, the canvas stops being a diagram of the process and becomes a diagram of a bug you cannot see.</p><h3>The build and buy line I hold</h3><p>Nobody should build an email sender. Same for the CRM and the enrichment vendor. The part worth building is the decision layer: the code that says this lead is qualified, this account gets the two-week nurture, this one goes straight to a rep. If that logic is genuinely yours and changes every few weeks, it belongs in a service you can test, not in a filter step somebody edits in a browser tab at 6pm.</p><p>Effort shapes, roughly. A Zapier motion is live in a week and costs about a day a month of someone reconnecting auth and replaying failed tasks. A self-hosted n8n build is three weeks of integration work plus a named owner for upgrades. A custom service over the CRM API is six to ten weeks for the first motion, after which the second motion takes days, because the state model already exists. Pricing shapes differ too. Zapier charges per task, so a chatty webhook is a line item. HubSpot charges per seat and per marketing contact, which changes how you feel about importing that list.</p><h3>Who owns which decision</h3><p>The revenue lead owns the process and the written definition of qualified. Ops owns fields, picklists, deduplication rules, and who is allowed to edit them. Engineering owns retries, idempotency, backfills, and where the state lives. The founder owns exactly one call: which piece of the qualification logic is proprietary enough to keep in code you control rather than in a vendor’s workflow builder.</p><p>The common drift is engineering defining what qualified means, because they wrote the filter and no one else read it. You fix that with one artifact: a list of every state and its exit condition, owned by the revenue lead, referenced by name in the code.</p><h3>Test these before you sign</h3><p>During the trial, with real data rather than three test rows:</p><ul><li>Load 200 genuine leads, messy ones included, and count the duplicates that come out the other side.</li><li>Revoke the CRM OAuth token mid-run and watch whether the tool retries, queues, or silently drops the record.</li><li>Push an out-of-office autoreply into a sequence and see if it registers as a reply.</li><li>Rename a pipeline stage.</li></ul><p>That last one costs nothing and tells you the most. Rename “Demo Booked” to “Discovery Booked” and see what survives. Filters bound to the stage ID keep running and your reports keep counting. Filters that matched the label go quiet, the enrollment stops, and nobody notices until the weekly number is short by eleven deals and someone starts reading run logs on a Friday.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=286ad6878fba" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Contact Center Automation: Which Parts of a Support Workflow to Hand Over, and Who Owns the…]]></title>
            <link>https://alex-on-ai.medium.com/contact-center-automation-which-parts-of-a-support-workflow-to-hand-over-and-who-owns-the-90f78b3c7a39?source=rss-93cf9798cdda------2</link>
            <guid isPermaLink="false">https://medium.com/p/90f78b3c7a39</guid>
            <category><![CDATA[product-management]]></category>
            <category><![CDATA[software-development]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[automation]]></category>
            <category><![CDATA[healthcare]]></category>
            <dc:creator><![CDATA[Alex Pavlov]]></dc:creator>
            <pubDate>Thu, 30 Jul 2026 22:32:30 GMT</pubDate>
            <atom:updated>2026-07-30T22:32:30.000Z</atom:updated>
            <content:encoded><![CDATA[<h3>Contact Center Automation: Which Parts of a Support Workflow to Hand Over, and Who Owns the Fallback</h3><p>Someone in a planning meeting has asked what percentage of your support workflow you can automate. That is the wrong unit of measure. A ticket is not one thing. It is six or seven steps, and contact center automation succeeds or fails on each step separately.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*v6HewdM56wcu4QmI.jpg" /></figure><p>The step that reads an order status and the step that issues the refund carry completely different risk. One is a read. The other moves money and cannot be un-moved by a retry. Most of the scoping work I do on <a href="https://highcraft.io/contact-center-automation/?utm_source=medium&amp;utm_medium=article&amp;utm_campaign=contact-center-automation">automation for support and contact center workflows</a> is pulling those two categories apart inside a process that currently treats them as one job. Once they are separated, the useful question appears: which steps get write access to your systems of record, and what does the workflow do when one of them fails halfway through.</p><h3>A ticket is six steps and only two of them are hard</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*qvUkNeHCYK2FbRTz.jpg" /></figure><p>Break any support interaction down and you get roughly this:</p><ul><li>intake and intent classification</li><li>identity resolution</li><li>entitlement check (is this person owed the thing they are asking for)</li><li>decision</li><li>execution against a system of record</li><li>confirmation, logging, and status change</li></ul><p>Classification is close to solved. A model sitting on top of your existing Zendesk macro taxonomy will route intents well enough on day one, and if you map its labels to the macros you already report on, your dashboards keep working through the transition. That part takes days.</p><p>Identity resolution and execution are where projects stall. Everything in between is bookkeeping.</p><h3>Containment is a vanity metric with a good PR team</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*Rsw2OJCaKLxt38S1.jpg" /></figure><p>Containment and deflection rates flatter whoever built the bot. Zendesk will move a solved ticket to closed after a set number of days, and anything a bot marks solved counts as contained even if the customer opened a fresh ticket about the same problem on Tuesday. Reopen rate inside 72 hours tells you far more, and it is cheap to instrument.</p><p>Watch the second number too: average handle time on tickets that a bot touched before escalating. It is usually higher than on tickets that came straight to an agent, because the agent now reads a transcript before doing anything. If your automation is quietly adding ninety seconds to every escalated contact, containment of 40 percent is buying you less than the slide claims. (Every support dashboard has one number that is politely lying. It is normally the biggest one.)</p><p>If you are evaluating Intercom Fin or anything else priced per resolution, read the vendor’s definition of resolution before you sign. That definition becomes your internal KPI whether you intended it or not.</p><h3>Your CRM data model decides what you can automate</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*nmh_pD48X8D1jnly.jpg" /></figure><p>This is the part that gets underestimated in every kickoff I have sat in. The bot cannot answer “where is my order” if orders live in the ERP and your CRM only knows about contacts and tickets. It cannot check entitlement if the subscription state lives in the billing system and nobody has agreed which one wins when they disagree.</p><p>HubSpot will let you model this through associations and custom objects, but custom objects are gated by tier, and the association labels you need for a parent account to child contact relationship are fiddly to maintain once support agents start creating contacts by hand. Salesforce gives you a stronger Case object and record types, and then rejects your API write with a validation exception because a required field was added by an admin last quarter and nobody told the integration team. Your automation now fails on a subset of tickets for reasons that live in a config screen you do not own.</p><p>Then there is identity. Half of B2B support email arrives from shared aliases. Three humans share billing@, the CRM has two contacts on the same address, and your workflow needs to decide which account it is acting on before it acts. You are not integrating an API here. You are reconciling two identity models that were never designed to agree.</p><p>Budget for it honestly. Prompt tuning for intent handling is a couple of afternoons. Reconciling identity and entitlement across a CRM, a billing system, and an ops database is three weeks of integration work before anything customer-facing ships. On a healthcare product where entitlement lives behind an HL7v2 feed or a FHIR R4 endpoint on an Epic instance, add the negotiation time for the interface itself, which is measured in calendar weeks, not sprints.</p><h3>Zapier gets you a demo, n8n gets you an on-call rota</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*ixGHP7GT1gRn2DPG.jpg" /></figure><p>Both of these belong in a support stack. They belong in different parts of it.</p><p>Zapier is excellent for the first version. Connectors are broad, and you will have a working triage path in an afternoon. Live with it for a month and the edges show up. Polling triggers on lower plans check on an interval rather than firing on a webhook, so “instant” is a plan feature. Billing is per task, which means a chatty loop is a budget event. Branching state across many steps gets awkward, error handling is configured Zap by Zap, and version control is whatever discipline the person editing it had that morning.</p><p>n8n self-hosted gives you the things you want when the workflow starts touching money. Queue mode with Redis and separate workers so one slow HTTP call does not stall the rest. Error workflows that catch failures centrally. Real versioning, because the workflow JSON sits in your repository. The cost is that you now run it. Execution data goes into Postgres and grows until you configure pruning, and I have seen a support automation database outgrow the app database it was helping. Upgrades need pinned versions and a staging instance, same as any service you actually depend on.</p><p>Neither of them is your routing layer. Amazon Connect or Genesys still owns queues, voice, and agent state. The automation platform orchestrates around it.</p><h3>Model the failure paths before you write a prompt</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*h_XDX34XPQUyfd1a.jpg" /></figure><p>Draw the process properly. <a href="https://www.bpmn.org/">BPMN</a> is worth the hour it takes to learn the notation, mostly because it forces you to place two things you would otherwise skip: timeout boundary events and compensation.</p><p>A timeout is what happens when the billing API takes eleven seconds and your workflow gave it eight. Compensation is what happens when step five succeeded and step six did not. That specific case is the one that hurts. The refund goes through in Stripe, the ticket update fails, the retry runs the whole branch again, and the customer gets paid twice. Add an idempotency key derived from the ticket ID and the amount. The double refund stops being possible rather than becoming rare.</p><p>Decide up front where a half-finished execution lands. A dead-letter queue that a named human reviews every morning is fine. A retry loop with no ceiling is not, and neither is a workflow that silently swallows the exception and marks the ticket solved.</p><h3>The escalation handoff is where trust is won or lost</h3><p>When automation hands a contact to a person, it should hand over a packet: detected intent, confidence, the entities it extracted, every action it attempted, and what it changed. Put that in an internal note, never in the customer-visible thread.</p><p>On voice this is mechanical and people still get it wrong. Amazon Connect carries contact attributes through to the agent workspace, Genesys has participant data, and if you transfer without populating either, the agent starts from zero and asks the customer to verify again. That is the moment the customer decides your automation cost them time. The reverification ask is more damaging than the bot failing outright, because failing fast at least routes them to a human with their patience intact.</p><p>Set the escalation confidence threshold in configuration and give your support lead the ability to move it without a deploy. They will tune it weekly for the first month and then leave it alone.</p><h3>Who owns which decision</h3><p>This is the part that is genuinely a management decision rather than a technical one, and it goes badly when it is left implicit.</p><p>The support lead owns which intents are allowed to auto-resolve, the tone, and the escalation threshold. Product owns which write actions exist as callable tools at all, because that list is the actual surface area of the system. Engineering owns idempotency, rate limits, secrets, and the rollback path. Finance owns the refund ceiling and should be the one to state the number out loud.</p><p>If you take one decision away from this piece, take that middle one. Retrieval and triage can be fully automated now. Write actions should only be automated where a single command reverses them and the value sits under an agreed ceiling. Anything touching money, entitlement, or identity keeps a human in the loop until the reopen data says otherwise.</p><p>Keep that ceiling in a config value that goes through code review, not in a prompt, and log the config version alongside every decision the system makes. When someone asks in March why a particular refund went out in January, the answer should be a row you can query.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=90f78b3c7a39" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Automating Outbound Lead Generation Strategies: Where n8n Beats Zapier, and Where HubSpot Should…]]></title>
            <link>https://alex-on-ai.medium.com/automating-outbound-lead-generation-strategies-where-n8n-beats-zapier-and-where-hubspot-should-df21a312e9d6?source=rss-93cf9798cdda------2</link>
            <guid isPermaLink="false">https://medium.com/p/df21a312e9d6</guid>
            <category><![CDATA[software-development]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[technology]]></category>
            <category><![CDATA[healthcare]]></category>
            <category><![CDATA[automation]]></category>
            <dc:creator><![CDATA[Alex Pavlov]]></dc:creator>
            <pubDate>Thu, 30 Jul 2026 11:59:58 GMT</pubDate>
            <atom:updated>2026-07-30T11:59:58.302Z</atom:updated>
            <content:encoded><![CDATA[<h3>Automating Outbound Lead Generation Strategies: Where n8n Beats Zapier, and Where HubSpot Should Own the Record</h3><p>The decision in front of you is narrower than it looks. Not “should we automate outbound,” but which specific steps a machine runs unattended, and which ones stop in a human’s inbox. Everything after that is plumbing, and plumbing has a bill.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*HGwvTD3fgc4U9cJp.jpg" /></figure><p>Most automated outbound lead generation strategies come apart at the join between three systems that each believe they own the contact. The sequencer has its own lead object. The CRM has a contact and maybe a separate lead. The enrichment tool has a person keyed on an email that changed jobs in March. If you want the longer version of the argument, I wrote up <a href="https://highcraft.io/blog/outbound-lead-generation-strategies/?utm_source=medium&amp;utm_medium=article&amp;utm_campaign=blog-outbound-lead-generation-strategies">how outbound lead generation strategies break at the data layer</a> separately. Here I want to get specific about the tools and who signs off on what.</p><h3>The sequencer sends, the CRM remembers</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*AoPprPmiB_Yt8yI6.jpg" /></figure><p>Pick your system of record before you connect a single webhook. This sounds obvious until you look at how these products actually store people.</p><p>Instantly and Smartlead keep their own lead objects, scoped to a campaign. Upload the same person into two campaigns and you have two records with two independent statuses. HubSpot deduplicates contacts primarily on email address, so a prospect who moves from one company to another arrives as a brand new contact with an empty history. Salesforce splits the problem differently, with Leads and Contacts as separate objects and a conversion step that rewrites ownership and drops some of the activity linkage you were counting on for reporting.</p><p>So the rule I hold to: the sequencer is a sending appliance, not a database. It gets a name, an email, a few merge fields, and a campaign. Every state change it produces (replied, bounced, unsubscribed, meeting booked) gets pushed back into one system and only one. If your revenue lead pulls reply rates from the sequencer dashboard and your board deck pulls them from HubSpot, those two numbers will disagree by the end of the quarter, and you will spend a Friday finding out which one lied.</p><h3>Zapier charges per event, n8n charges per hour of your attention</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*CsFV6CFhehr7u0pJ.jpg" /></figure><p>Zapier is excellent at exactly one shape of work: a single trigger, a handful of steps, low volume, no loops. Form submission comes in, contact gets created, Slack message fires. Ten minutes of setup and it holds.</p><p>Where it stops being the cheap option is when your motion needs to iterate. Backfilling four hundred accounts through an enrichment call, then batching the results into a CRM upsert, is not a Zap. You end up with a Looping by Zapier step, a chain of storage hacks, and a task counter that grows with every record. Zapier’s error behavior is also worth knowing before you rely on it. A step that keeps failing does not queue politely. The Zap gets held, you get an email, and whatever was mid-flight sits in a state nobody wrote down.</p><p>n8n solves the shape problem. Real branching, batching with a Split In Batches node, a Code node when the transformation is genuinely fiddly, and cron triggers you can reason about. Self-hosted, running in queue mode with Postgres behind it, it will chew through the volumes that make Zapier’s pricing page look like a dare.</p><p>The cost is not money. It is that you now run a service. Postgres backups, version upgrades that occasionally change node behavior, credential rotation, and a workflow that silently stops firing because a container restarted and nobody was watching the executions list. When one of those workflows becomes load-bearing for pipeline, expect it to absorb a real slice of an engineer’s week, permanently. This decision belongs to whoever gets the call when it breaks at 7am on a Tuesday. If that person is your only technical hire and they are also shipping the product, use Zapier and pay the tollbooth.</p><h3>Enrichment is a cache with a time to live</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*X7E7-9iYtNssMYxp.jpg" /></figure><p>Treat Clay and Apollo as caches, never as live lookups. I have watched teams call enrichment inline at send time, which means their sending pipeline now inherits the availability of a third-party API and the latency of a waterfall that tries six providers before it gives up.</p><p>What works better: a scheduled job that enriches on a cadence, writes the raw provider payload into your own store, stamps it with a fetched-at timestamp, and only then maps fields into the CRM. Keep the raw JSON. When you decide in month four that you actually needed headcount-by-department, you will not want to re-buy nine thousand records you already paid for.</p><p>Field mapping is where this bites. HubSpot enumeration properties reject values that are not in the option list, so an enrichment provider returning “Head of Revenue Operations” against a seniority dropdown gives you a 400 and a skipped record, not a helpful warning. Normalize titles into your own seniority buckets in the automation layer, before the write. Do the mapping in code you can read, not in a mapping UI three people have edited.</p><p>Also make every write idempotent. I have shipped FHIR R4 integrations where each write carried an idempotency key because duplicates were unacceptable, and an outbound pipeline deserves the same care for a cheaper reason: a retried batch that creates 1,200 duplicate contacts will also happily email all of them.</p><h3>Deliverability is infrastructure and it needs a named owner</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*doq0u7SI31v3wH99.jpg" /></figure><p>Sending domains are not a growth setting. They are infrastructure, and they degrade in ways you cannot undo with a support ticket.</p><p>The non-negotiables are SPF, DKIM, and DMARC on every sending domain, plus a warm-up ramp that does not go from zero to two hundred a day in a week. Google’s own <a href="https://support.google.com/a/answer/81126">Gmail sender guidelines</a> put the spam complaint threshold at under 0.10 percent and treat 0.30 percent as a line you should never touch, and they require authentication for bulk senders. That is the floor, not a target.</p><p>Never send cold volume from the domain that carries your invoices, password resets, and customer support. One domain per motion, one warm-up schedule, one owner. A burned domain does not send you a resignation letter. It just quietly stops arriving, and you find out from a customer who never got their renewal notice.</p><p>One more trap in this layer. Sequencers detect replies by parsing the inbox, and out-of-office autoresponders look exactly like engagement. If your automation moves anyone with a reply into a “warm” stage, your pipeline will inflate every August. Filter on common autoresponder headers and subject patterns before you let a reply change CRM state.</p><h3>Your list is the strategy, the copy is just how it arrives</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*B-fIxWfjjOiT-F-l.jpg" /></figure><p>Segment definitions belong in version control, not in a saved CRM view that somebody adjusted at 4pm to hit a weekly number. I keep them as SQL or as a documented query in the automation layer, and I snapshot membership at send time. Otherwise you cannot answer the only question that matters in month three: which segment actually replied.</p><p>The join people skip is suppression. Before any contact enters a sequence, it needs to clear a check against current customers, open opportunities, recently churned accounts, active support escalations, partner contacts, and anyone who unsubscribed anywhere. (The sequencer will happily send to all of them. It has no idea who your customers are.) In HubSpot you can approximate this with an active list, but active lists refresh asynchronously and the lag is real, so a contact created and sequenced in the same run can slip through. Do the suppression check as an explicit lookup in the workflow, immediately before the enrollment call.</p><p>Cost of getting this wrong is not a metric. It is your best account’s VP of Engineering receiving a cold pitch for the product they already pay for, forwarded to your CEO with three question marks.</p><h3>Reply handling is where the automation should stop</h3><p>Classification is a good job for a machine. Response drafting is not, at least not unattended.</p><p>What I build: an LLM step that reads the inbound reply and tags it into a small set of buckets, something like interested, not now, wrong person, referral, unsubscribe, autoresponder. That tag routes the thread and updates the CRM. It does not send anything. The bucket set stays small on purpose, because a fifteen-category taxonomy will drift and nobody will notice until the routing is wrong.</p><p>Routing itself has a licensing footprint worth checking before you design around it. Round robin assignment and meeting-link routing in HubSpot sit behind paid seat tiers, and if you have two SDRs on free seats, your elegant assignment logic ends up as a Slack message and a race condition. I would rather write the assignment in n8n against a simple ownership table than discover the plan limit after the workflow is live.</p><h3>Buy the sending, build the join</h3><p>Sequencers, enrichment, calendars, all bought. The layer nobody sells you is the one that reconciles them: a small owned store with contacts, accounts, segment membership with timestamps, suppression state, and every event the sequencer emitted. Two tables and a couple of workflows is a fine starting size. It exists so that when you switch sequencers (and you will, roughly every eighteen months, usually over deliverability), you are migrating campaigns, not history.</p><p>One last instrumentation note, because it changes what you optimize. Open rates are close to meaningless now. Apple’s Mail Privacy Protection pre-fetches tracking pixels, so opens inflate and the inflation is not evenly distributed across your list. Measure reply rate and meeting rate, per segment, per variant, against the segment snapshot taken on the day of send. If your dashboard cannot express that sentence, the dashboard is decoration and the numbers in your weekly review are describing a campaign that no longer exists.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=df21a312e9d6" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[What Every Healthcare Founder Asks Before Hiring an AI Automation Agency]]></title>
            <link>https://alex-on-ai.medium.com/what-every-healthcare-founder-asks-before-hiring-an-ai-automation-agency-f45d98ff6dd1?source=rss-93cf9798cdda------2</link>
            <guid isPermaLink="false">https://medium.com/p/f45d98ff6dd1</guid>
            <category><![CDATA[automation]]></category>
            <category><![CDATA[healthcare]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[machine-learning]]></category>
            <dc:creator><![CDATA[Alex Pavlov]]></dc:creator>
            <pubDate>Thu, 16 Jul 2026 18:29:53 GMT</pubDate>
            <atom:updated>2026-07-16T18:29:53.419Z</atom:updated>
            <content:encoded><![CDATA[<p>Your third analyst just quit this month, worn down by doing prior authorization lookups by hand, and now you are asking the same question every healthcare founder eventually asks: build the automation yourselves, or bring in an ai automation agency to do it. Nobody asks that question when things are going well. I have spent eleven years building software, a good chunk of it inside FinTech compliance rules that make HIPAA look relaxed by comparison, and this is the call I get most often these days.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*F_d6EHM_PSZaXKkv.png" /></figure><p>My answer is almost never a straight yes or no. It depends on whether the workflow is actually a workflow, or just a chore wearing a workflow’s clothes. I wrote up how we think about that distinction on the <a href="https://highcraft.io/ai-automation-agency/?utm_source=medium&amp;utm_medium=article&amp;utm_campaign=ai-automation-agency">ai automation agency</a> page, and I want to go further here, into the parts that do not fit on a services page: the failure modes, the trade-offs, and the moment in a project where you find out if the vendor actually understood healthcare or just nodded a lot in the discovery call.</p><p>Honestly, the discovery call is where most of this gets decided, long before anyone opens an IDE.</p><h3>Healthcare workflows are not spreadsheets with better manners</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*tYg84qkt5iQNAjzn.png" /></figure><p>Every healthcare ops team I have worked with describes their process as simple right up until I ask what happens on the exception path. Insurance eligibility check, fine. Except when the payer’s API times out. Except when the patient has two policies. Except when the diagnosis code changed last quarter and nobody updated the mapping. The exceptions are the job. The happy path is a rounding error.</p><p>An automation built for the happy path looks great in a demo and falls over in week three, usually on a Friday, usually right before a compliance audit. I have seen teams try to patch this with more prompts. That is like trying to fix a leaky roof with a nicer umbrella. The fix is mapping the actual process, branches included, before a single model gets called.</p><p>This is where <a href="https://www.bpmn.org/">BPMN</a> earns its keep. Not because a diagram is exciting, nobody has ever framed a swimlane chart and hung it in the lobby, but because it forces someone to answer the exception questions on paper instead of in production. If your vendor skips this step and goes straight to a slide about their AI stack, that is a tell.</p><h3>The build-versus-buy question is really a build-versus-audit question</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*_UoG_0Mh7JZ4B7up.png" /></figure><p>Founders frame this as build in-house versus hire an agency. The real fork is narrower: can you audit what gets shipped, or are you trusting it. In regulated healthcare, an automation that denies a claim, mis-routes a referral, or drops a consent flag is not a bug ticket. It is an incident report, possibly a legal one.</p><p>Building in-house does not automatically solve this. Plenty of in-house teams ship automations nobody can explain six months later, because the one engineer who understood the edge cases left for a startup with a ping pong table. Hiring an agency does not automatically solve it either. Plenty of agencies ship a demo, invoice you, and vanish, leaving behind a black box with your logo on it.</p><p>What actually solves it is an audit trail: who approved the workflow, what data it touches, what happens when the model is unsure, and who signs off on changes. If a vendor cannot describe their release approval process in one paragraph, they do not have one. Ask early. Ask twice if the first answer sounds rehearsed.</p><h3>Hallucination is not the scary failure mode, silent drift is</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*NS4RGL4NWtrfdoRR.png" /></figure><p>Everyone worries about the AI making something up. Fair, that happens, and in clinical or billing contexts it is genuinely bad. But hallucination is loud. It produces an obviously wrong answer that a human usually catches, the way a smoke alarm is annoying but at least tells you something is on fire.</p><p>The failure mode that actually gets teams in trouble is quieter. A model performs fine for months, then a payer changes a policy format, or an EHR vendor tweaks a field name, and the automation keeps running, confidently, on stale assumptions. Nobody notices because nothing crashed. The queue just quietly fills with wrong outputs that look plausible enough to pass a tired reviewer’s glance.</p><p>The fix is boring: monitoring, sampling, and a human review threshold that is not just theater. I have watched teams install a review step, treat it as a checkbox for the compliance binder, and rubber stamp everything through it. That is not a control, that is a control-shaped object. A real ai automation agency builds the review step to actually catch drift, with sampling rates and escalation paths, not a button someone clicks without reading.</p><h3>The vendor conversation that actually matters</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*zqK4fEcYLQwniEZ-.png" /></figure><p>Most sales conversations with automation vendors spend time on the model. GPT-this, fine-tuned-that. I care less about which model and more about three things: how they handle protected health information in transit and at rest, how they roll back a bad deployment, and how they price change requests after launch.</p><p>On the first point, ask for specifics on data residency and access logging, not a slide with a HIPAA badge on it. A badge is not a control either.</p><p>On the second, ask what a rollback actually looks like at 2am when a workflow starts misclassifying referrals. If the answer involves someone SSHing into a server and hoping, that is not a plan.</p><p>On the third, and this is the one nobody wants to talk about, ask what happens six months in when the payer changes their portal and the automation needs an update. Some agencies price that as a new project. Others treat it as maintenance included in the relationship. Know which one you are signing up for before you sign anything.</p><p>I would also ask how much of the team is senior. Not because junior engineers cannot write good code, they can, but because healthcare automation punishes assumptions, and assumptions are cheaper to catch with more experience in the room. My own team runs this way for a reason, and it shows up in fewer 2am rollbacks, not in the sales deck.</p><h3>When you should not hire anyone at all</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*w_0QiHGzWs4m7_mx.png" /></figure><p>If your workflow touches under twenty transactions a week and a junior analyst can clear it in an hour, you do not need an agency, a model, or a consulting engagement. You need a checklist and maybe a better spreadsheet. I have talked founders out of automation projects more than once, and every one of those conversations saved them a budget line and, more importantly, six months of maintaining something they did not need.</p><p>The volume and complexity threshold is where you actually start needing engineering. Below that line, buy a template or hire a part-time ops person. Above it, and especially where the process branches into exceptions the way healthcare always does, that is where automation, done properly, pays for itself.</p><h3>What a real engagement costs and how long it takes</h3><p>A scoped MVP for a single high-volume workflow, prior auth intake or eligibility verification are common ones, generally runs a focused engineering effort measured in weeks, not a year-long platform rebuild. The quickest path is narrow scope: one workflow, one exception map, one review gate, shipped and monitored before anyone talks about scaling it to five more processes. The slowest path, and the one I have watched burn budgets, is trying to automate the whole department at once with a single sweeping AI initiative. It reads well in a board deck. It ships nothing for a year.</p><p>The cleanest option I have seen work, across a US digital-health platform’s intake process and a mid-size payer’s claims triage, is the same pattern every time: map the process with <a href="https://www.bpmn.org/">BPMN</a> or an equivalent, automate the highest-volume branch first, put a real human review gate on it, then expand once the metrics hold up for a full billing cycle. Boring. Repeatable. Works.</p><h3>The agency question, answered honestly</h3><p>So, build it yourself or bring in an ai automation agency. If you have senior engineers with healthcare domain experience and spare bandwidth, build it, and hold yourself to the same audit-trail standard you would demand of a vendor. If you do not have that bandwidth, or the domain experience, hiring outside help is not a shortcut around rigor, it is a way to buy rigor you do not currently have in house.</p><p>I wrote more on how we scope that first engagement on the ai automation agency page, including what a discovery sprint actually covers before any code gets written. Read it before your next vendor call. It will save you from nodding along to a slide about AI models when what you actually needed to ask about was rollback plans.</p><p>Either way, map the exceptions before you automate the happy path. The happy path was never the problem. It was just the only part anyone bothered to draw.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=f45d98ff6dd1" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Hiring an SEO Lead Generation Firm Won’t Fix a Healthcare Product Nobody Has Scoped]]></title>
            <link>https://alex-on-ai.medium.com/hiring-an-seo-lead-generation-firm-wont-fix-a-healthcare-product-nobody-has-scoped-f62fc6f75194?source=rss-93cf9798cdda------2</link>
            <guid isPermaLink="false">https://medium.com/p/f62fc6f75194</guid>
            <category><![CDATA[automation]]></category>
            <category><![CDATA[software-development]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[healthcare]]></category>
            <dc:creator><![CDATA[Alex Pavlov]]></dc:creator>
            <pubDate>Mon, 13 Jul 2026 06:14:26 GMT</pubDate>
            <atom:updated>2026-07-13T06:15:09.375Z</atom:updated>
            <content:encoded><![CDATA[<p>Founders building regulated healthcare products keep asking me some version of the same question, almost every month. Should we hire seo lead generation services and let the pipeline handle itself while engineering ships the product. Honestly, the question is usually a symptom. The real problem is almost never traffic. It is that nobody has decided what a qualified lead looks like before the ad spend or the content calendar gets approved.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*DWmL25YNgcFOvSix.png" /></figure><p>I wrote up the longer version of my thinking on this at <a href="https://highcraft.io/seo-lead-generation/?utm_source=medium&amp;utm_medium=article&amp;utm_campaign=seo-lead-generation">a breakdown of how healthcare teams should evaluate seo lead generation services</a>, and I still send that link to founders before we get on a call, because it saves us both twenty minutes of me explaining why “just get us more leads” is not a spec.</p><p>The short version: healthcare content marketing has a compliance tax that generic SEO advice does not account for, and most vendors selling seo lead generation services have never had to run a piece of copy past a compliance officer who has read the FTC’s health claim guidance twice and trusts nobody, including themselves. That changes the math on build versus buy more than most founders expect.</p><h3>The build decision starts before the backlog does</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*o-BlteNVBKDYMitR.png" /></figure><p>Every build-versus-buy conversation I have run for a client starts with the same exercise. Define the lead. Not the visitor, not the email capture, the lead. In consumer SaaS that is easy. In healthcare it is a small research project. Is a lead a licensed provider who filled out a demo form. A health system procurement contact who downloaded a whitepaper. A patient who is not actually your buyer at all, no matter how many blog posts you rank for their symptoms. I have watched teams spend two quarters optimizing for search volume around a condition, then discover their actual buyer is a hospital IT director who never searches that term because he already knows what the condition is, he just wants to know if your API talks to Epic.</p><p>Get that definition wrong and no amount of ranking fixes it. You just get faster, cheaper, more convincing wrong leads. That is not a growth problem. That is a targeting problem wearing a growth costume.</p><h3>Healthcare content has a compliance tax generic SEO playbooks ignore</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*0MVTPN8qLlsFZk3J.png" /></figure><p>Most SEO playbooks assume you can publish fast, iterate on what ranks, and prune what doesn’t. In a regulated healthcare business, every piece of content that touches clinical claims, outcomes data, or anything that reads like medical advice needs a review pass before it ships. That review pass is not optional and it is not fast. I have seen founders budget for content velocity as if their compliance team does not exist, then wonder why the agency’s promised twelve posts a month turned into three.</p><p>The trade-off here is real and worth naming out loud instead of discovering it in month two. You can move fast on content that stays firmly in the operational and product lane (integration guides, security posture, workflow explainers) and slow down deliberately on anything that touches clinical outcomes or patient-facing claims. Vendors who don’t draw that line for you are not being efficient, they are being reckless with your legal exposure. Ask any agency pitching you seo lead generation services what their review process looks like for a claim like “reduces readmissions.” If the answer is “we cite a source,” that is not a compliance process, that is a hope.</p><h3>Most vendors sell traffic, not qualified pipeline</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*F9a-uqmH77rujZ8q.png" /></figure><p>Here is the failure mode I see most often, and it is almost boringly consistent. An agency ranks a page. Traffic goes up. Everyone celebrates in the monthly report. Three months later, sales complains that none of the form fills convert, because the page ranked for a broad symptom term, and the people filling out the form are patients, family caregivers, or students writing a paper, not the buyer who signs a contract.</p><p>This is not a bad-actor problem, most of the time. It is an incentive problem. Agencies get paid, or get renewed, on traffic and rankings because those are easy to report and easy to screenshot. Qualified pipeline is harder to measure and slower to show, so it gets deprioritized unless you force the contract to price it that way. If you are evaluating seo lead generation services for a healthcare product, price the engagement on qualified opportunities reaching a sales conversation, not on keyword position. It changes which vendors want the work, and it changes what they build.</p><h3>The trade-off is speed versus defensibility</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*uxiApLjSbiBqpMzF.png" /></figure><p>Building an internal content and SEO function is slower to start and more defensible over time. You own the editorial calendar, you own the compliance relationship, and you are not renegotiating scope every quarter with an outside team that has five other healthcare clients competing for the same senior writer. Buying gets you output faster, but you inherit someone else’s process, and in a regulated space, someone else’s process is now your risk if it is thin.</p><p>We have shipped internal content operations for a US digital-health platform where the constraint was not writing talent, it was clinical review bandwidth. The fix was not more writers. It was a shorter, tighter content brief template that flagged anything requiring clinical sign-off before a single word got drafted, so the review queue stopped being a surprise. That is a workflow fix, not a marketing fix, and no seo lead generation services contract will hand you that unless you build it into the scope yourself.</p><h3>Email still decides whether your leads ever see your product</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*luMjHX2MfRNDT9fK.png" /></figure><p>Ranking and converting a form fill is only half the job. What happens after the form fill is where I see healthcare teams lose leads they already paid to earn. Nurture sequences in this space get flagged, bounced, or ignored more than in general SaaS, partly because healthcare buyers are cautious clickers and partly because sloppy sending patterns get you throttled. If your outbound volume or formatting drifts from what mailbox providers expect, deliverability degrades quietly, and nobody notices until pipeline dries up for reasons that look like a top-of-funnel problem but are actually a plumbing problem. It is worth reading the <a href="https://support.google.com/a/answer/81126">Gmail sender guidelines</a> before you scale any nurture sequence past a few hundred contacts, because the rules changed the economics of cold-adjacent outreach for everyone, healthcare included.</p><h3>A rule of thumb for buying versus building</h3><p>If your compliance review process is mature and your bottleneck is genuinely writer bandwidth, buying makes sense, and it can move fast. If your bottleneck is that nobody has agreed internally on what counts as a claim versus a fact, buying will not fix that, it will just produce more content that gets stuck in the same review queue. I tell founders: fix the definition of a qualified lead and the compliance review path first, then decide who executes. Skipping that order is the single most common way I have seen healthcare teams burn a two-quarter SEO budget and end up with a nicer blog and the same sales pipeline.</p><h3>The honest numbers nobody puts in a deck</h3><p>I am not going to hand you a stat about average conversion lift, because anyone quoting one to you for a regulated healthcare product without knowing your buyer, your review process, and your sales motion is making it up. What I can tell you is the shape of the trade-off: content velocity goes down when compliance review is real, qualified lead volume goes up when you price vendors on pipeline instead of rankings, and deliverability quietly caps your nurture funnel long before your landing pages do. Those three levers matter more than which agency has the flashiest case study.</p><p>I put more of this reasoning, with the questions I actually ask vendors on a discovery call, at my longer piece on seo lead generation services for healthcare software teams. If you are choosing between building this in-house or buying it, read that before your first vendor call, not after your third disappointing quarterly report.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=f62fc6f75194" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Workflow Automation Tools Look Identical Until Someone Asks for an Audit Trail]]></title>
            <link>https://alex-on-ai.medium.com/workflow-automation-tools-look-identical-until-someone-asks-for-an-audit-trail-6dc602342566?source=rss-93cf9798cdda------2</link>
            <guid isPermaLink="false">https://medium.com/p/6dc602342566</guid>
            <category><![CDATA[software-development]]></category>
            <category><![CDATA[telemedicine]]></category>
            <category><![CDATA[automation]]></category>
            <category><![CDATA[healthcare]]></category>
            <dc:creator><![CDATA[Alex Pavlov]]></dc:creator>
            <pubDate>Sun, 12 Jul 2026 07:06:35 GMT</pubDate>
            <atom:updated>2026-07-12T07:06:35.351Z</atom:updated>
            <content:encoded><![CDATA[<p>Every workflow automation tool demo looks the same. A form gets submitted, a webhook fires, a Slack message lands in the right channel, and everyone in the room nods like they just watched a card trick. Then someone in procurement asks what happens when the insurance eligibility check times out mid-transaction, and the room gets quiet in a way that tells you a lot about how the tool was actually tested.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*VO1KWm2Y3Rb9Yh-a.png" /></figure><p>I have sat in that quiet room more than once. The pattern is always the same: a healthcare client has outgrown their spreadsheets, someone on the team found n8n or Zapier over a weekend, and the pitch to leadership is that automation will save the intake team a few hours a week. It usually does, for the cases that go the way the diagram says they go. Healthcare workflows do not go the way the diagram says they go. If you want a longer version of how I think about the <a href="https://highcraft.io/workflow-automation/?utm_source=medium&amp;utm_medium=article&amp;utm_campaign=workflow-automation">workflow automation</a> build-versus-buy call, I wrote one down, but the short version is this: the keyword everyone searches is workflow automation tools, and the question nobody searches is what happens when the tool is wrong.</p><p>A telehealth platform we built ran into this in its first month. A prescription refill request looked like a simple state change: requested, reviewed, approved, sent to pharmacy. Zapier could move that ticket in its sleep. Then a refill request arrived for a controlled substance in a state with different e-prescribing rules than the prescriber’s home state, and the “simple” workflow needed a human, a licensing check, and an audit entry that could survive a state board inquiry two years later. That branch was not an edge case we forgot to diagram. It was the actual job.</p><h3>The demo works because nobody demos an insurance denial</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*A05CZRQsSLcoBdOu.png" /></figure><p>Off-the-shelf automation platforms are built to sell a demo, and the demo is always the happy path. Form submitted, webhook fires, record updated, done in four seconds flat. Nobody in a sales call shows you a claim getting denied for a missing modifier code, a patient consent that expired between visits, or a lab result that arrives out of order because the reference lab batches overnight. Those are not rare events in healthcare operations. They are Tuesday.</p><p>The honest failure mode I have watched teams walk into is treating a workflow automation tool as if it were the process, instead of a thin layer sitting on top of one. Zapier and n8n are genuinely good at connecting two systems that already agree on what a “customer” or a “lead” is. HubSpot workflows are good at nurturing a contact through a funnel where every branch ends in either a sale or a polite unsubscribe. None of that maps cleanly onto a clinical intake process where the same human might be a patient, a guarantor, and a caregiver for a dependent, all in the same visit, with different consent rules attached to each role.</p><h3>A generic tool assumes a straight line</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*Odbh-itg_Z6JK_07.png" /></figure><p>Most commercial automation platforms model a workflow as a sequence of triggers and actions. That works fine until your process needs a real state machine, with a task that can be paused, escalated, reassigned, or rolled back without losing its history. Clinical prior authorization is a state machine. So is credentialing. So is a referral that bounces between two health systems that do not share a patient identifier.</p><p>This is exactly the kind of complexity people describe with <a href="https://www.bpmn.org/">BPMN</a>, and there is a reason BPMN diagrams for healthcare processes tend to look like subway maps instead of the tidy left-to-right arrows in a Zapier tutorial. Gateways, parallel branches, boundary events for timeouts, compensation flows for when a step needs to be undone. A generic automation tool can fake some of this with enough conditional logic bolted on, the way duct tape can fake a bumper. It holds until the first pothole.</p><p>The honest trade-off is speed against shape. A no-code tool gets you a working automation in an afternoon. A custom-built workflow engine, hand-rolled around your actual state model, takes real weeks of integration work before it does anything visible. I tell founders to spend that time on the ten percent of the workflow that actually branches, and buy or borrow the rest.</p><h3>HL7v2 and FHIR do not automate anything by themselves</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*Wd1fS0aG-LX2QygQ.png" /></figure><p>A recurring misunderstanding I run into is the assumption that adopting FHIR R4 or getting an HL7v2 feed from Epic is the automation. It is not. Those are data formats and interface specs, not decision engines. I have seen teams spend a full quarter of engineering time standing up an HL7v2 listener, celebrate when the ADT messages start flowing, and then realize they still have to write every rule about what happens when a patient is discharged, readmitted, and discharged again inside six hours.</p><p>The interface work is necessary. It is also the smaller half of the problem. The harder half is encoding the actual clinical and administrative logic: who gets notified, what counts as a duplicate encounter, which fields are legally required before a claim can be submitted. A workflow automation tool that only knows how to move a message from queue A to queue B has done none of that thinking for you. It has just moved the paperwork faster to the desk of a human who still has to decide.</p><h3>Audit trails are not a feature you bolt on</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*REGX4uaYcjmImlh6.png" /></figure><p>Here is the detail that separates a real build decision from a weekend integration project. In a regulated healthcare workflow, every state transition needs a reason, a timestamp, and an actor, and that record needs to survive an audit that might happen years after the transition occurred. Most commercial automation tools log that a step ran. Far fewer log why it ran, what data it read to make that decision, and who could have overridden it.</p><p>Retrofitting that kind of audit trail onto a tool that was not built for it is one of the more expensive mistakes I have watched teams make. It usually surfaces during a security review or a HIPAA risk assessment, right when the team least wants to discover it. The fix is rarely a patch. It is closer to rebuilding the automation layer with logging as a first-class citizen instead of an afterthought, which is a nice way of saying: start over, but keep the parts that already worked.</p><h3>What actually decides the build call</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*jzery_llS_uwpw_e.png" /></figure><p>The honest way to make this decision is to separate the workflow into two piles before choosing any tool at all. Pile one: steps where the rule is stable, the exception rate is low, and getting it wrong costs a resend, not a compliance incident. Appointment reminders, intake form nudges, internal task routing. Buy the tool for that pile. Zapier, n8n, or a HubSpot workflow will do fine, and arguing for a custom build there is arguing for a more expensive way to send the same email.</p><p>Pile two: steps where the exception is the actual clinical or financial event, the kind of thing an auditor will ask about by name. Prior authorization, controlled substance handling, anything touching consent across a family account. That pile needs a system that treats the exception as a normal path through the state machine, not a try-catch block wrapped around someone else’s automation. I have never seen that pile built cheaply, and I have seen it built badly plenty of times by teams who assumed the same tool that handled pile one could stretch to cover pile two.</p><p>The part that saves a team real money is not skipping the custom build. It is figuring out early, before a quarter of engineering time gets spent, which pile a given workflow actually belongs to. Most teams guess wrong in the direction of “this looks simple,” right up until the first exception lands on someone’s desk with a patient’s name attached to it.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=6dc602342566" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Robotic Process Automation in Healthcare: Why the Bot Breaks at the First Payer Edge Case]]></title>
            <link>https://alex-on-ai.medium.com/robotic-process-automation-in-healthcare-why-the-bot-breaks-at-the-first-payer-edge-case-2c0b3e559384?source=rss-93cf9798cdda------2</link>
            <guid isPermaLink="false">https://medium.com/p/2c0b3e559384</guid>
            <category><![CDATA[telemedicine]]></category>
            <category><![CDATA[hipaa]]></category>
            <category><![CDATA[healthcare]]></category>
            <category><![CDATA[automation]]></category>
            <category><![CDATA[software-development]]></category>
            <dc:creator><![CDATA[Alex Pavlov]]></dc:creator>
            <pubDate>Fri, 10 Jul 2026 08:05:44 GMT</pubDate>
            <atom:updated>2026-07-10T08:05:44.213Z</atom:updated>
            <content:encoded><![CDATA[<p>A technical founder emails me with a familiar shape of problem. Intake staff retype the same demographics from a fax into three systems. Claims sit in a queue because someone has to eyeball an EOB before it posts. Someone asks about robotic process automation in healthcare like it’s a single button, and I have to explain that the button has opinions about HL7v2 and none of them are friendly.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*elLZKFZlw7kuzbmn.png" /></figure><p>Robotic process automation in healthcare is not new. It is screen scraping and rules engines wearing a nicer jacket. The pitch is real: fewer humans retyping fields, faster claims, less burnout on the ops team. I have watched it work and I have watched it quietly rot a workflow from the inside, and the difference usually comes down to how the team scoped the exceptions before they wrote a line of code. If you want the fuller version of how I think about that scoping, I wrote it up on the <a href="https://highcraft.io/healthcare-automation/?utm_source=medium&amp;utm_medium=article&amp;utm_campaign=healthcare-automation">healthcare automation</a> page, but the short version is here, and it is less flattering to the tooling than most vendor decks.</p><p>Here is the decision a buyer actually faces. Do you automate the happy path and staff around the rest, or do you build enough intelligence into the bot that it can handle the messy 20 percent without paging a human at 2 a.m.? Most teams answer this by accident, not on purpose, and they find out which answer they gave about six months in.</p><h3>The bot works until the claim form changes</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*vDyCQz8mKOepnJb3.png" /></figure><p>RPA tools are pattern matchers with a keyboard. UiPath, Power Automate, a custom script pointed at a browser, it does not matter which one you pick. They all assume the screen they were trained on stays the screen they see in production. A payer portal redesigns its form, a field moves three pixels to the left, and the bot either fails loudly or, worse, fails quietly and submits garbage. I have seen a claims bot post the wrong procedure code for two weeks before anyone noticed the denial rate had crept up. Nobody was watching the bot. Everyone was watching the dashboard the bot was supposed to update.</p><p>This is not an argument against RPA. It is an argument for treating every automated screen interaction as a dependency you now own, the same way you own a third-party API. If the payer changes their UI without telling you, and payers do this constantly, your bot needs a canary check before it needs a new feature.</p><h3>HL7v2 feeds do not forgive shortcuts</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*tYahfX-_9GhuD16S.png" /></figure><p>If your data is already moving through HL7v2 or FHIR R4, the automation conversation changes. You are no longer scraping a screen, you are parsing a message, and that message has a spec. This is better in almost every way except one: the exceptions get subtler. A malformed HL7v2 segment does not always throw an error, it sometimes just drops a field silently, and your bot inherits that silence. I’ve spent three weeks of integration work on a project where the actual bug was a missing OBX segment that only showed up for one lab vendor’s odd interpretation of the standard.</p><p>FHIR R4 helps here because the schema validation is stricter and the tooling around it is better, but only if your vendor actually implements it consistently. Epic’s FHIR endpoints are solid. Plenty of smaller EHR vendors ship something they call FHIR R4 that is really HL7v2 with a JSON wrapper and a a marketing team. Test against the real endpoint before you commit a build plan to it.</p><h3>The build decision is a staffing decision in disguise</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*WcxG3kkKCmhAg57u.png" /></figure><p>Every RPA proposal I’ve reviewed skips the maintenance line item. The build is the easy part. Someone has to own the bot’s exception queue, the credential rotations, the UI-drift monitoring, forever, and that someone needs healthcare domain knowledge, not just Zapier or n8n fluency. I like n8n for internal ops glue precisely because it’s transparent about its own limits: you can see the workflow, you can see where it breaks, and you are not paying a per-task fee to a vendor who benefits from your bot staying complicated. HubSpot-adjacent automations for patient outreach are fine for reminders and low-stakes messaging. They are not fine for anything touching PHI without a real conversation about the <a href="https://www.hhs.gov/hipaa/for-professionals/security/index.html">HIPAA Security Rule</a> and how your BAA actually covers the tool.</p><p>A quarter of a team’s time, ongoing, for a mid-sized automation footprint, is a reasonable planning number. If that sounds like a lot for a bot that was supposed to save headcount, that is the actual math, not the vendor’s math.</p><h3>Compliance is not a checkbox on a slide</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*iJNLA7FM9M4kRQOn.png" /></figure><p>A bot that reads PHI off a screen and writes it into a database is a bot that needs an audit trail, encryption in transit and at rest, and someone who can explain its access pattern to an auditor without opening the source code live in the meeting. I’ve sat through that meeting. It’s not fun for anyone, and it’s especially not fun when the answer to “who approved this credential” is a shrug.</p><p>The fix is boring: log every action the bot takes as if a human did it, because eventually you will need to explain a human-shaped decision that a script actually made. Store the logs somewhere immutable. Rotate the service account credentials on the same schedule you’d rotate a human employee’s, which is to say, more often than anyone wants to.</p><h3>Where RPA earns its keep</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*4MQJQBleT6DkZOhz.png" /></figure><p>The honest cases where I’d greenlight this without much hesitation: eligibility checks against a stable payer API, appointment reminder sequencing, prior authorization status polling, document intake for structured PDFs. These are high-volume, low-ambiguity, and the failure mode is a retry, not a wrong diagnosis code sitting in a claim for a month.</p><p>Where I’d push back hard: anything touching clinical decision support, anything where the bot is the last check before a claim submits with no human review, anything where the vendor can’t tell you what happens when their scraping target changes its layout. A telehealth platform I worked on automated intake forms but kept a human review step before anything hit the clinical record, and that one seam saved more incidents than the automation itself created.</p><h3>The failure mode nobody schedules a retro for</h3><p>The incidents that actually hurt are not the dramatic outages. They are the slow, silent drift: a bot that’s technically running, technically producing output, and quietly wrong in a way that takes a denial-rate spike or an angry patient call to surface. Build your monitoring around output correctness, not just uptime. A bot that’s up and wrong is worse than a bot that’s down, because down gets paged and wrong gets ignored until someone finally reads the fax by hand and asks why the numbers don’t match.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=2c0b3e559384" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[We Almost Lost a BAA to AI Marketing Tools, and I’d Still Use Them Again]]></title>
            <link>https://alex-on-ai.medium.com/we-almost-lost-a-baa-to-ai-marketing-tools-and-id-still-use-them-again-62064445b891?source=rss-93cf9798cdda------2</link>
            <guid isPermaLink="false">https://medium.com/p/62064445b891</guid>
            <category><![CDATA[ai-marketing]]></category>
            <category><![CDATA[ai-automation]]></category>
            <category><![CDATA[ai-agent]]></category>
            <category><![CDATA[lead-generation]]></category>
            <dc:creator><![CDATA[Alex Pavlov]]></dc:creator>
            <pubDate>Wed, 08 Jul 2026 11:28:55 GMT</pubDate>
            <atom:updated>2026-07-08T11:28:55.319Z</atom:updated>
            <content:encoded><![CDATA[<p>A founder I was advising last year asked a simple question. Can we just wire our patient onboarding emails through a normal marketing platform instead of building our own notification service. Fair question. Every roadmap has a line item that says “automate patient comms” and every vendor demo makes it look like an afternoon of work. I looked into a good breakdown of where <a href="https://highcraft.io/blog/ai-marketing-tools/?utm_source=medium&amp;utm_medium=article&amp;utm_campaign=blog-ai-marketing-tools">AI-driven marketing tools actually fit in a product stack</a> before answering him, and the short version is: marketing automation tools are excellent at the thing they were built for, and healthcare keeps asking them to do something else.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*Dwvj4EPecR4h0wLu.png" /></figure><p>The honest answer is that it works, right up until a support rep builds an email segment using a diagnosis field because it was sitting right there in the contact record. Nobody flags it. Nobody meant harm. It just happened because the tool made it easy, and easy is the whole business model.</p><h3>The first drip campaign taught me where the line sits</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*7-0zhadN076IjmXD.png" /></figure><p>Picture a generic US digital-health platform running a welcome sequence for new patients. Day one, account confirmation. Day three, how to book your first visit. Day seven, a nudge if you haven’t scheduled. Standard stuff, the kind any marketing automation tool ships out of the box.</p><p>The trouble started when the product team wanted the day-seven email to reference the specific service line the patient signed up for, because a generic nudge converts worse than a specific one. Reasonable growth instinct. Except now the marketing platform’s contact record holds a condition category, and that record lives in a system whose audit logging was designed to prove someone opened an email, not to prove who accessed a health attribute and when. The vendor’s base tier didn’t cover a signed BAA either. The enterprise tier did, for a price that made the CFO ask why we were paying that much to send an email.</p><h3>Marketing automation tools were built for a different risk model</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*jJMwVk6GQFRkxv9e.png" /></figure><p>These platforms are optimized for lead scoring, attribution, and campaign velocity. Great tools for that job. Nine years ago I’d have said use whichever one your growth team already knows. I still mostly believe that, for the top of the funnel.</p><p>What changes in healthcare is the risk model underneath. A regulated product needs to prove minimum necessary access, keep a defensible audit trail on anything touching a health attribute, and show data residency and incident response terms that match a signed BAA, not a marketing addendum. Most automation platforms were never built to prove any of that, because their buyer never asked. Their buyer wanted open rates.</p><h3>The build versus buy decision is really a boundary decision</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*TP1xe4L2D_G5fozt.png" /></figure><p>Nobody actually needs to choose between “buy a marketing automation tool” and “build the whole comms stack from scratch.” That framing wastes a quarter of planning. The real decision is where the boundary sits between content that is fine inside a commercial platform and anything keyed to a patient identifier or a clinical fact.</p><p>Generic newsletters, product announcements, non-clinical onboarding tips: fine in the platform your marketing team already owns. Appointment reminders that reference a visit type, refill nudges tied to a prescription, anything with a diagnosis or a lab result anywhere near it: that belongs behind your own service, one that only ever hands the marketing tool an opaque identifier and a template slot, never the fact itself.</p><p>We shipped exactly that pattern for a telehealth client. A small internal notification service owned every trigger tied to a clinical event. It called out to the commercial email and SMS platform only with a template ID and a first name. The marketing platform never saw a diagnosis, a visit reason, or a lab value in its life. It just sent pretty emails, which is what it’s good at.</p><h3>Diagramming the workflow exposes the trade-off before the code does</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*-smD-uPByeWENKWv.png" /></figure><p>Before any of that gets built, I want it on a whiteboard, or better, in a proper process diagram using <a href="https://www.bpmn.org/">BPMN</a>. Not because compliance officers like diagrams, though they do, but because a BPMN swim lane makes the PHI boundary visible in a way a Slack thread never will. You can see, right there in the diagram, the exact moment a patient record crosses from your service into a third-party tool. If that crossing point makes you wince when you draw it, it’ll make an auditor wince more when he finds it in production. Draw it early. It’s cheaper on paper than in an incident report.</p><h3>Three failure modes I keep seeing</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*X0FL_KSJYNSlen9h.png" /></figure><p>The vendor’s marketing BAA covers the platform’s core tier, not the automation or CDP add-on the team quietly turned on six months later. Nobody re-checks the paper trail when a feature flag flips.</p><p>A segmentation field meant for “users who clicked pricing” ends up populated with a clinical tag, because the field existed and building it felt faster than filing a ticket with engineering.</p><p>An integration syncs the full contact object instead of a filtered subset, so PHI rides along with every list export whether anyone asked for it or not. This one is boring and it is also the one I’ve seen most often. Integrations default to “sync everything,” and everything usually includes the field you didn’t want anywhere near a marketing dashboard.</p><h3>The trade-off buyers actually face</h3><p>A commercial marketing automation tool is faster to stand up, cheaper per month at low volume, and comes with a marketing team who already knows the interface. A thin custom notification service costs real engineering time up front and needs its own on-call rotation. Nobody wants to hear that a Tuesday reminder email needs an on-call rotation, but if it’s the one sending refill nudges, it does.</p><p>The deciding factor isn’t company size or funding stage. It’s whether the content itself, not just the recipient, ever needs to reference something clinical. If your comms are pure marketing content to a list of leads, buy the tool and don’t overthink it. If any message is triggered by a clinical event or contains a clinical fact, keep that trigger and that content inside your own boundary, and let the commercial tool handle delivery mechanics only.</p><h3>What I’d tell a technical founder evaluating this today</h3><p>Start by listing every automated message your product sends, then mark which ones are triggered by a health event. That list alone tells you most of what you need to know. Ask your vendor for the specific BAA language covering the automation or CDP module, not just the base platform, because those are often sold and insured separately. Model the actual data flow before you sign anything, in BPMN or in whatever notation your team already reads, and make someone from compliance sign off on the diagram, not just the contract.</p><p>I still send people to that breakdown of AI marketing tooling when they’re picking a platform, because the tool landscape moves fast and most of the differentiation now is in how well a vendor handles exactly this boundary question. Some do it well. Most still assume their buyer runs an e-commerce store.</p><p>Honestly, the fix is rarely dramatic. Keep the clinical trigger logic in a system you control, let the polished marketing platform handle the parts it was actually built for, and check that boundary every time someone adds a new field to a form. The alternative is explaining to a very serious person in a very serious meeting why a diagnosis code ended up in a Tuesday newsletter segment. I’d rather draw the diagram.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=62064445b891" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[How To Become A Badass .NET Developer. Part 1]]></title>
            <link>https://alex-on-ai.medium.com/how-to-become-a-badass-net-developer-part-1-3ecec0f2ddf9?source=rss-93cf9798cdda------2</link>
            <guid isPermaLink="false">https://medium.com/p/3ecec0f2ddf9</guid>
            <category><![CDATA[software-development]]></category>
            <category><![CDATA[programming]]></category>
            <category><![CDATA[software-engineering]]></category>
            <category><![CDATA[microsoft]]></category>
            <category><![CDATA[coding]]></category>
            <dc:creator><![CDATA[Alex Pavlov]]></dc:creator>
            <pubDate>Tue, 21 Mar 2023 00:00:04 GMT</pubDate>
            <atom:updated>2023-03-21T00:02:14.282Z</atom:updated>
            <content:encoded><![CDATA[<h3>How To Become A Badass .NET/C# Developer</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1000/1*Dqf4hc9yuxjjZTjd-_L2dg.jpeg" /></figure><blockquote>Raise your quality standards as high as you can live with, avoid wasting your time on routine problems, and always try to work as closely as possible at the boundary of your abilities. Do this, because it is the only way of discovering how that boundary should be moved forward.<br>Edsger Dijkstra</blockquote><p>This article will describe a roadmap to becoming a badass .NET Software Engineer. I will start with the C# language itself and walk you through the frameworks and technologies you need to master on your journey.</p><p>This roadmap is pretty subjective as I describe recommendations primarily based on my own experience. However, I recommend only hand-picked topics and materials.</p><h3>Step 1. Learn C#</h3><p>Find one good book and stick to it. In parallel, you can watch Youtube and video courses on platforms such as Pluralsight and Udemy, but only as an addition to reading the book.</p><p>I recommend Andrew Troelsen’s book <strong><em>Pro C# 10 with .NET 6: Foundational Principles and Practices in Programming</em></strong>. I learned C# from the old edition of this book 12 years ago and new editions still have excellent ratings on Amazon. Feel free to do your own research and find a good book for yourself.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/200/1*tdMHAmPlFp1MRIbiVkubKA.jpeg" /><figcaption><strong><em>Pro C# 10 with .NET 6: Foundational Principles and Practices in Programming</em></strong></figcaption></figure><p>I think that reading a book is still the best way to learn a new programming language if you are a beginner. Nowadays we have myriads of good resources on the Internet, but you need to stay focused and not be distracted by every new shiny course or video. Do not get stuck in the tutorial hell when you keep switching from one tutorial to another looking for the best one without learning or coding anything. The book gives structure to your learning and a feeling of progress with every page.</p><p>Reading doesn’t mean learning. Learning doesn’t mean memorizing. The best way to learn and memorize is to practice what you read. Write simple programs, and play with the language, debugger, the IDE. Write your own console utility that lists all the running processes or removes all the files from the folder just for the sake of it. Create a simple WinForms app that keeps track of your weight. Write your own web scraper. Code something useful or useless, just practice.</p><h4>Here is a list of good resources for learning C#:</h4><ol><li><a href="https://www.amazon.com/C-Players-Guide-5th/dp/0985580151">The C# Player’s Guide</a></li><li>Microsoft’s <a href="https://learn.microsoft.com/en-us/dotnet/csharp/tour-of-csharp/tutorials/">Introduction to C#</a></li><li>Microsoft’s <a href="https://learn.microsoft.com/en-us/dotnet/csharp/">C# Documentation</a></li><li><a href="https://dotnet.microsoft.com/en-us/learn/csharp">Popular learning materials from Microsoft</a></li></ol><h3>Step 2. Learn ASP.NET framework</h3><p>Now it is time to learn the real stuff.</p><h4>ASP.NET Core</h4><p><strong>ASP.NET Core </strong>is a framework for creating web applications.</p><p>There are two parts to it:</p><ol><li>ASP.NET Core Web API is used for creating the back-end for the applications.</li><li>ASP.NET Core Blazor is a framework for building client-side web applications.</li></ol><p>I recommend concentrating only on the first item — <strong>ASP.NET Core Web API.</strong> It is popular in the industry and you will see it used in production. As for <strong>ASP.NET Core Blazor</strong>, I do not recommend spending too much time on it as it’s not widely adopted and other front-end frameworks such as React or Angular are used in real-world applications.</p><p>There are more technologies in <strong>ASP.NET</strong>, but I will omit them because they are older or not widely adopted.</p><p>I recommend starting with the book <strong><em>Pro ASP.NET Core 6: Develop Cloud-Ready Web Applications Using MVC, Blazor, and Razor Pages </em></strong>by<strong> </strong>Adam Freeman to get a good overview of the platform.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/200/1*LWjtrOLWv1l80SIXYEAUgg.jpeg" /><figcaption><strong><em>Pro ASP.NET Core 6: Develop Cloud-Ready Web Applications Using MVC, Blazor, and Razor Pages</em></strong></figcaption></figure><p>Additionally, it’s worth learning <a href="https://swagger.io/tools/swagger-ui/">Swagger UI</a> and <a href="https://www.openapis.org/">OpenAPI</a> for documenting your APIs and <a href="https://www.postman.com/">Postman</a> for testing them and trying out 3rd party APIs.</p><p><strong>Other good resources hand-picked by me:</strong></p><ol><li><a href="https://code-maze.com/ultimate-aspnetcore-webapi-second-edition/">ULTIMATE ASP.NET CORE WEB API</a></li><li>Microsoft’s <a href="https://learn.microsoft.com/en-us/aspnet/core/tutorials/first-web-api?view=aspnetcore-6.0&amp;tabs=visual-studio">Tutorial: Create a web API with ASP.NET Core</a></li></ol><h3>Step 3. Learn SQL</h3><p>Every software program works with data. The most popular and powerful language to manipulate data is SQL. It’s used in almost every project and you will get questions about it in every job interview.</p><p>As a software engineer, you don’t need to know all the ins and outs of SQL and SQL Server internals, but you should at least know the basics.</p><p><strong>Make sure you know the following topics as they often are asked in the interviews:</strong></p><ol><li><a href="https://www.w3schools.com/sql/sql_join.asp">SQL Joins</a></li><li><a href="https://learn.microsoft.com/en-us/sql/t-sql/data-types/data-types-transact-sql?view=sql-server-ver16">SQL Data Types</a> (e.g. nvarchar vs varchar)</li><li><a href="https://learn.microsoft.com/en-us/sql/relational-databases/indexes/clustered-and-nonclustered-indexes-described?view=sql-server-ver16">Clustered and nonclustered indexes</a></li><li><a href="https://blog.devart.com/how-to-optimize-sql-query.html">SQL Optimization</a></li></ol><p><strong>I recommend picking a book and sticking to it. You can find the best books in the list below:</strong></p><ol><li>SQL Cookbook by Anthony Molinaro</li><li>SQL in 10 Minutes, Sams Teach Yourself by Ben Forta</li><li>SQL Practice Problems by Sylvia Moestl Vasilik</li><li>SQL Antipatterns by Bill Karwin</li><li>Learning SQL by Alan Beaulieu</li><li>SQL Pocket Guide by Jonathan Gennick</li><li>The Art of SQL by Stephane Faroult</li><li>SQL Performance Explained Everything Developers Need to Know about</li><li>SQL Performance by Markus Winand</li></ol><h3>Step 4. Learn ADO.NET framework and Entity Framework</h3><p>ADO.NET allows you to use SQL in .NET to access a database. I recommend learning to use <a href="https://github.com/DapperLib/Dapper">Dapper</a> as an object mapper instead of raw ADO.NET.</p><p><strong>Entity Framework</strong> is an object-database mapper that allows you to access a database without writing SQL and work with objects instead of relational tables. It is widely used, but you should know SQL anyway to understand how to optimize the queries and troubleshoot performance issues.</p><p><strong>Recommended resources on ADO.NET:</strong></p><ol><li><a href="https://learn.microsoft.com/en-us/dotnet/framework/data/adonet/ado-net-overview">Microsoft’s docs</a></li></ol><p><strong>Recommended resources on Entity Framework:</strong></p><ol><li><a href="https://learn.microsoft.com/en-us/ef/core/">Microsoft’s docs</a></li><li><a href="https://www.manning.com/books/entity-framework-core-in-action-second-edition">Entity Framework Core in Action</a></li><li><a href="https://www.pluralsight.com/authors/julie-lerman?utm_source=google&amp;utm_medium=paid-search&amp;utm_campaign=upskilling-and-reskilling&amp;utm_term=ssi-emea-xyz-dynamic&amp;utm_content=free-trial&amp;gclid=CjwKCAjw_MqgBhAGEiwAnYOAenDmbQyPl95Hhtmtk6olxpxlx695pTxBSigkX3IcYmvplRFfwEG29BoCQ5EQAvD_BwE">Julie Lerman — Entity Framework Author | Pluralsight</a></li></ol><h3>Step 5. Deepen your knowledge of .NET and C#</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/612/1*LMvpv3e75qZ1Jo95wCKvSw.jpeg" /></figure><p>To become an absolute beast in .NET you should extend your knowledge even further than I described before. I recommend two books that I always have on my bookshelf and review before every interview.</p><p>The first one is <strong>CLR via C# by Jeffrey Richter</strong>.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/200/1*OFFZ_qLMcL5VbvDfDMx-6Q.jpeg" /><figcaption>CLR via C# by Jeffrey Richter</figcaption></figure><p>This book is probably known to every .NET developer. It describes the internals of the framework, the mechanisms behind the asynchronous programming in .NET, how Garbage Collector works, and a lot more. It covers the topics that are often asked in interviews. It is a must-read if you want to be an expert.</p><p>The second book is <a href="https://www.amazon.com/C-9-0-Nutshell-Definitive-Reference/dp/1098100964"><strong>C# In A Nutshell by Joseph Albahari</strong></a><strong>.</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/200/1*PVUUbKofz8SgtRyl6UJddA.jpeg" /><figcaption><a href="https://www.amazon.com/C-9-0-Nutshell-Definitive-Reference/dp/1098100964"><strong>C# In A Nutshell by Joseph Albahari</strong></a></figcaption></figure><p>It has a broad coverage of a wide range of topics. It is one of the best reference books on .NET you can have on your shelf. I like this book because it has one of the best materials on asynchronous programming, concurrency, parallelism, and LINQ. By the way, the popular interactive .NET scratchpad LINQPad is created by the book’s author.</p><h3>Step 6. Learn Design Patterns</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*LzKppsHIMabQhnCCdpdQXg.jpeg" /></figure><p>Design patterns are reusable solutions to common reoccurring problems in software design.</p><p>I recommend two books. The first one is <a href="https://www.amazon.com/Head-First-Design-Patterns-Brain-Friendly/dp/0596007124">Head First Design Patterns by Eric Freeman</a>. It is best for beginners because it is written in simple language with good examples.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/200/1*uh_SOpXTG3juRVUX8wxsDA.png" /><figcaption>Head First Design Patterns</figcaption></figure><p>The second one is Read Design Patterns: Elements of Reusable Object-Oriented Software. It’s more hardcore than the previous one and better suited for seasoned developers. Some claim that it is too abstract and complex, but actually, it is a pretty concise and thin book with clear pattern descriptions and examples. The only downside is that examples are written in C++, but you can skip them because the descriptions and diagrams give a very good overview of the patterns.</p><h3>Step 7. Learn Data Structures and Algorithms</h3><p>Data structures and algorithms are something that you may not encounter in your everyday job, but they are crucial for passing the interviews in companies like FANG. Anyways, knowledge of this subject makes you a better developer capable of creating better software and solving complex problems.</p><p>If you are a novice, I certainly recommend the book <a href="https://www.amazon.com/Grokking-Algorithms-illustrated-programmers-curious/dp/1617292230">Grokking Algorithms</a>. It’s a friendly guide to algorithms with great illustrated examples.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/200/1*JTgBqdKnRDzxoaxwyWCTBg.jpeg" /><figcaption><a href="https://www.amazon.com/Grokking-Algorithms-illustrated-programmers-curious/dp/1617292230">Grokking Algorithms</a></figcaption></figure><p>For seasoned programmers, I recommend the book <a href="https://www.amazon.com/Algorithms-4th-Robert-Sedgewick/dp/032157351X">Algorithms by Robert Sedgewick</a>.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/201/1*hL9yzVdvY9Ej0N6zsKyI3Q.jpeg" /><figcaption>Algorithms by Robert Sedgewick</figcaption></figure><p>Additionally, the author has a great course on Coursera with the same name. Examples and assignments are in Java, but it should not be a problem as Java and C# are pretty similar.</p><h3>Step 8. Learn Software Architecture</h3><p>Learn the following concepts:</p><ol><li>Three-layer architecture</li><li>Microservice architecture</li><li>Message brokers</li><li>Domain Driven Design</li><li><a href="https://learn.microsoft.com/en-us/azure/architecture/patterns/cqrs">CQRS</a></li><li>API Gateways</li><li>CI/CD</li><li>Clean Architecture</li><li>RESTful</li></ol><p>Resources:</p><ol><li><a href="https://learn.microsoft.com/en-us/dotnet/architecture/microservices/">.NET Microservices: Architecture for Containerized .NET Applications</a></li><li><a href="https://learn.microsoft.com/en-us/dotnet/architecture/modern-web-apps-azure/">Architect Modern Web Applications with ASP.NET Core and Azure</a></li><li><a href="https://learn.microsoft.com/en-us/dotnet/architecture/dapr-for-net-developers/">Dapr for .NET Developers</a></li></ol><h3>Step 9. Learn Microsoft Azure Cloud</h3><p>Cloud technologies are omnipresent. If you stick to the Microsoft tech stack, I recommend learning Azure Cloud.</p><p>Make sure you are familiar with the following technologies:</p><ol><li><a href="https://learn.microsoft.com/en-us/azure/azure-functions/functions-overview">Azure Functions</a></li><li><a href="https://learn.microsoft.com/en-us/azure/service-bus-messaging/service-bus-messaging-overview">Azure ServiceBus</a></li><li><a href="https://learn.microsoft.com/en-us/azure/key-vault/general/basic-concepts">Azure Key Vault</a></li><li><a href="https://learn.microsoft.com/en-us/azure/azure-monitor/app/app-insights-overview?tabs=net">Application Insights</a></li><li><a href="https://docs.docker.com/get-started/">Docker</a></li><li><a href="https://kubernetes.io/docs/concepts/overview/">Kubernetes</a></li><li>Horizontal scaling</li><li>Virtual machines</li><li><a href="https://learn.microsoft.com/en-us/azure/api-management/api-management-key-concepts">Azure API Management</a></li></ol><p><strong>Resources:</strong></p><ol><li><a href="https://learn.microsoft.com/en-us/dotnet/architecture/cloud-native/">Architecting Cloud Native .NET Applications for Azure</a></li></ol><h3>Step 10. Learn NoSQL Databases</h3><p>I recommend learning some base concepts of NoSQL databases such as the <a href="https://en.wikipedia.org/wiki/CAP_theorem">CAP theorem</a>, clustering, sharding, and eventual consistency.</p><p>I recommend getting acquainted with the following DBs:</p><ol><li><a href="https://www.mongodb.com/docs/manual/tutorial/getting-started/">MongoDB</a> — the most popular NoSQL document-oriented database.</li><li><a href="https://learn.microsoft.com/en-US/azure/cosmos-db/introduction">CosmosDB</a> — highly scalable NoSQL database from Microsoft.</li><li>Redis — key-value in memory storage used as cache. To be frank, now it is not only in memory and not only key-value :)</li></ol><p>As for books, I highly recommend <a href="https://www.amazon.com/NoSQL-Distilled-Emerging-Polyglot-Persistence/dp/0321826620">NoSQL Distilled: A Brief Guide to the Emerging World of Polyglot Persistence</a>. So far, it is the best book on NoSQL I read.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/196/1*PS6VgSQyqlIzYiii542nbg.jpeg" /><figcaption>NoSQL Distilled</figcaption></figure><h3>Step 11. 3rd party .NET libraries that are de facto standards in the industry</h3><p>There are a number of great .NET libraries that became a standard in lots of projects and definitely worth checking out.</p><p>I came up with a list of hand-picked ones.NET libraries that I find the most useful and used extensively over the years.</p><ol><li><a href="https://docs.fluentvalidation.net/">FluentValidation </a>— library for building strongly-typed validation rules. It is the most useful when you have complex validation logic for API requests.</li><li><a href="https://serilog.net/">Serilog</a> — best library for logging. It enables logging to a file, database, or console with just a few configuration lines.</li><li><a href="https://automapper.org/">Automapper</a> — library for mapping objects. With this library, you can get rid of repetitive code that maps your request models to DTOs or domain models.</li><li><a href="https://dapperlib.github.io/Dapper/">Dapper</a> — a simple object mapper for simplifying mappings from SQL query results to strongly typed models.</li><li><a href="https://github.com/App-vNext/Polly">Polly</a> — .NET resilience library mostly used to retry and rate limit some operations. Especially useful for high-load microservice architecture.</li><li><a href="https://github.com/jbogard/MediatR">Mediatr</a> — library for in-process messaging with support of request/response, commands, queries, notifications and events.</li><li><a href="https://playwright.dev/">Playwright</a> — modern library from Microsoft for end-to-end testing of web apps. Can be used for web scraping of client-rendered apps.</li><li><a href="https://www.quartz-scheduler.net/">Quartz.NET</a> — job scheduling system for .NET.</li><li><a href="https://spectreconsole.net/">Spectre.Console</a> — .NET library for creating beautiful console applications.</li><li>Moq — the most popular and friendly mocking library for .NET</li><li><a href="https://docs.hangfire.io/">Hangfire</a> — library for background job processing with embedded monitoring dashboard and persistent storage.</li></ol><h3>Step 12. Prepare for the interview</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/612/1*ZonQkV-0nXltrKrRQcBtsg.jpeg" /></figure><ol><li>Be ready to describe your last project or prior experience</li><li>Write down all questions after every interview and process them to fill the gaps in your knowledge</li><li>Be comfortable with writing code on the whiteboard and explaining your thought process</li></ol><p>One of the best books on interview preparation is <a href="https://www.amazon.com/Cracking-Coding-Interview-Programming-Questions/dp/0984782850">Cracking the Coding Interview by Gayle Laakmann McDowell.</a></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/188/1*MFbUvmhfeRaVAoYCAXbZ_w.jpeg" /><figcaption>Cracking the Coding Interview</figcaption></figure><p>Thanks for reading and good luck!</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=3ecec0f2ddf9" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>