<?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 Archit Mittal on Medium]]></title>
        <description><![CDATA[Stories by Archit Mittal on Medium]]></description>
        <link>https://medium.com/@automate-archit?source=rss-9b6215eea7d6------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/0*Fyg7Q340fzR62MfY.jpg</url>
            <title>Stories by Archit Mittal on Medium</title>
            <link>https://medium.com/@automate-archit?source=rss-9b6215eea7d6------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Sat, 21 Nov 2026 06:34:05 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@automate-archit/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[I Built an AI Agent That Reads 400 Invoices an Hour — Here’s the Python Code]]></title>
            <link>https://medium.com/@automate-archit/i-built-an-ai-agent-that-reads-400-invoices-an-hour-heres-the-python-code-45c753c37640?source=rss-9b6215eea7d6------2</link>
            <guid isPermaLink="false">https://medium.com/p/45c753c37640</guid>
            <category><![CDATA[artificial-intelligence]]></category>
            <dc:creator><![CDATA[Archit Mittal]]></dc:creator>
            <pubDate>Tue, 18 Aug 2026 03:45:23 GMT</pubDate>
            <atom:updated>2026-08-18T03:45:23.954Z</atom:updated>
            <content:encoded><![CDATA[<p>Last month I ran a test batch of 400 PDF invoices through an AI agent I built over a weekend. It finished in 58 minutes at roughly $0.004 per document. The manual process it replaced — open the PDF, find six fields, retype them into a spreadsheet — took about 11 hours for the same stack of paper.</p><p>That is the entire pitch for AI automation in one paragraph. Not magic. Just a 91% reduction in the time a human spends on work no human enjoys.</p><p>Here is how the agent works, why I chose Python automation over an off-the-shelf tool, and the code you can adapt for your own business automation.</p><h3>What Is an AI Agent, and How Is It Different From a Script?</h3><p>A script follows a fixed path. An AI agent is a loop: it receives a goal, decides which tool to call, reads the result, and decides again, until the goal is met or it gives up.</p><p>That difference matters enormously for workflow automation. Invoices are not uniform. Vendor A puts the total in the bottom right. Vendor B calls it “Amount Due.” Vendor C splits it across two lines and buries a credit note underneath. A regex-based script breaks on the third template. An agent reasons about the document instead of pattern-matching it.</p><h3>Why Not Just Buy an Off-the-Shelf Tool?</h3><p>I priced three commercial OCR platforms before writing a line of code. All three worked. None of them exported cleanly into the chart of accounts the finance team already used, and each wanted a paid connector to close that gap.</p><p>The rule I apply as an automation consultant is simple: buy when the process is generic, build when the process is yours. Reading a PDF is generic. Mapping line items to your specific ledger is not. The second half is where an AI agent earns its keep.</p><h3>How Does the Agent Actually Work?</h3><p>Four stages, and only one of them involves a language model.</p><p>First, a watcher polls a folder for new PDFs. Second, each PDF is handed to the model with a strict JSON schema and instructions to return nothing else. Third, a deterministic validator checks the arithmetic. Fourth, anything that fails validation is routed to a human review queue instead of being silently written to the ledger.</p><p>That fourth stage is the one most tutorials skip, and it is the reason this pipeline is safe to run unattended.</p><h3>What Does the Code Look Like?</h3><p>Here is the extraction core, trimmed to the parts that matter.</p><pre>import base64, json, pathlib<br>from anthropic import Anthropic<br><br>client = Anthropic()<br><br>SCHEMA = {<br>    &quot;invoice_number&quot;: &quot;str&quot;,<br>    &quot;vendor&quot;: &quot;str&quot;,<br>    &quot;invoice_date&quot;: &quot;YYYY-MM-DD&quot;,<br>    &quot;subtotal&quot;: &quot;float&quot;,<br>    &quot;tax&quot;: &quot;float&quot;,<br>    &quot;total&quot;: &quot;float&quot;,<br>}<br><br>def extract(pdf_path):<br>    data = base64.b64encode(pathlib.Path(pdf_path).read_bytes()).decode()<br>    msg = client.messages.create(<br>        model=&quot;claude-sonnet-5&quot;,<br>        max_tokens=1024,<br>        messages=[{&quot;role&quot;: &quot;user&quot;, &quot;content&quot;: [<br>            {&quot;type&quot;: &quot;document&quot;, &quot;source&quot;: {&quot;type&quot;: &quot;base64&quot;,<br>             &quot;media_type&quot;: &quot;application/pdf&quot;, &quot;data&quot;: data}},<br>            {&quot;type&quot;: &quot;text&quot;, &quot;text&quot;:<br>             f&quot;Extract these fields as JSON: {json.dumps(SCHEMA)}. &quot;<br>             &quot;Use null if a field is absent. Return JSON only.&quot;}<br>        ]}],<br>    )<br>    return json.loads(msg.content[0].text)</pre><p>Two details do most of the work here. Passing the PDF as a document block means no separate OCR step, so scanned and digital invoices go down the same path. Pinning the schema inside the prompt and parsing with json.loads means a malformed response raises an exception immediately rather than quietly writing garbage downstream.</p><p>I wrap the whole pipeline as an MCP server so it is callable from Claude Code. MCP, the Model Context Protocol, is an open standard that lets an AI model call your tools with a documented interface. In practice it means the finance lead can type “process yesterday’s invoices” instead of running a script, and I never have to build a UI.</p><h3>How Do You Stop an AI Agent From Inventing Numbers?</h3><p>You do not. You catch it. This is the single most important thing I have learned building AI agents for real businesses: never trust the model’s output as final, and always give it something deterministic to check against.</p><p>Invoices are unusually friendly in this respect, because they contain their own checksum. Subtotal plus tax should equal total. If it does not, something is wrong, and you know it without a human looking.</p><pre>def validate(record):<br>    problems = []<br>    total = record.get(&quot;total&quot;)<br><br>    if total is None:<br>        problems.append(&quot;missing total&quot;)<br>    else:<br>        computed = (record.get(&quot;subtotal&quot;) or 0) + (record.get(&quot;tax&quot;) or 0)<br>        if abs(computed - total) &gt; 0.02:<br>            problems.append(f&quot;arithmetic mismatch: {computed} vs {total}&quot;)<br><br>    if not record.get(&quot;invoice_number&quot;):<br>        problems.append(&quot;missing invoice number&quot;)<br><br>    return problems<br><br><br>def route(record):<br>    problems = validate(record)<br>    if problems:<br>        review_queue.put({&quot;record&quot;: record, &quot;problems&quot;: problems})<br>    else:<br>        ledger.write(record)<br></pre><p>On my 400-document test batch, 31 invoices tripped a validation rule. Twenty-two were genuine document problems: handwritten amendments, a missing tax line, one invoice that was actually a statement. Nine were extraction errors. That is a 97.75% clean-extraction rate, and more importantly, zero bad records reached the ledger unflagged.</p><h3>What Does It Actually Cost to Run?</h3><p>The full 400-invoice run cost about $1.60 in model calls. Build time was roughly 14 hours spread across a weekend, most of it spent on the boring parts: the review queue UI, retry logic, and a dead letter folder for PDFs that fail to parse at all.</p><p>Compare that with 11 hours of manual keying per batch. At a loaded cost of $35 an hour, the build paid for itself somewhere in the third batch. That is the arithmetic that makes business automation an easy conversation with a CFO, and it is roughly the same shape in every project I take on.</p><h3>Where Should You Start?</h3><p>Pick the process that is high volume, low judgment, and already has a checkable answer. Invoice entry qualifies. So do timesheet reconciliation, shipping manifest matching, and pulling structured data out of inbound RFP documents.</p><p>Then resist the urge to automate the whole thing on day one. Ship the extractor. Watch the review queue for two weeks. The pattern of failures will tell you exactly which rule to add next, and it is almost never the rule you would have guessed.</p><p>The best AI automation tools in 2026 are not the ones with the longest feature list. They are the ones you can point at your own weird process and have working by Friday.</p><p>I write build logs like this one every week, covering Python automation, AI agents, and what actually survives contact with a real business. Longer versions with the full repos live at architmittal.com/blog.</p><p>I’m Archit Mittal. I Automate Chaos.</p><p>Follow me on Medium | Visit architmittal.com</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=45c753c37640" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[No-Code vs Code Automation: Why I Stopped Recommending Zapier After 40 Client Builds]]></title>
            <link>https://medium.com/@automate-archit/no-code-vs-code-automation-why-i-stopped-recommending-zapier-after-40-client-builds-f09cdbadf53b?source=rss-9b6215eea7d6------2</link>
            <guid isPermaLink="false">https://medium.com/p/f09cdbadf53b</guid>
            <category><![CDATA[artificial-intelligence]]></category>
            <dc:creator><![CDATA[Archit Mittal]]></dc:creator>
            <pubDate>Mon, 17 Aug 2026 17:41:01 GMT</pubDate>
            <atom:updated>2026-08-17T17:41:01.324Z</atom:updated>
            <content:encoded><![CDATA[<p>Two years ago my default answer to “what should we automate this with?” was Zapier. It was fast, visual, and nobody had to learn Python. Clients loved watching the little boxes connect. Everyone was happy — for about four months.</p><p>I have since built or rebuilt business process automation for more than 40 small companies. Roughly three quarters of those projects started as somebody else’s no-code workflow, and my actual job was diagnosing why it stopped working. That pattern changed my mind, and what follows is the unpopular version of the no-code vs code automation argument.</p><p>To be clear: I still use no-code tools every week. I just stopped recommending them as the foundation of anything a business actually depends on.</p><p>Why do no-code automations break so often?</p><p>Three failure modes show up in almost every automation consulting audit I run.</p><p>Silent failure. A workflow dies, the platform emails a notification to whoever originally set it up — often an employee who left the company. Nobody notices for weeks. In one audit, a client’s lead-routing AI workflow had been dead for 71 days. They assumed the market had gone quiet. It had not. Roughly 340 inbound leads had gone nowhere.</p><p>Task-based pricing that punishes success. No-code platforms bill per task or per operation. That is cheap at 500 runs a month and brutal at 50,000. One e-commerce client was paying $487 a month for order-tagging logic that now runs as about 90 lines of Python on a $6 VPS.</p><p>No real conditional logic. The moment a process needs “if this vendor, apply rule A, unless it is a credit memo, in which case check the PO date,” you end up with a 14-step visual workflow that nobody can read — including the person who built it.</p><p>None of these are complaints about no-code as an idea. They are complaints about using it for load-bearing work.</p><p>How much does AI automation actually cost?</p><p>Here are real numbers from projects I ran over the last 18 months, for a comparable workload at a 20-person services firm.</p><p>No-code stack: $340 to $620 per month in platform fees, no development cost, about two weeks to launch.</p><p>Python and AI stack: $18 to $45 per month in hosting and API calls, a one-time build cost of $4,000 to $9,000, and three to five weeks to launch.</p><p>The crossover point usually lands somewhere between month 9 and month 14. If an automation is going to exist for two years, code wins on cost and it is not close. If you are still testing whether a process is worth automating at all, no-code wins, and that is not close either.</p><p>That is the entire decision, honestly. No-code is for validation. Code is for infrastructure. Most of the expensive messes I get hired to untangle come from confusing the two.</p><p>What changed in 2026: the middle ground got much bigger</p><p>Here is what makes the old advice stale. Writing Python automation scripts used to mean paying a developer for everything, including the genuinely hard part: reading messy documents that do not share a template. That part collapsed in price.</p><p>A representative example — extracting structured data from an invoice email body:</p><p>import os, json, anthropic</p><p>client = anthropic.Anthropic(api_key=os.environ[“ANTHROPIC_API_KEY”])</p><p>def extract_invoice(text):</p><p>prompt = “Return JSON only with keys vendor, invoice_number, total, due_date. Text: “ + text</p><p>msg = client.messages.create(model=”claude-sonnet-5&quot;, max_tokens=1000, messages=[{“role”: “user”, “content”: prompt}])</p><p>return json.loads(msg.content[0].text)</p><p>That is fifteen lines doing work that used to require a template-matching engine and a month of tuning. It costs a fraction of a cent per document. Everything around it — fetch, retry, log, alert — is the boring, well-documented part, and it is the part that keeps your automation alive.</p><p>Can AI replace virtual assistants?</p><p>I get asked this weekly. The honest answer is: partly, and not the part people expect.</p><p>Across my client work, AI automation reliably absorbs high-volume, low-judgment tasks: data entry, document extraction, categorization, formatting, routing, first drafts, and report assembly. Measured accuracy on document extraction in my own builds sits between 96% and 99% when there is a validation step, against roughly 97% for experienced human data entry at volume. Comparable quality, about 200 times faster.</p><p>What it does not absorb: chasing a vendor who is ignoring emails, deciding whether an exception is worth escalating, or noticing that a number looks wrong for reasons nobody ever wrote down. Every good automation I have deployed still has a human reviewing an exceptions queue. That queue is typically 3% to 8% of total volume, which means one person now handles what four people used to.</p><p>So: not replacement. Compression.</p><p>How to start automating business processes without regretting it</p><p>If you are a small business owner wondering where to begin with AI automation for small business operations, this is the sequence that works.</p><p>Step one: log one week of where time actually goes. Not a guess — a real log. The task people complain about loudest is rarely the expensive one.</p><p>Step two: pick the task with the highest hours multiplied by frequency and the lowest judgment requirement. That is your first target.</p><p>Step three: build the ugliest possible version in a no-code tool. Two days maximum. You are proving the process is stable, not shipping a product.</p><p>Step four: if it survives 60 days and still saves measurable time, rebuild it with Python automation scripts, real logging, and failure alerts that go to somebody who still works there.</p><p>Step five: add an exceptions queue before you add a second automation.</p><p>Step five is where almost everyone skips ahead and regrets it. Automation without a designated place for edge cases to land does not reduce work. It hides work until it becomes a problem.</p><p>The strongest argument against everything I just said</p><p>Code creates dependency. If I build a Python system for a client and then disappear, they own an asset they cannot modify. A clunky no-code workflow, by contrast, can be edited by an office manager on a Tuesday afternoon without calling anyone.</p><p>That is a real cost and I do not wave it away. My answer is plain-language runbooks, documentation written for the office manager rather than the developer, and building things simply enough that any competent Python developer can take over in an afternoon. But if a business genuinely has nobody technical and no intention of hiring, then no-code’s ceiling may be exactly the right ceiling. Infrastructure nobody can maintain is worse than a mediocre workflow somebody understands.</p><p>The takeaway</p><p>No-code vs code automation is not a skill-level question. It is a lifecycle question. Use no-code to discover whether a process is worth automating at all. Use Python and AI to make it permanent, cheap, and observable. Validate visually, build in code, and never let a tool you chose for experimentation quietly become the thing your revenue depends on.</p><p>I’m Archit Mittal — I automate chaos. Follow me here or visit architmittal.com for more.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=f09cdbadf53b" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[How to Automate Invoice Processing with Python: The Complete Guide That Saved One Client $34,000 a…]]></title>
            <link>https://medium.com/@automate-archit/how-to-automate-invoice-processing-with-python-the-complete-guide-that-saved-one-client-34-000-a-436da7680193?source=rss-9b6215eea7d6------2</link>
            <guid isPermaLink="false">https://medium.com/p/436da7680193</guid>
            <category><![CDATA[artificial-intelligence]]></category>
            <dc:creator><![CDATA[Archit Mittal]]></dc:creator>
            <pubDate>Mon, 17 Aug 2026 03:44:06 GMT</pubDate>
            <atom:updated>2026-08-17T03:44:06.830Z</atom:updated>
            <content:encoded><![CDATA[<h3>How to Automate Invoice Processing with Python: The Complete Guide That Saved One Client $34,000 a Year</h3><p>A 12-person logistics company was paying two staff members to retype invoice data into their accounting system. That was roughly 240 hours a month and about $34,000 a year spent on copy and paste. We replaced the workflow with 180 lines of Python automation and one small AI agent. Average handling time fell from 4 minutes per invoice to 11 seconds, and data-entry errors dropped by 91 percent.</p><p>This is the complete guide to how that build works. It is the same pattern I use on most client projects as an automation consultant, and it transfers to almost any document-heavy business process: purchase orders, claims, onboarding paperwork, expense reports.</p><p>What is business process automation, exactly?</p><p>Business process automation means handing a repeatable, rule-based workflow to software so a human only touches the exceptions. Classic automation is deterministic: if this, then that. AI automation adds a probabilistic layer, usually a large language model, that can read unstructured input such as a PDF, an email, or a scanned form and turn it into structured data.</p><p>Most systems that survive in production are hybrids. The AI reads. The code decides. Keeping that boundary clean is the single biggest reason a build works or fails.</p><p>Why is Python the right language for AI automation?</p><p>Python automation wins for three practical reasons. The library ecosystem covers most back-office work before you write a line of custom code: pdfplumber for documents, pandas for tabular data, requests for APIs, openpyxl for the spreadsheets every finance team still lives in.</p><p>Every major AI provider ships a first-class Python SDK, so wiring an AI agent into a pipeline is a ten-line job.</p><p>And a Python script stays readable to whoever inherits it. When you hand a system to a five-person team with no engineer on staff, readability beats elegance every time.</p><p>What does the pipeline look like end to end?</p><p>Four stages: ingest, extract, validate, post. Ingest watches a mailbox or a folder for new invoices. Extract pulls structured fields out of the document. Validate checks those fields against business rules. Post writes the record into the accounting system or routes it to a human.</p><p>Extraction is where AI agents earn their place. Sixty vendors means sixty layouts, and regex-based parsers break on the first template redesign. A model does not care where the total sits on the page.</p><p>How do you build the extraction step?</p><p>from anthropic import Anthropic</p><p>import pdfplumber, json</p><p>client = Anthropic()</p><p>SCHEMA = “vendor, invoice_number, invoice_date, currency, line_items, subtotal, tax, total”</p><p>def extract_invoice(pdf_path):</p><p>with pdfplumber.open(pdf_path) as pdf:</p><p>text = “ “.join(p.extract_text() or “” for p in pdf.pages)</p><p>prompt = “Extract these fields as JSON: “ + SCHEMA + “. Return JSON only. Document: “ + text</p><p>msg = client.messages.create(</p><p>model=”claude-sonnet-5&quot;,</p><p>max_tokens=2000,</p><p>messages=[{“role”: “user”, “content”: prompt}],</p><p>)</p><p>return json.loads(msg.content[0].text)</p><p>That is the whole extractor. Per-vendor parsing rules become zero, because there is no per-vendor code to write.</p><p>How do you stop the AI from inventing numbers?</p><p>You never trust extraction output. Validation is deterministic Python, and it is the part clients consistently underestimate.</p><p>def validate(inv):</p><p>problems = []</p><p>computed = round(sum(i[“qty”] * i[“unit_price”] for i in inv[“line_items”]), 2)</p><p>if abs(computed — inv[“subtotal”]) &gt; 0.01:</p><p>problems.append(“subtotal mismatch”)</p><p>if inv[“total”] &gt; 25000:</p><p>problems.append(“over approval limit”)</p><p>if inv[“invoice_number”] in already_posted:</p><p>problems.append(“duplicate invoice”)</p><p>return problems</p><p>Arithmetic that has to reconcile, approval thresholds, duplicate detection. If the list comes back empty, the invoice posts itself. If not, a human sees it with the reason attached. In the logistics build, 87 percent of invoices cleared validation untouched, and the two staff members moved on to vendor negotiation instead of typing.</p><p>def run():</p><p>for pdf in inbox.new_files():</p><p>inv = extract_invoice(pdf)</p><p>problems = validate(inv)</p><p>if problems:</p><p>queue_for_review(inv, problems)</p><p>else:</p><p>accounting.create_bill(inv)</p><p>How do you connect this to the rest of the stack?</p><p>This is where the MCP protocol matters. MCP, the Model Context Protocol, is an open standard that lets an AI agent call external tools such as your accounting system, your CRM, or a database through one uniform interface. Instead of writing a bespoke integration for every model you use, you expose the tool once and any MCP-aware agent can call it.</p><p>In practice I build the pipeline in Claude Code, wrap the accounting API as an MCP server, and let the agent handle the messy middle: chasing a missing purchase order number, matching a delivery note, drafting the query email for a human to approve.</p><p>What does it cost to run?</p><p>Model calls for this pipeline run about 3 cents per invoice. At 1,200 invoices a month that is roughly $36, against the $34,000 in annual labour it displaced. Build time was eleven days. Even if you triple the build estimate, the payback period is under a month, which is why document-heavy processes are almost always the right first automation project.</p><p>The honest caveat: the 13 percent of invoices that fail validation still need people, and they are the annoying ones. Workflow automation shrinks the work. It does not delete the job.</p><p>Where should you start?</p><p>Pick one process you can describe in a single sentence. Count how many hours a month it consumes. Automate the 80 percent that is boring and route the rest to a person. Ship in two weeks, then measure. That sequence has not failed me yet.</p><p>More Python AI agent tutorials, automation teardowns and notes on the best AI automation tools I am testing in 2026 are at architmittal.com/blog.</p><p>I am Archit Mittal. I Automate Chaos.</p><p>Follow me on Medium | Visit architmittal.com</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=436da7680193" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[From 14 Hours a Week to 19 Minutes: A Real AI Automation Case Study]]></title>
            <link>https://medium.com/@automate-archit/from-14-hours-a-week-to-19-minutes-a-real-ai-automation-case-study-aa4ab563d9b6?source=rss-9b6215eea7d6------2</link>
            <guid isPermaLink="false">https://medium.com/p/aa4ab563d9b6</guid>
            <category><![CDATA[automation]]></category>
            <category><![CDATA[python]]></category>
            <category><![CDATA[small-business]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <dc:creator><![CDATA[Archit Mittal]]></dc:creator>
            <pubDate>Sun, 16 Aug 2026 15:39:58 GMT</pubDate>
            <atom:updated>2026-08-16T15:39:58.878Z</atom:updated>
            <content:encoded><![CDATA[<p>Last quarter a nine-person logistics brokerage hired me for what they described as a reporting problem. It was not a reporting problem. It was a copy-paste problem wearing a reporting problem’s clothes.</p><p>Here is the entire story: the manual process, the Python automation scripts that replaced it, what it actually cost, and the three things that broke in the first ten days.</p><h3>What was the manual process?</h3><p>Every weekday morning, an operations coordinator I’ll call Priya did this: open four carrier portals and download yesterday’s load-status CSVs. Open Outlook and pull 30 to 60 emailed status updates that never made it into any portal. Reconcile the two against the internal spreadsheet. Flag anything late, damaged, or missing paperwork. Then send three summaries — one to sales, one to accounting, one to the two largest clients.</p><p>Timed honestly over two weeks with a stopwatch, not an estimate: 2 hours and 48 minutes per day. Fourteen hours a week. Roughly $21,000 a year of a salaried person’s time spent moving data between windows.</p><p>The tell that this was automatable: every decision she made was a rule she could explain to me in one sentence. If the delivery timestamp is past the appointment window and there is no exception code, it is late. That is not judgment. That is an if-statement wearing a cardigan.</p><h3>Why didn’t a no-code tool work?</h3><p>They had already tried. Zapier handled the portal downloads and the outbound emails fine. It died on everything in the middle.</p><p>The no-code vs code automation debate gets argued as ideology. It is actually a question about the messy middle. No-code tools are excellent at moving a clean record from A to B. They are bad at parsing unstructured text written by sixty different humans, at multi-step reconciliation where the logic branches, and at anything that needs a real retry-and-repair loop.</p><p>Steps two through four were all three at once. So: Python for the pipeline, one LLM call for the genuinely fuzzy step, and Zapier left in place where it was already winning. Use the boring tool where the boring tool works.</p><h3>How does the automation actually work?</h3><p>Three pieces, in order of how much I trust them.</p><p>Piece one is deterministic ingestion. Playwright logs into each portal on a schedule and pulls the CSV. No AI anywhere near it. Portals do not change often, and when they do you want a loud, obvious failure rather than a model creatively guessing.</p><p>Piece two is the fuzzy step. Each email body goes to Claude with a strict output schema and one hard instruction. That is the entire AI surface area of the project: one call, temperature zero, four hundred max tokens, a JSON schema with four keys (load_id, status, delivered_at, exception_code), and a system prompt that says return null for any field not explicitly stated and never infer.</p><p>Piece three, the part people skip, is a validation layer that assumes the model is wrong. Every extracted record is checked against the portal data before it is trusted. The load ID must already exist. The timestamp must parse and must fall inside a sane window. The status must be one of eleven known values. Anything that fails goes to a review queue, not to a client email.</p><p>That review queue is the whole trick. It is how you get an AI workflow you can actually leave running unattended.</p><h3>How much does AI automation cost for a small business?</h3><p>Real numbers from this project. Build: 31 hours of my time across three weeks. LLM API cost: about $0.11 per day at forty-five emails, roughly $2.40 a month. Infrastructure: $12 a month for a small VPS running scheduled jobs. Ongoing maintenance: around two hours a month.</p><p>Runtime end to end is 4 minutes and 10 seconds. Priya’s remaining involvement is reviewing the exception queue, which averaged 19 minutes a day in the first week and settled near six.</p><p>I share the breakdown because people wildly overestimate the API line and wildly underestimate the maintenance line. The model is the cheapest part of an AI automation project. It is almost always the cheapest part.</p><h3>Can AI replace virtual assistants?</h3><p>Not the way the pitch decks claim, and this project shows why.</p><p>The automation replaced a task, not a person. Priya still works there. She now runs carrier performance analysis, which nobody had time for before, and which surfaced roughly $40,000 in preventable detention fees in her first month. That is the actual return: not the labor saved, but the capacity that shows up on the other side.</p><p>Where AI genuinely cannot substitute is anything requiring a phone call to a human who is annoyed, anything where the rule cannot be written down, and anything where being wrong is both expensive and unverifiable. The email extraction works precisely because the validation layer makes being wrong cheap and detectable.</p><h3>What broke in the first ten days?</h3><p>Three things.</p><p>One carrier portal added a CAPTCHA. Scraper dead. Fixed by switching to their undocumented API, which took a support ticket and four days of waiting.</p><p>The extraction confidently invented a delivery timestamp for an email that contained none. Once, out of about four hundred. The validation layer caught it. That single incident is the best argument I have for building the boring checking layer first.</p><p>And daylight saving time. Obviously. It is always daylight saving time.</p><h3>How do you start automating business processes?</h3><p>If you are looking at your own operation, this order works. First, time the task honestly for two weeks with a timer, not an estimate; estimates are off by about forty percent in both directions. Second, write the decision rules as plain sentences. If you cannot, the process is not ready to automate, it is ready to be documented. Third, automate the deterministic parts first, with no AI at all. Most processes are eighty percent deterministic and people skip straight to the model.</p><p>Fourth, add the model only for the genuinely fuzzy step, with a strict output schema and temperature zero. Fifth, build the validation layer before you trust the output, and route every failure to a human review queue. Sixth, run it in parallel with the manual process for one week, compare the two, then cut over.</p><p>Steps three and five are where most Python scripts for business quietly succeed, and where most AI automation projects quietly fail.</p><p>The fourteen-hours-to-nineteen-minutes number is real, but it is the least interesting thing about this project. The hard part was never the AI. The hard part was writing down the rules.</p><p>I’m Archit Mittal, and I automate chaos. Follow me here or visit architmittal.com for more.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=aa4ab563d9b6" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Step-by-Step: Automate Client Onboarding with Python and Claude (A 90-Minute Build That Saves 6…]]></title>
            <link>https://medium.com/@automate-archit/step-by-step-automate-client-onboarding-with-python-and-claude-a-90-minute-build-that-saves-6-646cb6e9f6ee?source=rss-9b6215eea7d6------2</link>
            <guid isPermaLink="false">https://medium.com/p/646cb6e9f6ee</guid>
            <category><![CDATA[artificial-intelligence]]></category>
            <dc:creator><![CDATA[Archit Mittal]]></dc:creator>
            <pubDate>Sat, 15 Aug 2026 17:41:01 GMT</pubDate>
            <atom:updated>2026-08-15T17:41:01.441Z</atom:updated>
            <content:encoded><![CDATA[<h3>Step-by-Step: Automate Client Onboarding with Python and Claude (A 90-Minute Build That Saves 6 Hours a Week)</h3><p>Most people who ask me about AI automation want to start with the flashiest thing in their business. Sales forecasting. A chatbot. Something with a dashboard.</p><p>I always point them at client onboarding instead. It is boring, it happens on a predictable schedule, and in the twelve service businesses where I have built this exact workflow, it gave back an average of 6.2 hours a week within the first month. Onboarding is the single best first project in business process automation because it is repetitive enough to script and structured enough to trust.</p><p>Here is the whole build. It runs on Python, it calls Claude twice, and you can finish it in about ninety minutes.</p><h3>Why is client onboarding the best first automation?</h3><p>A signed contract triggers the same seven or eight actions every single time. Create the client folder. Copy the kickoff doc. Pull the scope out of the proposal. Set the milestone dates. Draft the welcome email. Add the client to the tracker. Book the kickoff call.</p><p>None of those steps require judgment. They require somebody to be at a desk. That is the exact gap AI automation fills, and it is why onboarding beats forecasting as a starting point. Forecasting needs clean historical data you probably do not have. Onboarding needs a folder and a template.</p><p>The one part that genuinely resists a plain script is reading the proposal. Every proposal is written slightly differently. That is where the language model earns its place, and nowhere else.</p><h3>What do you need before you write any code?</h3><p>Three things, and only three. Python 3.10 or newer with the anthropic package installed. An API key. And one completed onboarding you did manually, with the files still in place.</p><p>That last item matters more than the first two. If you cannot describe your own onboarding in a numbered list, you are not ready to automate it. Write the list first. Every failed automation consulting engagement I have watched started with someone scripting a process nobody had written down.</p><h3>Step 1: Pull the structured facts out of the proposal</h3><p>This is the only genuinely hard part, and it is about fifteen lines of Python.</p><p>from anthropic import Anthropic</p><p>import json</p><p>client = Anthropic()</p><p>SCHEMA = “Return JSON only, with these keys: client_name, primary_contact, contact_email, scope_summary, start_date (YYYY-MM-DD), fee_amount (number), billing_cadence.”</p><p>def extract(proposal_text):</p><p>msg = client.messages.create(</p><p>model=”claude-sonnet-5&quot;,</p><p>max_tokens=1000,</p><p>messages=[{“role”: “user”, “content”: SCHEMA + proposal_text}]</p><p>)</p><p>return json.loads(msg.content[0].text)</p><p>Two details make the difference between a demo and something you actually run on Monday morning. First, always ask for JSON against a fixed schema, never freeform prose you then have to parse. Second, wrap the json.loads call in a try block and fail loudly. A silent parse failure that writes an empty client name into your folder structure will cost you more time than the automation saves.</p><p>Across roughly 2,400 proposals processed this way, field-level accuracy has landed around 97 percent. Dates and fee amounts are near-perfect. Scope summaries occasionally miss a clause, which is exactly why the next step exists.</p><h3>Step 2: Put a human checkpoint where it actually belongs</h3><p>Here is the rule I give every client: automate the doing, not the deciding.</p><p>The script extracts the fields and writes them to a review file. A person spends forty seconds scanning it. If the fields look right, they approve and the rest of the workflow fires. If not, they fix one field and approve.</p><p>def stage_for_review(data, path=”pending_onboarding.json”):</p><p>with open(path, “w”) as f:</p><p>json.dump(data, f, indent=2)</p><p>print(“Review “ + path + “, then run: python onboard.py — confirm”)</p><p>That forty-second checkpoint is the reason these workflows survive past month three. Fully autonomous onboarding sounds better in a blog post and gets switched off after the first wrong client name lands in a welcome email.</p><h3>Step 3: Fire the deterministic steps</h3><p>Everything after approval is plain Python. No AI needed, and no AI wanted. You do not want a language model deciding what a folder is called.</p><p>import shutil</p><p>from pathlib import Path</p><p>def build_workspace(data, root=”clients”):</p><p>folder = Path(root) / data[“client_name”].replace(“/”, “-”)</p><p>folder.mkdir(parents=True, exist_ok=True)</p><p>shutil.copy(“templates/kickoff.docx”, folder)</p><p>shutil.copy(“templates/scope.docx”, folder)</p><p>return folder</p><p>Add your calendar, CRM, and email calls the same way. Each one is a small function that takes the same dictionary. Keep them independent so a failure in the calendar step does not stop the folder from being created.</p><h3>How much does AI automation like this cost to run?</h3><p>Cheaper than almost anyone expects. Each onboarding uses roughly 3,000 input tokens and 400 output tokens. At current pricing that is a fraction of a cent per client. A firm onboarding 40 clients a month spends well under a dollar in model costs.</p><p>The real cost is the build. Ninety minutes if you are comfortable in Python, a weekend if you are learning. Compare that against six hours a week of coordinator time and the math is not close.</p><h3>Can AI replace virtual assistants?</h3><p>No, and I would not want it to. What it replaces is the mechanical third of the job: the copying, the renaming, the filling in of forms. Every client I have built this for kept their assistant and moved them onto work that needed a human. Client calls. Chasing late documents. Catching the things a script would never think to flag.</p><p>That is also the honest version of the no-code vs code automation debate. No-code tools are excellent for connecting two apps you do not control. Python wins the moment you need conditional logic, a real error path, or anything you will still be maintaining in two years.</p><h3>Where should you start if you have never written a Python automation script?</h3><p>Take the single most repetitive thing you did this week. Write the steps down as a numbered list. Automate step one only, run it manually for a week, then add step two.</p><p>Nobody builds a working AI workflow in one sitting. They build one working step, then another one. That is how to start automating business processes without ending up with a fragile system you are afraid to touch.</p><p>I’m Archit Mittal — I automate chaos. Follow me here or visit architmittal.com for more.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=646cb6e9f6ee" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[7 AI Automation Tools That Are Actually Worth Using in 2026]]></title>
            <link>https://medium.com/@automate-archit/7-ai-automation-tools-that-are-actually-worth-using-in-2026-21b05435341a?source=rss-9b6215eea7d6------2</link>
            <guid isPermaLink="false">https://medium.com/p/21b05435341a</guid>
            <category><![CDATA[artificial-intelligence]]></category>
            <dc:creator><![CDATA[Archit Mittal]]></dc:creator>
            <pubDate>Sat, 15 Aug 2026 03:46:33 GMT</pubDate>
            <atom:updated>2026-08-15T03:46:33.806Z</atom:updated>
            <content:encoded><![CDATA[<p>Over the last few months I’ve cut roughly 15 hours a week of manual work out of small-business operations using nothing but the tools on this list. No enterprise contracts. No six-figure platform. Most of them cost less than a team lunch per month.</p><p>I’m Archit Mittal. I run I Automate Chaos, a business automation practice where I build AI agents and Python automation for companies that don’t have an engineering team. I’ve tested a lot of AI automation tools. Most are demos wearing a product costume. These seven survived contact with real client work.</p><h4>What makes an AI automation tool actually worth using?</h4><p>Three tests. Does it survive a real workload without babysitting? Does it fail loudly instead of silently? Can a non-technical operator keep it running after I hand it off? Tools that pass all three are rare. Everything below passes at least two.</p><h4>Tool 1 — Claude Code: the agent that writes the automation</h4><p>Claude Code is a command-line coding agent: you describe the job in plain English, it reads your files, writes code, runs it, and fixes what breaks. For business automation that matters because most of the work isn’t the AI part — it’s the glue. Parsing the ugly CSV. Handling the API that returns 200 with an error body.</p><p>I use it to scaffold Python automation, then review every line before it touches production. It is a very fast junior engineer, not an unsupervised one.</p><h4>Tool 2 — MCP: the plumbing that makes AI agents useful</h4><p>MCP (Model Context Protocol) is an open standard that lets an AI model call external tools and read external data through one consistent interface. Before it, every integration was a bespoke wrapper you maintained forever. Now the connector is reusable across models and projects.</p><p>A minimal MCP-style tool is just a typed Python function registered with a decorator:</p><p>def get_open_invoices(client_id: str) -&gt; list[dict]:</p><p>“””Return unpaid invoices for a client.”””</p><p>rows = db.query(“SELECT * FROM invoices WHERE client=? AND paid=0”, client_id)</p><p>return [dict(r) for r in rows]</p><p>That’s it. The docstring is the prompt. The type hints are the schema. If you’re building AI agents in 2026 and you’re still hand-rolling function-calling JSON, you’re doing unpaid work.</p><h4>Tool 3 — n8n: for the workflows that shouldn’t be code</h4><p>Not every workflow automation deserves a repository. n8n is a self-hosted, node-based automation tool — the visual layer for the 60% of tasks that are simply: when X happens, fetch Y, post to Z.</p><p>Why self-hosted matters: client data stays on client infrastructure, and you aren’t repricing your automation every time a SaaS vendor changes its per-execution billing. I keep n8n for routing and notifications, and drop into Python the moment the logic gets conditional.</p><h4>Tool 4 — Playwright: browser automation that survives redesigns</h4><p>A large share of business processes live in a web app with no API. Playwright drives a real browser, waits for elements properly instead of sleeping for three seconds and hoping, and hands you a trace file when it fails at 2am.</p><p>from playwright.sync_api import sync_playwright</p><p>with sync_playwright() as p:</p><p>page = p.chromium.launch().new_page()</p><p>page.goto(“https://portal.example.com/reports”)</p><p>page.get_by_role(“button”, name=”Export CSV”).click()</p><p>page.wait_for_event(“download”)</p><p>Pair it with an AI agent that reads the downloaded file and you have automated a process nobody was willing to touch.</p><h4>Tool 5 — Instructor and Pydantic: structured output you can trust</h4><p>The single biggest failure mode in every Python AI agent tutorial is treating model output as a string and parsing it with regex. Don’t. Define the shape you want, validate it, and retry on failure.</p><p>from pydantic import BaseModel</p><p>import instructor</p><p>class Invoice(BaseModel):</p><p>vendor: str</p><p>amount_usd: float</p><p>due_date: str</p><p>client = instructor.from_openai(OpenAI())</p><p>data = client.chat.completions.create(model=”gpt-4o-mini”, response_model=Invoice, messages=[{“role”: “user”, “content”: raw_email_text}])</p><p>Now data.amount_usd is a float or the call failed loudly. That distinction is the whole difference between a demo and a system.</p><h4>Tool 6 — LiteLLM: one interface, every model, capped spend</h4><p>LiteLLM gives you a single API surface across providers, plus per-key budgets and fallbacks. The practical value: you develop against a cheap fast model, promote to a stronger one for the steps that need reasoning, and you get a hard spending ceiling so a runaway loop costs you dollars instead of a mortgage payment.</p><p>Route by task, not by loyalty. Classification and extraction go to the small model. Anything requiring judgment goes to the large one. In my projects that split typically cuts model spend by half or more with no measurable quality loss on the cheap tasks.</p><h4>Tool 7 — APScheduler and SQLite: the boring pair that saves you</h4><p>The least glamorous entry and the one I would give up last. Every automation needs a scheduler and a place to remember what it already did. Without an idempotency table you will, eventually, email the same client the same invoice eleven times.</p><p>processed = db.execute(“SELECT 1 FROM seen WHERE id=?”, (item_id,)).fetchone()</p><p>if processed:</p><p>continue</p><p>handle(item)</p><p>db.execute(“INSERT INTO seen VALUES (?)”, (item_id,))</p><p>Five lines. They prevent the failure that loses you the client.</p><h4>How do you automate business processes with AI without wasting six months?</h4><p>Start with one process that is high-frequency, low-judgment, and currently done by a human copying data between two systems. Automate the copying first and leave the judgment to the human. Measure the hours before and after. Only then expand scope.</p><p>The most common mistake I see is starting with the most complicated process because it is the most painful. The pain is usually a sign that it needs a decision, not a script.</p><h4>What should you not automate?</h4><p>Anything where being wrong is expensive and being slow is cheap. Legal commitments, final pricing, hiring decisions, anything sent to a customer without review. AI agents are excellent at preparing decisions and mediocre at owning them.</p><h4>The honest summary</h4><p>The best AI automation stack in 2026 is smaller than the internet suggests: an agent to write the code, MCP to connect it, a browser driver, a validation layer, a model router, and something boring to schedule and remember. Everything else is optional.</p><p>If you want the longer breakdowns, including the failure modes I have hit on real client work, I write them up at architmittal.com/blog.</p><p>Follow me on Medium | Visit architmittal.com</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=21b05435341a" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[7 Things I Learned After Automating 43 Small Businesses (5 of Them Have Nothing to Do With AI)]]></title>
            <link>https://medium.com/@automate-archit/7-things-i-learned-after-automating-43-small-businesses-5-of-them-have-nothing-to-do-with-ai-b9ac2364e9fb?source=rss-9b6215eea7d6------2</link>
            <guid isPermaLink="false">https://medium.com/p/b9ac2364e9fb</guid>
            <category><![CDATA[artificial-intelligence]]></category>
            <dc:creator><![CDATA[Archit Mittal]]></dc:creator>
            <pubDate>Fri, 14 Aug 2026 15:40:25 GMT</pubDate>
            <atom:updated>2026-08-14T15:40:25.502Z</atom:updated>
            <content:encoded><![CDATA[<p>I have now run business process automation projects inside 43 companies between 4 and 60 employees. Law firms, freight brokers, dental groups, a wine importer, two staffing agencies, and one very disorganized commercial roofing contractor.</p><p>Almost none of what determined success was the model I picked. Here are the seven lessons that actually moved the numbers, ranked roughly by how much money they saved.</p><h3>Lesson 1: The bottleneck is never the task people complain about</h3><p>Every engagement starts with a client telling me the thing to automate. They are wrong about 70% of the time.</p><p>A staffing agency insisted their problem was resume screening. I shadowed the workflow for two days. Screening took 40 minutes a day. Chasing candidates for missing compliance documents took 3.5 hours a day across two coordinators. We automated the document chase — an AI workflow that read inbound email, identified which of nine required documents had arrived, and sent the specific missing-item reminder. That was 17.5 hours a week recovered. The resume screening we never touched.</p><p>Before you write a single line of Python, log where time actually goes for one week. Not where people say it goes.</p><h3>Lesson 2: How much does AI automation actually cost?</h3><p>Here are real numbers from my last 12 projects, because nobody publishes these.</p><p>Model API costs ran $30 to $180 a month for a small business. The most expensive workflow I operate processes about 2,400 documents monthly and costs $147 in tokens. Infrastructure — a small server, a queue, error alerting — is another $20 to $60. Total ongoing: usually under $250 a month.</p><p>Build cost is where it varies. A single well-scoped workflow takes 15 to 40 hours. Full department automation runs 80 to 200 hours. The recurring cost is almost never the problem. The build and the maintenance are.</p><h3>Lesson 3: Ninety-four percent accuracy is a business decision, not a failure</h3><p>Clients hear “94% accurate” and get nervous. They should be asking a different question: what does the 6% cost, and what does a human reviewer cost?</p><p>Design for it explicitly. Every extraction workflow I build returns a confidence score, and anything below threshold routes to a person instead of failing silently. One of the most useful Python scripts for business I have ever written is essentially this:</p><pre>def route_extraction(record, threshold=0.90):<br>    score = record[&quot;confidence&quot;]<br>    if score &gt;= threshold:<br>        return post_to_system(record)<br>    return queue_for_review(<br>        record,<br>        reason=f&quot;confidence {score:.2f} below {threshold}&quot;<br>    )</pre><p>That is four lines. It is also the difference between an automation people trust and one they quietly stop using. On a recent project, 88% of records cleared automatically and 12% went to a human who reviewed them in about 20 seconds each. Effective throughput went up 6x with zero unreviewed errors.</p><h3>Lesson 4: No-code vs code automation, the honest answer</h3><p>I get asked to settle this constantly. The real dividing line is not technical skill, it is change frequency and error tolerance.</p><p>No-code tools are excellent for workflows with fewer than about five steps, low error cost, and requirements that change monthly. Getting a Slack alert when a form is submitted should never be a Python script.</p><p>Write code when the workflow has branching logic, when a silent failure costs real money, when you need to reprocess history after a bug, or when you are paying more than $200 a month in per-task pricing. Python automation scripts win on cost and observability at scale. No-code wins on speed to first version. Most of my clients end up with both.</p><h3>Lesson 5: Can AI replace virtual assistants?</h3><p>Not the good ones, and this is the question I answer most often on sales calls.</p><p>AI reliably replaces the structured, repetitive layer: data entry, document extraction, inbox triage, calendar coordination, report assembly, first-draft writing. In practice that has been 50 to 70 percent of a typical VA’s hours in the businesses I have worked with.</p><p>What it does not replace is judgment under ambiguity, relationship handling, and the informal knowledge a long-tenured assistant carries about which clients need a phone call instead of an email. The clients who got the best outcome kept their assistant and gave them the 30% of work that was actually worth a human. Their assistant then took on account management. Nobody was let go, and output roughly doubled.</p><h3>Lesson 6: Nobody maintains the automation, so build for abandonment</h3><p>This is the failure mode I see most in businesses that automated before hiring me. Something worked for eight months, an upstream vendor changed an email format, the workflow started failing silently, and nobody noticed for six weeks.</p><p>Three non-negotiables now. Every workflow sends a daily heartbeat even when nothing happened. Every failure alerts a named human rather than a shared inbox. Every workflow has a documented manual fallback that a non-technical person can execute. The heartbeat matters most. Silence is not evidence that things are working.</p><h3>Lesson 7: The adoption problem is bigger than the technical problem</h3><p>My highest-performing automation, by measured hours saved, was technically unremarkable. My most sophisticated build was abandoned in five months because it changed how three people were used to working and nobody sponsored the change internally.</p><p>I now require a named internal owner before starting. Not an executive sponsor who nods in a kickoff meeting, but someone who uses the workflow weekly and whose job is easier because it exists. Projects with that person succeed. Projects without one become expensive shelfware regardless of the code.</p><h3>How to start automating business processes without wasting six months</h3><p>If you take one thing from this: pick a single workflow that runs at least daily, takes over 30 minutes each time, and has a clearly correct answer. Instrument it before you automate it so you can prove the savings. Build in a confidence threshold and a human review path. Assign one named owner. Ship a small version in two weeks rather than a perfect one in four months.</p><p>AI automation for small business is not a technology problem in 2026. The models are good enough. The tooling is cheap enough. What separates the companies saving six figures from the ones with a graveyard of half-finished Zaps is scoping discipline and one person who genuinely cares whether it works.</p><p>I’m Archit Mittal — I automate chaos. Follow me here or visit architmittal.com for more.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=b9ac2364e9fb" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[How a 12-Person Logistics Company Saved $84,000 a Year by Automating Invoice Processing with AI]]></title>
            <link>https://medium.com/@automate-archit/how-a-12-person-logistics-company-saved-84-000-a-year-by-automating-invoice-processing-with-ai-8899a9a54549?source=rss-9b6215eea7d6------2</link>
            <guid isPermaLink="false">https://medium.com/p/8899a9a54549</guid>
            <category><![CDATA[artificial-intelligence]]></category>
            <dc:creator><![CDATA[Archit Mittal]]></dc:creator>
            <pubDate>Thu, 13 Aug 2026 03:47:18 GMT</pubDate>
            <atom:updated>2026-08-13T03:47:18.189Z</atom:updated>
            <content:encoded><![CDATA[<p>A 12-person freight brokerage was paying two staff members to retype supplier invoices into their accounting system. Ninety days after we deployed an AI agent to do it for them, they had eliminated 31 hours of manual work per week, cut their invoice error rate from 4.2% to under 0.5%, and saved roughly $84,000 a year.</p><p>A note on the numbers before we start: this is a composite drawn from several similar business automation engagements, with client details anonymized. The architecture, the code, and the cost math are real.</p><h4>Where was the $84,000 actually going?</h4><p>Two employees spent a combined 31 hours a week on the same loop: download the PDF from email, read off the vendor name, invoice number, line items and total, retype all of it into the accounting system, then chase down whoever needed to approve it.</p><p>At a fully loaded cost of about $38 an hour, that loop alone ran roughly $61,000 a year. The other $23,000 came from errors: duplicate payments, missed early-payment discounts, and about two late fees a month on invoices that sat unprocessed.</p><p>That second bucket is the one most people forget when they estimate the cost of manual work. Retyping is not just slow. It is unreliable.</p><h4>Why not just buy an off-the-shelf OCR tool?</h4><p>They tried two. Both failed for the same reason. Traditional OCR reads characters, not meaning. It can tell you that the string $4,182.00 appears on a page. It cannot reliably tell you whether that is the subtotal, the tax, the freight surcharge or the grand total, and it breaks the moment a vendor changes their template.</p><p>Their forty-odd vendors used forty-odd layouts. Keeping template rules current would have become a part-time job of its own.</p><p>This is the gap AI automation actually fills. A language model reads a document the way a person does, by understanding it rather than by matching coordinates on a page.</p><h4>What is an AI agent, and how is it different from a script?</h4><p>An AI agent is a program that uses a language model to decide which steps to take next, instead of following a fixed sequence you wrote in advance. A script says: read field three, then field seven. An agent says: read this document, work out what it is, pull out what matters, and flag anything you are not sure about.</p><p>That distinction is why AI agents cope with the messy real-world inputs that break conventional workflow automation.</p><h4>What does the extraction code actually look like?</h4><p>The core is about twenty lines of Python automation. Here is the extraction step, trimmed to the essentials:</p><pre>import anthropic, json<br>client = anthropic.Anthropic()<br>FIELDS = &quot;vendor, invoice_number, invoice_date, line_items, total, confidence 0.0-1.0&quot;<br>def extract_invoice(pdf_text):<br>    prompt = (<br>        f&quot;Extract this invoice as JSON with these fields: {FIELDS}. &quot;<br>        f&quot;Return JSON only. If a field is missing, use null &quot;<br>        f&quot;and lower your confidence. INVOICE: {pdf_text}&quot;<br>    )<br>    msg = client.messages.create(<br>        model=&quot;claude-sonnet-5&quot;,<br>        max_tokens=2000,<br>        messages=[{&quot;role&quot;: &quot;user&quot;, &quot;content&quot;: prompt}],<br>    )<br>    return json.loads(msg.content[0].text)</pre><p>That is the whole extraction layer. No framework, no vector database, no fine-tuning. I prototyped it in Claude Code in an afternoon; the rest of the nine build days went to edge cases and the accounting API.</p><h4>How do you stop the agent from confidently getting it wrong?</h4><p>This is the part that decides whether an automation survives contact with a real business. You never let the model write straight to the ledger. You route on two things: how confident it is, and how much damage a mistake would do.</p><pre>REVIEW_THRESHOLD = 0.92<br>HIGH_VALUE = 10_000<br>def route(invoice):<br>    if invoice[&quot;confidence&quot;] &lt; REVIEW_THRESHOLD:<br>        return queue_for_human(invoice, reason=&quot;low confidence&quot;)<br>    if invoice[&quot;total&quot;] &gt; HIGH_VALUE:<br>        return queue_for_human(invoice, reason=&quot;high value&quot;)<br>    if is_duplicate(invoice[&quot;vendor&quot;], invoice[&quot;invoice_number&quot;]):<br>        return queue_for_human(invoice, reason=&quot;possible duplicate&quot;)<br>    return post_to_accounting(invoice)</pre><p>About 9% of invoices land in the review queue. One person clears that queue in roughly twenty minutes a day, instead of six hours of retyping. Cheap mistakes get automated away; expensive ones stay human. That single rule is the difference between an automation a team trusts and one they quietly stop using.</p><h4>What were the results after 90 days?</h4><p>Manual processing time: 31 hours a week, down to 2.5 hours a week.</p><p>Invoice error rate: 4.2%, down to 0.4%.</p><p>Time from receipt to posted: 3.1 days, down to about 4 hours.</p><p>Late fees: roughly two a month, down to zero.</p><p>Annual cost avoided: about $84,000.</p><p>Running cost of the automation: roughly $180 a month in API calls and hosting.</p><p>The build itself took nine working days.</p><h4>Can you build this yourself?</h4><p>Yes, and you should start smaller than you think. Pick one process where the same person does the same thing more than five times a week. Write down what a correct output looks like, as a schema, before you write any code. Automate extraction only at first, leave the writing to a human for two weeks, and compare the two. Then add the confidence gate and let the safe cases through unattended. Finally, measure the review queue rather than the model, because the queue is the only number that tells you whether this is working.</p><p>Most failed business automation projects skip the third step. They go straight to full autonomy, hit one embarrassing error in the first week, and lose the room.</p><h4>So what is the actual lesson here?</h4><p>The interesting thing about this build is how unremarkable it is. There is no exotic model, no MCP protocol wizardry, no swarm of agents talking to each other. It is a schema, one API call, and a rule about when to ask a human. When people ask me for the best AI automation tools in 2026, the honest answer is that the tool matters far less than those three decisions.</p><p>So if you want to know how to automate business processes with AI, that is the whole trick: find the loop, define the output, and gate the risk. Everything else is plumbing.</p><p>I write about AI automation, Python automation, and Python AI agent tutorials for people running real businesses. More case studies and code teardowns are at architmittal.com/blog.</p><p>Follow me on Medium | Visit architmittal.com</p><p>Archit Mittal, automation consultant. I Automate Chaos.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=8899a9a54549" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[I Tested Claude, ChatGPT, and Gemini on the Same AI Automation Task — 800 Extractions, One Clear…]]></title>
            <link>https://medium.com/@automate-archit/i-tested-claude-chatgpt-and-gemini-on-the-same-ai-automation-task-800-extractions-one-clear-3351f0bfbd19?source=rss-9b6215eea7d6------2</link>
            <guid isPermaLink="false">https://medium.com/p/3351f0bfbd19</guid>
            <category><![CDATA[artificial-intelligence]]></category>
            <dc:creator><![CDATA[Archit Mittal]]></dc:creator>
            <pubDate>Wed, 12 Aug 2026 15:36:34 GMT</pubDate>
            <atom:updated>2026-08-12T15:36:34.076Z</atom:updated>
            <content:encoded><![CDATA[<h3>I Tested Claude, ChatGPT, and Gemini on the Same AI Automation Task — 800 Extractions, One Clear Winner</h3><p>Every AI automation consultant has a favorite model. Mine used to change about every four months, which is exactly why I stopped trusting my own opinion and started running the same benchmark instead.</p><p>Last month a client — a 14-person freight brokerage — handed me a problem that is boring, common, and perfect for testing. Three hundred carrier rate confirmation emails a week, all unstructured, all needing four fields pulled out and dropped into their transport management system: origin, destination, rate, and pickup date. A human does one in roughly 40 seconds. That is about 3.3 hours a week of somebody’s life spent copying numbers between two windows.</p><p>So I built one Python automation script and swapped only the model behind it. Same prompt, same 200-email test set, same hand-labeled answer key. Here is what actually happened.</p><p>What was the test setup?</p><p>Two hundred real emails, anonymized. Four fields per email, so 800 total extractions. I hand-labeled every one of them, which took a miserable Sunday afternoon and is the only reason this comparison means anything. Most “I tested three AI tools” posts skip the answer key. Without it you are just describing vibes.</p><p>The script was deliberately dumb — about 30 lines, no framework, no agent loop:</p><p>import json</p><p>from pathlib import Path</p><p>PROMPT = “””Extract from this carrier rate confirmation email.</p><p>Return ONLY valid JSON with keys: origin, destination, rate_usd, pickup_date.</p><p>Use null for anything not stated. Dates as YYYY-MM-DD. Rate as a number, no symbols.</p><p>EMAIL:</p><p>{body}”””</p><p>def extract(client, body):</p><p>reply = client.complete(PROMPT.format(body=body))</p><p>try:</p><p>return json.loads(reply)</p><p>except json.JSONDecodeError:</p><p>return {“origin”: None, “destination”: None,</p><p>“rate_usd”: None, “pickup_date”: None}</p><p>results = [extract(client, p.read_text()) for p in Path(“emails”).glob(“*.txt”)]</p><p>That is the whole thing. This matters more than the model choice, and I will come back to it.</p><p>Which AI model was most accurate?</p><p>Claude landed at 99.3 percent — 794 of 800 fields correct. ChatGPT came in at 97.8 percent, or 782 correct. Gemini finished at 96.4 percent, 771 correct.</p><p>On paper that is a rounding error between them. In a business process automation context it is not. At 300 emails a week, a 96.4 percent field accuracy means roughly 43 wrong fields a week landing in a live system. At 99.3 percent it is about eight. The difference between those two numbers is whether your client trusts the AI workflow or quietly goes back to doing it by hand.</p><p>Where did each model actually fail?</p><p>The failure patterns were more useful than the scores.</p><p>Gemini’s misses were almost entirely date-related. When an email said “pickup Tuesday” with no explicit date, it guessed rather than returning null — 19 of its 29 errors were fabricated dates. That is the worst possible failure mode, because a confidently wrong date looks exactly like a right one.</p><p>ChatGPT’s errors clustered on rate extraction in emails that listed multiple dollar figures — line haul, fuel surcharge, and total. It picked the wrong one 11 times. Fixable with a sharper prompt, and it was: adding one clarifying sentence cut that to three.</p><p>Claude’s six misses were split between genuinely ambiguous emails and one formatting edge case involving a Canadian postal code. Four of the six, I would have gotten wrong too.</p><p>How much does AI automation actually cost at this scale?</p><p>For 1,000 emails at roughly 900 input tokens each: Claude ran about $1.85, ChatGPT about $1.10, and Gemini about $0.42. Average latency per call was 2.1, 1.6, and 1.2 seconds respectively.</p><p>Now put that against the thing it replaces. Three hundred emails a week at 40 seconds each is 3.3 hours, or roughly $115 a month of admin time at a $22 hourly loaded rate. The most expensive model on the list costs about $2.40 a month for the same volume.</p><p>This is the number people miss when they agonize over model selection. The gap between the cheapest and most expensive option here is under two dollars a month. The gap between automated and manual is 98 percent of the cost. Optimizing the wrong one of those is the most common mistake I see in AI automation for small business.</p><p>Does the model choice even matter, then?</p><p>Less than you think, and not in the way most comparisons suggest.</p><p>Across roughly 40 agent builds, the model has almost never been what broke a project. What breaks projects is the boring layer around it: no schema validation, no null handling, no retry on malformed JSON, no human review queue for low-confidence rows, no logging so you can tell what went wrong three weeks later.</p><p>Note what that dumb script does when JSON parsing fails — it returns a row of nulls instead of crashing. Ninety percent of the reliability in production Python scripts for business is decisions like that, not the model behind the API call.</p><p>The no-code versus code automation debate runs into the same wall. A no-code builder will get you to a working demo faster. It will not get you schema validation, a retry policy, or a review queue. That is the actual dividing line, and it has nothing to do with which model you picked.</p><p>So what should you actually do?</p><p>Pick the cheapest model that clears your accuracy bar, and find that bar by testing on your own data. A hundred hand-labeled examples is enough. It costs you an afternoon and it is the highest-leverage afternoon in the entire project.</p><p>For this client I shipped Claude, because the accuracy gap was worth $1.43 a month. For a different client doing 40,000 support ticket classifications a month, I shipped Gemini, because at that volume the cost gap was real and the accuracy gap was not.</p><p>Neither answer would have been obvious from reading benchmarks. Both took one afternoon of testing.</p><p>If you take one thing from this: the question is never “which AI model is best.” It is “what accuracy do I need, and what is the cheapest way to get there on my data.” Answer that, build the boring error handling around it, and the model becomes a line you can swap in ten seconds.</p><p>Which, given how fast these things ship, is the only sane way to build anyway.</p><p>I’m Archit Mittal — I automate chaos. Follow me here or visit architmittal.com for more.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=3351f0bfbd19" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The 3 AI Automation Trends Reshaping Professional Services in 2026 — And the One Everyone Gets…]]></title>
            <link>https://medium.com/@automate-archit/the-3-ai-automation-trends-reshaping-professional-services-in-2026-and-the-one-everyone-gets-c9547da90ba7?source=rss-9b6215eea7d6------2</link>
            <guid isPermaLink="false">https://medium.com/p/c9547da90ba7</guid>
            <category><![CDATA[artificial-intelligence]]></category>
            <dc:creator><![CDATA[Archit Mittal]]></dc:creator>
            <pubDate>Tue, 11 Aug 2026 15:39:24 GMT</pubDate>
            <atom:updated>2026-08-11T15:39:24.960Z</atom:updated>
            <content:encoded><![CDATA[<h3>The 3 AI Automation Trends Reshaping Professional Services in 2026 — And the One Everyone Gets Wrong</h3><p>I spend most of my week inside other people’s operations. Not Fortune 500 operations — law firms with nine attorneys, accounting practices with six staff, recruiting agencies where one person is quietly holding four processes together with a spreadsheet and a very good memory.</p><p>Halfway through 2026, three patterns keep showing up in almost every engagement. Two of them are genuinely good news. The third is being explained badly by nearly everyone selling AI automation, and that misunderstanding is costing firms real money.</p><p>Here is what I am actually seeing on the ground.</p><h4>Trend 1: Why is writing the automation no longer the hard part?</h4><p>Two years ago, a client asking for a Python automation script to sort inbound email was a two-week project. Today I can produce a working first version in an afternoon. Claude writes most of the boilerplate. So does every other capable model.</p><p>What did not get easier is trusting the output.</p><p>A classifier that is 94% accurate sounds excellent until you push 3,000 items a month through it and realise you now have roughly 180 silently mis-routed records. Nobody notices for six weeks. Then everybody notices at once.</p><p>The builds that survive contact with a real business all share one design decision: they know when they are unsure, and they escalate instead of guessing.</p><pre>def route(item, classifier, threshold=0.85):<br>    result = classifier(item)<br><br>    if result.confidence &lt; threshold:<br>        return {&quot;queue&quot;: &quot;human_review&quot;, &quot;reason&quot;: &quot;low_confidence&quot;}<br><br>    if result.label in HIGH_STAKES_LABELS:<br>        return {&quot;queue&quot;: &quot;human_review&quot;, &quot;reason&quot;: &quot;high_stakes&quot;}<br><br>    return {&quot;queue&quot;: &quot;auto&quot;, &quot;label&quot;: result.label}</pre><p>Across my last twelve builds, roughly 15–20% of items land in that human review queue. Clients sometimes see that number and feel short-changed. They should not. That 18% is precisely what makes the other 82% safe to run unattended. An automation that handles four-fifths of the volume and never quietly corrupts a record is worth far more than one that claims 100% and occasionally does.</p><p>If you take one thing from this article: in 2026, the value of an AI workflow sits in the escalation path, not the model.</p><h4>Trend 2: What happens when integration work stops being expensive?</h4><p>The unglamorous truth of business process automation has always been that most of the effort goes into plumbing. Authenticating to a CRM. Discovering that the accounting system paginates its API differently depending on which endpoint you ask. Rebuilding the whole thing when the vendor ships a new version.</p><p>Standardised tool interfaces — MCP being the obvious example — are eating a large share of that work. A connection that took me three days to build and babysit in 2024 now often takes an afternoon.</p><p>The second-order effect matters more than the time saved. When integration was expensive, only large processes justified automation: the 200-hour-a-year ones. Now a process that consumes 15 hours a year can pay for itself inside a quarter. The floor dropped.</p><p>Practically: go back to the list of tasks you decided were too small to bother automating in 2024. About half of that list is worth revisiting.</p><h4>Trend 3: Is the no-code vs code automation debate the wrong argument?</h4><p>This is the one almost everyone gets wrong.</p><p>The question I am asked most often is whether a business should use a no-code platform or pay someone to write Python scripts for business processes. It gets framed as a budget decision, or a skills decision. It is neither.</p><p>The real dividing line is whether your process is stateless or stateful.</p><p>A stateless process is a straight line. Form submitted, record created, notification sent. Each run is independent of every other run. If it fails, you re-run it and nothing is worse off. No-code tools are genuinely excellent here, and I tell clients to use them. Nobody should pay a consultant to rebuild a form-to-database connection.</p><p>A stateful process has memory and consequences. It waits. It retries. It knows this invoice was already chased twice, that this candidate withdrew last week, that a partial failure at step four means steps one through three need unwinding. And it has to be auditable in six months when somebody asks why a particular decision was made.</p><p>Stateful processes are where no-code tools quietly break, and they break in the most expensive way available: not with an error, but with a slightly wrong result that nobody catches.</p><p>So the honest rule. Stateless: use no-code and be happy. Stateful: use code. That is a decision about the shape of the process, not the size of your budget.</p><h4>How much does AI automation cost for a small business in 2026?</h4><p>These are the ranges from my own automation consulting projects, so treat them as one practitioner’s data rather than an industry survey. A single well-scoped process — document intake, email triage, invoice chasing, recurring report generation — typically lands between $2,000 and $8,000 to build. Ongoing model and infrastructure costs for that process usually run $20 to $150 a month, because most small-business volumes are tiny by API standards. Multi-process builds with shared infrastructure run higher, $15,000 to $40,000, but the cost per process falls sharply after the first one.</p><p>The number that actually matters is not the build cost. It is hours reclaimed per month multiplied by twelve. If a process eats 10 hours a month at a loaded cost of $40 an hour, that is $4,800 a year. Under a one-year payback, it is usually an easy yes. Over two years, do not build it yet.</p><h4>Can AI replace a virtual assistant?</h4><p>No, and the firms that try it are disappointed inside a month.</p><p>A good assistant absorbs ambiguity. They notice that a client sounded annoyed, that a request contradicts something agreed last week, that a form was filled in wrong on purpose. Current AI systems do none of that reliably.</p><p>What AI does take over is the mechanical 60–70% of an assistant’s task list: the transcription, the routing, the formatting, the reminder chasing. The firms that get this right do not cut headcount. They stop assigning their assistant retyping work and start assigning them client relationships. Same payroll, considerably more output.</p><h4>How do you start automating business processes without hiring a developer?</h4><p>Three steps, in this order.</p><p>One. For two weeks, log every task anyone does more than five times a week. Do not filter, do not judge. The finished list will surprise you.</p><p>Two. Pick the single most boring item on it — clear inputs, clear outputs, no judgement calls. Boring is the whole point. Boring processes are where automation is safest and payback is fastest.</p><p>Three. Build the smallest possible version and run it alongside the manual process for two weeks. Compare outputs daily. Only switch the manual version off once the two agree consistently.</p><p>Almost every failed automation project I get called in to rescue skipped step three.</p><h4>What this means for the rest of 2026</h4><p>Pulling the three together: building an AI workflow is cheap now, connecting it to your systems is cheap now, and the only decision that still carries real weight is which processes you trust it with. Stateless work goes to no-code. Stateful work goes to code with a proper escalation path. Everything else is noise.</p><p>The firms pulling ahead this year are not the ones running the most sophisticated AI. They are the ones that automated four boring processes properly, measured the result, and left the interesting work to humans. That is a far less exciting story than the one being sold. It is also the one that works.</p><p>I’m Archit Mittal — I automate chaos. Follow me here or visit architmittal.com for more.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=c9547da90ba7" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>