<?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 Maharshi B on Medium]]></title>
        <description><![CDATA[Stories by Maharshi B on Medium]]></description>
        <link>https://medium.com/@jobs.maharshi?source=rss-d58d7b0b71c6------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/0*ZlLnFXpCAgo8eiv7</url>
            <title>Stories by Maharshi B on Medium</title>
            <link>https://medium.com/@jobs.maharshi?source=rss-d58d7b0b71c6------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Thu, 19 Nov 2026 14:11:41 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@jobs.maharshi/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 AI Product Managers Actually Need in 2027: Evals, Failure Modes, and Model Choice]]></title>
            <link>https://medium.com/@jobs.maharshi/what-ai-product-managers-actually-need-in-2027-evals-failure-modes-and-model-choice-76a606884a45?source=rss-d58d7b0b71c6------2</link>
            <guid isPermaLink="false">https://medium.com/p/76a606884a45</guid>
            <category><![CDATA[ai-product-management]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[careers]]></category>
            <category><![CDATA[product-management]]></category>
            <dc:creator><![CDATA[Maharshi B]]></dc:creator>
            <pubDate>Tue, 17 Nov 2026 16:03:33 GMT</pubDate>
            <atom:updated>2026-11-17T16:03:33.200Z</atom:updated>
            <content:encoded><![CDATA[<h4>A practical AI product manager skills roadmap for senior BAs and Product people moving into AI product work</h4><p>The AI product manager skills that matter most in 2027 are not prompt writing. They are knowing when AI is the wrong answer, knowing how it will break, and owning the evals that define what “good” looks like before anything ships.</p><p>I have watched strong BAs and Product people stall on AI work because they studied the wrong things. They learned model names and tool lists. What teams actually needed from them was judgment: which approach to pick, what could go wrong, and how to prove it works.</p><p>This piece is a skills roadmap, not a job-hunt guide. If you already write solid requirements, run discovery, and manage trade-offs with engineering, you are closer than you think. You need three more skills, and you can practise all of them in 90 days.</p><h3>AI Product Manager vs AI Product Owner</h3><p>The AI Product Owner lives close to the backlog and the delivery team. They turn AI work into stories, acceptance criteria, and sprint decisions.</p><p>The AI Product Manager owns the bigger questions. Should we use AI here at all? Which approach? What is the risk if it is wrong? How will we measure quality over time? Many people do both jobs. The skills below are the PM side.</p><h3>Skill 1: Model and approach selection logic</h3><p>Most bad AI products I have seen started with “we should add AI” instead of “what problem are we solving.” Your first job is to choose the simplest approach that works.</p><p>I use a plain decision path:</p><ul><li><strong>No AI.</strong> If the rules are clear and stable, write the rules. If a lookup table, a search filter, or a form fixes it, use that. It is cheaper, faster, and easier to test. Saying “we do not need a model for this” is a real product decision, and teams respect the person who makes it.</li><li><strong>Prompt-only.</strong> If the task is general (summarising, rewriting, classifying, drafting) and does not depend on your company’s private facts, a well-written prompt on a capable model is often enough. Start here before building anything heavier.</li><li><strong>RAG (retrieval plus generation).</strong> If the answer must come from your own documents, policies, or data, and those change often, you need retrieval. The model reads the right sources at request time and answers from them. Most “chat with our knowledge base” products belong here.</li><li><strong>Fine-tuning or a custom model.</strong> Only when you have a narrow, repeated task, a lot of good examples, and proof that prompting plus retrieval fall short. This is rarely the first move.</li></ul><p>Then add three checks that are pure product thinking:</p><ul><li><strong>Cost per request.</strong> Does the unit cost still work at real volume?</li><li><strong>Latency.</strong> Will users wait for this answer, or does the flow need to be instant?</li><li><strong>Data sensitivity.</strong> Can this data leave our environment? Who approves that?</li></ul><p>If you can walk a stakeholder through this path in ten minutes, you are already doing AI product management.</p><h3>Skill 2: Failure-mode thinking</h3><p>Traditional software fails loudly. AI fails quietly and confidently. That changes how you write requirements.</p><p>These are the four failure modes I ask every team to plan for:</p><ul><li><strong>Hallucination.</strong> The model invents a fact, a policy, or a source. Ask: what happens if the answer is made up? In a brainstorming tool, not much. In a benefits or compliance tool, a lot.</li><li><strong>Wrong but confident.</strong> Worse than hallucination, because the tone hides the error. Users trust fluent answers. Your design needs a way to show sources, signal uncertainty, or route edge cases to a human.</li><li><strong>Latency.</strong> A correct answer that takes too long is a failed feature. Set a time budget early and design around it: streaming, smaller models for simple steps, or async flows.</li><li><strong>Prompt injection.</strong> Users, documents, or web pages can contain instructions that hijack the model. The <a href="https://genai.owasp.org/llmrisk/llm01-prompt-injection/">OWASP Top 10 for LLM Applications</a> lists it as the first risk for good reason. As a PM, you do not need to build the defence, but you must ask: what can this system do if someone tricks it? If it can send emails, change records, or reveal private data, that needs limits.</li></ul><p>My simple habit: for every AI feature, write a short “how this breaks” section next to the user stories. One line per failure mode, the impact, and the guardrail. It takes 30 minutes and saves weeks. If you want the organisation-level view of the same idea, I covered it in <a href="https://medium.com/@jobs.maharshi/ai-governance-is-not-a-policy-document-a-practical-enterprise-ai-governance-framework-18ee2fbc7db7">AI Governance Is Not a Policy Document</a>.</p><h3>Skill 3: Owning evals and “what good looks like”</h3><p>This is the skill that separates people who talk about AI from people who ship it.</p><p>An eval is a repeatable test of output quality. Engineers can build the harness. But someone has to decide what a good answer is, and that is a product decision. If no one owns it, quality becomes opinion, and every demo turns into an argument.</p><p>Here is how I approach it:</p><ol><li><strong>Collect real examples.</strong> Pull 30 to 50 real inputs from users or support logs. Include the messy ones.</li><li><strong>Write the expected answer or the rules for a good one.</strong> Correct facts, right tone, cites a source, refuses when it should.</li><li><strong>Define pass and fail in plain language.</strong> “Answers using only the policy document” is testable. “Sounds helpful” is not.</li><li><strong>Set a bar before launch.</strong> Agree with the team what level is good enough to ship, and what is a hard stop (for example, any answer that exposes private data).</li><li><strong>Re-run on every change.</strong> New prompt, new model, new documents. Same test set. Track the results over time.</li></ol><p>Senior BAs are well suited to this. Writing acceptance criteria for fuzzy outputs is still writing acceptance criteria. You are just testing behaviour instead of buttons. For the leadership side of measuring AI value, see <a href="https://medium.com/@jobs.maharshi/how-to-measure-ai-roi-in-2027-a-one-page-scorecard-for-cios-d78d54512a8b">How to Measure AI ROI in 2027</a>.</p><h3>A 90-day practice plan for AI product manager skills</h3><p>You do not need a new job title to build these skills. Use your current role or a side project.</p><ul><li><strong>Days 1 to 30: Selection.</strong> Pick three problems at work or in your own life. For each, write a one-page decision: no AI, prompt-only, RAG, or fine-tune, with cost, latency, and data notes. Build one prompt-only prototype with any general AI tool. Notice where it fails.</li><li><strong>Days 31 to 60: Failure modes.</strong> Take that prototype and try to break it. Feed it bad inputs, missing context, and a document with hidden instructions. Write a “how this breaks” page with guardrails. Share it with an engineer and ask what you missed.</li><li><strong>Days 61 to 90: Evals.</strong> Build a 30-example test set in a spreadsheet. Define pass and fail for each. Score your prototype, change one thing (the prompt, the model, or the source documents), and score it again. Write a one-page summary: what improved, what got worse, and whether you would ship.</li></ul><p>At the end you have three artefacts you can show in any interview or internal pitch: a selection memo, a failure-mode page, and an eval report. That is a stronger portfolio than a certificate.</p><h3>The AI product manager checklist</h3><ul><li>I can explain when not to use AI, with a real example</li><li>I can choose between prompt-only, RAG, and fine-tuning and defend the choice</li><li>I check cost per request, latency, and data sensitivity before committing</li><li>I write a “how this breaks” section for every AI feature</li><li>I have a plan for hallucination and wrong-but-confident answers</li><li>I ask what the system can do if prompt injection succeeds</li><li>I own a written definition of “what good looks like”</li><li>I have built and scored a test set of real examples</li><li>I agree a launch bar and hard stops with the team before release</li><li>I re-run evals after every prompt, model, or data change</li></ul><p>If you can tick most of these, you are doing the job, whatever your title says.</p><h3>FAQ</h3><h4>What skills does an AI product manager need?</h4><p>Three matter most: choosing the simplest approach that works (no AI, prompt-only, RAG, or fine-tuning), planning for failure modes like hallucination and prompt injection, and owning evals that define what a good answer looks like before launch.</p><h4>What is the difference between an AI Product Manager and an AI Product Owner?</h4><p>The AI Product Owner works close to the backlog, turning AI work into stories and acceptance criteria. The AI Product Manager owns the bigger questions: whether to use AI at all, which approach, what the risk is, and how quality is measured over time.</p><h4>What is an AI eval?</h4><p>An eval is a repeatable test of output quality. You collect real examples, define pass and fail in plain language, set a launch bar, and re-run the same test set after every change to the prompt, model, or documents.</p><h3>A note from Aarambh</h3><p>At Aarambh, we work with senior BA, Product, and Data people who bring exactly this kind of judgment to AI delivery teams. If you are building these skills and want to be on our radar for upcoming roles, you can reach us at <a href="https://www.aarambhconsulting.ca/">aarambhconsulting.ca</a>.</p><p>Maharshi B, founder of Aarambh Consulting Group, a Canadian-owned and operated firm.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=76a606884a45" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[How to Measure AI ROI in 2027: A One-Page Scorecard for CIOs]]></title>
            <link>https://medium.com/@jobs.maharshi/how-to-measure-ai-roi-in-2027-a-one-page-scorecard-for-cios-d78d54512a8b?source=rss-d58d7b0b71c6------2</link>
            <guid isPermaLink="false">https://medium.com/p/d78d54512a8b</guid>
            <category><![CDATA[roi]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[ai-strategy]]></category>
            <category><![CDATA[ai-governance]]></category>
            <category><![CDATA[digital-transformation]]></category>
            <dc:creator><![CDATA[Maharshi B]]></dc:creator>
            <pubDate>Tue, 10 Nov 2026 16:10:49 GMT</pubDate>
            <atom:updated>2026-11-10T16:10:49.949Z</atom:updated>
            <content:encoded><![CDATA[<h4>A practical way to measure AI ROI by use case: a baseline, four numbers that matter, and a stop rule your CFO will trust.</h4><p>To measure AI ROI in 2027, measure one use case at a time. Capture a baseline before launch, then track four things every month: the business outcome, real adoption, quality and risk, and the full cost, including people and governance. Agree a stop rule in advance. Token counts and seat licences are activity, not ROI.</p><p>I keep hearing the same question from delivery leaders and CIOs this fall: “We have spent real money on AI. What did we get?” Most teams can show usage dashboards. Few can show a number their CFO would sign.</p><p>The pressure is about to get louder. On October 14, 2026, <a href="https://www.gartner.com/en/newsroom/press-releases/2026-10-14-gartner-says-over-us-dollars-three-and-a-half-trillion-of-artificial-intelligence-infrastructure-spending-could-be-exposed-to-impairment-by-the-end-of-2030">Gartner said</a> AI data-centre owners would need a cumulative $8.5 trillion in revenue through 2030 to earn their target return, against no more than $3.6 trillion it expects AI to generate. That is a provider problem first. But when the people selling AI are under pressure to prove value, the people buying it will be asked to prove it too.</p><p>Here is the one-page approach I use to measure AI ROI without a six-month consulting project.</p><h3>Why is measuring AI ROI so hard right now?</h3><p>Because most organizations track AI activity instead of AI outcomes, and they skip the baseline. Without a “before” number, any “after” number is a guess.</p><p>There are three common traps:</p><ul><li>Measuring usage. Logins, prompts and tokens show that people touched the tool. They do not show that work got cheaper, faster or better. <a href="https://kpmg.com/us/en/articles/2026/new-economics-of-ai.html">KPMG makes the same point</a>: AI activity is not a proxy for AI value.</li><li>Measuring the portfolio, not the use case. “Our AI program saved 4% across the company” cannot be checked. “Claims triage cut handling time on simple claims” can.</li><li>Counting only the licence. Integration, evaluation, human review, security and governance work are real costs. Leave them out and ROI looks great until finance looks closer.</li></ul><p>The cost of getting this wrong is not theoretical. MIT’s Project NANDA report on generative AI in business, <a href="https://fortune.com/2025/08/18/mit-report-95-percent-generative-ai-pilots-at-companies-failing-cfo/">as reported by Fortune</a>, found that most of the enterprise pilots it studied produced no measurable impact on profit and loss. People still debate the method. The lesson holds: if you do not design for measurement, you will not have evidence.</p><h3>What should you measure before you launch an AI use case?</h3><p>Measure the current process exactly as it runs today, for at least two to four weeks, using the same definitions you will use after launch.</p><p>For each use case, write down:</p><ul><li>The unit of work. One claim, one ticket, one requirement document, one contract review.</li><li>Volume. How many units per week.</li><li>Cycle time. Start to finish, per unit.</li><li>Cost per unit. People time multiplied by a loaded rate, plus tools.</li><li>Quality. Error rate, rework rate, escalations or customer complaints.</li><li>The decision you want to change. Faster approval? Fewer handoffs? Better first answer?</li></ul><p>If you cannot define the unit of work, the use case is not ready for an AI budget yet.</p><h3>How do you measure AI ROI on one page?</h3><p>Four lines, each with a named owner. If a line has no owner, nobody will defend the number.</p><ul><li>Outcome: Did the business result change? Example metrics: cost per case, cycle time, revenue per rep, error rate. Owner: process or business owner.</li><li>Adoption: Are people really relying on it? Example metrics: share of eligible work that goes through AI, override rate. Owner: product owner or ops lead.</li><li>Quality and risk: Is it safe and accurate enough? Example metrics: evaluation pass rate, incidents, complaints, policy breaches. Owner: technical owner with risk owner.</li><li>Full cost: What does it really cost to run? Example metrics: licences, inference, integration, review time, governance effort. Owner: finance partner.</li></ul><p>McKinsey <a href="https://www.mckinsey.com/capabilities/quantumblack/our-insights/from-promise-to-impact-how-companies-can-measure-and-realize-the-full-value-of-ai">published in April 2026</a> a more detailed five-layer model that ties technical performance to financial impact with clear owners. For a mid-sized company, four lines on one page will get you most of the way.</p><p>Watch the override rate. If people accept AI output and then quietly redo it, adoption looks high and value is close to zero.</p><h3>How do you count the full cost of an AI use case?</h3><p>Add up everything you would not be spending if the use case did not exist, including people time. Then compare it with the value you measured, not the value in the original pitch.</p><p>A practical cost checklist:</p><ul><li>Licences, seats and API or inference usage</li><li>Integration and data preparation work</li><li>Evaluation and testing, before launch and ongoing</li><li>Human review and exception handling</li><li>Security, privacy and vendor review</li><li>Monitoring, incident handling and retraining or prompt updates</li><li>Change management and training</li></ul><p><a href="https://www.cio.com/article/4230264/3-best-practices-for-cios-to-assess-ai-roi.html">CIO.com’s guidance for CIOs</a> makes a point I agree with: factor in governance costs early, before an initiative is funded, especially for high-risk applications.</p><h3>When should you stop or scale an AI use case?</h3><p>Decide the rule before launch: scale when the outcome line beats baseline by the agreed margin for two review cycles, and stop when it does not move after a fixed window.</p><p>A simple rule I use:</p><ul><li>Scale if the outcome metric beats baseline by your target for two months in a row, adoption is rising and quality is within tolerance.</li><li>Fix if adoption is low but the users who rely on it see results. That is usually a workflow or training problem.</li><li>Stop if, after 90 days, the outcome line has not moved, or the full cost is above the value.</li></ul><p>The guidance from <a href="https://www.uts.edu.au/news/2026/09/ai-roi-a-practical-framework-for-assessing-the-return-on-ai-investments/contentassets/uts_hti_kpmg-ai-roi-framework_web.pdf">UTS and KPMG’s AI ROI framework</a> says the same in board language: be willing to stop if benefits are not being realised or costs climb above them. Stopping is not failure.</p><h3>How does AI governance connect to AI ROI?</h3><p>They use the same evidence. The inventory, owners, risk tier and monitoring that governance needs are exactly what you need to measure value.</p><p>The <a href="https://www.nist.gov/itl/ai-risk-management-framework">NIST AI Risk Management Framework</a> organizes AI risk work into four functions: Govern, Map, Measure and Manage. The “Measure” work you do for risk, such as evaluation results, incidents and drift, sits on the same dashboard as your ROI numbers. When I wrote that <a href="https://medium.com/@jobs.maharshi/ai-governance-is-not-a-policy-document-a-practical-enterprise-ai-governance-framework-18ee2fbc7db7">AI governance is not a policy document</a>, this is what I meant: governance that produces evidence also produces your value story. If you are starting from zero, my <a href="https://medium.com/@jobs.maharshi/ai-governance-in-90-days-a-starter-guide-for-mid-sized-organizations-f43fad99ca51">90-day governance starter guide</a> covers the inventory and owners step by step.</p><h3>What does a 30-day plan to start measuring AI ROI in 2027 look like?</h3><p>Pick your top three AI use cases by spend and give each one a baseline, a scorecard and a review date within a month.</p><ul><li>Week 1: List every AI use case with spend attached. Pick the three largest.</li><li>Week 2: Define the unit of work and pull a baseline for each one.</li><li>Week 3: Build the four-line scorecard and name an owner for every line.</li><li>Week 4: Hold the first review. Write down the scale, fix or stop rule and the date you will apply it.</li></ul><h3>FAQ: measuring AI ROI</h3><h4>What is a good ROI for an AI project?</h4><p>There is no universal number. A good ROI beats your organization’s normal hurdle rate for technology investments once the full cost, including human review and governance, is counted.</p><h4>How long does it take to see ROI from AI?</h4><p>It depends on the use case, but you should see movement in adoption and the outcome metric within one to three months. If nothing moves after about 90 days, review the use case before spending more.</p><h4>Should we measure generative AI ROI differently from traditional machine learning?</h4><p>The scorecard is the same. Generative AI usually needs more attention on the quality and risk line, because outputs vary and human review costs can be large.</p><h4>Is time saved a valid AI ROI metric?</h4><p>Only if the saved time turns into something measurable, such as more cases handled, faster delivery or lower overtime. Hours “saved” that disappear into the day are not ROI.</p><h4>Who should own AI ROI measurement?</h4><p>The business owner of the process owns the outcome. Finance validates the cost and value, and the technical owner provides the quality evidence.</p><p>A note from Aarambh: At Aarambh Consulting Group, our AI Governance and Production Readiness assessment helps teams build the inventory, owners and evidence that both governance and ROI reporting depend on. If you are being asked to prove AI value this budget season, <a href="https://www.aarambhconsulting.ca/">we are happy to compare notes</a>.</p><p>Maharshi B, founder of Aarambh Consulting Group, a Canadian-owned and operated firm.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=d78d54512a8b" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[How to Become a Forward Deployed Engineer in 2027: The Complete Roadmap]]></title>
            <link>https://medium.com/@jobs.maharshi/how-to-become-a-forward-deployed-engineer-in-2027-the-complete-roadmap-4a16c563bb94?source=rss-d58d7b0b71c6------2</link>
            <guid isPermaLink="false">https://medium.com/p/4a16c563bb94</guid>
            <category><![CDATA[ai-career]]></category>
            <category><![CDATA[career-development-tips]]></category>
            <category><![CDATA[software-engineering-job]]></category>
            <category><![CDATA[forward-deployed-engineer]]></category>
            <category><![CDATA[generative-ai-consulting]]></category>
            <dc:creator><![CDATA[Maharshi B]]></dc:creator>
            <pubDate>Tue, 03 Nov 2026 16:03:25 GMT</pubDate>
            <atom:updated>2026-11-03T16:03:25.286Z</atom:updated>
            <content:encoded><![CDATA[<h4>What a forward deployed engineer really does, how the role differs from solutions engineering and consulting, and a 6–12 month plan to get there.</h4><p>A forward deployed engineer (FDE) is a software engineer who embeds with one customer and ships production AI inside their systems, data, and workflows. To become one in 2027, you need strong production coding, data skills, LLM and agent integration, and customer-facing problem solving, plus a portfolio of real deployments that people actually use.</p><p>A few months ago I wrote <a href="https://medium.com/@jobs.maharshi/how-to-become-an-ai-product-owner-in-2026-the-complete-roadmap-9415317aef74">How to Become an AI Product Owner: The Complete Roadmap</a>. The most common question I got afterwards came from engineers: “What’s the equivalent path for someone who wants to keep writing code?”</p><p>For a lot of them, the answer is the Forward Deployed Engineer. The AI Product Owner decides <em>what</em> to build and why. The FDE sits at the customer’s site (or in their Slack, or their data warehouse) and makes the thing actually work.</p><h3>What Is a Forward Deployed Engineer?</h3><p>The role started at Palantir. Its early customers were government and intelligence agencies whose data and workflows couldn’t be understood from a sales call, so Palantir sent engineers in to learn the problem by building inside it. Internally those engineers were called “Deltas” and were paired with domain experts called “Echos.” Palantir’s own framing, <a href="https://newsletter.gtmengineering.ai/p/forward-deployed-engineering-101">quoted widely</a>, is the clearest definition I know: <em>“A Dev’s focus is one capability for many customers, while a Delta’s focus is many capabilities for one customer.”</em></p><p>Heading into 2027, the same model shows up at AI labs and AI startups. <a href="https://openai.com/careers/forward-deployed-engineer-%28fde%29-sf-san-francisco/">OpenAI’s FDE posting</a> describes FDEs who “lead complex end-to-end deployments of frontier models in production alongside our most strategic customers,” owning discovery, scoping, system design, build, and rollout. <a href="https://job-boards.greenhouse.io/anthropic/jobs/5423029008">Anthropic’s Applied AI FDE role</a> asks engineers to embed with strategic customers and build production applications on Claude. Scale AI, Google Cloud, Databricks, and a long tail of vertical AI startups hire for the same shape of job, often under titles like “Applied AI Engineer” or “Deployment Engineer.”</p><p>Put simply, an FDE:</p><ul><li>Writes production code against one customer’s messy data and systems</li><li>Wires LLMs and agents into real workflows</li><li>Owns whether the thing gets used, not just whether it compiles</li><li>Feeds what they learn back to the product team</li></ul><h3>Why Is the Forward Deployed Engineer Role Growing?</h3><p>Two reasons, and they’re connected.</p><p><strong>Models are easy to call and hard to deploy.</strong> Anyone can get a demo running in an afternoon. Getting an agent to work against a 15-year-old ERP, with access controls, evals, and users who don’t trust it, is the hard part. Gartner <a href="https://www.gartner.com/en/newsroom/press-releases/2024-07-29-gartner-predicts-30-percent-of-generative-ai-projects-will-be-abandoned-after-proof-of-concept-by-end-of-2025">predicted in 2024</a> that at least 30% of generative AI projects would be abandoned after proof of concept by the end of 2025. FDEs exist to close that gap.</p><p><strong>Hiring has followed.</strong> Indeed data reported by Business Insider showed FDE postings rising from about 643 in April 2025 to more than 5,300 in April 2026, <a href="https://fdeinterviews.com/guide/forward-deployed-software-engineer-fdse">as summarized here</a>. Treat any single number with care, but the direction is not in doubt.</p><p>On pay, I’ll only point to what’s published: OpenAI’s San Francisco FDE posting lists a base range of <a href="https://openai.com/careers/forward-deployed-engineer-%28fde%29-sf-san-francisco/">$185K–$300K USD plus equity</a>. Canadian and non-lab roles will look different, so check the posting in front of you.</p><h3>FDE vs. Solutions Engineer vs. Consultant: What’s the Difference?</h3><p>These get confused constantly. Here’s how I separate them:</p><ul><li><strong>Solutions engineer (sales engineer):</strong> Works pre-sale. Runs demos, answers technical questions, builds proofs of concept to win the deal. Success = deal closed.</li><li><strong>Consultant:</strong> Advises, designs, and often manages delivery. May or may not write the code. Success = engagement delivered to scope.</li><li><strong>Forward Deployed Engineer:</strong> Works post-sale (or mid-sale on large deals). Writes and ships production code inside the customer’s environment. Success = the system is live and people use it. OpenAI’s posting literally says success is measured through “production adoption, measurable workflow impact, and eval-driven feedback.”</li></ul><p>The test I use: if you’re not on the hook when it breaks in production, you’re probably not doing FDE work.</p><h3>What Skills Does a Forward Deployed Engineer Need?</h3><p>Reading the OpenAI and Anthropic postings side by side, the overlap is clear:</p><ul><li><strong>Production coding.</strong> Python is the default; TypeScript or Java helps. OpenAI asks for production-grade code “across frontend and backend.” You’ll build internal tools, APIs, and UIs, not just notebooks.</li><li><strong>Data work.</strong> SQL, data cleaning, joining systems that were never meant to talk. Most deployments live or die on data access.</li><li><strong>LLM and agent integration.</strong> Anthropic lists “advanced prompt engineering, agent development, evaluation frameworks, and deployment at scale,” plus building MCP servers. Know RAG, tool calling, structured outputs, and how to write evals.</li><li><strong>Customer-facing problem solving.</strong> Discovery, scoping, saying “no” politely, and turning a vague request into a two-week build plan.</li><li><strong>Enterprise basics.</strong> SSO, permissions, security reviews, logging. Anthropic calls enterprise IT experience “a plus.” In my experience it’s what gets you through the customer’s security review. It also helps to understand how the customer governs AI; I wrote about that in <a href="https://medium.com/@jobs.maharshi/ai-governance-is-not-a-policy-document-a-practical-enterprise-ai-governance-framework-18ee2fbc7db7">AI Governance Is Not a Policy Document</a>.</li></ul><p>Note the experience bar: OpenAI asks for 5+ years, and Anthropic’s London role asks for 4+ years in a technical, customer-facing role. Former technical founders are explicitly encouraged to apply. Palantir does hire new grads, so that’s the main entry-level door.</p><h3>The 2027 Forward Deployed Engineer Roadmap: 6–12 Months</h3><p>This assumes you can already code. If you can’t, add six months of fundamentals first.</p><h4>Step 1 (Months 1–2): Close Your Gaps</h4><p>Be honest about which side you’re weaker on.</p><ul><li><strong>Strong engineer, little customer exposure?</strong> Volunteer for internal stakeholder work. Sit in on support calls. Run a requirements session.</li><li><strong>Strong with customers, rusty code?</strong> Ship one full-stack app end to end: API, database, simple UI, deployed somewhere real.</li></ul><h4>Step 2 (Months 2–4): Learn LLM Engineering Properly</h4><p>Go past tutorials.</p><ul><li>Build a RAG app over a document set you know well, then measure retrieval quality instead of eyeballing it</li><li>Build an agent that calls three or more tools, with retries and error handling</li><li>Write an MCP server that exposes one internal-style system (a ticket queue, a CRM export)</li><li>Write an eval suite: 50–100 test cases with expected outputs, run on every change</li></ul><h4>Step 3 (Months 4–7): Build a Deployment Portfolio</h4><p>Hiring loops screen for evidence that you’ve shipped something to a named user and owned whether it worked. Tutorials won’t show that. Real users will.</p><p>Portfolio ideas that read like FDE work:</p><ul><li><strong>Invoice or claims triage agent</strong> for a small business you know: reads PDFs, extracts fields, routes exceptions to a human</li><li><strong>Internal knowledge assistant</strong> for a non-profit or community group, with access controls and an eval report</li><li><strong>CRM clean-up agent</strong> that proposes merges and fixes, with a human approval queue</li><li><strong>Legacy system bridge:</strong> an MCP server over a CSV-and-email workflow, so an LLM can query it safely</li></ul><p>For each, write a one-page case study: the problem, the messy part, what you built, how you measured it, and what you’d do next. That page matters as much as the code.</p><h4>Step 4 (Months 6–9): Get Customer-Facing Reps</h4><p>Join an implementation project at work, or do two or three small deployments for local businesses. Real users teach you what side projects can’t: unclear requirements and people who don’t want the tool.</p><h4>Step 5 (Months 8–12): Position and Apply</h4><ul><li><strong>Titles to search:</strong> Forward Deployed Engineer, Forward Deployed AI Engineer, Applied AI Engineer, Deployment Engineer, AI Solutions Engineer (read the description carefully)</li><li><strong>Resume framing:</strong> lead with deployments and outcomes, not tech stacks. “Shipped X to Y users, cut Z from N hours to M” (only with numbers you can defend)</li><li><strong>Interview prep:</strong> expect a coding round, a system design round on a realistic customer scenario, and a case-style round where you scope an ambiguous problem out loud. Practice talking while you think.</li><li><strong>Travel and location:</strong> OpenAI’s postings mention up to 50% travel and a hybrid office model. Decide upfront whether that works for your life.</li></ul><h3>Who Is the FDE Role Not For?</h3><p>Bob McGrew, Palantir’s early head of product engineering and later at OpenAI, has said the person not suited for it is someone who “loves ensuring their abstractions are perfect and their code is maintainable for 12 years” (<a href="https://zerofuturetech.substack.com/p/fde-the-hottest-engineering-job-youve">quoted here</a>). If “we’ll probably rewrite this” makes you anxious, platform or core product roles may suit you better. That’s a different career, not a lesser one.</p><h3>FAQ</h3><h4>What does a forward deployed engineer do?</h4><p>An FDE writes and ships production code inside one customer’s environment, wiring LLMs and agents into their real data and workflows. Success is measured by whether the system goes live and people actually use it, and the FDE feeds what they learn back to the product team.</p><h4>Is a forward deployed engineer the same as a solutions engineer?</h4><p>No. A solutions engineer works pre-sale, running demos and proofs of concept to win the deal. An FDE works post-sale (or mid-sale on large deals) and is on the hook when the system breaks in production.</p><h4>How long does it take to become a forward deployed engineer?</h4><p>If you can already code, plan on roughly 6–12 months: close your skill gaps, learn LLM engineering properly, build a deployment portfolio, get customer-facing reps, then position and apply. If you can’t code yet, add about six months of fundamentals first.</p><h3>Final Thought</h3><p>The AI Product Owner decides what’s worth building. The FDE makes it real in someone else’s building. If you like code <em>and</em> people, and you’d rather see a tool used on Monday than polished for months, this is one of the best seats in AI right now.</p><p><em>At </em><a href="https://www.aarambhconsulting.ca/"><em>Aarambh Consulting Group</em></a><em>, we help teams with AI Governance and Production Readiness assessments, and we place senior BA, Product, and Data talent. If you’re a senior practitioner moving toward deployment-style work and want to hear about contract roles, feel free to reach out.</em></p><p><em>Maharshi B, founder of Aarambh Consulting Group, a Canadian-owned and operated firm.</em>How to Become a Forward Deployed Engineer in 2027: The Complete Roadmap</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=4a16c563bb94" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[AI Governance in 90 Days: A Starter Guide for Mid-Sized Organizations]]></title>
            <link>https://medium.com/@jobs.maharshi/ai-governance-in-90-days-a-starter-guide-for-mid-sized-organizations-f43fad99ca51?source=rss-d58d7b0b71c6------2</link>
            <guid isPermaLink="false">https://medium.com/p/f43fad99ca51</guid>
            <category><![CDATA[risk-management]]></category>
            <category><![CDATA[it-leadership]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[responsible-ai]]></category>
            <category><![CDATA[ai-governance]]></category>
            <dc:creator><![CDATA[Maharshi B]]></dc:creator>
            <pubDate>Tue, 27 Oct 2026 14:55:45 GMT</pubDate>
            <atom:updated>2026-10-27T14:55:45.737Z</atom:updated>
            <content:encoded><![CDATA[<h4>You don’t need an AI governance department. You need an inventory, risk tiers, and names next to every AI system. Here’s a 30/60/90-day plan for organizations with 100 to 1,500 people.</h4><p>How do you start AI governance in a mid-sized organization? Give it 90 days. Spend the first 30 finding every AI tool in use, including vendor AI features. Spend the next 30 sorting them into three risk tiers with proportionate controls. Spend the last 30 naming an owner for every medium- and high-risk system.</p><p>In a separate piece, I argue that <a href="https://medium.com/@jobs.maharshi/ai-governance-is-not-a-policy-document-a-practical-enterprise-ai-governance-framework-18ee2fbc7db7">AI governance is not a policy document</a>. It’s a lifecycle: <strong>Discover, Assess, Control, Own, Monitor.</strong></p><p>The most common reply I get from IT leaders at mid-sized organizations is: “Fine. But where do we start on Monday?”</p><p>This is that answer. It’s written for organizations big enough to have real AI risk (people pasting client data into chatbots, vendors switching on AI features in tools you already pay for) but too small to hire a Chief AI Officer and a team of five.</p><h3>What Do Mid-Sized Organizations Actually Need for AI Governance?</h3><p>Not everything the big banks have. You need five things:</p><ol><li><strong>A short acceptable-use policy</strong> people will actually read</li><li><strong>A model and use-case inventory</strong>, including AI inside vendor tools</li><li><strong>Risk tiers</strong>, so a meeting-notes summarizer and a hiring screener don’t go through the same process</li><li><strong>Named owners</strong> for every system above the lowest tier</li><li><strong>A review rhythm</strong>, so the inventory doesn’t go stale by spring</li></ol><p>Everything else can wait.</p><h3>Which AI Governance Frameworks Matter (and How Much)?</h3><p>You’ll hear three names. Here’s how I’d use each.</p><p><strong>NIST AI RMF.</strong> The <a href="https://www.nist.gov/itl/ai-risk-management-framework">NIST AI Risk Management Framework</a> is voluntary and free, and it’s organized around four functions: Govern, Map, Measure, Manage. For generative AI, NIST’s <a href="https://nvlpubs.nist.gov/nistpubs/ai/NIST.AI.600-1.pdf">Generative AI Profile (AI 600–1)</a> lists the specific risks to think about. Use it as your vocabulary and checklist. NIST has said the RMF is being revised, so check for updates before you build on it heavily.</p><p><strong>ISO/IEC 42001.</strong> This is the certifiable AI management-system standard. It matters if your customers or regulators are likely to ask for audit-grade evidence. For most mid-sized orgs, the right move in year one is to <em>align</em> with 42001’s structure, not to certify.</p><p><strong>EU AI Act.</strong> If you sell into the EU or your AI outputs are used there, it applies to you. The timeline changed this year. The <a href="https://eur-lex.europa.eu/legal-content/EN/TXT/?uri=celex:32026R1744">Digital Omnibus on AI (Regulation (EU) 2026/1744)</a> entered into force on 27 July 2026 and moved the main high-risk obligations to <strong>2 December 2027</strong> for standalone (Annex III) systems such as recruitment and credit assessment, and <strong>2 August 2028</strong> for AI embedded in regulated products. Prohibited practices have applied since February 2025, and Article 50 transparency obligations still apply from August 2026, <a href="https://www.secureaudit.nl/en/knowledge-base/eu-ai-act-digital-omnibus-deadlines">as this summary explains</a>. A delay is not an exemption. It’s time to prepare.</p><h3>Does Canada Have an AI Law? The Canadian Context</h3><p>Canada does not have a federal AI law. The Artificial Intelligence and Data Act died with the last Parliament. In June 2026 the federal government launched <a href="https://www.canada.ca/en/innovation-science-economic-development/news/2026/06/minister-solomon-highlights-canadas-national-artificial-intelligence.html">AI for All, Canada’s national AI strategy</a>, which focuses on trust, opportunity, and sovereignty, and tabled <a href="https://www.canada.ca/en/innovation-science-economic-development/news/2026/06/government-of-canada-tables-new-legislation-to-protect-childrens-data-strengthen-privacy-and-build-trust-in-the-digital-economy.html">Bill C-36, the Protecting Privacy and Consumer Data Act</a> to replace the private-sector part of PIPEDA. It’s still a bill, so watch it.</p><p>“No AI law” doesn’t mean “no rules.” Privacy law (PIPEDA, and Quebec’s Law 25 with its <a href="https://crewdle.com/blog/canada-ai-rules-after-aida/">automated-decision provisions</a>), human-rights law, and sector regulators already apply. If you’re a federally regulated financial institution, OSFI’s <a href="https://www.osfi-bsif.gc.ca/en/guidance/guidance-library/guideline-e-23-model-risk-management-2027">Guideline E-23 on model risk management</a> takes effect <strong>May 1, 2027</strong> and explicitly covers AI and machine learning models.</p><p>My practical read for Canadian mid-sized orgs: build on NIST AI RMF, structure it so it maps to ISO 42001 later, and check EU exposure separately.</p><h3>Days 1–30: Discover</h3><p>Goal: know what AI you’re running and who’s using it.</p><ul><li><strong>Send a short survey to team leads.</strong> Three questions: What AI tools does your team use? What data goes into them? What decisions do they influence?</li><li><strong>Check your SaaS stack.</strong> Your CRM, HR system, help desk, and office suite have probably shipped AI features. List them.</li><li><strong>Ask IT for logs.</strong> Network or SSO logs show which AI services people actually reach.</li><li><strong>Publish an interim acceptable-use policy.</strong> One page. Approved tools, data that must never go into a public AI tool (client data, personal information, credentials), and who to ask.</li><li><strong>Name an executive sponsor</strong> and one working owner, often someone in IT, risk, or privacy.</li></ul><p>Output: a first-draft inventory in a spreadsheet. Mine usually starts with columns for system, vendor, business owner, purpose, data types, users, and “makes or influences decisions about people? (Y/N).”</p><h3>Days 31–60: Assess and Control</h3><p>Goal: sort the inventory by risk and put proportionate controls on each tier.</p><p><strong>Use three tiers.</strong> More than three and people argue about the boundaries instead of doing the work.</p><ul><li><strong>Tier 1 (Low):</strong> Internal productivity, no personal or confidential data, a human reviews output. <em>Example: drafting internal emails, summarizing public research.</em> Control: approved tool list and the acceptable-use policy.</li><li><strong>Tier 2 (Medium):</strong> Touches confidential or personal data, or output reaches customers. <em>Example: a support chatbot, a RAG assistant over client contracts.</em> Controls: privacy review, vendor data terms checked, human review of outputs at launch, logging.</li><li><strong>Tier 3 (High):</strong> Influences decisions about people’s jobs, money, health, or access to services. <em>Example: CV screening, credit or claims decisions.</em> Controls: everything in Tier 2, plus documented testing for accuracy and bias, a human who can override, an explanation process for affected people, and executive sign-off. These are also the use cases most likely to fall under EU Annex III.</li></ul><p><strong>Add a lightweight intake step.</strong> New AI use cases fill in a short form, which tells you the tier, and the tier tells you which review it needs. A Tier 1 request should be approved in days, not weeks. If governance makes the easy things slow, people go around it.</p><p><strong>Check your vendors.</strong> For every Tier 2 and 3 system: Is your data used to train the vendor’s models? Where is it stored? Can you get logs? Can you switch the AI feature off?</p><h3>Days 61–90: Own and Monitor</h3><p>Goal: make governance something that runs, not something that happened once.</p><ul><li><strong>Put a name against every Tier 2 and Tier 3 system.</strong> A person, not a department. They own its performance, its incidents, and its annual review.</li><li><strong>Set a review rhythm.</strong> Quarterly for Tier 3, twice a year for Tier 2, annually for Tier 1.</li><li><strong>Define what counts as an AI incident</strong> (a harmful output, a data leak, a wrong decision affecting a person) and route it into your existing incident process. Don’t build a new one.</li><li><strong>Report up.</strong> A one-page quarterly summary to leadership: how many systems per tier, new requests, incidents, open actions.</li><li><strong>Train by role.</strong> Everyone gets 30 minutes on the policy. Owners of Tier 2 and 3 systems get more.</li></ul><h3>A Practical AI Governance Starting Checklist</h3><p>Print this and tick it off:</p><ul><li>☐ Executive sponsor and working owner named</li><li>☐ One-page acceptable-use policy published</li><li>☐ AI inventory started, including vendor AI features</li><li>☐ Every system assigned a risk tier (1, 2, or 3)</li><li>☐ Intake form for new AI use cases live</li><li>☐ Vendor data-use terms checked for Tier 2 and 3 systems</li><li>☐ Named owner for every Tier 2 and 3 system</li><li>☐ Human oversight defined for every Tier 3 system</li><li>☐ AI incidents routed into the existing incident process</li><li>☐ EU exposure checked against the 2027/2028 dates</li><li>☐ Sector rules checked (privacy, OSFI E-23 if applicable)</li><li>☐ Quarterly review date in the calendar</li></ul><h3>FAQ</h3><h4>How long does it take to set up AI governance?</h4><p>For an organization with 100 to 1,500 people, about 90 days gets you a working foundation. That means an AI inventory, three risk tiers with proportionate controls, and a named owner for every medium- and high-risk system.</p><h4>Which AI governance framework should a mid-sized organization use?</h4><p>Build on the NIST AI Risk Management Framework, which is voluntary and free, and structure it so it maps to ISO/IEC 42001 later. In year one, align with 42001 rather than certifying, and check EU AI Act exposure separately.</p><h4>Does Canada have a federal AI law?</h4><p>No. The Artificial Intelligence and Data Act died with the last Parliament. Privacy law, human-rights law, and sector regulators still apply, including OSFI’s Guideline E-23, which takes effect May 1, 2027 for federally regulated financial institutions.</p><h3>Final Thought</h3><p>Ninety days won’t make you mature. It will take you from “we don’t know what AI we’re running” to “we know, we’ve sorted it, and someone owns it.” For most mid-sized organizations, that’s a bigger jump than any framework certificate.</p><p><em>At </em><a href="https://www.aarambhconsulting.ca/"><em>Aarambh Consulting Group</em></a><em>, we help teams with AI Governance and Production Readiness assessments, and with senior BA, Product, and Data talent when they need extra hands to do the work.</em></p><p><em>Maharshi B, founder of Aarambh Consulting Group, a Canadian-owned and operated firm.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=f43fad99ca51" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[AI Governance Is Not a Policy Document: A Practical Enterprise AI Governance Framework]]></title>
            <link>https://medium.com/@jobs.maharshi/ai-governance-is-not-a-policy-document-a-practical-enterprise-ai-governance-framework-18ee2fbc7db7?source=rss-d58d7b0b71c6------2</link>
            <guid isPermaLink="false">https://medium.com/p/18ee2fbc7db7</guid>
            <category><![CDATA[enterprise-ai]]></category>
            <category><![CDATA[leadership-development]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[risk-management]]></category>
            <category><![CDATA[ai-governance]]></category>
            <dc:creator><![CDATA[Maharshi B]]></dc:creator>
            <pubDate>Tue, 20 Oct 2026 15:03:04 GMT</pubDate>
            <atom:updated>2026-10-20T16:46:44.092Z</atom:updated>
            <content:encoded><![CDATA[<p>Enterprise AI adoption is moving faster than most governance programs.</p><p>Teams are experimenting with copilots, RAG applications, predictive models, AI agents, third party AI services, and increasingly autonomous workflows.</p><p>Many organizations have responded by creating AI policies, responsible AI principles, review committees, and acceptable use guidelines.</p><p>Those are <strong>important</strong>.</p><p>But having an AI policy does not mean you have <strong>operational AI governance</strong>.</p><p>The real test is much simpler.</p><p>Can your organization answer these questions today?</p><ol><li><strong>What AI systems are we actually using?</strong></li><li><strong>Who owns each system?</strong></li><li><strong>What risks does each system introduce?</strong></li><li><strong>What controls should apply to those risks?</strong></li><li><strong>How do we know those controls continue working after deployment?</strong></li></ol><p>If answering those questions requires emails, spreadsheets, meetings, and detective work across multiple teams, the organization may have AI governance <em>documentation</em> without having an AI governance <em>operating system</em>.</p><p><strong>That distinction matters.</strong></p><h3>AI Governance Has to Become Operational</h3><p>A useful way to think about AI governance is as a lifecycle rather than a policy.</p><p>At Aarambh Consulting, we have been exploring this problem through five operational questions:</p><blockquote><strong>Discover → Assess → Control → Own → Monitor</strong></blockquote><p>These aren’t intended to replace established frameworks such as NIST AI RMF or ISO/IEC 42001.</p><p>Instead, they represent the operational questions organizations need to answer when translating governance principles into day-to-day delivery.</p><h3>1. Discover — What AI Do We Have?</h3><p>Governance begins with visibility.</p><p>Organizations increasingly have AI embedded across:</p><ul><li>internally developed machine-learning models</li><li>generative AI applications</li><li>RAG systems</li><li>SaaS products containing AI capabilities</li><li>employee copilots</li><li>APIs consuming external foundation models</li><li>automated decision systems</li><li>experimental AI projects</li><li>AI agents capable of taking actions</li></ul><p>This creates a basic governance problem:</p><p><strong>You cannot govern systems you don’t know exist.</strong></p><p>An AI inventory should therefore be more than a list of model names.</p><p>For each AI system, an organization should be able to identify information such as:</p><ul><li>business purpose</li><li>business owner</li><li>technical owner</li><li>model or AI provider</li><li>data being processed</li><li>users affected</li><li>decisions influenced</li><li>deployment environment</li><li>external integrations</li><li>current lifecycle status</li></ul><p>This inventory becomes the foundation for everything that follows.</p><p>Without it, governance starts from assumptions rather than evidence.</p><h3>2. Assess — How Risky Is It?</h3><p>Not every AI system deserves the same governance process.</p><p>A developer using an internal coding assistant presents a very different risk profile from an AI system influencing:</p><ul><li>healthcare decisions</li><li>employment screening</li><li>financial eligibility</li><li>insurance decisions</li><li>public services</li><li>customer-facing recommendations</li></ul><p>Treating every AI use case identically creates bureaucracy.</p><p>Treating every use case casually creates exposure.</p><p>The better approach is <strong>risk-based governance</strong>.</p><p>Organizations can evaluate <strong>factors </strong>such as:</p><p><strong>Impact</strong></p><p>What happens if the system is wrong?</p><p><strong>Data sensitivity</strong></p><p>Does it process personal, confidential, regulated, or proprietary information?</p><p><strong>Decision authority</strong></p><p>Does AI recommend, influence, or directly make decisions?</p><p><strong>External exposure</strong></p><p>Is the output internal or customer-facing?</p><p><strong>Autonomy</strong></p><p>Can the system take actions without human approval?</p><p><strong>Reversibility</strong></p><p>Can an incorrect decision easily be corrected?</p><p>The result should be a practical risk tier.</p><p>For example:</p><p><strong>Low Risk → Moderate Risk → High Risk → Critical</strong></p><p>The classification then determines the governance requirements.</p><p>Governance becomes proportional rather than universal.</p><h3>3. Control — What Must We Do About the Risk?</h3><p>Identifying risk is only useful if it changes how the system is designed, deployed, or operated.</p><p>This is where governance must connect directly to engineering and operations.</p><p>A high-risk AI system might require controls covering:</p><ul><li>data privacy</li><li>cybersecurity</li><li>access control</li><li>model evaluation</li><li>hallucination testing</li><li>bias or fairness assessment</li><li>explainability</li><li>human oversight</li><li>prompt and output logging</li><li>vendor assessment</li><li>fallback mechanisms</li><li>approval gates</li><li>incident response</li><li>monitoring</li></ul><p>This creates an important relationship:</p><p><strong>Risk → Control → Evidence</strong></p><p>Suppose an organization identifies hallucination as a material risk for a customer-facing RAG application.</p><p>Writing <em>“the system should provide accurate responses”</em> in a governance document is not a control.</p><p>Operational controls could instead include:</p><ul><li>groundedness evaluation</li><li>citation verification</li><li>confidence thresholds</li><li>restricted knowledge sources</li><li>human escalation</li><li>production monitoring</li></ul><p>And governance should capture evidence showing whether those controls actually operate.</p><p>That is the difference between governance as intention and governance as execution.</p><h3>4. Own — Who Is Accountable?</h3><p>AI systems cross organizational boundaries.</p><p>A single production AI application might involve:</p><ul><li>product</li><li>engineering</li><li>data science</li><li>security</li><li>privacy</li><li>legal</li><li>compliance</li><li>risk management</li><li>business operations</li><li>third-party vendors</li></ul><p>Without explicit ownership, governance becomes everyone’s responsibility and therefore nobody’s responsibility.</p><p>Every AI system should have identifiable accountability.</p><p>At minimum:</p><p><strong>Business owner</strong></p><p>Why does this AI system exist, and what business outcome does it support?</p><p><strong>Technical owner</strong></p><p>Who operates and maintains the system?</p><p><strong>Risk owner</strong></p><p>Who accepts or escalates the residual risk?</p><p><strong>Control owner</strong></p><p>Who is responsible for ensuring specific controls operate?</p><p>This does not mean creating another committee for every AI application.</p><p>It means making accountability explicit.</p><h3>5. Monitor — Is the System Still Safe and Effective?</h3><p>Traditional governance often concentrates heavily on approval.</p><p>A system goes through assessment, receives approval, enters production — and governance largely stops.</p><p>AI systems make that model insufficient.</p><p>Models change.</p><p>Data changes.</p><p>Prompts change.</p><p>Knowledge bases change.</p><p>Vendors change their models.</p><p>Users discover unexpected behaviours.</p><p>Agents gain new tools and permissions.</p><p>An AI system that was acceptable six months ago may not have the same risk profile today.</p><p>Governance therefore needs a continuous feedback loop.</p><p>Organizations may need to monitor:</p><ul><li>model performance</li><li>hallucination rates</li><li>retrieval quality</li><li>policy violations</li><li>abnormal usage</li><li>human overrides</li><li>incidents</li><li>data drift</li><li>model or vendor changes</li><li>cost anomalies</li><li>control failures</li></ul><p>When thresholds are breached, governance should trigger action.</p><p>That might mean investigation, human review, remediation, rollback, reapproval, or — in serious cases — shutting down the AI capability.</p><p>Governance therefore doesn’t end at deployment.</p><p><strong>Deployment is where operational governance begins.</strong></p><h3>The AI Governance Lifecycle</h3><p>Put together, the model looks straightforward:</p><p><strong>DISCOVER</strong></p><p>What AI systems exist?</p><p>↓</p><p><strong>ASSESS</strong></p><p>What risks do they create?</p><p>↓</p><p><strong>CONTROL</strong></p><p>What safeguards are required?</p><p>↓</p><p><strong>OWN</strong></p><p>Who is accountable?</p><p>↓</p><p><strong>MONITOR</strong></p><p>Are the system and controls still operating appropriately?</p><p>↓</p><p><strong>REASSESS</strong></p><p>Has anything materially changed?</p><p>The final step creates the loop.</p><p>AI governance should not be a one-time approval process.</p><p>It should operate throughout the AI system lifecycle.</p><h3>Where AI Governance and AI Operations Meet</h3><p>This is also where the boundary between <strong>AI Governance</strong> and <strong>AI Operations</strong> becomes interesting.</p><p>Governance defines:</p><p><strong>What should happen.</strong></p><p>Operations determines:</p><p><strong>Whether it is actually happening.</strong></p><p>For example:</p><p>Governance might require that a production RAG system maintain an acceptable groundedness threshold.</p><p>AI Operations provides the evaluation pipeline, telemetry, alerts, and incident workflow needed to determine whether that requirement continues to be satisfied.</p><p>The two disciplines therefore cannot remain disconnected.</p><p>The future enterprise AI operating model will increasingly need to connect:</p><p><strong>Governance + Engineering + Operations + Risk</strong></p><p>That connection is where I believe much of the difficult enterprise AI work is going to happen.</p><h3>Start With Visibility, Not a Giant Governance Program</h3><p>Organizations beginning their AI governance journey do not necessarily need a large governance platform or dozens of committees.</p><p>Start by answering five questions:</p><p><strong>What AI do we have?</strong></p><p><strong>How risky is it?</strong></p><p><strong>What controls apply?</strong></p><p><strong>Who owns the risk and controls?</strong></p><p><strong>How will we know when something changes or fails?</strong></p><p>If those answers are visible, traceable, and connected to actual engineering and operational processes, governance starts becoming real.</p><p>If they exist only inside policies and presentations, there is still work to do.</p><p>I’m exploring these problems through <strong>Aarambh Consulting</strong>, with a particular focus on turning AI governance principles into practical operating models, controls, evidence, and AI Operations workflows.</p><p>Next week: AI Governance in 90 Days, a practical, week-by-week plan to take these ideas from principle to working controls. Follow along so you don’t miss it.</p><p>Maharshi B, founder of Aarambh Consulting Group, a Canadian-owned and operated firm.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=18ee2fbc7db7" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[We Are Entering the Age of Exaggerated Promises: Agreeing with Jeff Bezos on the New AI Bubbles.]]></title>
            <link>https://medium.com/@jobs.maharshi/we-are-entering-the-age-of-exaggerated-promises-agreeing-with-jeff-bezos-on-the-new-ai-bubbles-b015a0477e59?source=rss-d58d7b0b71c6------2</link>
            <guid isPermaLink="false">https://medium.com/p/b015a0477e59</guid>
            <category><![CDATA[ai-product-strategy]]></category>
            <category><![CDATA[ai-trends]]></category>
            <category><![CDATA[ai-bubble-2026]]></category>
            <category><![CDATA[jeff-bezos]]></category>
            <dc:creator><![CDATA[Maharshi B]]></dc:creator>
            <pubDate>Fri, 21 Nov 2025 22:05:09 GMT</pubDate>
            <atom:updated>2025-11-21T22:05:09.183Z</atom:updated>
            <content:encoded><![CDATA[<p>In one of his most interesting recent reflections, Jeff Bezos said something that instantly caught my attention:</p><blockquote><strong>“We’re in AI bubbles right now: in 90’s we had a bio-tech bubble, a banking bubble in 2008, and an industrial AI bubble right now. The technology is real — but the promises are exaggerated.”</strong></blockquote><p>When Jeff Bezos recently declared that we’re in an “<a href="https://ca.finance.yahoo.com/news/jeff-bezos-says-ai-bubble-145251428.html">industrial bubble</a>” driven by artificial intelligence, he wasn’t warning of doom so much as spotlighting the <strong>shape</strong> of our investments and hype.</p><p>He contrasted this with classic <strong>financial bubbles</strong> like those that triggered the 2008 banking crisis (“the banking bubble… that’s just bad”) and earlier episodes of health / biotech over-promise (“bio health big pharma bubble”)</p><p>Here’s why I agree — and why the real story today is the industrial AI bubble.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*LkjXYj7TRWQwReg59Lnjeg.png" /><figcaption>Gemini generated image</figcaption></figure><h3>The Past: Two Bubble Archetypes</h3><p><strong>1. </strong><a href="https://en.wikipedia.org/wiki/2008_financial_crisis"><strong>Banking/Financial Bubble </strong></a><strong>(2008)</strong><br> The housing-credit bubble, subprime mortgages, complex derivatives — these were primarily financial in nature: speculative assets, leverage, systemic risk. The bursting triggered a global crisis.</p><p><strong>2.</strong><a href="https://drug-dev.com/management-insight-the-biotech-bubble-what-goes-up-must-come-down/"><strong> Biohealth / Big Pharma Bubble</strong></a><br> In the life sciences sector we have seen a surge of funding, rising valuations, but questionable returns. For example: large numbers of biotech / pharma companies restructuring during downturns; funding flows racing ahead of commercialization.<br> These bubbles were driven by hope and breakthrough promise — and in many cases delivered long-term innovations — but not without enormous risk, wasted capital, and many failures.</p><h3>The Present: Industrial AI Bubble</h3><p>This is where Bezos’s phrase “<a href="https://fortune.com/2025/10/04/jeff-bezos-amazon-openai-sam-altman-ai-bubble-tech-stocks-investing/">industrial bubble</a>” comes in. He argues that the hype around AI is indeed bubble-like — <strong>every experiment gets funded</strong>, investors struggle to distinguish good ideas from bad. <br> But the key distinction: this is <strong>industrial infrastructure + transformation</strong>, not purely speculative finance. The underlying technology (LLMs, RAG, agents, data infrastructure) has a shot at generating real value across industries.</p><p>So yes: there is hype. But also potential.</p><h3>Why This Matters</h3><ul><li>If you view this simply as another “dot-com bubble” or “subprime” moment, you might miss the fact that<a href="https://thefinanser.com/2025/10/financial-bubbles-destroy-everything-industrial-bubbles-make-the-future-jeff-bezos"> <strong>some infrastructure will survive</strong></a> and become foundational (as with fiber-optic build-outs, as Bezos noted).</li><li>If you treat it like the biotech boom, you see where promise outpaces delivery — many AI startups will fail, many experiments won’t scale. But the winners will reshape industries.</li><li>For product owners, engineers, investors: this means being <strong>very selective</strong>. Focus on fundamentals — problem, data, model, deployment — not just hype.</li></ul><h3>My Take: How to Navigate It</h3><ul><li><strong>Recognize the bubble elements:</strong> inflated valuations, hype, “AI label” for everything.</li><li><strong>Distinguish the industrial characteristics:</strong> infrastructure, operationalization, enterprise deployment, cost savings, transformation.</li><li><strong>Learn from past bubbles:</strong> oversupply of capital doesn’t guarantee outcomes. The biotech sector saw waves of failures even though underlying science was real.</li><li><strong>Be part of the winners: </strong>work on problems that matter, align with strong fundamentals, emphasize measurable value, and design for scalability.</li></ul><h3>In Summary , my overall take is :</h3><p>Yes, the investment environment feels frothy. Yes, we might be in a bubble. But as Jeff Bezos argues, this may be a <a href="https://www.youtube.com/live/E0x3UZDKSNo?si=GTxzXEmIuJ7MCYxL"><strong>good kind of bubble</strong></a> — if the foundational work is done right, the infrastructure built, and the transformation across industry realized.</p><p>For those of us building and owning AI-powered products, this is a moment. We just need to steer smartly, learn from the past, and focus on what endures.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=b015a0477e59" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[How to Become an AI Product Owner in 2026: The Complete Roadmap]]></title>
            <link>https://medium.com/@jobs.maharshi/how-to-become-an-ai-product-owner-in-2026-the-complete-roadmap-9415317aef74?source=rss-d58d7b0b71c6------2</link>
            <guid isPermaLink="false">https://medium.com/p/9415317aef74</guid>
            <category><![CDATA[ai-product-management]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[ai-product-owner]]></category>
            <dc:creator><![CDATA[Maharshi B]]></dc:creator>
            <pubDate>Fri, 21 Nov 2025 21:18:54 GMT</pubDate>
            <atom:updated>2025-11-21T21:18:54.622Z</atom:updated>
            <content:encoded><![CDATA[<p>Artificial Intelligence is no longer a “future skill.” By 2026 and beyond, AI Product Owners (AI-POs) will be at the center of value creation in every industry — healthcare, retail, finance, logistics, government, and even small-business tech.</p><p>Companies don’t just want AI tools.<br> They want <strong>AI that solves real business problems</strong>.<br> They want <strong>owners</strong>, not just <strong>builders</strong>.</p><p>This is why the AI Product Owner role is exploding — and why now is the perfect time to enter.</p><p>In this guide, I’ll break down exactly how to become an AI Product Owner in 2026, even if you’re starting from a traditional BA/PO/PM background or coming from engineering/data roles.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*04ENXBxLMDau0zXvtYRaJA.jpeg" /><figcaption>Photo by <a href="https://unsplash.com/@aideal?utm_source=unsplash&amp;utm_medium=referral&amp;utm_content=creditCopyText">Aideal Hwa</a> on <a href="https://unsplash.com/photos/man-in-black-jacket-sitting-on-white-chair-OYzbqk2y26c?utm_source=unsplash&amp;utm_medium=referral&amp;utm_content=creditCopyText">Unsplash</a></figcaption></figure><h3>What Is an AI Product Owner?</h3><p>An AI Product Owner sits at the intersection of:</p><ul><li><strong>Business strategy</strong></li><li><strong>User needs</strong></li><li><strong>Data &amp; machine learning</strong></li><li><strong>AI engineering</strong></li></ul><p>Your job is to translate business problems into AI features, manage the AI lifecycle, and ensure that the solution actually delivers measurable outcomes.</p><p>You are the <strong>voice of the business</strong> AND the <strong>translator for AI teams</strong>.</p><p>In 2026, this role now includes:</p><ul><li>Managing <strong>RAG pipelines</strong>, <strong>LLM agents</strong>, and <strong>prompt engineering</strong></li><li>Working with <strong>vector databases</strong>, <strong>embeddings</strong>, and <strong>AI evaluation metrics</strong></li><li>Ensuring <strong>AI governance</strong>, compliance, and data security</li><li>Defining acceptance criteria for <strong>GenAI-powered user experiences</strong></li><li>Driving <strong>AI product experiments</strong> using Lean/Agile</li></ul><h3>Why AI Product Ownership Is the Hottest Career of 2026</h3><h3>Talent shortage</h3><p>Companies cannot find people who understand <strong>product + AI + business</strong>. Engineers know AI. PMs know business. <strong>AI-POs know both.</strong></p><h3>Highest comp growth</h3><p>AI Product roles rose <strong>31% YoY in compensation globally (2024–2025)</strong>.</p><h3>Every company wants AI</h3><ul><li>Banks -&gt; AI copilots for analysts</li><li>Healthcare -&gt; AI triage, documentation assistants</li><li>Retail -&gt; forecasting, recommendations, chatbots</li><li>Government -&gt; automation and citizen-facing agents.</li></ul><h3>Low barrier if you already have BA/PO/PM skills</h3><p>You can transition without becoming a data scientist.</p><h3>The 2026 Roadmap to Becoming an AI Product Owner</h3><h3>Step 1 — Strengthen Core Product Ownership Skills</h3><p>Before stepping into AI, you need to master the fundamentals:</p><h3>✔ Backlog management</h3><h3>✔ User story writing</h3><h3>✔ Acceptance criteria</h3><h3>✔ Stakeholder facilitation</h3><h3>✔ Prioritization frameworks (WSJF, MoSCoW, RICE)</h3><h3>✔ Data-driven decision making</h3><p><strong>If you’ve been a BSA, BA, PM, or PO, you already have 60% of what’s needed.</strong></p><h3>Step 2 — Learn the AI &amp; ML Fundamentals (Non-Technical but Powerful)</h3><p>You do NOT need to code models.<br>But you DO need to understand how they work.</p><h3>Key Concepts to Learn in 2026:</h3><ul><li>What is an LLM?</li><li>Embeddings &amp; vector search</li><li>Retrieval-Augmented Generation (RAG)</li><li>Agentic AI &amp; multi-step reasoning</li><li>Prompt engineering lifecycle</li><li>Evaluation metrics (BLEU, ROUGE, hallucination rate, factual accuracy)</li><li>Data labeling, preprocessing, quality</li><li>MLOps (high-level understanding)</li></ul><h3>Recommended Fast Learning Stack:</h3><ul><li><strong>Azure AI Studio / Azure ML</strong> (enterprise standard)</li><li><strong>OpenAI / Anthropic API basics</strong></li><li><strong>LangChain / Azure Prompt Flow</strong></li><li><strong>Pinecone / Weaviate / FAISS (vector DBs)</strong></li></ul><p><strong>Your goal:</strong><br> Speak the language of AI engineers and guide business teams with confidence.</p><h3>Step 3 — Learn How AI Products Are Built (Not Just AI Technology)</h3><p>AI products require a different delivery lifecycle:</p><h3>AI product lifecycle you must know:</h3><ol><li><strong>Define business problem</strong></li><li><strong>Evaluate if AI is even required (often it isn’t)</strong></li><li><strong>Data discovery &amp; feasibility</strong></li><li><strong>Model selection / RAG pipeline / agent design</strong></li><li><strong>Human-in-the-loop workflows</strong></li><li><strong>Bias, fairness, governance</strong></li><li><strong>Pilot → A/B test → evaluate accuracy</strong></li><li><strong>Production deployment &amp; monitoring</strong></li></ol><h3>Why this matters:</h3><p>AI products fail not because models fail, but because:</p><ul><li>No real business case</li><li>No data</li><li>No user adoption</li><li>Poor accuracy evaluation</li><li>Compliance issues</li></ul><p>A strong AI-PO prevents this.</p><h3>Step 4 — Build a Real AI Portfolio (The 2026 Requirement)</h3><p>Recruiters in 2026 expect <strong>portfolio + GitHub + case studies</strong>… even for non-technical candidates.</p><h3>Build 3–5 projects that include:</h3><ul><li>Business problem statement</li><li>Personas &amp; user journeys</li><li>Requirements &amp; story mapping</li><li>Architecture diagram (high level)</li><li>RAG / LLM concept flow</li><li>Acceptance criteria</li><li>Metrics &amp; ROI calculation</li></ul><h3>Portfolio Project Ideas:</h3><ol><li><strong>AI Customer Support Agent</strong> for a retail store</li><li><strong>Healthcare RAG triage system</strong> for symptoms</li><li><strong>Real estate pricing prediction model</strong> (API + dashboard)</li><li><strong>AI assistant for PMO / backlog grooming</strong></li><li><strong>Financial statement summarizer with agents</strong></li></ol><p><strong>Important:</strong> You don’t have to code.<br> Use tools like Azure AI Studio, Low-Code RAG, Replit AI agents, or ChatGPT Builders.</p><h3>Step 5 — Learn AI Governance &amp; Risk (The #1 differentiator in 2026)</h3><p>Every major employer demands this.</p><h3>What you must understand:</h3><ul><li>Responsible AI frameworks</li><li>Data privacy (GDPR, HIPAA, PHIPA, CPPA 2025)</li><li>Explainability (XAI)</li><li>Human oversight</li><li>Model drift and accuracy monitoring</li><li>Red-teaming, hallucination reduction</li></ul><p>This turns you into <strong>a safe pair of hands</strong> — critical for enterprise hiring.</p><h3>Step 6 — Master AI Communication &amp; Storytelling</h3><p>The best AI-POs win because they can answer:</p><ul><li>“What problem does this solve?”</li><li>“How much cost does this eliminate?”</li><li>“How accurate is the model?”</li><li>“Why AI, not automation?”</li><li>“How do we monitor hallucinations?”</li><li>“What is the ROI in 90 days?”</li></ul><p>Become the person who can make AI simple for C-suite, SMEs, and engineers.</p><h3>Step 7 — Position Yourself on LinkedIn for AI Product Roles</h3><p>In 2026, hiring managers search for:</p><blockquote><em>“AI Product Owner”<br> “AI Product Manager”<br> “LLM Product Owner”<br> “RAG Product Lead”<br> “AI Business Analyst”</em></blockquote><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=9415317aef74" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>