[{"data":1,"prerenderedAt":599},["ShallowReactive",2],{"insight-blog-ai-doesnt-reward-risk-it-punishes-bad-scoping":3},{"post":4,"related":580},{"id":5,"title":6,"author":7,"body":11,"category":556,"description":557,"extension":558,"featuredImage":559,"featuredImageAlt":560,"meta":561,"navigation":562,"path":563,"publishedAt":564,"relatedSlugs":565,"seo":568,"stem":571,"tags":572,"type":578,"__hash__":579},"insights\u002Finsights\u002Fblog\u002Fai-doesnt-reward-risk-it-punishes-bad-scoping.md","AI Doesn't Reward Risk. It Punishes Bad Scoping.",{"name":8,"role":9,"initials":10},"Dusan Stamenkovic","Founder & Senior AI Strategy Consultant","DS",{"type":12,"value":13,"toc":534},"minimark",[14,22,25,28,31,34,46,49,54,57,60,64,67,72,81,84,87,92,107,111,116,119,122,125,138,141,153,157,160,163,170,181,185,188,191,194,206,210,213,216,219,225,231,243,248,254,258,261,265,268,271,279,292,296,303,306,309,320,324,327,333,336,348,352,355,359,362,365,385,388,400,404,407,410,420,423,435,439,443,519,522,525,528],[15,16,17,18],"p",{},"In traditional investing, risk and return move together. You take on more risk because you expect a higher upside. ",[19,20,21],"strong",{},"AI doesn't work like that.",[15,23,24],{},"In AI, risk is not something you consciously \"take\" in exchange for higher return. The ceiling is fixed: a specific business metric, improved by a defined amount. A poorly scoped AI initiative does not create more upside than a well-scoped one. That doesn't change.",[15,26,27],{},"What changes is how likely you are to get there.",[15,29,30],{},"A clear scope lowers uncertainty: you know what you're building, what data it requires, and what success looks like. A weak scope does the opposite: it leaves gaps, assumptions, and unknowns that compound into failure.",[15,32,33],{},"The result is simple: same ceiling, worse odds.",[35,36,38,43],"insight-callout",{"label":37},"The reframe",[15,39,40],{},[19,41,42],{},"Risk tolerance is the wrong frame for AI investment decisions.",[15,44,45],{},"The real differentiator is scope quality — because scope quality is what sets the level of risk.",[15,47,48],{},"None of the ten signals below increase your upside. They only reveal where poor scoping is putting your outcome at risk.",[50,51,53],"h2",{"id":52},"how-to-use-this-scorecard","How to Use This Scorecard",[15,55,56],{},"Ten risk signals. Score each one 0 or 1. The first five — structural risks — are worth 2 points each. The remaining five — execution and capability risks — are worth 1 point each. Maximum score: 15.",[15,58,59],{},"A higher score doesn't mean you shouldn't build. Most of these signals are fixable. It means you have planning work to do before the budget is committed. The score tells you where to focus — not whether to proceed.",[50,61,63],{"id":62},"structural-risks-2-points-each","Structural Risks — 2 Points Each",[15,65,66],{},"These are flaws in how the project is defined and how accountability is designed into the workflow. If any of these are present, no amount of execution quality compensates. Each structural signal present adds 2 points to your score.",[68,69,71],"h3",{"id":70},"_1-you-have-a-metric-it-might-not-be-the-right-one","1. You have a metric. It might not be the right one.",[15,73,74,75,80],{},"I wrote about the risk inherent to the absence of a defined business metric in ",[76,77,79],"a",{"href":78},"\u002Finsights\u002Fblog\u002Fthe-hidden-roi-killer-in-ai-projects-starting-with-the-solution\u002F","The Hidden ROI Killer in AI Projects: Starting With the Solution",". This signal is subtler: a metric exists, everyone agrees on it, and it doesn't connect to any P&L line.",[15,82,83],{},"Proxy metrics feel real because they're measurable. Model accuracy. Feature adoption rate. \"AI suggestions accepted.\" Response time. NPS improvement. All trackable. None of them are what the business actually optimizes for. A 95%-accurate model that moves the wrong thing is worthless. An AI feature with 80% suggestion acceptance that can't be traced to revenue, cost, or risk reduction has no business case. Regardless of how the model performs.",[15,85,86],{},"The proxy metric failure is especially common in product companies because product metrics are the language of the roadmap. They are the right metrics for product decisions. They are not always the right metrics for AI investment decisions.",[15,88,89],{},[19,90,91],{},"The test: not \"do we have a metric,\" but \"does this metric have a direct, traceable path to a P&L line?\"",[35,93,95,101],{"label":94},"Score it — 2 points",[15,96,97,100],{},[19,98,99],{},"Score 1"," if your primary success metric doesn't connect directly to revenue, cost, or risk reduction.",[15,102,103,106],{},[19,104,105],{},"Score 0"," if it does.",[68,108,110],{"id":109},"_2-working-means-different-things-to-three-different-people-in-the-room","2. \"Working\" means different things to three different people in the room.",[15,112,113],{},[19,114,115],{},"When a project ships and nobody agrees it succeeded, the failure happened in planning.",[15,117,118],{},"The PM says it's live. The ML engineer says accuracy was hit. The business owner says the metric didn't move. All correct — because no one agreed what \"success\" meant upfront. Sometimes the target is even set after the fact, based on what the model achieved.",[15,120,121],{},"Same failure, different moment.",[15,123,124],{},"Define success before you build:",[126,127,128,132,135],"ul",{},[129,130,131],"li",{},"Model hits target",[129,133,134],{},"Feature is live and stable",[129,136,137],{},"Business metric moves as expected",[15,139,140],{},"If these aren't aligned early, alignment later gets expensive.",[35,142,143,148],{"label":94},[15,144,145,147],{},[19,146,99],{}," if you cannot state in one sentence what each of these three stakeholders considers \"success\", and if those definitions are not reconciled before build begins.",[15,149,150,152],{},[19,151,105],{}," if they are.",[68,154,156],{"id":155},"_3-a-single-option-plan-is-not-a-plan","3. A single option plan is not a plan.",[15,158,159],{},"The risk isn't that your approach is wrong. It's that if it is, you have no recovery path. If accuracy falls short, latency is too high, or it breaks in production — there is nothing to fall back on.",[15,161,162],{},"This usually happens because teams optimize for decisiveness over resilience. \"We're building X\" sounds better than \"we're testing options.\" The real cost — failing and starting over — shows up later.",[15,164,165,166,169],{},"Not every project needs multiple parallel tracks. But every project needs a clear answer to one question: ",[19,167,168],{},"If this approach fails, what's next?"," If there's no answer, you have a single point of failure.",[35,171,172,177],{"label":94},[15,173,174,176],{},[19,175,99],{}," if your project plan has no explicit answer to that question.",[15,178,179,106],{},[19,180,105],{},[68,182,184],{"id":183},"_4-a-business-case-that-only-works-at-100-adoption-is-fiction","4. A business case that only works at 100% adoption is fiction.",[15,186,187],{},"The math is consistent: [time saved per user] × [number of users] × [loaded hourly cost] = total saving. The math works. The assumption — that all intended users will adopt fully, immediately, and maintain adoption — almost never holds.",[15,189,190],{},"For internal AI tools, 40–60% of intended users adopt in the first 90 days. Of those, a meaningful fraction use the tool partially. Adoption resistance for AI tools is also structurally different from resistance to deterministic software: it requires trust in probabilistic outputs that are occasionally surprising. That trust gap doesn't exist for software that calculates a discount. It exists for any AI system that recommends, generates, or decides.",[15,192,193],{},"A business case approved on 100% adoption at 80% ROI becomes 40% adoption at 32% ROI. The business case was always wrong. Nobody identifies it as the cause at the annual review.",[35,195,196,201],{"label":94},[15,197,198,200],{},[19,199,99],{}," if your ROI model has no adoption sensitivity analysis, or if it assumes any adoption rate above your organization's historical baseline for new internal tool rollouts.",[15,202,203,205],{},[19,204,105],{}," if it accounts for realistic adoption curves.",[68,207,209],{"id":208},"_5-you-are-building-in-a-domain-where-you-are-legally-responsible-for-the-output","5. You are building in a domain where you are legally responsible for the output.",[15,211,212],{},"If a law firm ships an AI-assisted contract review tool where the workflow allows a lawyer to forward AI-generated analysis to a client without a mandatory review step — and the analysis is wrong — the firm is liable. The tool is not. \"The AI did it\" has never succeeded as a legal defense and it never will. The organization bears full professional and legal responsibility for every output that leaves under its name. The workflow design either reflects that or it doesn't — and fixing it after the product is live means rebuilding the product.",[15,214,215],{},"This applies wherever accountability sits with the organization regardless of what produced the output: legal, medical, financial, and any domain involving high-stakes decisions. AI genuinely improves speed and quality in these domains — a lawyer processing 200 contracts instead of 20 is more productive. None of that removes the accountability. What it changes is whether the workflow makes that accountability visible and enforceable.",[15,217,218],{},"Two mitigations — both non-negotiable:",[15,220,221,224],{},[19,222,223],{},"Human-in-the-loop as a workflow requirement, not an option."," The output cannot reach the next stage — delivery, signature, publication, decision — without an explicit human sign-off gate. Not \"you can optionally review this.\" A required step. If the workflow can be completed without a human reviewing the AI output, the accountability gap is open.",[15,226,227,230],{},[19,228,229],{},"Responsibility-acknowledging UX."," The interface must make clear to the professional using it — and to any downstream recipient — that a human professional has reviewed and is accountable for this output. The AI is the assistant. The professional is the author. If the interface makes the AI the visible actor and the human invisible, the accountability model is inverted. Organizations discover this inversion at the worst possible moment.",[35,232,233,238],{"label":94},[15,234,235,237],{},[19,236,99],{}," if your workflow operates in a domain where professional or legal accountability sits with your organization — and either there is no mandatory human review gate, or the UX does not surface the human professional as accountable.",[15,239,240,242],{},[19,241,105],{}," if both mitigations are in place.",[244,245],"score-tally",{"label":246,"max":247},"Structural subtotal","10",[35,249,251],{"label":250},"The red line",[15,252,253],{},"If your structural subtotal is 6 or higher — three or more structural signals present — treat the project as high risk regardless of your total score. Execution quality does not compensate for structural design failures.",[50,255,257],{"id":256},"execution-risks-1-point-each","Execution Risks — 1 Point Each",[15,259,260],{},"These are risks in how the project will be run. Present alongside structural risks, they compound. Present alone, they are manageable with planning. Each signal present adds 1 point.",[68,262,264],{"id":263},"_6-the-infrastructure-bill-wasnt-in-the-budget","6. The infrastructure bill wasn't in the budget.",[15,266,267],{},"Build cost is calculated. Run cost at MVP scale is often estimated. Run cost at production scale with real adoption is almost never calculated — which is where the project fails financially.",[15,269,270],{},"LLM API costs scale with usage, not with features. A system processing 10 documents per day costs almost nothing to run. Deploying it to 200 users processing 50 documents each is a 1,000× increase in API spend. If the ROI model was built on the prototype's cost figure, the project will be unprofitable at full adoption — the exact moment the ROI was supposed to materialize.",[15,272,273,274,278],{},"The organizations that completed the data readiness work described in ",[76,275,277],{"href":276},"\u002Finsights\u002Fblog\u002Fyour-data-isnt-ai-ready-neither-is-your-initiative\u002F","Your Data Isn't AI-Ready"," — specifically those that standardized their ETL pipelines — already know their per-unit processing cost. Organizations that skipped that work are most likely to discover it in the infrastructure bill.",[35,280,282,287],{"label":281},"Score it — 1 point",[15,283,284,286],{},[19,285,99],{}," if you haven't calculated per-unit AI cost and multiplied it by your full-adoption usage projection — and verified the result fits within the ROI model's assumptions.",[15,288,289,291],{},[19,290,105],{}," if you have.",[68,293,295],{"id":294},"_7-you-are-planning-this-ai-sprint-like-a-software-sprint","7. You are planning this AI sprint like a software sprint.",[15,297,298,299,302],{},"AI development does not have deterministic \"done.\" Software timelines can be estimated from velocity and story points because the work is compositional: more engineers means more features. ML development is not compositional in the same way. Three more engineers don't produce three times the accuracy improvement. Most ML improvements come from insights, not parallelization. Insights don't have delivery schedules. (This is the timeline version of the non-determinism argument covered in ",[76,300,301],{"href":78},"The Hidden ROI Killer in AI Projects",".)",[15,304,305],{},"The pattern: week 2, model is running, results are promising. Week 4, the accuracy plateau hits. Week 6, sprint ends — model is at 81%, business required 90%. Week 7, the PM is in a difficult conversation with the board. Nobody was incompetent. The planning model was wrong.",[15,307,308],{},"The workaround — shipping on the calendar regardless of model performance — is the most expensive version of this failure. Iteration in production is significantly harder than iteration in development.",[35,310,311,316],{"label":281},[15,312,313,315],{},[19,314,99],{}," if your timeline has a fixed delivery date with no explicit plan for what happens if the model hasn't reached required performance by that date.",[15,317,318,106],{},[19,319,105],{},[68,321,323],{"id":322},"_8-the-user-sees-your-ai-output-for-the-first-time-at-the-sprint-review","8. The user sees your AI output for the first time at the sprint review.",[15,325,326],{},"AI outputs are probabilistic. When they're wrong, they can be wrong in ways that are confidently stated, superficially plausible, and hard to identify without the right context. A user who sees an AI output for the first time — without that context — doesn't know when to trust and when to verify. Their implicit expectation is deterministic: if the system says X, X is correct. When they discover it isn't, the trust damage is significant and often permanent for that user.",[15,328,329,330,332],{},"The midnight auditor story from ",[76,331,301],{"href":78}," is exactly this failure at client scale. The document processing system performed exactly as designed. The executive who ran it twice and saw different phrasing had never been calibrated on what \"designed\" meant.",[15,334,335],{},"What calibration looks like before MVP: a working session, not a demo. Key users interact with the system, surface edge cases, and explicitly name which failure modes they can tolerate versus which ones make the tool unusable. The session produces requirements for the fallback paths that need to be in the MVP — not added later.",[35,337,338,343],{"label":281},[15,339,340,342],{},[19,341,99],{}," if the first time a non-technical stakeholder or intended end user will interact with the AI output is at or after the MVP launch.",[15,344,345,347],{},[19,346,105],{}," if calibration sessions are planned before launch.",[50,349,351],{"id":350},"capability-risks-1-point-each","Capability Risks — 1 Point Each",[15,353,354],{},"Risk multipliers about the team and the system's external exposure. Not disqualifiers — context-dependent signals that change what the MVP needs to include, not whether to proceed.",[68,356,358],{"id":357},"_9-the-first-time-in-production-is-the-most-expensive-classroom","9. The first time in production is the most expensive classroom.",[15,360,361],{},"A first-time team can ship a good AI product. What they can't do is anticipate the failure modes they've never hit — and those are the expensive, late-to-surface ones.",[15,363,364],{},"Three things first-time teams routinely underestimate:",[126,366,367,373,379],{},[129,368,369,372],{},[19,370,371],{},"Production hardening:"," latency, load, edge cases, and graceful degradation behave differently in production than in a notebook.",[129,374,375,378],{},[19,376,377],{},"Integration complexity:"," the gap between \"the model returns the right answer\" and \"the answer is wired in with proper error, loading, and fallback states\" is bigger than it looks.",[129,380,381,384],{},[19,382,383],{},"Monitoring & drift:"," the model that launched well will degrade; without monitoring, you learn it from the business metric months later instead of the system weeks earlier.",[15,386,387],{},"If this signal is present, it changes MVP scope: the infrastructure an experienced team builds by reflex, such as monitoring, fallback paths, and evaluation loops, has to be explicit in the plan, not assumed.",[35,389,390,395],{"label":281},[15,391,392,394],{},[19,393,99],{}," if no one on the core team has shipped an AI feature to production with real users and real adoption, and no one with that experience is shaping the architecture.",[15,396,397,399],{},[19,398,105],{}," if someone is.",[68,401,403],{"id":402},"_10-publishing-ai-at-scale-without-an-editorial-gate-makes-you-invisible","10. Publishing AI at scale without an editorial gate makes you invisible.",[15,405,406],{},"The quality risk is obvious. The competitive one isn't.",[15,408,409],{},"A single bad AI-generated post is embarrassing. A thousand of them, published without editorial review over six months, is a different kind of problem — and it isn't quality. AI-generated content is technically correct, stylistically coherent, and completely generic. It reads like every other AI piece in your space. The brand that maintains an editorial standard — where human judgment shapes what ships — stays distinguishable. The brand that publishes at AI scale without one becomes invisible, because it's indistinguishable from everything else.",[15,411,412,413,419],{},"This isn't hypothetical. Per SEO analyses ",[76,414,418],{"href":415,"rel":416},"https:\u002F\u002Fzkami.substack.com\u002Fp\u002Fhow-clickups-blog-lost-976-of-its",[417],"nofollow","using third-party traffic data",", ClickUp's blog fell from roughly 1.19 million monthly organic visits to about 28,000 in 15 months. A 97.6% collapse across nine consecutive Google updates. The content wasn't broken. It was scaled, templated, promotional, and interchangeable. And Google's Helpful Content System, with AI Overviews answering the query directly, stopped rewarding it.",[15,421,422],{},"You pay for this slowly: not with one bad post, but with months of compounding invisibility.",[35,424,425,430],{"label":281},[15,426,427,429],{},[19,428,99],{}," if any public-facing AI-generated content ships without human editorial review or a defined content\u002Fstyle policy.",[15,431,432,434],{},[19,433,105],{}," if an editorial gate is in place.",[244,436],{"label":437,"max":438},"Execution\u002FCapability subtotal","5",[244,440],{"label":441,"max":442},"Total score","15",[444,445,446,463],"table",{},[447,448,449],"thead",{},[450,451,452,457,460],"tr",{},[453,454,456],"th",{"align":455},"left","Score",[453,458,459],{"align":455},"Risk level",[453,461,462],{"align":455},"Recommended action",[464,465,466,480,493,506],"tbody",{},[450,467,468,474,477],{},[469,470,471],"td",{"align":455},[19,472,473],{},"0–3",[469,475,476],{"align":455},"Low",[469,478,479],{"align":455},"Proceed. Apply standard AI delivery discipline.",[450,481,482,487,490],{},[469,483,484],{"align":455},[19,485,486],{},"4–7",[469,488,489],{"align":455},"Moderate",[469,491,492],{"align":455},"Fix all structural signals before committing full budget.",[450,494,495,500,503],{},[469,496,497],{"align":455},[19,498,499],{},"8–11",[469,501,502],{"align":455},"High",[469,504,505],{"align":455},"Stop. Address structural signals. Re-scope before proceeding.",[450,507,508,513,516],{},[469,509,510],{"align":455},[19,511,512],{},"12–15",[469,514,515],{"align":455},"Critical",[469,517,518],{"align":455},"Do not fund at current scope. This is not a risk to manage — it is a budget to lose.",[15,520,521],{},"Most of the signals on this scorecard are planning and scoping failures — not technical ones. A proxy metric can be replaced. Success criteria can be defined. A scaling cost calculation can be run in an afternoon. A fallback approach can be added to the plan. The organizations that do this work before the sprint starts will have fewer projects that fail at the structural level, and more that compound.",[15,523,524],{},"The organizations that don't will keep running the same retrospective. The technology was fine. The team was capable. The project didn't deliver. The scorecard exists to make the reason visible before the budget is committed, not after.",[15,526,527],{},"If you want an external perspective on where your project sits across these dimensions — a structured assessment of the structural and execution risks before a budget decision is made — that is what Prosperaize's Prosperity Audit is.",[15,529,530],{},[531,532,533],"em",{},"Dušan Stamenković is the founder of Prosperaize, an AI Asset Management Consultancy. He advises organizations on whether, where, and how to invest in AI — reducing risk and maximizing return across the AI investment lifecycle.",{"title":535,"searchDepth":536,"depth":536,"links":537},"",2,[538,539,547,552],{"id":52,"depth":536,"text":53},{"id":62,"depth":536,"text":63,"children":540},[541,543,544,545,546],{"id":70,"depth":542,"text":71},3,{"id":109,"depth":542,"text":110},{"id":155,"depth":542,"text":156},{"id":183,"depth":542,"text":184},{"id":208,"depth":542,"text":209},{"id":256,"depth":536,"text":257,"children":548},[549,550,551],{"id":263,"depth":542,"text":264},{"id":294,"depth":542,"text":295},{"id":322,"depth":542,"text":323},{"id":350,"depth":536,"text":351,"children":553},[554,555],{"id":357,"depth":542,"text":358},{"id":402,"depth":542,"text":403},"Strategy","Before you fund it, score your project on these ten dimensions. Or accept the downside.","md","\u002Fimages\u002Finsights\u002Fai-doesnt-reward-risk-it-punishes-bad-scoping.png","A senior executive writing performance charts on a glass board, a colleague in the background",{},true,"\u002Finsights\u002Fblog\u002Fai-doesnt-reward-risk-it-punishes-bad-scoping","2026-06-16",[566,567],"the-hidden-roi-killer-in-ai-projects-starting-with-the-solution","your-data-isnt-ai-ready-neither-is-your-initiative",{"title":569,"description":570},"AI Risk Assessment: Score Your Project Before You Fund It","Score your AI project on 10 risk dimensions before you commit budget. Most AI failures are scoping failures, not technical ones. Here's how to catch them.","insights\u002Fblog\u002Fai-doesnt-reward-risk-it-punishes-bad-scoping",[573,574,575,576,577],"AI Strategy","AI Investment","Risk Assessment","Scoping","Feasibility","blog","9007G5Gk3r263JfB8mtC3ETnk3aBpHdUe6MqKSvdzsI",[581,587,592],{"title":582,"path":583,"featuredImage":584,"type":578,"publishedAt":585,"description":586},"If Your Data Isn't AI-Ready, Neither Is Your Initiative","\u002Finsights\u002Fblog\u002Fyour-data-isnt-ai-ready-neither-is-your-initiative","\u002Fimages\u002Finsights\u002Fblog-data-not-ready.png","2026-05-06","Data readiness — not data existence — determines whether an AI initiative ships. Five practices that turn an existing data environment into one AI systems can actually use.",{"title":79,"path":588,"featuredImage":589,"type":578,"publishedAt":590,"description":591},"\u002Finsights\u002Fblog\u002Fthe-hidden-roi-killer-in-ai-projects-starting-with-the-solution","\u002Fimages\u002Finsights\u002Fblog-roi-killer.jpg","2026-04-05","Most AI investment failures are decided in the first two weeks. This is how you ensure those two weeks are used wisely.",{"title":593,"path":594,"featuredImage":595,"type":596,"publishedAt":597,"description":598},"$200K, Six Months, and an AI Agent That Couldn't Survive Real Use","\u002Finsights\u002Fcase-study\u002F200k-failed-ai-agent-rescue-to-production-intelligence","\u002Fimages\u002Finsights\u002Fslika_za_case_study.jpeg","case-study","2026-04-20","From failed deployment to operational intelligence: how Prosperaize used audit, data refoundation, and iterative validation to deliver an agent that works in real environments.",1781619920589]