# Enterprise GenAI Use Cases and Value
In 2023, a global consumer bank stood up 47 generative AI proofs-of-concept in nine months. Eighteen months later, exactly three had reached production. The other 44 died in what the bank's CDO now calls "the demo graveyard" — a shared drive full of Streamlit apps that dazzled a steering committee once and were never touched again. The postmortem revealed the pattern: nearly every dead project had been chosen because it was *technically interesting* or *executive-visible*, not because anyone had modeled where the value would actually land or who would own it after the applause.
This is the CDO's core problem with GenAI. The technology demos beautifully and produces almost nothing without disciplined selection. Your job is not to run experiments — your fundamentals covered that. Your job is to build a portfolio that compounds. That requires a different set of judgments than any previous data initiative, because GenAI inverts several assumptions you've been operating on.
You already know how to prioritize analytics projects: estimate the value, estimate the cost, weight by feasibility, sequence by dependency. GenAI quietly violates three of the assumptions underneath that method.
Marginal cost is not zero — it's variable and often unpredictable. A traditional BIBITechnologies and processes that turn raw data into actionable insights via reporting, dashboards and analysis, so teams can decide based on facts rather than intuition. dashboard costs the same to serve 10 or 10,000 users. A GenAI feature costs per , per call, per retrieval. A summarization tool that looks cheap in a pilot with 50 users can become a six-figure monthly line item at enterprise scale — especially if you've built it on a frontier model where you're paying a premium for reasoning you don't actually need. Unit economics must be part of use-case selection *from day one*, not a scaling afterthought.
Accuracy is probabilistic, so the value equation includes an error-cost term you can't ignore. The relevant question is never "does it work?" It works often enough to demo. The question is: *what does a wrong answer cost, and who catches it?* A GenAI tool that drafts marketing copy has a trivial error cost — a human reviews it, and the worst outcome is a discarded draft. The same underlying capability applied to summarizing a patient's medication history has a catastrophic error cost. Same technology, radically different value profile, driven entirely by the consequence of being wrong.
The moat is rarely the model. Anyone can call the same APIAPIApplication Programming Interface: a standardised interface that lets applications communicate and exchange data without knowing each other's internal workings.Voir la définition complète → you can. Durable advantage comes from what you wrap around the model: proprietary data, domain-specific evaluation, workflow integration, and the accumulated feedback that lets you tune and route intelligently. When you evaluate a use case, ask what part of it a competitor *couldn't* replicate by next quarter. If the answer is "nothing," you're building a feature, not an asset.
Hold these three shifts — variable cost, error-cost weighting, and moatmoatA lasting edge over competitors: a resource, capability or position they cannot easily replicate, letting a firm earn above-average returns over time.Voir la définition complète → location — because every framework below is built on them.
Plot every candidate use case on two axes. The first is value concentration: does this use case attack a large, measurable cost or revenue pool, or does it sprinkle small conveniences across the org? The second is error tolerance: how much does a wrong output cost, and how contained is that cost?
This produces four quadrants, and each demands a different play.
This is the fertile ground, and it's where most enterprises *don't* start because it's less glamorous than the risky stuff. Think contact-center agent assist that drafts responses a human sends, developer code-completion, contract-clause extraction with a lawyer in the loop, or first-draft generation for high-volume knowledge work. The value pool is real and large; the human in the loop absorbs the error cost.
The classic reference point is Klarna's customer-service assistant, which the company said handled the equivalent of hundreds of agents' worth of volume. What made it work was not model sophistication — it was that the use case sat squarely in this quadrant: enormous, measurable ticket volume (concentrated value) with human escalation paths and a bounded domain (contained error cost). That is the shape of a first bet.
Autonomous financial reconciliation, clinical decision supportdecision supportTechnologies and processes that turn raw data into actionable insights via reporting, dashboards and analysis, so teams can decide based on facts rather than intuition.Voir la définition complète →, automated underwriting. The prize is large but a wrong answer is expensive or dangerous. These are *not* off-limits — they're where the biggest advantages eventually live — but they require investment in guardrails, evaluation harnesses, human-review gates, and often a "confidence routing" design where the system only acts autonomously above a calibrated certainty threshold and escalates everything else. Sequence these *after* you've built organizational muscle in the safer quadrant.
The internal "chat with our wiki" bot. The meeting summarizer nobody trusts enough to skip the meeting. These are cheap to build, safe to run, and easy to demo — which is precisely why they metastasize. They generate activity and consume roadmap without moving a number anyone reports to the board. Kill these quickly or bundle them into a platform play, but never let them anchor your portfolio.
High risk, low reward. If a use case lands here, the only correct action is to not do it, or to redesign the workflow so the error cost drops before you touch it.
The discipline this framework enforces: *you never justify a use case on feasibility alone.* "We can build it" belongs to no quadrant. Value and error cost are the only admission criteria.
Quadrants sort; they don't rank. To sequence a portfolio, translate each surviving candidate into four numbers your CFO and your board will recognize.
1. Value pool ($): The addressable cost or revenue the use case touches, multiplied by a realistic capture rate. Be brutal on capture — a tool that saves each analyst 30 minutes a day only creates value if that time converts to output or headcount, not if it evaporates into longer coffee breaks. Model the *realized* value, not the theoretical.
2. Unit economics at scale ($ per transaction): Estimate cost per useful output at projected volume, using the model you'd actually deploy — not the flagship model you used to prototype. Most production use cases run fine on smaller or open-weight models once retrieval and promptingpromptingPrompt engineering is the practice of designing and refining text inputs to guide large language models toward accurate, relevant, and reliable outputs.Voir la définition complète → are tuned.
3. Error-cost exposure: The expected cost of a wrong output — probability of error times consequence times volume, minus what your review layer catches. This is where you force the confidence conversation before building.
4. Moat contribution: Does this use case compound? Does it generate proprietary feedback data, deepen a workflow lock-in, or build reusable platform capability? A modest-value use case that builds a reusable retrieval-and-evaluation platform can outrank a higher-value one-off.
A simple weighted score makes the tradeoffs explicit and defensible in a governance forum:
priority_score = (
value_pool * capture_rate # realized annual value ($)
- annual_run_cost # tokens + infra at scale ($)
- expected_error_cost # P(error) * consequence * volume ($)
) * moat_multiplier # 1.0 baseline, up to ~1.5 for platform-buildingThe output isn't a precise dollar figure — treat it as a ranking instrument. Its real power is procedural: it makes the person championing a beloved pet project state their capture rate and error cost out loud, in front of peers. Half the demo graveyard would never have been dug if sponsors had been forced to name a capture rate.
One more sequencing rule: bias your first two or three production bets toward use cases that share infrastructure. If your top-ranked contact-center tool and your third-ranked contract-analysis tool both need a document-retrieval pipelinepipelineAll active sales opportunities across the stages of the sales process, together with their combined potential value and probability of closing.Voir la définition complète → and an evaluation harness, building them together means the second one is far cheaper and your platform moatmoatA lasting edge over competitors: a resource, capability or position they cannot easily replicate, letting a firm earn above-average returns over time.Voir la définition complète → starts compounding. Portfolio thinking beats project thinking precisely here.
Vérification des acquis
1. The lesson describes a 'demo graveyard' of abandoned GenAI proofs-of-concept. What is the underlying selection failure this metaphor illustrates?
2. Why does the lesson argue that GenAI unit economics must be part of use-case selection from day one rather than a scaling afterthought?
3. The lesson states that with GenAI 'the relevant question is never does it work?' What deeper reasoning does this reflect about evaluating GenAI value?
4. Select ALL correct answers. According to the lesson, which assumptions underlying traditional analytics prioritization does GenAI violate?
Sélectionnez toutes les réponses correctes.
5. Select ALL correct answers. What does the lesson imply is the CDO's actual job with respect to enterprise GenAI?
Sélectionnez toutes les réponses correctes.
Selecting well gets you a good roadmap. Realizing value requires three governance mechanisms that most organizations bolt on too late.
Stage-gate with kill criteria defined up front. Every use case advances through pilot → limited production → scale, and *each gate has a numeric threshold set before the work begins.* Not "does the demo look good" but "does it achieve X accuracy on the held-out evaluation set AND land within Y cost per transaction AND show measurable adoption from the target users." The bank in our opening had no kill criteria, so nothing ever died on schedule — projects lingered as zombies, consuming budget while producing nothing. Pre-committed kill criteria are the antidote to sunk-cost paralysis. The willingness to kill is itself a competitive capability: it frees capital to concentrate on winners.
Evaluation as a first-class asset, not an afterthought. For every production use case, you need a domain-specific evaluation set — a curated collection of inputs with known-good outputs, scored continuously. This is where your proprietary moatmoatA lasting edge over competitors: a resource, capability or position they cannot easily replicate, letting a firm earn above-average returns over time.Voir la définition complète → physically lives. Two competitors using the identical foundation model are differentiated entirely by whose evaluation harness is better, because that's what lets you swap models, tune prompts, and catch regressions without fear. Treat your eval sets like you treat your master data: owned, versioned, and governed. When a new model is released, the organization with a strong eval harness knows within a day whether to switch; the organization without one is guessing.
Value attribution wired in before launch. Decide *how you will prove* a use case created value before you ship it. That means instrumenting the baseline now — current handle times, current draft-to-publish cycle, current error rates — because once the tool is live you can no longer measure the world without it. Where possible, run a genuine holdout: a control group that doesn't get the tool. GitHub's productivity claims for its coding assistant carried weight precisely because they were backed by controlled measurement, not vibes. Vague "efficiency gains" get discounted by every CFO who has heard the phrase before; a defensible before-and-after with a control group is what unlocks the next round of funding.
These three mechanisms convert a portfolio from a slide into a machine. Stage-gates ensure capital flows to winners. Evaluation harnesses build the technical moatmoatA lasting edge over competitors: a resource, capability or position they cannot easily replicate, letting a firm earn above-average returns over time.Voir la définition complète →. Value attributionattributionA framework for assigning credit to the touchpoints that contributed to a conversion, so you can measure which channels and interactions actually drive results.Voir la définition complète → builds the *credibility* moatmoatA lasting edge over competitors: a resource, capability or position they cannot easily replicate, letting a firm earn above-average returns over time.Voir la définition complète → that keeps the board funding the program through the inevitable trough after the initial excitement fades.
The CDOs who will look prescient in three years are not the ones who ran the most experiments. They're the ones who were ruthless about the four quadrants, who forced capture rates into the open, who built shared infrastructure across their top bets, and who could prove — with a control group and a straight face — exactly what value each production system created.