Articles

    Corporate AI business modelling: why 11 pilots rarely add up to one return

    December 7, 2025
    8 min read
    By Tamsin Velkovane
    Updated October 26, 2026
    Share this article

    Say an insurer ends the year with 11 AI pilots. Nine report success on their own dashboards, and none can show a figure the CFO would book.

    The models work. What fails is judging each pilot on its own. The money only adds up across the whole set of projects, where running costs, answers that are sometimes wrong and changing skill needs have to be managed together. Corporate AI business modelling means deciding where AI creates value across that set, how to put it into daily operations, and how to measure a return that still holds at full scale.

    Treating AI bets as a portfolio with a single budget

    At product level, the question is whether one feature pays for itself. At company level it is wider. Which workflows does AI touch? Which systems does it change? Which data does it depend on? And what running costs and commitments does it quietly add along the way? Answering it takes one strategy and one budget for the whole set, because features justified one at a time tend to duplicate data work and hide the costs they share.

    The first move is getting every initiative onto one page, sorted by the return horizon that actually governs how it should be funded. Efficiency and automation (document processing, routing and classification, anomaly detection, customer-service automation) pays back in the short term and is where most enterprises should start. Experience and personalisation (assistants and copilots, dynamic recommendations, workflow augmentation) returns over a middle horizon. Strategic, AI-native products (new revenue lines, proprietary domain models, partner-and-data platforms) pay back slowly and carry the most risk. Zillow Offers is the cautionary case: a model-priced home-buying business that Zillow shut down in November 2021 after an inventory write-down of about $304 million.

    Mixing these on one list without their horizons is how enterprises end up starving a compounding bet to fund a quick win, or the reverse.

    AI can be deployed almost anywhere, so the useful question is where it makes an existing advantage harder to copy. For some enterprises that means proprietary data the market cannot access; for others it is operational excellence, where a few points of throughput compound across thousands of transactions; for others still, domain knowledge, a distinctive customer experience, or a partner ecosystem. AI amplifies whatever the company is already better at. Applied outside those areas it produces improvements competitors replicate within a quarter. Modelling the choice explicitly (in terms of value, risk and scenario outcomes across the portfolio) is what keeps investment concentrated where the advantage actually compounds.

    Portfolio metrics have to answer three separate questions. Mixing them up is the usual reason enterprise AI reports become unreadable. The money question asks whether the numbers work: output per workflow, cost per automated task, risk actually reduced. The usage question asks whether people really use the tool or it is merely installed. The technical question asks whether the model stays reliable over time. That one matters because a tool that slowly gets worse keeps showing healthy usage right up to the day users stop trusting it. Above all three sits contribution to the north-star metric, whether that is efficiency, throughput or engagement, which is what connects an individual initiative to a portfolio-level story.

    Getting from anecdote to a defensible return figure

    Finance will not sign a return figure built on adoption charts. AI ROI behaves unlike ordinary software ROI, because it drags in inference cost, retraining cycles, safety requirements and data-governance obligations that a licence-and-seats model never had to price.

    Cost-to-serve at enterprise scale is rarely just the number teams quote first. Inference cost per request is the visible driver, but context length and token generation decide how large it gets in practice.

    Looking up documents in a search database adds a second cost on every call, and small pilots tend to hide it. The data-centre region changes the price of some services, sometimes by more than the choice of model, so apply the right rate to each cost line. Two more costs appear only in production. Running and monitoring the models is a permanent expense. Retraining comes back again and again for as long as the tool is in use. Business cases that count all of these, and project costs several years ahead, are the ones that survive at scale.

    Cost-to-serve per request, decomposed

    cost_per_request =
        ((input_tokens + output_tokens) / 1e6) * price_per_1M_tokens   # generation
      + retrieval_cost_per_call                                        # vector-DB lookup
      + monitoring_and_mlops_allocation                                # amortised per request
      # the whole sum is then multiplied by the compute-region factor
    
    • input_tokens + output_tokens: context length and generation length are what actually move the bill; the model name sets only the rate.
    • price_per_1M_tokens: the visible driver teams quote first, and usually the smallest surprise.
    • retrieval_cost_per_call: the second per-call cost low-volume pilots hide; it scales with production traffic.
    • monitoring_and_mlops_allocation: a permanent operating cost spread across requests, with periodic retraining amortised the same way.
    • region factor: applied to the whole line, because the same service can be priced differently from one data-centre region to the next.

    A quick check: a 3,000-token call at $5 per 1M tokens is ~$0.015 of generation before retrieval, ops and region are added. The drivers a pilot ignores add up at portfolio volume, so carry the calculation through one workflow.

    Take claims-document triage at an illustrative 400,000 documents a year, two model calls per document. Per call: $0.015 of generation, $0.002 of retrieval and $0.004 of monitoring and MLOps allocation, times a region factor of 1.2, gives (0.015 + 0.002 + 0.004) × 1.2 ≈ $0.025. The annual run cost is 400,000 × 2 × $0.025 ≈ $20,000. Generation, the line everyone quoted in the pilot, is about 70% of the per-call figure; the other 30% and the region uplift are what the pilot budget missed.

    The saving side needs the same discipline. If triage saves three minutes per document, the headline is 20,000 hours. Suppose only a quarter of that turns into lower overtime and contractor spend at $40 an hour: 5,000 × $40 = $200,000 of bookable saving. Charge the workflow its share of the platform (an evaluation specialist and part of an MLOps team, say $120,000 a year) and the net is about $200,000 minus $20,000 minus $120,000, or $60,000. Positive, but thin. Put a second use case on the same platform and each carries $60,000 of shared cost instead of $120,000, which roughly doubles the net for both. That arithmetic, more than any architecture diagram, is the case for treating AI as a portfolio.

    The return itself accumulates in four layers worth keeping separate, because they arrive on different timelines and convince different audiences.

    Direct cost savings show up first: reduced labour hours, faster resolution, shrinking backlogs. Productivity gains follow (higher throughput, lower error rates), harder to book but usually larger. Strategic benefits (retention, cross-sell, reduced compliance risk) take longest to attribute yet are typically what justifies the whole programme. Incremental revenue, through premium upsell and eventually AI-native lines, arrives last. Whichever layer is being claimed, tie it to a specific operational KPI. An hour saved that never leaves the department never reaches the P&L.

    An A/B test alone cannot prove any of these figures, because what changes is a whole workflow. Checking starts offline: test the model on data it has not seen, before any real user is exposed to it. Then run controlled pilots in a small part of the organisation. After that, use real operating data to turn the model's behaviour into business terms, such as value per task. Watch for side effects too; the most common is a faster step that quietly pushes extra work to the next team. Finally, check that the effect is large enough and not just noise before scaling a promising pilot.

    The road from pilot to something the enterprise runs

    Standing up the capability that produces the return usually runs in three phases. The first is deliberately cheap: at this stage the company is paying for information, and returns come later. Companies identify the workflows amenable to AI at all, build proof-of-value pilots narrow enough to fail without consequence, and, usually the step that reshapes the whole plan, map what data actually exists and what condition it is in. Evaluating early model behaviour here means learning where the model fails and whether those failures are tolerable inside the workflow.

    Learning precedes scaling, and skipping this phase is how enterprises discover in month eight that the data cannot support the plan.

    The second phase turns one-off wins into shared tools: a common catalogue of models, a shared way to test them, one place for prepared data, data rules, alerts when a model's quality slips, and safety and compliance checks. This is where AI stops being a set of experiments and becomes routine. From here, the second and third use cases cost a fraction of the first.

    The third phase scales into products that could not exist without the model: AI embedded across business units, AI-native customer experiences, domain-specific platforms, multi-step workflows, and consolidated model families that business units reuse. This depends less on model quality than on product leadership clear enough to hold the portfolio together.

    The capability work that decides whether any of this lands

    The second phase above assumes someone can build the shared infrastructure and someone else can tell when it is drifting. That assumption is where many transformations break: organisations fund models and leave the skills and roles around them to chance, even though capability is what still sets a company apart once the models themselves are available to everyone.

    Product managers come first, because they decide which workflows enter the portfolio at all, and whoever admits a workflow is the person who has to know what the model will get wrong inside it. They need enough AI literacy to know a model's constraints, enough data fluency to question a dataset, and enough grasp of cost and inference economics to read the cost-per-request line above without help. Experimentation and evaluation, an eye for ethics and compliance, and the cross-functional orchestration that holds these together complete the role.

    What a product manager specifies, engineers have to keep running: models that cope with growing traffic, training, alerts when quality slips, automatic retraining, monitoring and several models working together. That work is the difference between a demo and a system that stays reliable at scale. Between the two sits a role many enterprises skip: the evaluation specialist who decides whether a model is good enough. This person keeps the test data accurate, sorts errors into clear types, sets quality thresholds and tests for bias and made-up answers. Without that gate, the quality alerts set up in the second phase have nothing to measure against.

    None of this lasts if it lives in a few heads. The organisation has to institutionalise what it learns, through internal labs, cross-functional guilds, shared repositories and scenario exercises, so the transformation does not depend on a handful of individuals who can walk out the door.

    Governance, and the risks you have to price in

    Governance belongs in the model from the first pilot, because the enterprise has to trust what those capabilities produce before it lets them compound. A common scaffold is NIST's AI Risk Management Framework 1.0, published on 26 January 2023, which sorts the work into four functions: Govern, Map, Measure and Manage.

    In practice it covers rules for data, written records for each model, when a person must check the output, risk scores, audit trails and version control for models. Done properly, it is what lets a regulated enterprise deploy a second and third use case without re-litigating trust each time.

    The risks that shape both pricing and product constraints are specific and worth putting numbers on: hallucination and inaccuracy, data privacy and residency, compliance violations, model degradation and drift, adversarial misuse, bias and fairness, and cost spikes from unpredictable load. Each of these is a line in the business case, because a risk left unpriced becomes a cost that arrives without a budget.

    Pricing one means expected cost. If, in the triage example, 1 in 2,000 automated documents needs a $400 manual remediation, 400,000 documents a year carry 200 × $400 = $80,000 of expected cost. That single line is four times the inference bill and larger than the whole net saving, which is why the error rate belongs in the business case next to token prices.

    When a portfolio office only slows things down

    Teams that have shipped AI successfully without any of this have a fair complaint. Central portfolio governance adds review boards, shared platforms become queues, and the business unit that knows its workflow best waits months for a platform team to prioritise it. Many early wins came from one team with a budget, an API key and a clear problem. On that view the portfolio office is overhead dressed as discipline.

    They are right when there is nothing to share. With one or two production use cases, different data sources and no common vendor contract, a central platform has no second customer to amortise against, and the review layer costs more than it saves. Let those teams ship, with only the cost-per-request and error-rate lines reported upward.

    The complaint stops holding at roughly the third use case that touches the same data, the same model vendor or the same regulator. From that point duplicated data work and unpriced shared risk grow faster than the speed a team gains by going alone, and the triage arithmetic above starts to favour the platform. The threshold is a judgement call; the trigger worth writing down is the first time two teams pay separately to clean the same dataset.

    Strategy, money, roadmap and risk in one view

    Governance and its priced risks are the last of six models an enterprise AI business case has to carry at once. The other five came up earlier: a strategic position (differentiation, data advantage, long-horizon capability bets); a capability architecture running from data to model to orchestration to experience; a financial model covering cost-to-serve, retraining cycles, unit economics, ROI and portfolio impact; an organisational-capability model tracking skills and maturity; and an experimentation model spanning offline evaluation, online impact testing and business-case validation. Governance, with compliance, safety and lifecycle documentation, completes the set.

    Keeping them in one view matters because changes travel between them. Take a cheaper model. On its own it is good news for the cost-per-request line, yet it also reopens the evaluation gate, may alter where and how the data it touches is processed, and can change which workflows clear the ROI bar at all. A new regulation or a data-quality problem propagates the same way. When the six models sit side by side, each change can be traced through to the others; when they live in separate decks, it surfaces months later as a surprise.

    Prioritisation follows from the same view: value, feasibility, risk, data quality and reuse potential, applied with one lens to every initiative so that money follows the portfolio logic and the loudest pilot loses its advantage. Reuse deserves more weight than most scoring sheets give it. In the triage numbers, a second use case on the shared platform halves the platform charge each one carries, from $120,000 to $60,000. A candidate that reuses the claims dataset and the evaluation harness can therefore beat a standalone pilot with a headline saving $50,000 higher, and only the portfolio view shows that.

    A practical starting point is the one-page portfolio view from the first section, extended with a column for each of the other models: cost-to-serve per request, the ROI layer being claimed, the governance controls in place, and the skills the initiative depends on. An initiative with blank cells is not ready for a scaling budget. Filling those gaps through capability mapping before the next pilot is funded is what turns a drawer of disconnected experiments into a portfolio that finance, risk and product leaders read the same way, and keeps AI spend from settling into a standing line of sunk cost.

    Related Articles