Governed AI is not the same as AI that delivers.
AI Value is about the return on AI. Is a use case worth doing? Is the one you built delivering? Where would the next investment return the most? AI already runs in decisions, processes and customer contact, so the harder question is what it has returned. The spending is real, and it recurs. Unless measurement was planned in from the start, the evidence often comes down to a demo, a usage chart and a strong opinion. We help close that gap in both directions: proving the value that's there, and finding where there's more.
- Written against
- ISO 31000:2018 · ISO/IEC 42001:2023 · ISO/IEC 38507:2022 · ISO 21502:2020 · ISO 9001:2026 · PMBOK Guide 8th ed.
- Last updated
How do we show that AI pays off?
Measure value at the process level first, against the process's own baseline, then roll it up to the financial figure. Typical process measures are first-time-right rate, cycle time, exception and rework rate, and on-time delivery. Realised value is then compared with the range approved at the gate, so the business case is checked after go-live and not only before.
What is a use-case gate?
A gate is the decision, before commitment, on whether a proposed AI use case is worth doing. Each proposal is assessed on five criteria: expected value (a range with a named owner), feasibility, risk, data readiness, and a kill criterion set before work begins. Rejecting a weak proposal counts as a success.
How do we manage an AI portfolio?
Keep one AI inventory and look at it through three lenses: value, risk and governance. Related use cases that share a data foundation or platform are managed together as a benefit case, with benefits tracked through handover and sustainment. Expected return, discounted for uncertainty and net of failure and control costs, is ranked alongside risk.
Value sits inside the definition of the work.
Project management doesn't treat value as an extra added on top of delivery. The current PMBOK Guide defines a project as a temporary initiative in a unique context undertaken to create value. It defines value as the excess of benefits over investment, financial and nonfinancial. And it names value per unit of investment as the ultimate indicator of success.
The same text treats projects as investments: the expected value has to clear a threshold before the work is justified, and the gate on this page applies that threshold. What a project hands over isn't yet value. Deliverables lead to outcomes, outcomes to benefits, and benefits to value, so we judge a project by the far end of that chain.
The definition scales. A portfolio is managed as a group to achieve strategic objectives. It's where an organization decides which value is worth pursuing, and measured project value is the data behind each decision to select, continue or stop. This page follows the same shape: value decided at the gate, read across the portfolio, protected in operation, and delivered under a discipline that can prove it. Quality management says the same from the other side: a process exists to deliver its intended outcome, and improvement is a duty, not an option.
The cause is structural. You buy AI as a capability and put it into a process that's already running, so from day one there's nothing to compare it with. What the process cost before is rarely on record. Where a process runs under a quality management system, its performance is already measured and recorded. That is the baseline the value case needs, and it is why we start with the processes that have one. Activity metrics fill the gap: queries served, seats licensed, documents summarised. None of them is a financial outcome.
The second problem is timing. A business case written after deployment gets built backwards from whatever happened to be measured. It reports what was easy to capture, and misses what the investment was meant to change. Then nobody can answer the honest question: would we fund this again?
Perception adds to it. People can feel faster without being faster, because the effort drops while the elapsed time doesn't. Nobody is being dishonest, but it's one more reason to measure the work instead of surveying people.
A quality management system has one purpose: that the organization's processes deliver their intended outcomes, consistently, and that the organization learns when they don't. Everything in ISO 9001 serves that: the process approach, the objectives and measures, risk-based thinking, audits, management review, improvement. The system is not the point; the outcomes are.
That is also where AI earns its keep. AI can improve the process itself and its results: fewer exceptions, shorter cycles, fewer escapes, more consistent decisions. Not the paperwork of the QMS, the processes it manages.
So value is first measured where the AI acts, at the process: the intended outcome, the measure the QMS already reports, the baseline it already holds, and the improvement the AI brings. The financial line is the roll-up, not the starting point. The gate in the next section then prices that improvement.
Take an order-handling process with a first-time-right rate and a cycle time. AI classifies and routes the incoming orders. The two measures move, and the QMS records the change as improvement, with the baseline beside it.
An organization that runs ISO 9001 already has what an AI value case usually lacks: a defined intended outcome, a measure, a baseline and an improvement duty. We start there.
The cheapest time to fix an AI investment is before you commit to it, and committing rarely means building. Far more often it means buying a licence, switching on a feature that shipped disabled, or extending a tool into a decision it was never bought for. Every proposal goes through the same gate, whichever of these it is. Some shouldn't get through, and stopping a weak proposal early counts as a success. The ones that pass go into the inventory, long before anything is live.
- 01
Expected value
Which financial line moves, by how much, and who owns that number: revenue, cost, cycle time or margin, named before you buy, enable or build. We state it as a range with a confidence level. A forecast of thirty percent that might really be eight is a different investment from one that might be forty.
- 02
Feasibility
Whether you can actually reach the capability with the data, platform and integrations you have. If you can't, what would have to change.
- 03
Risk
Both directions, both priced. On the downside: what the system could get wrong at the volume it runs at, what remediation would cost, and what its controls cost to run. On the upside: how much confidence the benefit estimate really deserves. A proposal can look strong on headline value and still fail here. That's the gate doing its job.
- 04
Data readiness
Whether the data exists with the quality, provenance and permissions the case needs. If you bought the system, whether you control enough of its input to be accountable for it. Data failures often arrive disguised as model failures.
- 05
Kill criterion
What would have to be true for this to stop, written down before it starts. It's a threshold, set while nobody is yet invested in the answer. A use case with no stated stopping point doesn't get killed when it underperforms. It gets defended, and then renewed.
Risk belongs in the ROI, in both directions
When a business case gives one number for the benefit and files the risk separately, it's really a forecast with the uncertainty taken out. Under the definition of risk we work to, the uncertainty in the benefit is part of the risk: a deviation from the expected, in the favourable direction as much as the adverse one.
Say the forecast is a thirty percent cut in handling time. If that assumes full adoption, clean inputs and no exceptions, it's a different proposition from the same thirty percent with those assumptions tested. The untested version gets approved faster, which is a large part of why realised value falls short of the paper.
The downside goes into the same calculation. That's the cost of wrong decisions at the volume the system runs at, remediation and redress, and the running cost of controls: oversight staff, logging, evaluation and assessment. The controls are the easiest cost to leave out. Their cost rises steeply if the system falls in a high-risk tier or puts you in the provider role. A high-risk system costs more to run than a minimal-risk one doing similar work, and that difference belongs in the case at the gate, before it shows up in year two.
Two use cases can promise the same return and still be very different investments. Picture one built on assumptions you can test, and another on assumptions you can't. The estimate built on testable assumptions deserves more confidence, and we reflect that when we compare AI investments. A portfolio ranked on expected value alone will pick the volatile ones more often than it should. So we rank on risk-adjusted return.
AI creates value when it improves what an organization achieves, and its consequences don't outweigh the benefit. The financial return stays the measure we build on. But decision quality, the consistency of the process outcome, resilience, human capability, customer trust and the effect on the people a system reaches decide whether that return is real.
Automation changes more than cost. A decision to automate has to weigh its effect on people, skills, accountability and resilience. Fifty people at ten hours a week is a number, not yet a case. Three questions belong in the same decision as the saving. Once nobody does the work by hand, can anyone still check it? Does the process still run if the model or its vendor becomes unavailable? And who carries the consequence when the output is wrong at full volume?
Where these effects carry a cost, they go into the ledger and get discounted like any other exposure. Where they don't, they become criteria: the conditions a use case has to meet to go ahead, and the ones that stop it.
You can't manage the value of AI nobody can see. Before anything can be ranked, every use case and every system has to be in one place. That inventory is the same one the risk and governance work uses, read from a different angle. Built once, it answers three different questions.
Value — the portfolio
What each system costs to run, what it returns against its case, and where the next investment earns the most. Also where to stop: a healthy portfolio retires systems that no longer justify their cost.
Risk — the exposure map
What each system touches, what it would cost if it went wrong at the volume it runs at, and how much confidence the original value estimate still deserves. It's the same rows, read for what could deviate in either direction.
Governance — the register
Who owns each system, which controls apply, and what evidence exists that they operate. This is the view ISO/IEC 38507 places with the governing body.
Every use case and every AI system in one place: proposals that passed the gate, next to systems already in production. Holding both makes it a portfolio, because you can only place the next investment well when new candidates and running systems sit side by side. ISO/IEC 42001's Annex A asks for the resources behind each AI system to be identified and documented (A.4.2). NIST AI RMF asks for the inventory outright (GOVERN 1.6).
Not every AI investment is a single project. Most arguments about AI value happen one level up. Say several use cases share a data foundation, a platform and the AI literacy to run them. Together they form one benefit case. The shared parts produce no benefit on their own, and the use cases can't produce theirs without them.
Managed as a group, the benefit case runs a life cycle instead of stopping at a delivery date. Benefits are identified, planned, monitored through delivery, handed over to the areas that use them, and sustained after the programme closes. The last two steps decide whether the benefit materialises, so model monitoring, retraining and adoption support are planned before handover. The life cycle wasn't invented for AI. It comes straight from programme management.
A benefits register makes this manageable. It starts from the benefits and maps each one to the components that produce it. So a shared data platform gets no line of its own. Its contribution shows up only through the benefits of the components it enables. That's the honest answer to "what did the platform return?", and it's why platform investment is hard to defend on its own.
Whether a benefit is still worth pursuing gets rechecked throughout. If its life-cycle cost overtakes it, or the window it depended on closes, the case is reassessed. One use case can then be stopped or reshaped without the wider benefit case failing. The benefits management plan has a process for removing benefits that are no longer needed, which lets you enforce the kill criterion later, when people have a stake in the work and stopping is harder.
Risk and return on one ledger
Risk and return reach a decision by different routes: different people, at different times, on different scales. We put them on the same ledger. Expected return is discounted for uncertainty, less the potential cost of failure and the cost of controls. Ranked that way, the inventory starts to work like a portfolio. Investment goes where the risk-adjusted balance is strongest, and attention goes where the risk no longer justifies the return. That gives you a comparison you can defend in front of a board.
Value has to be assured, not just ranked
Ranking decides where the money goes. It doesn't show that what you funded turned into anything. The chain is longer than it looks: deliverables produce outcomes, outcomes produce benefits, and benefits produce value. Each link can hold while the next one fails, so reporting that jumps from the first link to the last is reporting an intention.
Checking that the chain holds takes ordinary discipline: trace requirements to what was actually built, review at the points where real decisions sit, and keep acceptance criteria separate from completion criteria. Work can be finished and still not be acceptable. A portfolio that tracks only completion won't see the difference.
An empty benefit record is a worse failure. A deliverable accepted exactly as specified can still produce no outcome, and only measuring the next link will show it. So a portfolio can hold a full list of accepted deliverables and no realised benefits, with nothing in its reporting to flag the contradiction.
Measured value also needs a way of measuring that everyone agreed on before the numbers arrive. Value reporting is always contested, because it decides where the next budget goes. If the basis is agreed in advance, the conversation can be about the AI instead of the arithmetic.
One method, three levels
The same discipline works at every level, so you don't end up with three separate reporting regimes. A use case is a project, measured against its own baseline. A benefit case that spans several use cases is a programme, measured on benefits realised against benefits planned. The whole is a portfolio, measured on the value it adds against the strategy it was funded to serve. Only the unit changes.
The standards also settle who answers for the value. The sponsors of the portfolio own its value. The part of the organization that receives a benefit owns keeping it alive. So the value number can't be outsourced, to us or to anyone else. What we can hand over is the discipline that produces it.
The value you found only holds if the system keeps working. Models drift as the world moves away from their training data. They hallucinate when a question is ambiguous. And they struggle under context load: as the input grows, accuracy falls quietly, without a failure anyone would notice.
None of this announces itself. A degraded model answers with the same confidence as a healthy one, so a customer, an auditor or a headline often finds the problem first. Evaluation, monitoring and traceability make a silent failure visible to your own team.
Here value and risk meet. A model that has drifted is a business problem and a compliance problem at the same time. The value case erodes, and the obligation to show control is breached. One root cause does both.
What clears the gate still has to be delivered
Permalink to “What clears the gate still has to be delivered”Between the gate and the indicators sits the delivery work, and it decides whether the result still matches the case that was approved. A gate that releases work into an unmanaged pipeline hasn't really decided anything. The measurement at the far end will simply record what happened.
Delivery is a managed activity or it is an assumption
The disciplines that turn an approved case into a delivered one aren't specific to AI. Work runs in phases, each ending in a decision to continue, change or stop, against criteria set in advance. Decisions that belong to the client are named up front and stay with the client. Progress is reported against the approved baseline.
None of this is new, and that's the point. The usual argument for skipping it is that the technology is exploratory. But exploratory work needs more decision discipline than routine work, not less. The phase-gate model is decades old, and the project-management standards have carried it ever since.
Three things AI changes, and only three
We adapt the delivery method for AI where it needs it, and keep the rest. Some assumptions break, and a usable method names them openly instead of burying them in contingency.
Data is a design input, and usually the dominant one.
Classic design control assumes the specification governs the output. In a system that learns, the training data shapes behaviour more than the specification does. So data is a design decision, and a plan that treats it as a supply problem misses what most decides the result.
The output is uncertain by construction.
A conventional deliverable meets its specification or it doesn't. A model performs across a distribution, so what it does can only be stated statistically. A plan should promise a measured level of performance on a defined test set. A capability promised outright is something the technology can't deliver.
Validation is iterative and does not end at release.
You get evidence that a system works by evaluating it, learning that the target was wrong or the data wasn't enough, and evaluating again. That's a planned loop with a defined exit. When it gets treated as slippage, AI projects get a reputation for never finishing.
What "delivered" is allowed to mean
Delivery ends with a judgement: this is done. Left informal, that call gets made by whoever is most tired of the question. Afterwards, the record shows that something exists. It doesn't show that anyone verified it.
A quality plan makes that judgement explicit. Acceptance criteria for each deliverable are written before work starts and aren't edited afterwards. If one turns out to be wrong, or a requirement legitimately changes, it's superseded with a dated note. Each criterion can be answered yes or no against evidence someone can find, and the verification method is named in advance. Results are recorded as passed, failed and corrected, or failed and open, with who checked and when. Anything failed and open isn't released unless someone with the authority to carry the consequence accepts it in writing, on record.
This is ordinary quality practice. Machine learning changes some of its assumptions.
Where classic quality assumptions break on AI
The adaptations above change how the work is planned. The harder half is how the result is judged. Quality management built its methods in manufacturing and engineering, and four of its core assumptions break on a system that learns. Once they're named, a quality plan can deal with them.
Conformity stops being per-unit.
Established practice assumes a nonconforming output can be spotted and set aside. Every output of a generative system is new, and a wrong one, such as a made-up citation, can look exactly like a correct one without expert review. Quality management has met this before. Where inspection can't fully verify the output, as in welding or sterilisation, you validate the process and the conditions it runs under, and qualify the people who run it. Inference is that kind of process. So conformity is defined statistically, with sampling, review gates and output monitoring designed in.
The specification does not determine the behaviour.
As noted above, in a system that learns, the training data shapes more of the result than the specification does. So data quality, provenance and preparation have to go through design control as design inputs. Otherwise the control covers the part of the system that matters least.
Validation does not stay valid.
Validation assumes the conditions of intended use can be described once. A model valid at deployment degrades as the world moves on, and nothing inside the system changes to trigger a review. Periodic revalidation and performance monitoring are already ordinary practice. What changes is the trigger: a monitored measurement, where it used to be an engineering change request.
There is no calibration chain for model evaluation.
Measurement discipline traces instruments back to reference standards. Benchmarks aren't reference standards, and results move with the test set. The instrument still has to be fit for what it measures, so its suitability has to be argued for each system.
Once named, all four have quality-engineering answers. What a quality system doesn't give you is an account of what the AI system does to the people it touches. That's one of the things an AI management system adds, and it's why the two run together.
The shortfall is visible long before the end
Delivery reporting usually tells you whether work is moving. The sharper question is whether the work the value case depends on is getting done at the rate the case assumed, and you can answer that while the work is still running.
It takes three numbers and no new instrumentation: what the plan said would be finished by now, what has actually been finished, and what it cost to get there. The first two show the gap against schedule, the last two the gap against cost. Projected forward, they give a sound estimate of where the work will land if nothing changes. "Are we on track?" becomes arithmetic, while there's still budget left to act.
This is earned value management, and it fits AI well. The unit doesn't have to be money: work can be counted in hours or story points, which matters when AI work is hard to price. And it's defined for iterative delivery as well as fixed-scope plans. Both were already in the method.
The arithmetic tracks output against plan, and output isn't value. A project can earn every unit of its baseline and still return nothing. Here the question is narrower: is what the case was betting on being built at the rate it assumed? That's worth knowing early. Something was planned, something happened, and you can report the gap long before the end.
Value indicators answer one question: is the AI delivering what it was funded to deliver? They start where the AI acts, at the process, and then come in four groups: what has actually landed, what's still owed, what it took to get there, and whether the way you choose between use cases is working. We read them alongside the exposure indicators on the risk page and the control indicators on the governance page. A strong value number can hide weakening controls underneath.
- —First-time-right rate, against the process's own baseline.
- —Cycle time, against the process's own baseline.
- —Exception and rework rate.
- —Escapes and customer complaints attributable to the process.
- —On-time delivery.
- —The variation of the outcome, not only its average.
These are the measures a QMS already reports. The groups below roll them up.
- —Realised value per use case, measured against the range the case was approved on.
- —Cumulative value realised across the portfolio, against the cumulative forecast it was funded on.
- —Share of use cases with a defined and measured financial outcome. An activity metric standing in for one doesn't count.
- —Value still being realised twelve months after go-live, compared with value at first measurement. AI benefit decays as adoption drifts and processes move around it.
- —Adoption and utilisation of deployed AI against provisioned capacity.
- —Value approved but not yet realised: the committed pipeline, and how long each part of it has been waiting.
- —Value by stage: proposed, through the gate, in delivery, live but not yet measured, and measured. Where value piles up, the portfolio is stuck.
- —Realisation rate: measured value as a share of approved value. A gap that persists means either delivery is lagging or the forecasts didn't hold up in production.
- —Value written off: approved cases that were abandoned or quietly stalled, and what they were worth on paper. If nobody records this, it becomes the portfolio’s largest blind spot.
- —Total cost of ownership per system. That includes the running cost of oversight, logging, evaluation and assessment, on top of licence and build.
- —Cost per decision or per inference, and its direction as volume grows. Unit economics that work in a pilot don't always hold in production.
- —Share of total AI spend attributable to systems with no measured outcome.
- —Share of proposals that don't clear the gate. A gate that rejects nothing isn't a gate.
- —Forecast accuracy across the portfolio: realised value against the approved range, and whether the misses run consistently in one direction.
- —Time from commitment to first measured value. Going live is an engineering milestone. Measured value is the business one.
- —Decommission rate: systems retired because they no longer deliver value. A healthy portfolio does this on purpose.
A forecast only becomes evidence once it's compared with the outcome. The in-flight numbers are the easiest of the four groups to produce. A business case records its promised value once, at approval, and nothing later makes anyone look at it again. So a portfolio can end up owing more than anyone has noticed.
ISO 31000 — risk management creates and protects value
Value and risk belong in one conversation because the standard puts them there. ISO 31000 states that risk management exists to create and protect value. It defines risk as the effect of uncertainty on objectives, and an effect is a deviation from the expected, positive or negative. So opportunity is part of risk, and finding where AI could deliver more is part of risk management itself. That's how we treat it.
Why does AI value so often go unproven?
Baselines are often missing and the business case is written after deployment, so activity metrics stand in for financial outcomes. Planning the measurement before the commitment is what makes the return provable later.
Is value more than the work AI removes?
Yes. It also includes decision quality, consistency, resilience, human capability, trust, and effects on the people affected. These are harder to count than hours saved, so they need their own indicators.
How do we keep AI value from eroding after go-live?
Reliability protects the value: evaluation and monitoring make drift, hallucination and context-load degradation visible. Without them, failures can erode value silently. Going live is an engineering milestone; measured value is the business one.
How do we measure the cost side?
Look at total cost of ownership, including oversight, logging and evaluation, and at cost per decision. Also watch spend on systems that have no measured outcome.
Make the return on your AI provable.
Tell us where AI is deployed and what it was meant to change. We'll come back with a read on what can be measured today, what it would take to make the case provable, and a sensible first step.