Here is the uncomfortable number from KPMG's Global AI Pulse for Q3 2026: 89% of organisations say AI is delivering measurable value, yet only 12% consistently check that value against what AI costs them. Most enterprises, in other words, can describe AI benefits but cannot say whether those benefits are bigger than the bill.
That gap is now closing from the vendor side. On 28 October 2026, Anthropic made smart reports, in-app surveys and pooled team budgets generally available to Claude Enterprise customers, explicitly to show what AI spending buys. This guide explains what AI value measurement is, why usage dashboards alone will not satisfy your CFO, and a four-layer framework you can take into your next budget review.
What is AI value measurement?
AI value measurement is the discipline of linking what an organisation spends on AI to the work it produces and the business results that follow. It goes beyond counting logins or tokens. It compares cost with output, workflow change and outcomes, so leaders can decide where to expand AI, where to fix it and where to stop.
Think of it as the management-accounting layer for AI. Finance already knows how much the licences and tokens cost. Operations can see who uses which tool. What usually sits in nobody's job description is the connection between the two: which spend produced which result, and whether that result would have happened anyway.
--- Cost: licences, usage credits, integration and support.
--- Activity: who uses AI, how often and for which tasks.
--- Workflow change: which recurring processes now run differently.
--- Outcome: time, cost, quality, revenue or risk effects you can defend in a board paper.
Why has AI value measurement become urgent in late 2026?
AI value measurement has become urgent because spending is rising faster than evidence. KPMG reports average planned AI investment of US$210 million over the next 12 months, up from US$186 million in Q1 2026, while only one in eight organisations consistently compares value with cost. Vendors are now shipping measurement tools because buyers demand them.
According to Campus Technology’s summary of the KPMG survey, the findings come from 2,131 senior leaders across 20 countries, surveyed between July and August 2026. Another 26% assess value against cost for most initiatives, but not across the whole organisation.
The bill is also less predictable than it used to be. TechCrunch reported Jellyfish data showing that spending on Anthropic-powered coding tools grew roughly tenfold during 2026, to about 2% of engineering costs. As Jellyfish’s chief scientist put it, that is growth most CFOs did not plan for.
Usage-based pricing is part of the reason. When AI is billed by credits rather than seats, cost scales with behaviour, as UD explained in its guide to usage-based AI billing and Copilot Credits.
Why do usage dashboards fail to prove AI value?
Usage dashboards fail to prove AI value because they measure consumption, not consequence. High activity can mean a team found a great workflow, or that people are retrying poor answers. A dashboard tells you AI is being used. It cannot tell you whether the work got faster, cheaper or better, or whether anyone would notice if it stopped.
KPMG’s head of its US AI Center of Excellence summarised the problem bluntly: many organisations have made AI widely available and track its use, but consumption alone is not a strategy.
Vendor reports help, but they have a natural boundary. Anthropic’s smart reports group a team’s Claude activity into workstreams and show associated spend, deliverables and recurring failures such as missing integrations. That is valuable evidence. It still only covers one vendor’s tools, and it cannot see the downstream result in your ERP, CRM or finance system.
--- Activity can rise while value falls, for example when staff regenerate weak outputs.
--- Time saved is self-reported unless you compare against a baseline.
--- Saved hours are not saved money until they are redeployed or avoided as cost.
What should enterprises measure? A four-layer value framework
Enterprises should measure four layers in order: cost, activity, workflow change and business outcome. Each layer answers a different board question. Cost asks what you spend, activity asks whether people use it, workflow change asks what work now runs differently, and outcome asks what the organisation gained that it can defend financially.
Layer 1: Cost (what are we spending?)
Consolidate licences, usage credits, API tokens, integration work and internal support time into one view per business unit. Pooled team budgets, now offered by several vendors, make this easier to enforce.
Layer 2: Activity (are people actually using it?)
Track weekly active users by department and the share of seats unused for 60 days. Activity is a leading indicator, not proof of value.
Layer 3: Workflow change (what now runs differently?)
Count recurring workflows that have been redesigned around AI and are used by more than one person. This is where most value is created and where most programmes stop measuring.
Layer 4: Outcome (what did the business gain?)
Pick three to five workflows and measure cycle time, error rate, unit cost or revenue impact against a pre-AI baseline. These are the numbers that belong in a board paper.
How do you put a number on AI value?
To put a number on AI value, establish a baseline before rollout, measure the same workflow after, and apply a realisation rate that reflects how much freed time actually turns into saved cost or new output. Without a baseline and a realisation rate, most AI value claims overstate the return by a wide margin.
A worked example for a 300-person Hong Kong firm
The figures below are illustrative assumptions, not benchmarks. They show the arithmetic a CFO will expect.
--- Cost: 300 seats at US$30 a month is US$108,000 a year, about HK$842,000.
--- Activity: 70% of staff, or 210 people, use AI weekly.
--- Self-reported saving: two hours a week each across 46 working weeks, about 19,300 hours.
--- Gross value: at a loaded cost of HK$350 an hour, about HK$6.8 million.
--- Realisation rate: if only 25% of that time becomes avoided overtime, slower hiring or extra billable work, realised value is about HK$1.7 million.
--- Net: roughly HK$850,000 a year above licence cost, before integration and training.
The realisation rate is the number boards challenge most, so state it openly. KPMG found that organisations reporting established AI returns are four times as likely to measure value against cost consistently, which suggests the discipline itself is part of what produces the return.
How should Hong Kong enterprises handle employee data when measuring AI?
Hong Kong enterprises should treat AI measurement data as personal data whenever it can be linked to an employee. Collect only what the measurement needs, tell staff what is collected and why, aggregate at team level, and keep AI usage analytics out of individual performance reviews unless that purpose has been clearly notified.
The Privacy Commissioner for Personal Data published its checklist on employee use of generative AI in March 2025, covering permitted use, data protection and incident reporting. Usage analytics built on conversation transcripts raise the same questions under the Personal Data (Privacy) Ordinance: purpose, minimisation and transparency.
Vendors are drawing the same line. Anthropic states that its smart reports are meant to help organisations understand adoption and plan investment, not to evaluate individual employees, and that customer data from these features is not used for model training by default.
--- Aggregate: report by team or workflow, not by person.
--- Notify: update your employee privacy notice before switching on transcript-based analytics.
--- Separate: keep value measurement and performance management in different processes.
What does good AI value measurement look like in practice?
Good AI value measurement looks like a small number of named workflows, each with an owner, a baseline, a monthly cost line and one outcome metric, reviewed quarterly. It is deliberately narrow. Measuring five workflows well is more persuasive to a board than measuring adoption across fifty tools loosely.
Scenario: a regional logistics firm
The operations team uses AI to draft exception notices for delayed shipments. Baseline: 14 minutes per notice. After three months: 5 minutes, with the same complaint rate. Monthly AI cost for the team: HK$9,000. The COO can now show a cost per notice before and after, which is a far stronger claim than “usage is up 40%”.
Scenario: a mid-sized financial services group
Compliance uses AI to pre-screen marketing materials. The outcome metric is not hours saved but turnaround time and the number of issues caught before legal review. Because errors here carry regulatory risk, quality is measured alongside speed, and a human sign-off rate is tracked every month.
What are the most common AI value measurement mistakes?
The most common mistakes are measuring only activity, skipping the baseline, counting freed hours as cash, relying on one vendor’s report as the whole picture, and measuring everything instead of a few workflows. Each produces numbers that look impressive internally and collapse the first time a CFO or auditor asks how they were calculated.
No baseline. If you did not time the workflow before AI, every later saving is an estimate. Capture two to four weeks of baseline data before any major rollout.
Hours treated as money. Time saved only becomes financial value when it avoids a cost or creates billable output. Say which.
Vendor scope mistaken for enterprise scope. A single vendor’s console sees its own tools. Your measurement model should sit above all of them.
Measurement without cost control. Value tracking and spend control belong together, which is why it pays to pair this framework with the practices in UD’s guide to AI FinOps.
How do you report AI value to the board?
Report AI value to the board on one page each quarter: total AI cost by business unit, active-user share, the number of redesigned workflows, and before-and-after results for three to five priority workflows, with the realisation rate stated. Add a short list of what you stopped funding. Boards trust programmes that kill weak initiatives.
If your organisation is weighing a wider rollout, use the same page to show which teams have earned more licences. UD’s framework for company-wide AI licences explains how measurement should gate expansion rather than follow it.
--- Section 1: spend, by unit and by vendor.
--- Section 2: adoption, as a leading indicator.
--- Section 3: five workflows, baseline versus now.
--- Section 4: decisions: scale, fix or stop.
AI value measurement: key facts
--- The gap: 89% of organisations report measurable AI value; 12% consistently assess it against cost (KPMG Global AI Pulse, Q3 2026).
--- Spending: average planned AI investment of US$210 million over 12 months, up from US$186 million in Q1 2026.
--- Vendor response: Anthropic made smart reports, in-app surveys and pooled group budgets generally available to Claude Enterprise on 28 October 2026.
--- Framework: measure cost, activity, workflow change and outcome, in that order.
--- Hong Kong: treat employee-linked usage analytics as personal data under the PDPO.
What is the strategic takeaway on AI value measurement?
AI value measurement is quickly becoming the price of continued AI budget. The organisations that can show cost, workflow change and outcome on one page will keep funding their best initiatives. Those that can only show usage will find the CFO asking harder questions each quarter.
We understand the cold edges of AI and the hard parts of your work, and UD has walked with Hong Kong enterprises for twenty-eight years, making technology a partnership with warmth.
Now that you have the framework, the next step is finding out where your organisation can prove value first. We'll walk you through every step, from AI readiness assessment and baseline design to deployment, cost control and board-ready reporting.
Reviewed by the UD enterprise AI team.