AI ROI Leakage Registers Should Come Before More Licences

ROI & Cost Optimisation

4 September 2026 | By Ashley Marshall

Most AI budgets leak value after the invoice is paid. The fix is not another dashboard - it is a register that shows where promised benefits disappear.

Adoption is rising faster than operating discipline

UK leaders no longer need convincing that AI is moving into everyday work. The harder question is whether the benefits are being captured, measured and protected once the licences have been bought. The Office for National Statistics reported in July 2026 that AI use among UK businesses with 10 or more employees rose from around 12% in late 2023 to around 35% by June 2026. That is serious adoption momentum, but the same ONS analysis described usage as relatively shallow: the average adopting business moved only from around 1.4 to 1.6 AI technologies over the same period. In other words, many firms have entered the AI economy without yet building the management system that turns experimentation into repeatable value. The ONS data should make boards pause before approving another round of seats, agents or model credits. Adoption is not the same as value capture.

A practical ROI conversation starts with a simple admission: AI value leaks. It leaks when saved time is absorbed by more meetings, when teams keep manual checks because trust is low, when outputs need rework, when unused licences roll forward, and when pilots never become controlled workflows. A benefit leakage register is a lightweight operating document that names those losses before they become normal. It records the intended benefit, the place it can leak, the owner, the evidence source and the next corrective action. That may sound less exciting than a new agent platform, but it is exactly the kind of discipline that stops AI spend becoming another software line item with a vague productivity story attached.

The ROI problem is not spend, it is unclaimed value

The visible cost of AI is easy to spot. Finance can see licence fees, API spend, implementation services, cloud consumption and training budgets. The less visible cost is unclaimed value: minutes saved but not redeployed, response times improved but not linked to revenue, compliance effort reduced but not removed, and customer issues resolved faster but not reflected in service capacity. Lloyds Banking Group's 2026 Business Barometer findings show why this matters. Among UK businesses integrating AI into operations, 87% reported increased productivity and 48% reported higher profits over the previous 12 months. Of the firms reporting a profit boost, almost half recorded an uplift of 11% or more. That is encouraging evidence, but it also raises a sharper management question: if productivity is being created, who is banking it?

A leakage register forces the business to answer that question at workflow level. For a customer service assistant, the target benefit might be reduced average handling time, fewer escalations and better first contact resolution. Leakage could happen if agents paste AI outputs into legacy systems manually, if supervisors recheck every answer, or if the queue planning model does not change. For a finance automation, leakage might show up when month-end close is technically faster but review meetings remain unchanged. What this means in practice is simple: every AI investment needs a benefit owner who can change the process, not just a tool owner who can configure the software.

Measure the run, the rework and the handoff

Most AI dashboards still over-index on what vendors can measure easily: prompts, tokens, users, sessions, latency and cost per model call. Those are useful engineering signals, but they are not enough for business ROI. A board does not need to know that a workflow used 140,000 tokens last week unless that number is connected to cycle time, error rate, avoided labour, customer conversion or risk reduction. The better unit is the business run: one invoice processed, one support case resolved, one tender drafted, one compliance review completed, one sales follow-up sent. Cost per run, rework per run and handoff delay per run are the numbers that expose whether the AI system is improving work or simply adding a shiny step into the middle.

The ONS finding that improving business operations is the most common AI use, reported by close to three-fifths of businesses, supports this workflow-level view. It also found that around 63% of businesses using AI to improve operations reported no change in worker headcount, while 6% reported a decrease. That should cool down simplistic narratives about instant labour replacement. The more useful interpretation is that many firms are still in the process redesign phase. A benefit leakage register should therefore include three evidence fields for every significant workflow: baseline run time before AI, current run time after AI, and rework rate after human review. If the first two improve but rework rises, the benefit has leaked into quality control. If run time improves but handoff delay stays high, the bottleneck was never the AI step.

Governance costs belong inside the ROI case

A common misconception is that governance slows ROI. In reality, weak governance often hides the costs that later destroy it. Gartner warned that at least 30% of generative AI projects would be abandoned after proof of concept by the end of 2025 because of poor data quality, inadequate risk controls, escalating costs or unclear business value. Even where access to the full report is restricted, the pattern is familiar to anyone who has reviewed AI pilots: the demo works, the live workflow does not have enough evidence, ownership or control to survive production. The cost of governance is not a tax on value. It is the cost of making the value durable enough to count.

For UK businesses, that means ROI models should include assurance effort, data clean-up, permission reviews, security monitoring, evaluation packs, incident response planning and supplier change control. The National Cyber Security Centre's secure AI system development guidance tells providers and decision-makers to consider secure design, secure development, secure deployment and secure operation throughout the life cycle. Its 2026 agentic AI blog also stresses safeguards, sandboxing, oversight, observability, attribution and emergency shutdown for autonomous systems. That guidance is directly relevant to ROI because every uncontrolled exception creates hidden cost. What this means in practice: do not approve an AI agent business case that shows licence savings but excludes monitoring, review time, failure handling and rollback.

The register should be owned by operations, not innovation

Innovation teams are useful for discovery, but they are rarely the right long-term owner of AI benefits. Leakage happens in the daily operating rhythm: who changes the rota, who removes the redundant checklist, who updates the CRM field, who tells finance that a process now needs fewer external hours, who decides that an exception still needs human approval. Those decisions sit with operations, finance, risk, HR and departmental leaders. If the benefit register lives only in a transformation deck, it will become a reporting artefact. If it lives in the management cadence of the teams doing the work, it becomes a control surface for value.

A good register is deliberately plain. Each row should include the workflow, the AI capability, the expected benefit, the leakage risk, the metric, the evidence source, the owner, the review date and the corrective action. A support team might track minutes saved per case and whether staffing plans changed. A marketing team might track content cycle time and downstream conversion, not just output volume. A legal team might track first-pass review time, escalation rates and contract risk findings. The counterargument is that this creates admin overhead. It can, if the register tries to document everything. The better rule is to track only material workflows where the business has made a claim about value. If the benefit is important enough to justify spend, it is important enough to evidence.

A 30-day leakage review is enough to change decisions

This does not need to become a six-month consulting programme. A useful first review can be done in 30 days. Start with the ten most expensive AI licences, the five highest-volume AI-assisted workflows, and any agentic workflow with access to customer data, finance systems or production tools. For each one, write the original value claim in one sentence. Then ask four questions: what metric proves it, where could the value leak, who can change the process, and what decision will we make if the evidence is weak? That final question matters because measurement without decision rights becomes theatre. The register should drive actions such as removing unused seats, redesigning handoffs, adding evaluation tests, narrowing tool permissions, renegotiating vendor terms or moving a pilot into proper production ownership.

The result is a more honest AI budget conversation. Instead of asking whether AI is worth it in the abstract, leaders can compare workflows. Some will deserve more investment because the evidence shows real captured value. Some will need redesign because the model is good but the operating process is wasteful. Some should be stopped because the business case depends on benefits nobody can claim. That is not pessimism. It is how mature firms turn AI enthusiasm into compound operating advantage. The next phase of AI ROI will not be won by the companies with the most pilots. It will be won by the companies that can show, line by line, where the value went.

Frequently Asked Questions

What is an AI ROI leakage register?

It is a practical record of where expected AI benefits can disappear. It usually tracks the workflow, intended value, leakage risk, metric, evidence source, owner, review date and corrective action.

Why is this better than a normal AI dashboard?

A normal dashboard often shows usage, tokens and cost. A leakage register connects AI activity to business outcomes such as cycle time, rework, service capacity, quality, margin or risk reduction.

Who should own the register?

Operations or the accountable business function should own it, with finance and risk involved. Innovation or IT can support the tooling, but the owner must be able to change the workflow.

What should be reviewed first?

Start with the most expensive AI licences, the highest-volume AI-assisted workflows, and any autonomous workflow that touches customer data, finance systems or production tools.

Does this slow AI adoption down?

It may slow weak projects, but it speeds up serious adoption because leaders can see which workflows deserve more investment and which need redesign before scaling.

What metrics should UK businesses use?

Useful metrics include cost per run, run time, rework rate, handoff delay, exception rate, human review time, avoided external spend, customer response time and evidence of risk reduction.

How often should the register be updated?

For active pilots, review it weekly or fortnightly. For live AI workflows, include it in the normal monthly performance rhythm alongside financial and operational reporting.

What is the first practical step?

Pick five live AI workflows and write the original value claim for each in one sentence. Then identify the metric, owner and most likely place where the benefit is leaking.