Something peculiar is happening in enterprise AI adoption across the UK. Organisations have completed proof-of-concept deployments for agentic AI systems, seen them perform well, and then watched the project stall at the budget approval stage. The technology works. The demos impressed. Yet the business case keeps coming back rejected or deferred. The culprit, more often than not, is not scepticism about AI — it is the measurement framework being applied to it.
Agentic AI is not traditional software. It does not simply execute a defined function faster or cheaper than a human. It reasons, decides, and acts across sequences of tasks that previously required sustained human attention. Evaluating it using cost-per-transaction metrics borrowed from ERP implementations or RPA rollouts fundamentally misrepresents what it delivers. If your ROI framework cannot capture what is actually changing, it will always produce a weak business case — and that is precisely where a large number of UK enterprises find themselves right now.
The Problem with Cost-Per-Transaction Thinking
Traditional software ROI is built around efficiency: a process that cost £X per transaction now costs £Y, and the delta multiplied across volume yields a payback period. It is a model that works well when the work is homogeneous, measurable, and already well-understood. Agentic AI, however, is typically deployed against work that is heterogeneous, judgement-intensive, and often partially invisible — tasks that are currently absorbed by skilled employees as part of a broader role rather than appearing on any process map.
When you apply cost-per-transaction logic to an agentic deployment, you run into an immediate structural problem: the transactions are hard to price. How much does it cost a senior analyst to triage and respond to a complex supplier query? How long does a compliance officer spend synthesising regulatory updates into action items each week? These activities are real and consequential, but they rarely appear as line items in any cost model. The result is that the ROI calculation undersells the value, the business case looks marginal, and the project is either shelved or condemned to a permanent pilot status that serves nobody.
Reframing Around Task Deflection Rate and Human-Hour Recapture
The measurement shift that unlocks genuine business cases is moving from cost-per-transaction to two related metrics: task deflection rate and human-hour recapture. Task deflection rate measures the proportion of incoming tasks — queries, decisions, drafting requests, data lookups, escalations — that the agentic system handles end-to-end without requiring human intervention. Human-hour recapture measures the aggregate skilled time that is returned to the organisation as a result. Together, these metrics describe not just efficiency but capacity: what can your organisation now do that it could not do before, with the same headcount?
Consider a practical example. A UK financial services firm deploys an agentic assistant across its client onboarding function. Previously, relationship managers spent roughly 40% of their week on information gathering, document chasing, and status updates — work that required their involvement but not their expertise. After deployment, the agent handles 70% of those tasks autonomously. The task deflection rate is 70%. The human-hour recapture, across a team of 20 relationship managers, is approximately 56 hours per week. That is the equivalent of 1.4 additional full-time senior staff — without hiring. That number lands in a board presentation very differently from a per-transaction cost comparison.
Building a Measurement Framework That Survives Scrutiny
For this approach to work in practice, the measurement framework needs to be established before the pilot ends, not retrospectively constructed to justify a decision that has already been made. That means three things. First, baseline the work: conduct a structured time-audit across the roles the agent will support, categorising tasks by type, frequency, and average handling time. This does not need to be exhaustive — a two-week sample across representative staff is usually sufficient to build a credible baseline. Second, instrument the agent: ensure your deployment captures not just successful completions but also handoffs, partial completions, and escalations. Task deflection rate is only meaningful if you are counting the full denominator, including cases the agent attempted but could not resolve unaided.
Third, and critically, define what recaptured hours are worth in your specific context. This is where many frameworks go wrong by defaulting to blended salary cost. In most agentic deployments, the hours being recaptured are from senior, specialist, or revenue-generating staff. A relationship manager's recaptured time is worth more than an average hourly rate suggests — because it flows directly into client-facing activity, pipeline development, or complex problem-solving that the business has been starved of capacity to pursue. Build that context into your model explicitly, and your business case will reflect reality rather than an accounting abstraction.
From Pilot Purgatory to Production Confidence
There is a specific dynamic that afflicts many agentic AI pilots in larger UK organisations: they succeed technically but fail commercially, not because they do not deliver value but because the value is diffuse, qualitative, and poorly matched to the approval criteria of finance and procurement committees. The task deflection and human-hour recapture framework addresses this directly by producing numbers that are concrete, attributable, and translatable into operational and financial terms that boards and budget holders already understand.
It is also a framework that compounds well. Once you have established baseline measurements and a functioning instrumentation layer, each subsequent agentic deployment — whether in a new department, a new process, or a new capability — can be modelled and justified more quickly. The first deployment builds the measurement infrastructure; subsequent deployments leverage it. That cumulative dynamic is part of what separates organisations that move from pilot to scaled production from those that remain in perpetual proof-of-concept cycles, re-litigating the same business case questions with each new initiative.
If your organisation has an agentic AI pilot that has stalled at the business case stage, the right response is not to build a more elaborate cost model using the same underlying logic. It is to step back and ask whether the measurement framework you are using was designed for the kind of value agentic systems actually create. In most cases, it was not — and that is a solvable problem.
The practical starting point is straightforward: run a structured time-audit in the function where your pilot is operating, establish your task deflection baseline, and restate the business case in terms of human-hour recapture and what that capacity is operationally worth. For most organisations that do this rigorously, the numbers change materially — and so does the conversation with the people who control the budget. The technology is ready. The measurement framework just needs to catch up.
What exactly counts as a 'task deflection' in an agentic AI context?
A task deflection occurs when an agentic system handles a request or work item end-to-end without requiring a human to intervene, review, or complete it. This includes things like answering a query, gathering information, drafting a communication, or triggering a downstream action. Partial completions that still require human sign-off should be counted separately — they are not deflections, even if the agent did the bulk of the work.
How do we conduct a baseline time-audit without disrupting normal operations?
A structured sample is usually sufficient — typically a two-week period across a representative subset of the staff whose roles will be affected by the agentic deployment. Participants log time spent on task categories using a simple taxonomy, ideally through an existing tool such as a calendar or time-tracking system rather than a separate survey. The goal is a credible estimate, not a perfect accounting; most organisations find that even a rough baseline dramatically improves the quality of the ROI conversation.
Is task deflection rate applicable to all types of agentic AI deployments?
It is most directly applicable to deployments where the agent is handling identifiable, discrete tasks that currently consume human time — such as information retrieval, triage, drafting, or coordination. It is less straightforward for deployments focused on augmentation, where the agent enhances the quality of human decisions rather than replacing the task entirely. In those cases, you may need to supplement deflection metrics with quality or accuracy-based measures.
How should recaptured hours be valued if the affected staff are on fixed salaries?
Salary cost is a floor, not a ceiling. The more meaningful question is what those hours will be redirected towards. If recaptured time flows into revenue-generating activity, capacity-constrained services, or work the organisation has been deferring due to resource limits, the value of those hours should reflect the opportunity they unlock — not merely the salary denominator. Document the intended reallocation of recaptured time explicitly in your business case.
What is a realistic task deflection rate to expect from an early agentic deployment?
This varies significantly by use case, but initial deployments in well-structured domains — such as document processing, customer query handling, or internal knowledge retrieval — commonly achieve deflection rates of 50–75% within the first few months. More complex or judgement-intensive workflows typically start lower, in the 30–50% range, and improve as the system is refined. Setting realistic expectations during scoping is important; overestimating deflection rates in the business case creates credibility problems later.
How do we handle the cases the agent attempts but cannot resolve — should these be tracked differently?
Yes, and this is an important nuance. Attempted-but-escalated cases reveal where the agent's capability boundaries lie and should be tracked separately from both successful deflections and tasks the agent never attempted. This data is valuable both for improving the system and for ensuring your deflection rate calculation is honest — counting escalations as successful deflections will produce inflated metrics that undermine trust in the business case.
Can this measurement approach be applied retrospectively to an existing pilot?
It can, though retrospective baselining is less precise. If you have not captured pre-deployment time-use data, you will need to reconstruct it through interviews and estimates, which introduces more uncertainty. That said, even an approximate retrospective baseline is usually sufficient to reframe the business case meaningfully. Going forward, ensure any new pilot scoping includes a baselining phase before deployment begins.
How does task deflection rate interact with headcount planning and workforce strategy?
Task deflection does not automatically translate into headcount reduction, nor should it be positioned that way in most organisations. The more productive framing for workforce planning is capacity creation — the same team can handle more volume, take on more complex work, or improve service quality without additional hiring. Positioning the metric this way also tends to reduce resistance from staff and line managers, which is important for adoption.
What governance or audit requirements should UK organisations be aware of when deploying agentic AI in regulated sectors?
UK regulated sectors — including financial services, healthcare, and legal — will need to ensure that autonomous agent actions are logged, auditable, and attributable. This means your instrumentation layer serves both a measurement function and a compliance function. The FCA, ICO, and sector-specific regulators are increasingly focused on AI decision trails, so any deployment should include a clear record of what the agent decided, what it escalated, and on what basis — independent of the ROI tracking you put in place.
How do we prevent the business case from being undermined if deflection rates dip after initial deployment?
Deflection rates often fluctuate in the first few months as edge cases emerge, user behaviour adapts, and the agent is refined. Build this into your reporting by tracking trends over time rather than presenting a single-point metric. A declining deflection rate that is being actively investigated and improved tells a very different story from one that is simply falling without explanation. Quarterly reporting with a commentary on trajectory is usually more credible than monthly snapshots in the early stages.
Get in touch today
Book a call at a time to suit you, or fill out our enquiry form or get in touch using the contact details below