AI productivity starts with the completed workflow

Adoption is not the same as bankable productivity

A business can use AI every day without knowing whether it has improved the economics of a single completed workflow. The distinction matters as adoption becomes ordinary: a faster draft, a shorter search or a more polished presentation is useful, but none alone establishes that the organisation can deliver more accepted work with the same resources. Oakhampton's assessment is that the next management task is to measure that conversion explicitly.

The UK's AI adoption survey, published on 29 October 2026, reports that 32% of businesses used at least one AI technology, against 16% in 2025. Half of adopters reported an overall productivity increase, generally described as modest; 64% reported no immediate operating-cost effect. These are UK survey findings, not an Australian adoption estimate or a causal test of a particular product. [1]

The commercial question therefore moves beyond whether staff have access to a tool. It becomes whether a defined process produces an acceptable result sooner, with less total effort and no unacceptable deterioration in quality. That question can be answered in a small operational trial, without turning a software purchase into a company-wide transformation programme.

Measure the completed job, including review

Oakhampton recommends choosing a workflow with a clear beginning and a clear acceptance point. Preparing a supplier comparison, extracting shipment details or assembling a management briefing can each qualify. Define the input, the required output and the person who decides whether the result is usable. Avoid measuring an isolated generation step when the business actually depends on several subsequent checks.

Record the existing process before changing it. Include preparation time, document retrieval, drafting, review, correction and final handover. Separate active staff effort from elapsed waiting time: they answer different questions. A process can reduce hands-on work while leaving customer turnaround unchanged because approval remains the bottleneck. Conversely, earlier completion can be valuable even when staff effort changes little.

Run the assisted version against comparable work, retaining ordinary cases and awkward exceptions. Apply the same acceptance criteria to both versions. Count rejected outputs and corrections rather than excluding them from the result. An attractive average that conceals repeated failures on the most commercially important cases is not a sound basis for expansion.

Make the evidence specific to the operating context

The OECD's April 2026 D4SME paper draws on a non-representative sample of more than 2,000 SMEs across 12 countries. It describes uneven strategic, targeted and secure integration, with time constraints, maintenance costs and skills gaps hindering implementation. Its findings identify practical obstacles; the sample should not be treated as a national adoption benchmark. [2]

For a management trial, those obstacles belong in the measurement boundary. Record who maintains templates, resolves access problems and updates source material. If the workflow requires a specialist to repair it each week, that effort is part of the operating model. Treat employee training and initial setup separately from recurring support so the decision does not confuse a launch expense with a permanent requirement.

Keep a simple record of the configuration and source set used for each test. When either changes, preserve enough information to explain a performance difference. The purpose is not an elaborate experiment; it is to prevent a favourable result from becoming detached from the conditions that produced it.

Quality is part of productivity, not a separate appendix

NIST's AI Risk Management Framework 1.0 provides a complementary discipline. Its Measure function calls for evaluation in context, testing before deployment and regular testing in operation, with documented results and performance benchmarks. This is established voluntary guidance, not a new statutory requirement or a productivity certification. [3]

Oakhampton recommends a short quality rubric appropriate to the task: factual accuracy, completeness, source traceability and suitability for the intended recipient. Give serious errors their own category rather than averaging them into a general satisfaction score. A supplier comparison that omits a decisive restriction is not rescued by fluent writing or quick delivery.

Record reviewer effort alongside output quality. If the assisted process merely shifts work from the preparer to a more expensive reviewer, the apparent saving may disappear. Equally, if checking becomes easier because sources and exceptions are presented clearly, that benefit deserves to be measured rather than assumed.

Scale what the evidence actually supports

The final decision should state what changed: accepted work per staff hour, elapsed turnaround, correction burden and recurring support. Do not translate released minutes automatically into cash savings. Capacity creates commercial value only when it supports additional useful work, avoids a real constraint or permits a genuine change in expenditure.

Set a review point and an owner before expanding. Retain a route back to the established process if quality deteriorates or the workload changes. The strongest AI business case is not the largest estimated saving. It is a repeatable improvement in a named workflow, supported by evidence that includes the people who prepare, check and use the result.