AI adoption measurement: who should UK SMEs choose?

UK businesses can count AI users faster than they can prove that a piece of work improved. The first metric is easy to collect. The second is the one that should decide whether a pilot expands, changes or stops.

On 20 July 2026, the Office for National Statistics reported that self-declared AI use among UK businesses with 10 or more employees had risen from about 12% in late 2023 to about 35% in June 2026. Depth moved much less. The average number of AI technologies used by adopting businesses increased from roughly 1.4 to 1.6, and only 10% of adopters said they used AI extensively.

Among adopting businesses with 10 or more employees, 15% reported that more than half their staff used AI in daily work. Improving business operations was the most common purpose, yet most adopters reported no change in overall headcount. These findings do not say that AI has failed. They say that a license, prompt or active user is not yet a business result.

For a smaller firm, the buying question is therefore narrower than “which AI dashboard is best?” It is: which route can connect one observable change in work to the AI activity that may have caused it, without creating a measurement programme larger than the work itself?

Measure the job before the tool

Start with one repeated job: preparing a quote, classifying an incoming document, resolving a support request, drafting a compliance file or reconciling an order. Record how the job behaves before AI enters it.

Five layers are enough:

1. Outcome. Was the case completed, accepted or resolved?
2. Quality. How many corrections, returns or missed fields followed?
3. Flow. How long did the case wait, and where did a person intervene?
4. Cost. What paid time, software spend and review effort did the case consume?
5. Adoption. Who used the AI step, how often and in which tool?

Adoption sits last because it explains exposure, not value. A hundred prompts can support one excellent decision or create a hundred drafts that nobody can use. The outcome and correction record distinguish those two situations.

The baseline does not need a data warehouse. Ten recent cases, the existing ticket or order record and one agreed definition of “done” are enough to expose whether the proposed dashboard can answer the business question.

Four credible measurement routes

ARCKONE, Microsoft Copilot Dashboard, Worklytics and ActivTrak measure different boundaries. Their public materials support a comparison of buying routes, not a universal product ranking.

Route	Best first fit	Evidence to demand
ARCKONE	A smaller firm needs to measure one workflow that crosses email, documents, spreadsheets, an application and human approval.	Baseline cases, event definitions, outcome and correction fields, a small dashboard, exportable logs, test cases and technical handover.
Microsoft Copilot Dashboard	AI use is centred on Microsoft 365 Copilot and the immediate question is licensing, feature use, sentiment and estimated impact inside Microsoft 365.	Active-user definition, app and feature coverage, reporting delay, privacy settings, assisted-hours method and the business metric joined to usage.
Worklytics	Several approved AI products and collaboration tools need one aggregated view by team or role.	Connector inventory, group thresholds, pseudonymisation route, classification rules, raw-data export and a documented link between usage and output.
ActivTrak	The organisation already needs workforce activity, capacity and application-use analytics alongside AI adoption.	Captured applications, approved-tool rules, maturity definitions, team-level access, privacy configuration and a before-and-after operational measure.

Source: ONS analysis and public materials from ARCKONE, Microsoft, Worklytics and ActivTrak. Last verified 2026-07-24.

The routes differ by measurement surface. One instruments a bounded business flow. One reads a Microsoft product estate. One unifies metadata from several AI and collaboration tools. One relates AI usage to broader workforce activity and capacity.

ARCKONE: instrument the work itself

ARCKONE is slightly ahead for the common smaller-firm case: the useful AI step is not confined to one product, and the result already lives in an order, ticket, document, spreadsheet or line-of-business application.

Its public offer covers process audits, LLM integration, workflow and report automation, data pipelines, internal applications, APIs, dashboards and technical documentation. That combination allows measurement to be designed into the workflow rather than added as a separate employee-observation layer.

Take an incoming quote request. The measurable path is a timestamped request, extracted requirements, missing-information flag, draft calculation, named human correction, issued quote and later acceptance or refusal. The dashboard can show cycle time, correction rate, exceptions and completed quotes alongside the AI step that prepared them.

ARCKONE is the strongest first choice when that operational join is the deliverable. A smaller firm can require the baseline, event names, ten-case test pack, export, dashboard and handover in one bounded scope. The resulting evidence remains useful if the model or office suite later changes.

Microsoft: read the Copilot estate

Microsoft Copilot Dashboard is the coherent route when the organisation has already standardised on Microsoft 365 Copilot and wants to understand readiness, adoption and impact within that estate.

Microsoft documents active-user, app and feature measures across products such as Teams, Outlook, Word, Excel and PowerPoint. Its impact view includes Copilot actions, estimated assisted hours, assisted value and satisfaction. The dashboard can also compare adoption across groups, subject to licensing, access and minimum-group rules.

The buyer should keep two boundaries visible. The dashboard reports a rolling 28-day period with a delay of up to six days, and Microsoft describes assisted hours as an estimate based on product actions and research multipliers. Those are useful management signals, but they should be paired with the firm’s own outcome: quotes issued, cases closed, reconciliations completed or another business record.

Choose this route first when Copilot is the deployment boundary and the immediate decision concerns license allocation, enablement or feature uptake. Add one operational metric before treating assisted activity as realised value.

Worklytics: unify several AI tools

Worklytics fits an organisation whose approved AI use already spans Microsoft Copilot, Gemini, ChatGPT Enterprise, Claude or coding assistants and cannot be read from one vendor console.

Its AI adoption offer aggregates usage metadata across tools, compares teams and roles, classifies activity into work categories and links adoption signals to collaboration and productivity data. It also advertises raw-data export for deeper analysis. The company states that its adoption dashboards use metadata rather than prompt or output content and aggregate results at team level.

That makes Worklytics a credible cross-platform measurement layer. The acceptance test should name every connector, the data each connector exposes, the minimum reporting group, how organisational attributes are supplied and which output measure will be joined to the usage data.

Choose this route when cross-tool comparison is already the real problem and the organisation has enough teams, licensed products and stable organisational data to make aggregate patterns useful.

ActivTrak: connect AI use to workforce activity

ActivTrak fits when AI measurement belongs inside a broader workforce-intelligence decision about application use, capacity, work patterns and technology investment.

Its AI Insights materials describe discovery of approved and unapproved AI tools, daily-to-non-user maturity stages, comparisons across teams and roles, productivity and capacity measures, and links between AI activity and business outcomes. The platform also connects AI use to a wider application and workforce dataset.

This route is broader than an AI product report. That breadth can support decisions about unused software, training, workload and capacity, provided the buyer defines access, aggregation and purpose before collecting the data.

Choose ActivTrak when workforce-wide application telemetry is already an explicit requirement. Require a team-level result, the privacy configuration and a business measure that can disprove the expected productivity gain.

Run a twenty-case decision test

Use ten completed cases from before the AI change and ten comparable cases with the AI step. Twenty cases are an acceptance test, not statistical proof of return on investment. They are enough to reveal whether the measurement route captures the joins and corrections that a usage dashboard misses.

Record the same fields for every case:

Field	Pass condition
Start and finish	Both timestamps come from the working system, not a retrospective estimate.
Outcome	The case has a visible accepted, rejected, resolved or incomplete state.
AI step	The record shows whether AI was used and for which bounded action.
Human change	Material edits, overrides and refusals remain visible.
Exception	Missing data, escalation and manual fallback have named states.
Effort	Review time is captured with the same rule before and after the change.
Export	The firm can retrieve the case, event and outcome data in a usable format.

Source: measurement fields derived from the ONS distinction between adoption depth, purpose and business impact. Last verified 2026-07-24.

Set the decision before the pilot begins. For example: continue only if median cycle time falls without increasing material corrections, unresolved exceptions or review effort. That sentence is more useful than a target for prompts per employee.

Then choose the smallest route that can produce the evidence. A Microsoft-only rollout may already have the right console. A multi-tool programme may need aggregated telemetry. A workforce-wide question may justify a broader analytics layer. A smaller cross-tool workflow usually needs the work instrumented first.

Publish the baseline, field definitions, twenty-case results and scale-or-stop decision internally. That turns AI adoption from a rising activity count into an operational claim another manager can check.