The foundation: engage.re
AI and Data 30 July 2026 9 min read

AI in a UK Organisation: What Works and What Does Not

UK SME adoption of AI reached 54% in 2026, and only 12% of those users report higher revenue. That gap is not a mystery. Some tasks work reliably today and others fail in a predictable way, and one property decides which list a task falls into. This article gives both lists, names the property, and sets out the order to do things in.

Connect with us about where to start · · Replies within 24 hours

54%
of UK SMEs used AI in 2026, up from 35% in 2025
12%
of those users report higher revenue so far
46%
of small firms say they lack the knowledge to use AI

What does AI do well in a UK organisation today?

Tasks that work on text you supply, and where a person checks the output before it matters.

  • Drafting. A first version of a letter, a report section or a job advert.
  • Summarising. A long document reduced to its points, with the source in front of you.
  • Classifying. Sorting incoming messages into categories you defined.
  • Extracting. Pulling named fields out of an invoice or a form.
  • Translating. Moving text between languages for a first pass.
  • Explaining. Turning a technical passage into plain terms.

Each task shares three properties. The input arrives with the request, the output is text, and a person reviews it.

What does AI do badly?

Tasks that need a fact about your own organisation.

  • Answering questions about your records. How many active clients do we hold, and what is our current occupancy?
  • Producing figures for a report. Any number that a person will act on without checking it.
  • Applying your business rules. Rules nobody wrote down cannot be applied.
  • Deciding on a case. A decision with a legal or significant effect carries duties you must evidence.
  • Acting across systems. Each system holds its own meaning, and the agent must reconcile them.
One property separates the two lists. Where the task needs a fact about your organisation, the model must supply that fact from somewhere. Your database holds the structure of your data and not the meaning, so the model invents the business rules nobody recorded.

How large is the effect?

A 2026 benchmark measured it directly. Three frontier models answered 99 business questions on a public dataset, twice each. The second run added a 4KB document describing the measures and the conventions.

Condition Accuracy range
Schema only45.5% to 50.5%
Schema and a meaning document67.7% to 68.7%

Every model gained 17 to 23 percentage points. With the document present, the three models became statistically indistinguishable from each other.

The conclusion in the researchers' words. "the most consequential architectural decision is not which frontier model to use but whether the system is grounded in authoritative business semantics at all." From Semantic Layers for Reliable LLM-Powered Data Analytics.

Two conclusions follow for a buyer. Recording your meaning matters more than choosing a model, and 68.7% is not yet a figure to act on unchecked. We give the whole evidence base in why AI gets your own business data wrong.

Why does the adoption gap exist?

The Office for National Statistics reports the numbers in its 2023 to 2026 series. UK SME adoption stands at 54%, and 12% of those users report higher revenue.

The same series names the obstacles that firms report. 46% say they lack the knowledge, 49% raise privacy concerns, and 47% cite data quality or integration.

RAND examined failure causes in The Root Causes of Failure for Artificial Intelligence Projects. Within its sample, 84% of industry interviewees named leadership-driven issues as the primary cause. We cover that ground in why do AI projects fail.

Adoption is therefore easy and value is not. A tool that drafts a letter needs no readiness, and a tool that answers a question about your records needs all of it.

What order should you do things in?

  1. Start with the working list. Drafting and summarising deliver time savings this month, with a person in the loop.
  2. Write down your five most-used measures. One page. This is the cheapest high-value step available.
  3. Test one question about your own data. Ask a tool for a figure you already know, and compare.
  4. Name an owner. RAND's leadership finding applies whatever your size.
  5. Write a use policy. Name the approved tools and the permitted data categories.
  6. Then fix the foundation. Only after the first five steps have shown you which questions matter.

Step two carries the most value per hour. A 4KB document produced most of the gain in the benchmark, and a page describing your measures is that document.

What ESRE Media offers

We build your system on engage.re, and the platform moves tasks from the second list to the first.

  • Your records carry their own meaning. A shared dictionary holds the definition of every record type, and the server enforces it as a condition of storing data.
  • Your figures are current. A change reaches your reported numbers within 60 seconds, so an answer describes today.
  • Every answer is traceable. Each access is an event in a signed chain, so you can show which records produced a number.
  • Agents are accountable. One identity model covers your people, your applications and your agents, and every agent holds narrow permissions you can cancel on the next call.

Sense Future built engage.re, and it has run in production since December 2025.

What the law expects

The ICO's guidance on AI and data protection applies from your first use of a tool on personal data. Three duties matter most at the start.

  • A lawful basis. You need one for each purpose, and convenience is not one.
  • A record of the decision. Where a decision affects a person, you must be able to explain it.
  • Data minimisation. A tool should receive what a purpose needs, and no more.

HM Government sets out the wider national position in its AI Opportunities Action Plan.

The same order at two sizes

A council starts with drafting for its communications team, and it writes one page defining a household, a case and an active claim. Three services then produce the same figures, and the council has done more with one page than with a twelve-month pilot.

A care home group starts with summarising for its registered manager, and it writes one page defining an occupied bed, an incident and a resident. Two systems then agree on occupancy for the first time.

The care home software and charity software pages set out the sector detail. Our guides to charity data management and care home software costs cover the practical starting points.

What we do not claim

  • The two lists move over time. A task that fails today may work in two years, and the property that separates them will not change.
  • Recorded meaning does not make AI correct. The benchmark reached 68.7% with the document present.
  • Adoption figures measure use rather than value. The 54% and the 12% describe two different things.

What to do next

  1. Pick one task from the working list and use it this week.
  2. Write one page defining your five most-used measures.
  3. Ask a tool for a figure you already know, and compare the answers.
  4. Name the person who owns AI use in your organisation.
  5. Then decide whether that page stays a document, or becomes a declaration your system enforces.

Common questions

What does AI do well for a UK organisation?

Drafting, summarising, classifying, extracting fields, translating and explaining. Each task shares three properties: the input arrives with the request, the output is text, and a person reviews it before it matters. These deliver time savings without any data readiness work.

What does AI do badly?

Anything that needs a fact about your own organisation. Questions about your records, figures for a report, applying business rules nobody wrote down, deciding a case, and acting across systems that each hold their own meaning. The model must supply the missing fact, so it invents it.

Why do only 12% of UK AI users report higher revenue?

Because adoption is easy and value is not. A tool that drafts a letter needs no readiness, and a tool that answers a question about your records needs all of it. Firms report the obstacles: 46% lack the knowledge, 49% raise privacy concerns, and 47% cite data quality or integration.

What is the cheapest step with the highest value?

Writing one page that defines your five most-used measures. In a 2026 benchmark, a 4KB document describing what the data means raised model accuracy from 45.5% to 68.7%, a gain of 17 to 23 points for every model tested. That document is a page of definitions.

Should we choose a model or fix our data first?

Fix the data. With a meaning document present, three frontier models became statistically indistinguishable from each other. Without it, they also did. The document accounted for nearly all the variance, so the model choice barely mattered.

What does ESRE Media build?

We build your system on engage.re, which moves tasks from the failing list to the working one. Your records carry declared meaning that the server enforces, a change reaches your figures within 60 seconds, every access is a traceable event, and your agents hold narrow permissions under one identity model.

Connect with us about where to start · · Replies within 24 hours

Sources and further reading