Vendors now describe almost any dataset as AI-ready. The published definition is stricter, and it sets five conditions that must hold at the same time. Most organisations meet one or two of them. This article gives the five tests in plain terms, shows you how to run each one against your own estate, and names the condition that defeats most organisations.
Connect with us about your data foundation · · Replies within 24 hours
Data qualifies as AI-ready when it meets five conditions at the same time. It must be discoverable with business context. It must be accessible in real time or close to it. It must carry governance end to end. It must be high quality and certified. It must reach users through reusable interfaces.
Readiness describes what a machine can do with your data without a person in the middle. A dataset can therefore be clean, complete and current, and still fail the test.
The gap between adoption and result shows why this matters. The Office for National Statistics tracks AI use across UK business in its 2023 to 2026 series. Adoption among UK SMEs reached 54% in 2026, from 35% the year before, and only 12% of those users report higher revenue.
Discoverable means more than findable. Your data must carry metadata, a named owner, its lineage and its business context, so that a machine can learn what a field means without asking a person.
To run the test, pick three fields from three systems and ask four questions of each. Who owns it? Where did the value come from? What does it mean in business terms? Where is that meaning written down?
Most estates fail on the fourth question. The value exists, the owner exists, and everybody knows the meaning. Nobody has recorded that meaning in a form a machine can read.
Accessible means available in real time, or close to it. An agent that works from last night's extract works from a picture of yesterday.
To run the test, change one value in a source system and measure how long that change takes to reach the place a model reads. Count in seconds rather than hours.
A nightly batch fails this test, and so does a weekly export into a reporting warehouse. Both were reasonable designs for human reporting, and both break when a machine must act on the answer.
End-to-end governance needs three things at once: one identity model, one policy engine, and an audit trail.
To run the test, count your identity systems, then count the places where somebody decided who may read a record. If either number is greater than one, you hold several governed systems next to each other rather than end-to-end governance.
Certified means somebody has made a specific, checkable promise about the data. A data contract states what a field contains, what shape it takes, and who answers if it changes. A quality service level states how often the data may be wrong.
To run the test, ask for the contract behind one dataset. If nobody can produce one, your data carries no certification, whatever its quality.
Quality and certification are separate properties. A machine cannot inspect your intentions, and it can only read what you declared.
Provisioned as a product means the data has a defined interface that others can use again. Not a bespoke extract per consumer, and not a file that somebody emails.
To run the test, count the routes into one dataset. Where each consumer built its own route, somebody provisions that data by exception every time.
Gartner expects 75% of gateway vendors and 10% of integration platform providers to carry Model Context Protocol features by 2026. That protocol gives a model a standard way to reach a tool or a dataset, and it solves the connection rather than the meaning.
| Test | Typical score | Reason |
|---|---|---|
| Discoverable | Partial | Owners exist, and written business meaning does not. |
| Accessible | Fail | Reporting runs on batches designed for people. |
| Governed end to end | Fail | One identity model per supplier, not one per organisation. |
| Quality and certified | Partial | Data is often accurate and rarely contracted. |
| Provisioned as products | Partial | Interfaces exist per consumer, built on request. |
The usual response is a readiness programme. A team catalogues the data, writes the business meaning down, negotiates contracts and builds interfaces. That work delivers real gains, and it has to be repeated for every system. It also decays, because nothing breaks when a catalogue falls out of date.
A 2026 benchmark ran three frontier models over 99 business questions, twice each. The second run added a short document describing what the data means. Accuracy rose by 17 to 23 percentage points for every model, and the three models then became statistically indistinguishable from each other.
Readiness therefore decides the outcome, and it does not follow model selection. We set out the full evidence in why AI gets your own business data wrong.
We build your system on engage.re, and your system meets the five conditions as properties rather than as a programme.
| The test | How your system satisfies it |
|---|---|
| Discoverable with business context | A shared dictionary holds the meaning of every record type, and your application declares its vocabulary before it writes anything. |
| Accessible in real time | A change reaches your reported figures within 60 seconds. Reads inside your own domain run under 200 milliseconds at the 95th percentile. |
| Governed end to end | Identity is a key, so one model covers your people, your applications and your agents. One gate decides every crossing, and every access is a recorded event. |
| Quality and certified | Your declared vocabulary is the contract, and the server refuses any write outside it. |
| Provisioned as products | Six interfaces carry everything, over JSON, from any language. |
The third row is the row that defeats readiness programmes. One identity model and one policy engine are consequences of how a platform was built, and no supplier can add them later.
engage.re holds everything in five database tables, and those five tables never change. A new record type adds a dictionary entry instead of a table. Sense Future built the platform, and it has run in production since December 2025.
A council runs housing, revenues, social care and waste as separate systems, and each one holds a person. A readiness programme must catalogue four systems, negotiate four contracts and keep four descriptions current. On a shared dictionary the same concept carries one identifier in all four, so the description exists once.
A care home group runs a records system, a rota tool and a family portal, and each one holds a resident. The arithmetic is the same with three systems as with forty, and the registered manager still cannot get one occupancy figure without a person reconciling sources.
The care home software and charity software pages set out the sector detail. Our guides to the Data Security and Protection Toolkit and funder-ready impact reporting cover the obligations that sit on top.
Data is AI-ready when it meets five conditions at the same time. It must be discoverable, with metadata, a named owner, lineage and business context. It must be accessible in real time or close to it. It must carry governance end to end, with one identity model, one policy engine and an audit trail. It must be high quality and certified, with data contracts. It must reach users through reusable interfaces.
Governance end to end. It needs one identity model, one policy engine and one audit trail across the whole estate. A council may run forty systems from twenty suppliers, and each one holds its own idea of a user with its own permission model. The condition is architectural, so only architecture satisfies it.
No. Data can be accurate, complete and current, and still fail the test. Readiness describes what a machine can do with the data without a person in the middle. Accurate data with no written business meaning, no contract and no shared identity model is clean and not ready.
No. Accessible means available in real time or close to it, and an agent that works from last night's extract works from a picture of yesterday. Change one value in a source system, then measure how long that change takes to reach the place a model reads. Count in seconds.
Fix the data first. A 2026 benchmark ran three frontier models over the same questions, with and without a description of the data. Accuracy rose 17 to 23 points with the description, and the three models then became statistically indistinguishable. The data decided the outcome, and the model barely mattered.
We build your system on engage.re, which satisfies the five conditions as properties. Your vocabulary is declared and the server enforces it, a change reaches your figures within 60 seconds, one identity model and one gate cover your whole estate, and six interfaces carry everything over JSON.
Connect with us about your data foundation · · Replies within 24 hours