The prerequisite that decides whether AI works.
Modelled, documented, access-controlled and current data — the condition your records have to be in before any assistant, agent or automation can be trusted with them.
Why AI projects fail before the model is chosen
Retrieval quality is a data engineering result. If records are inconsistent, undocumented, or unreachable behind an application boundary, no amount of model selection compensates. Most stalled AI initiatives are data initiatives that were never funded as such.
- The information an assistant would need is spread across systems and documents.
- Records are inconsistent enough that a machine could not reconcile them.
- There is no way to tell what a given field means without asking a person.
- Permissions live in the applications, so any extract loses them.
What we build
The data conditions AI depends on: entities modelled consistently, fields documented, permissions carried through, and refresh on a cadence that matches how fast the business changes.
Read: data readiness is the real prerequisite →- Entities modelled consistently across the source systems that describe them
- A field-level dictionary: what each field means and who owns it
- Permission metadata carried into the layer rather than stripped on extract
- Refresh on a stated cadence, with freshness visible to consumers
- Quality checks that stop bad data reaching AI systems
- A readiness assessment naming what is usable now and what is not
How it runs
Driven by the questions you want answered, so the work is scoped rather than boiling the estate.
- 01Start from the questions
The questions the business wants answered determine which data has to be ready. Everything else waits.
- 02Assess the sources
For each required fact: where it lives, how consistent it is, and whether it can be read with its permissions intact.
- 03Model and document
Consistent entities and a field dictionary, so both people and retrieval systems can interpret the data.
- 04Carry permissions
Access rules travel with the data, so an AI system cannot become a route around them.
- 05Refresh and check
Scheduled refresh with quality gates, so what AI reads reflects the current state and passes basic sanity.
What changes once it is running
What readiness changes about every AI initiative that follows.
Retrieval finds the right thing
Consistent, documented data is what allows a retrieval system to return the correct passage.
Answers can be permissioned
Because access rules came with the data, AI can serve different users different answers correctly.
Confident wrong answers become rarer
Quality gates stop the inconsistent records that produce plausible, incorrect responses.
The next initiative is cheaper
Readiness is done once and reused by every assistant, agent and automation after it.
How an engagement is shaped
The readiness assessment is deliberately honest, including about what is not worth doing yet.
Readiness assessment
Two to three weeks against your intended use cases, returning what is usable now, what needs work, and what is not economic to prepare.
Prepare the first domain
Modelling, documentation and permission handling for the domain behind the highest-value question.
Extend
Further domains as use cases justify them, rather than preparing an estate for hypothetical future needs.
Common questions
The things buyers ask before they commit. If yours is not here, it is a good first question for the assessment.
- Can we not just point AI at our documents?
- You can, and it will work for simple lookups. It fails where an answer needs reconciling records across systems, respecting permissions, or knowing which of two conflicting documents is current — which is most enterprise questions.
- How much data preparation is enough?
- Enough for the questions you actually want answered. Preparing everything is a multi-year programme with no payback date. Scoping by use case is what makes it fundable.
- Do we need this before any AI at all?
- Not for general drafting assistance, which needs no company data. For anything answering questions about your business, this is the work, and skipping it is the most common reason those projects fail.
Bring the question you want AI to answer.
We will trace what data answering it reliably would require — and tell you honestly whether you have it.