The situation
The demo was convincing. A model answered questions about a folder of documents, or an agent updated records in a test system, and everyone could see the potential. Then someone tried it on real data.
Real data lives in a document store, two SaaS tools, a database nobody has documented and a set of spreadsheets maintained by one person. Some sources are updated daily; some have not changed in a year but are still treated as current. Formats disagree. Customer IDs do not match between systems. Most importantly, nobody can say clearly who is allowed to see what, and the demo ignored that question entirely.
The model is rarely what is missing. What is missing is the foundation underneath it: known sources, clear owners, enforceable permissions and data that stays fresh.
What the work involves
Inventory what the use case actually needs. Not every system in the organisation, only the sources this AI use case depends on. For each, I sample real data with its owner and record what it holds, its format and volume, how often it changes, known quality problems and who is responsible for it.
Map ownership and permissions. For each source I trace how access works today: groups, folder permissions, record ownership, client or region codes, informal rules. Then I work out how those rules can travel with the data, as metadata a retrieval system or agent can enforce at query time. This follows the same principle as the memory layer of the Substrate Pattern I have published: what an AI system can see is scoped by identity in a layer it cannot bypass. Where rules exist only as habits, they are written down and confirmed by the owner.
Add freshness and quality checks. Each source gets checks that alert when updates stop arriving, volumes change unexpectedly or the schema shifts. Answers built on stale data are one of the most common reasons users stop trusting AI systems.
Build the first pipelines. The priority sources get ingestion pipelines in your environment, carrying permission metadata and running the checks, so the AI system can be tested on real data under real rules.
The signature deliverable
You receive the source inventory, ownership and permission map, freshness checks and ingestion plan. Illustrative example of an inventory extract:
| Source | Owner | Changes | Access rule | Status |
|---|---|---|---|---|
| Policy library (document store) | HR operations | Weekly | Group-based; contractors excluded | Pipeline built; freshness check live |
| Customer contracts (shared drive) | Legal | Ad hoc | Per-client folders | Rule confirmed; pipeline next |
| Pricing sheet (spreadsheet) | No confirmed owner | Unknown | Anyone with link | Gap: needs owner before use |
| Support tickets (SaaS) | Support lead | Continuous | Per-team queues | API access pending approval |
Illustrative example showing the format, not a client deliverable.
How acceptance is judged
Acceptance measures are agreed after the inventory: every source in scope has a confirmed owner or a recorded gap; every source’s access rule is documented and, for built pipelines, enforced in the ingested data; freshness and quality checks run and alert correctly in a test; and the AI system can run on the ingested data in a test environment. The data engineering owner signs off.
Ownership and handover
Pipelines, checks and the inventory live in your repositories and tools. The owner receives the runbook for responding to alerts, adding a source through the same steps and handling access-rule changes. Gaps needing business decisions are handed to named owners. I do the work personally; any specialist help is disclosed and approved by you first.
Boundaries
Once the foundations are in place, the AI system itself is built through RAG and enterprise search engineering, AI workflow automation or document workflows. If you have not yet chosen which use case to pursue, start with the AI readiness and implementation roadmap.