Enterprise RAG Development & Knowledge AI
The answer usually exists and the employee cannot find it. A retrieval layer over SharePoint, Drive, Confluence, Jira and your CRM that respects the permissions those systems already enforce and shows the source behind every answer.
- Who
- Delivered by a senior team assembled for the engagement, against a defined scope.
What you're seeing
- People ask a colleague where a document lives before they ask what it says.
- Usually means Findability has failed, not authorship. The material exists and the cost is the interruption — usually paid twice, by the person asking and the person who knew.
- Two departments are working from different versions of the same policy.
- Usually means There is no owned source, so local copies became authoritative by convenience. Retrieval over that estate will answer confidently from whichever copy ranked highest, which is worse than the current situation because it looks authoritative.
- The same expert answers the same question every week.
- Usually means A senior person is operating as a lookup service. It is the most expensive form of documentation and it disappears when they change jobs.
- A new joiner's ramp depends on who they sit near.
- Usually means The knowledge that matters is transmitted socially rather than written down, so ramp time varies by team and nobody can improve it systematically.
- A general-purpose assistant was rolled out and people stopped using it.
- Usually means It has no access to the systems where the answers are, so it produces plausible generic text. Usage of that shape always falls after the first fortnight, once people learn it cannot answer the questions they actually have.
Fragmentation is a running cost
Policies in SharePoint, technical documentation in Confluence, the reasoning behind a decision in a Slack thread, customer context in the CRM, project history in Jira. Nothing is missing. It is distributed across five systems with five search boxes, and none of them knows about the other four.
The cost is paid in interruptions and in decisions made without context that existed. It is invisible in any single instance and substantial in aggregate, which is why it is rarely on anyone’s list until someone tries to measure ramp time.
What gets built
A retrieval layer over the systems you already run. Hybrid search combining keyword and semantic matching, reranking tuned against measured results rather than a default, and retrieval evaluation as a permanent fixture rather than a launch activity.
Around it: permission-aware filtering against the source systems, citations on every answer, ingestion and synchronisation with change detection and deletion propagation, and a visible indication of how fresh the material is. The interface is wherever people already work — Slack, Teams, an internal portal, or inside an existing application.
The part that is not engineering
Two things decide whether this works, and neither is technical.
Someone has to own each subject. Where two documents disagree, retrieval will pick one and sound certain, so the reconciliation has to happen before indexing. This is usually the longest pole and it is work the client does, with us naming the conflicts we find.
And the evaluation set has to be graded by people who know the right answer. That is a day of expert time, spent before launch, and it is what makes every subsequent change measurable.
Where it sits
This is the knowledge layer of AI & Automation, and it is frequently the first thing built because so much else depends on it. A role-shaped assistant needs it to be useful — see AI copilots — and a service agent needs it before it can answer a policy question.
Where the concern is which models see this material, where it is processed and what is retained, that is governance rather than retrieval, and it lives in private enterprise AI.
What improves, and how you would know
The claims made for knowledge systems are usually unfalsifiable — better decisions, faster onboarding, less duplicated effort. Each of those has a measurable proxy, and agreeing them before the build is what stops the project being judged on enthusiasm.
Time to find an answer is measurable by instrumenting the searches people already run, before anything changes. Ramp is measurable if the organisation already has a definition of when a new joiner is productive; where it does not, that definition is worth writing regardless of whether anything gets built. Repeat questions to named experts can be counted from Slack and ticket data. Answer quality is measured against the graded set, continuously, and it is the only one that tells you whether the system is degrading as the corpus changes.
Retrieval cost is the fifth, and it is the one that surprises people. Sending ten documents into context when two would do is the single largest driver of inference spend on this kind of system, and it is a design decision rather than a vendor price. It is set at build time by measuring how many results the answer actually needed.
What this covers
Each of these is a capability with its own page, its own order of work and its own outputs.
| Capability | What it means |
|---|---|
| Backend & API | Server-side systems and the interfaces other things depend on. Built with the contract written first, because an API is a promise and changing one after it has callers is expensive. |
| Database & Data Modeling | The schema the rest of the system is built on. Ownership, constraints and a migration path, decided before the model has a hundred callers and becomes expensive to change. |
How we decide
Every answer carries the source it came from
Costs Answers get longer, and a visible citation invites the reader to check it, which sometimes shows the system was wrong.
An answer nobody can verify cannot be relied on for anything consequential, so it gets used for trivia and abandoned. Citations also make the system self-correcting: a wrong answer traced to a stale document turns into a fix in the document, which improves everything downstream.
Freshness is a property of the pipeline, not a re-index button
Costs The connectors cost more to build and there is ongoing operational surface to own.
The failure that destroys trust is a confident answer from a superseded document. Change detection, deletion propagation and a visible last-synchronised time are what make the answer safe to act on. A system that is re-indexed when someone remembers is one stale answer away from being switched off.
Contradictory sources are reconciled before they are indexed
Costs It delays the build and hands the client work that is theirs to do.
Retrieval cannot arbitrate between two documents that disagree; it can only pick one and sound certain. Deciding which source owns which subject is a governance question, and doing it first is the difference between a knowledge layer and a faster way to spread a stale policy.
Four ways to make company knowledge findable
The first three are worth trying before the fourth. What separates them is how much of your estate they can see and what they do at the permission boundary.
| Approach | Reaches | Permissions | When it wins |
|---|---|---|---|
| Custom enterprise RAG | Any system with an API, including internal ones | Enforced at retrieval against the source system | Knowledge spread across systems no single vendor covers |
| A suite-native assistant | That vendor's own estate, well | Inherited from the suite, and solid inside it | Most of the knowledge already lives in one vendor's tools |
| A commercial enterprise search product | Many systems, through supported connectors | Connector-dependent, and the depth varies | A standard estate and no appetite to own the plumbing |
| Per-team chat subscriptions | Whatever people paste in | None, which is the problem | Nothing, at organisation scale |
The fourth row is where most companies already are without having decided to be. It is also the reason the first security review of an AI programme tends to be uncomfortable.
Is this you?
- Employees ask colleagues where information lives before they ask what it says
- Experts answer the same internal question every week
- Onboarding depends on sitting near the person who knows
- A single well-maintained document set that search already handles
- Organisations unwilling to reconcile documents that contradict each other
- A knowledge base nobody has owned for two years — that is a content problem first
How we run it
Permission is enforced at retrieval
The index is filtered by the asking user's own entitlements before anything is ranked, rather than the answer being checked afterwards. Filtering after generation means the restricted document was already read by the model and can leak through a paraphrase that cites nothing.