An Internal AI Knowledge Assistant That Knows Your Company

An internal AI knowledge assistant answers staff questions from your own material: standard operating procedures, contracts, past reports, ticket history, the compliance file nobody can find. The difference between one that gets used daily and one that is abandoned in three weeks is not the model. It is whether every answer is retrieved from a named source document, filtered by who is asking, and refused outright when the documents do not support an answer.

The problem is retrieval, not intelligence

Atlassian’s State of Teams 2025, a double-blind survey of 12,000 knowledge workers and 200 executives, reported that teams waste 25% of their time searching for answers. That figure is self-reported, so read it as a measure of how the day feels rather than a stopwatch.

The structural reason sits in Okta’s Businesses at Work 2025, published 12 March 2025 from anonymised data across its customer base: the average company now runs 101 distinct applications, crossing 100 for the first time. Your company’s knowledge is not missing. It is distributed across a hundred logins, and no single person holds all of them.

Microsoft’s Work Trend Index of 17 June 2025 measured the interruption side, using Microsoft 365 telemetry plus a survey of 31,000 workers across 31 markets: employees are interrupted every two minutes during core hours, 275 times a day. A meaningful share of those interruptions is one person asking another a question that an existing document already answers.

Why the generic chatbot did not close the gap

A public chat model has read the internet and none of your files. Ask it about your refund policy and it will produce a fluent, plausible, wrong answer, because a general-purpose model with no retrieval is guessing from a statistical average of every refund policy ever written.

Adoption is not the bottleneck. Atlassian’s 2026 edition, fielded January to February 2026 across 12,035 workers and 173 Fortune 1000 executives, found 85% of knowledge workers use AI at work while only 29% have it embedded in their flow of work. Near-universal usage, almost no integration into the actual job.

Meanwhile the vacuum fills itself. The KPMG and University of Melbourne global study published in April 2025, covering more than 48,000 respondents in 47 countries, found that 57% of employees hide their AI use and present AI-generated work as their own, and only 40% say their workplace has any policy on generative AI at all. We wrote separately about whether company data belongs in a public chat tool. The short version: staff will paste it in regardless, so the sanctioned option had better be better.

Grounding is the technical answer, and there is a clean measurement of it. OpenAI’s GPT-5 system card, published 13 August 2025, benchmarked the same models with retrieval switched on and off. On FActScore, OpenAI o3 recorded a 5.7% hallucination rate with browsing enabled against 24.2% with browsing disabled. That test used web browsing rather than a private document store, so treat it as directional rather than a promise about your wiki. The direction is not ambiguous.

Two things called the same name

	Generic AI chat	Grounded internal assistant
Source of the answer	Model training data	Your documents, retrieved per question
Who can see what	Everyone sees the same model	Filtered by the asker’s own permissions
When it does not know	Produces a fluent guess	Says so, and escalates to a named human
Citations	None	Every claim points at a source document
Failure mode	Confident invention	Stale answer from an old document
Audit trail	The user’s chat history	A run log per question and per update

What the agent does, step by step

Building one is less about the model call than about the six steps around it. In practice the sequence runs like this.

1. Ingest across categories rather than one folder: procedures, compliance records, incident reports, general reference entries. Real questions cross those boundaries.
2. Run two ingestion paths, not one. A scheduled sync pulls from the shared company drive; a manual entry path covers the knowledge that only ever existed in someone’s head. Both feed the same store.
3. Authenticate before retrieving. The assistant requires a permissioned session and fails closed: if it cannot verify who is asking, it does not answer. Permission filtering happens before ranking, never after.
4. Retrieve, then draft an answer against what was retrieved, with a lower-confidence mode when grounding is partial. Degrade gracefully instead of pretending.
5. Escalate what it cannot answer to a named human, and turn that human’s reply into a new entry. The gap closes once rather than being rediscovered monthly.
6. Run a separate maintenance job that checks entries for expiry and raises follow-up tasks. Nothing auto-publishes and nothing is silently overwritten: an outdated entry is flagged and queued for a human owner, and deleting one needs sign-off.

The failure modes are boring, which is why they win

Nobody’s internal assistant dies from a bad model. It dies from staleness. Because ingestion runs on a schedule rather than continuously, a broken sync does not throw a visible error. It quietly serves last month’s answer, and the first person who acts on it stops trusting the tool permanently.

The other three are just as unglamorous. Two versions of the same procedure, both plausible, neither marked current. Knowledge that was never written down, so retrieval has nothing to find. Scanned PDFs that are pictures of text, which look ingested and contain nothing.

Every one of those is a documentation problem wearing an AI costume, and every one is why pilots stall after the demo. We covered that pattern in why AI pilots die. The fix is ownership: someone named owns each knowledge domain, review is a queue rather than a good intention, and separate domains stay separate even when one team curates them all.

What it costs to install

This is a standard AI Workforce Sprint, from $9,500 fixed, four to six weeks, built on your accounts and handed over with documentation. A typical build touches four to six integrations: a document source, your identity provider for permissions, the store the entries live in, your task system for review and escalation, and wherever people actually ask questions.

What you get is one tomte, our word for one production agent with one defined job, answering from your own files with a citation attached to every claim. It runs under the same discipline as the rest of our agent fleet: registered before it runs, fenced to what it may touch, and writing a log entry for every action, which as of July 2026 covered 41 registered agents and 5,450 runs with zero failures over the 13 days measured.

Before you build anything, run the cheap test. Take the ten questions your team actually asks each week, and check whether each one has a current, findable, written answer. Every question that fails that test is a documentation job, not an AI job, and no assistant will rescue it.

Send us the ten questions and where the answers live today. You get a written fixed-price scope, a delivery window and the exact definition of what the assistant will and will not answer, within one business day. No meeting required. Start async.