An AI assistant over your own documents

An AI assistant over your own documents — usually built as retrieval-augmented generation, or RAG — searches your firm's files for relevant passages and has a language model answer from those passages rather than from its general training. It works well when your documents are accurate and current, and it fails quietly when they are not. For a firm of 5–50 people it is buildable in weeks; the real project is the state of your documentation.

What does an assistant over your documents actually do?

The distinction that matters: a plain chatbot answers from whatever its model absorbed during training, which includes nothing about your firm. An assistant over your documents adds a retrieval step. Before answering, it searches an index of your SOPs, proposals, runbooks and meeting notes, then grounds its answer in what it finds — ideally with a citation back to the source file.

In practice this turns "ask the person who knows" into "ask the system". A new account manager asks how you handle a client's overdue invoice; the assistant quotes your credit-control SOP rather than offering the model's generic opinion. That is the whole trick. It is modest, and it is one of the few applications I would place firmly in the "works today" column of what AI automation actually delivers for B2B firms.

How does retrieval actually work?

The mechanism in plain terms. When you add a document, the system splits it into chunks and stores a mathematical fingerprint of each chunk in an index. When someone asks a question, the question gets the same fingerprint treatment; the system pulls the chunks whose fingerprints sit closest to the question's, then hands those chunks to the model with an instruction that amounts to "answer only from these passages". When retrieval finds nothing close enough, a well-built assistant says it does not know, rather than improvising.

Every failure mode follows from that chain. When the right passage is not in the index, the answer is wrong. When the passage is in the index but out of date, the answer is confidently wrong — which is worse. When a question is vague, retrieval pulls vaguely related chunks and the answer drifts. Understand the chain and none of the behaviour is mysterious.

What is it genuinely good at?

Three uses I would stand behind for a small B2B service firm:

• Process questions. "How do we onboard a new client?" "What goes in a kickoff email?" Anything with a written SOP behind it gets answered in seconds instead of interrupting whoever wrote the SOP.
• New-starter onboarding. The first month of questions a new hire asks are almost all answerable from documents nobody remembers to share. An assistant makes that self-serve.
• Institutional memory. "Have we quoted for this kind of work before?" "What did we agree with that supplier?" Past proposals and contracts become searchable in plain English.

A concrete example from our own operation: when bounce rates spike mid-campaign, the answer lives in a runbook — check verification, check sending volume, check authentication records — most of which mirrors the practical guide to cold email deliverability. An assistant retrieves the relevant procedure faster than anyone rereads the full document.

Where does it fail?

Four places, reliably. First, stale documents. An assistant over out-of-date SOPs is a distribution system for wrong answers, delivered with fluent confidence. Someone has to own keeping the source documents current, or the tool decays within months.

Second, judgement calls. "Should we take this client?" is not a retrieval problem. The assistant can surface your qualification criteria; it cannot weigh them for you, and it should not pretend to.

Third, synthesis across many documents. Asking it to reconcile pricing across forty old proposals typically produces a plausible-sounding partial answer. Retrieval pulls a handful of chunks, not the whole corpus.

Fourth — the one nobody mentions in the sales demo — it cannot retrieve what was never written down. If your processes live in people's heads, there is nothing to index. The assistant is downstream of documentation, always.

What does it cost, and when is it worth building?

Hedging deliberately, because pricing moves: off-the-shelf options built into the office suites are commonly priced per seat per month at the level of a modest software subscription, and for most ten-person firms that is the right starting point. A custom build — your own index, your own access controls — is typically a low-thousands project plus ongoing maintenance, and only earns its keep once the off-the-shelf version has proven the habit. I have set out the wider tooling picture in an AI stack for a ten-person firm, with costs.

The test for whether it is worth building at all is unglamorous: do documented processes exist, and does someone own keeping them current? If yes, an assistant compounds their value. If no, write the SOPs first — they pay twice, because documented processes are also the raw material for automation proper, which is the subject of from SOP to automation: the promotion rule.

---

Next step: the Growth System Audit — £450, seven days, credited against any build — maps where your growth system leaks and what to build first.