AI assistant trained on company documents: what it takes to build one that answers accurately

What does "trained on company documents" actually mean?

The phrase is loose, and the difference between the two common approaches matters more than most vendors admit.

In almost every business case we see, the right approach is retrieval: your documents stay in your own storage, the assistant searches them when a question is asked, and the answer is built from the specific passages it found. Nothing is baked into a model. You update a policy document, and the next answer reflects it.

The alternative, fine-tuning a model on your content, changes how the model writes rather than what it knows, and makes corrections slow and expensive. For a business that wants correct answers about its own pricing, SLAs or procedures, retrieval is the sensible default.

If you are weighing the last row against a build, read custom AI chatbot vs off-the-shelf and custom software vs SaaS before you commit to either.

How much does an AI assistant trained on company documents cost?

We do not run a flagship package. Every build is scoped around what the business needs and already has, then priced from that written scope.

Payment is 50% upfront and 50% on launch by default, the audit is paid in full, and retainers are monthly in advance. The signed scope controls. Scope, tax treatment, currency and service allowances are confirmed in writing.

Running costs sit outside that. Hosting, software and AI usage are paid directly to the providers from your own accounts, so you see what the assistant actually costs to run and you can change provider without asking us.

What drives the price up

• Document mess. Five versions of the same policy across email, a shared drive and someone's laptop. Cleaning and deciding the single source of truth is work.
• Permissions. If a salesperson must not see HR files and a contractor must not see margins, the assistant needs role-aware retrieval, not one shared index.
• File types. Clean text is easy. Scanned PDFs, spreadsheets with meaning in the layout, and tables inside slide decks all take extra handling.
• Who it talks to. An internal assistant for staff is a smaller risk than one answering customers in your name.
• Integrations. Pulling live data from a CRM or job system alongside documents turns a search tool into a system.

How would Agentryx build it?

We follow the same six steps on every build, described in full on our approach page.

1. Audit. We map how the business runs, the tools it uses and where work gets stuck. For a document assistant, that means finding which questions people actually repeat and where the current answers live.
2. Design. A written scope covering systems, integrations, permissions and success measures. We agree which documents are approved sources and which are explicitly out of scope.
3. Build. Inside your own accounts, keeping what already works. You own the code and the data.
4. Test. Real cases, edge cases and failure paths before anything touches a customer. That includes questions the documents do not answer, so we can confirm the assistant says so instead of inventing something. Our notes on testing an AI agent before customers see it cover this in more detail.
5. Launch. In stages, with the team trained and documentation handed over.
6. Improve. Review the numbers, agree the next changes. Optional, as a monthly retainer.

Our advice on sequencing: start with the one journey that causes the most chasing, connect only what that journey needs, and add the rest once it is in use. A document assistant that answers the twelve questions your team asks every week is worth more than one that technically covers everything and is trusted by nobody.

What goes wrong, and the guardrails we build in

The failure mode people fear is a confident wrong answer. The failure mode we see more often is quieter: the assistant is right most of the time, nobody can tell which times, so the team stops using it.

• Answers with no source. If a person cannot click through to the paragraph the answer came from, they have to verify it anyway, which removes the saving.
• Stale documents. An assistant trained on last year's pricing will quote last year's pricing. Someone has to own the source folder.
• No "I don't know". Our AI systems are limited to approved information, say when they do not know, hand over to a person, and tell people they are automated.
• Permission leaks. Documents that one team can see and another cannot have to be enforced at retrieval, not by asking the assistant nicely.
• Legal exposure on customer-facing answers. In Moffatt v Air Canada (2024), the British Columbia Civil Resolution Tribunal held the airline liable for incorrect fare information its website chatbot gave a customer, and rejected the argument that the chatbot was responsible for its own actions. What your assistant says, you said.

So we keep a person on decisions about money, legal commitments, complaints, messages sent in your name and regulated advice. More on where that line sits in decisions to keep human in AI workflows and how to get accurate answers from an AI knowledge assistant.

If the assistant will also make outbound contact, note that under UK PECR, automated marketing calls need the person's specific prior consent, and general marketing consent is not enough.

When an AI assistant trained on company documents is not the right choice

Three situations where we will tell you not to build one.

• An existing tool already fits without workarounds. If your help desk or CRM has document search that your team will actually use, use it. A custom build that duplicates a feature you already pay for is a bad trade.
• The process is still changing week to week. If the documents themselves are being rewritten every few days, you are automating a moving target. Settle the process first, then build.
• You need it next week. We never promise a delivery time before a scope is signed. For comparison, client portals take three weeks to three months from a signed scope. If the deadline is immovable and close, a shared document with the ten answers written out by a human will beat anything we can ship.

Cheaper alternatives worth trying first

• A single, genuinely current source of truth: one folder, one owner, old versions archived. Most "the AI gave the wrong answer" problems start as "there were four versions of the document".
• A written FAQ page for the twenty questions customers repeat, published where they can find it.
• Lightweight automation to route questions to the right person. Our guide to what to automate first in a service business covers the ordering, and Zapier vs custom automation covers the point where connectors stop being cheap. Worth knowing: Zapier counts each successful action step as a task, triggers, filters and paths do not count, and at the plan limit new runs are held unless extra tasks are paid for.

What we have built with document handling and AI inside it

Two client builds are relevant, because both involved documents, roles and AI with a person in the loop.

MRM Group, a UK property development and investment company, has a public funnel connected to GoHighLevel plus an authenticated internal command centre for clients, deals and tasks. It includes client and deal workspaces with agreement-before-payment flows, document handling, task workflows, reminders and audit history, and operational reporting with human approval at sensitive financial-promotion and AI decision points. Read the MRM Group case study.

TMMB Academy, a TikTok Shop academy, came to us with a fragmented funnel, rising software subscriptions and no single system. We built one portal with role-specific workspaces for admins, coaches, creators, sales, assistants and brands, covering onboarding, creator performance, subscriptions and payments, course modules, calls, messages, community resources, brand deals, deliverables and operational alerts. The implementation ran for more than 60 days. TMMB was able to take on greater capacity, improve scalability and retention, and reports generating six figures within roughly two to three months (client-reported, shared with permission). Read the TMMB Academy case study.

On our own Agentryx platform, the AI turns a call summary into tasks and a draft follow-up email for a person to check. That is the pattern we recommend: AI drafts, a human sends.

From Trustpilot: "When it comes to building systems to acquire clients, Quam is your guy." (Denzil Jones, five stars). "Quam introduced me to the idea of having a solid infrastructure and helped me build it." (Stefano Ezeokaka, five stars).