Take an accounts payable clerk with a $40,000 software renewal in the queue, on the second day of the payment run. She asks the company's AI assistant whether it needs the CFO's approval. The assistant replies clearly and politely, based on how approval limits commonly work in companies of this kind. Your delegation of authority says something different, and the assistant has never seen it.
Nothing about that answer looks wrong. That is the problem AI with company context solves. A general model knows a great deal about finance in general and nothing about your finance: your definitions, your policies, the way your chart of accounts is used and the decisions you have already made.
What company context means in finance
Company context is the material a new finance hire spends their first six months absorbing. Most of it is written down somewhere. Some of it is not.
| Context | Example question it answers | Where it usually lives | Typical owner |
|---|---|---|---|
| Definitions | What counts as recurring revenue in the board pack? | KPI glossary, board pack notes, someone's memory | FP&A lead |
| Policies | Is this laptop purchase capitalised? | Accounting policy manual, expense policy | Controller |
| Chart of accounts | Which account does a cloud hosting invoice go to? | ERP account list, coding guide, past postings | Controller |
| Approval matrix | Who must approve a $40,000 renewal? | Delegation of authority, ERP workflow settings | CFO |
| Past decisions | How did we treat the lease modification last year? | Accounting memos, audit adjustments, board minutes | Controller and CFO |
Two of these deserve a word. The chart of accounts is only useful to an AI tool with descriptions of when each account is used, which many ledgers do not carry. Past decisions matter because finance is consistent by design. The auditor will ask why this year's treatment differs from last year's, and a tool that cannot see last year's memo cannot help answer.
Why most rollouts skip it
The context layer is unglamorous work. It means finding the current version of each policy, discovering that there are two, and asking someone to decide which one stands. It means writing account descriptions that were never written. None of that appears in a vendor demonstration.
There is demand for it all the same. Gartner's 2025 survey of finance leaders, with 183 respondents, found the top AI use case was knowledge management (49%), ahead of accounts payable process automation (37%) and error and anomaly detection (34%). Knowledge management only works when the knowledge has been curated. An assistant pointed at a shared drive full of drafts and superseded versions will find an answer. Whether it finds the right one is a matter of luck.
Supplying AI with company context safely
The approach that works in finance is plain. Instead of trying to teach a model your policies, you let it look up the approved documents at the moment a question is asked, much as a careful analyst would open the policy manual before answering. Four rules keep that safe.
Retrieve from approved sources only. Build a library of the documents the tool may use. Each has an owner, a version and an effective date. Drafts, email threads and personal folders stay out unless someone has approved them as a source.
Permissions follow the user. If the person asking cannot open the payroll file, the assistant cannot use the payroll file to answer them. This needs to be enforced by the system, not by an instruction in the prompt. A tool with broader access than the people using it becomes a way round the access controls you already have.
Every answer cites its source. The answer names the document, the section and the version it came from, so the person can check it in a few seconds. Where no approved source covers the question, the tool should say so rather than fall back on general knowledge.
Figures come from the system of record. Numbers in an answer should be read from the ledger or the reporting system, with a link back to the transactions. The model explains the figures. It should not produce them.
These rules make answers checkable. They do not make them infallible. A tool can still pick the wrong passage or misread a well-written policy, which is why the citation matters: it turns a review from a research task into a quick check.
Keeping the context current
A context library goes stale faster than people expect. A new approval limit is signed off in March, the old policy stays in the library, and the assistant keeps quoting it.
Give each document an owner. The owner is responsible for replacing it when it changes and retiring the old version the same day.
Tie updates to the events that cause them. A board resolution changing authority limits, a new accounting policy, an audit adjustment. Each should trigger a library update as part of the same approval.
Test with known questions. Keep a short list of questions with agreed answers, and run it after each significant change. If an answer moves, find out why before people notice.
Where to start with AI with company context
Start with one team's questions rather than the whole company's documents. Ask the AP team, or the people who email the controller most, for the questions they ask every week. Coding, approval limits, vendor setup rules and cut-off dates usually dominate.
Then gather only the documents that answer those questions. Resolve the contradictions you find, which is useful work whether or not AI is involved. Write the missing account descriptions. Load that small, clean set, test it against the questions, and widen it from there.
This is also where company context meets the rest of an adoption programme. Enterprise AI Adoption: Why Giving Everyone a Chatbot Is Not a Strategy sets out the other three pieces: use cases by role, guardrails and measurement.
What company context will not fix
It will not correct a policy that is wrong, or decide between two policies that disagree. The tool will faithfully quote whatever you approve, so the quality of the answers is capped by the quality of the library.
It will not capture judgment nobody has written down. Some of the most valuable context in a finance team is the reasoning behind past decisions, held by one or two experienced people. Getting that into memos is a finance task first, and an AI task second.
Frequently asked questions
What does AI with company context mean?
It means an AI tool that answers from your organization's own definitions, policies, chart of accounts, approval matrix and past decisions, rather than from general knowledge. The tool looks up approved documents when a question is asked and cites where the answer came from.
Do we need to train our own AI model on company data?
No, for most finance use cases the tool can look up approved documents at the time of each question instead. That keeps sources current, lets permissions apply per user, and means each answer can point to the document behind it.
How do you stop an AI assistant showing people data they should not see?
Make permissions follow the user, enforced by the system rather than by instructions to the model. If a person cannot open a document or a ledger, the assistant should not be able to use it to answer them.
Can an AI assistant give accurate financial figures?
Only if the figures are read from the ledger or reporting system and traced back to transactions. The model should explain numbers taken from the system of record, not produce them itself.
What company context should a finance team prepare first?
Start with the documents that answer the questions one team asks most, usually the approval matrix, the coding guide with account descriptions, and the expense and capitalisation policies. Resolve contradictions before loading them.
Where CFOLogic fits
Supplying company context is one part of the AI adoption approach we take with CFOs. We design tools that run read-only on top of the systems already in place. Our Navigator demo shows the pattern on fictional data: it reads QuickBooks Online, Xero or a hosted ledger read-only and answers in plain English, with figures traced to the transactions behind them. The CFO Success Partners page describes how we work.