Business ·
AI knows only what a company can name
A request to “train AI on our company data” often assumes that documents can be uploaded to produce a digital employee. A folder of files is not yet knowledge. It may contain two versions of a policy, an expired price and a spreadsheet with no owner. Before connecting a model, the company must decide which fact is current, who owns it and which task permits AI to use it.
Begin with a question rather than the whole archive
A company rarely needs an AI system that knows everything. It needs help with specific work: finding a contract term, preparing an answer from a catalogue, matching a request to a policy or drafting a report. The task determines which data matters and how the outcome can be checked.
Uploading the entire archive adds volume without necessarily adding clarity. A three-year-old presentation may contradict the current price list, while a private conversation may contain an exception that the model treats as a rule. The broader the context, the harder it becomes to trace an answer.
The first loop needs one process and a limited set of sources. If AI answers product questions, it needs approved specifications, application rules and update dates. Company celebrations and draft advertising copy do not support that task.
A source of truth needs an owner
A file becomes operational knowledge when someone is responsible for its content. A document needs a status, date and scope. Without them, the new and old versions appear equally convincing to a model.
A simple register is useful: which knowledge is required, where it lives, who updates it, how often it changes and what happens when sources conflict. This exercise exposes problems before integration. Employees may already be using different prices or disagreeing about when a process is complete.
AI does not resolve the contradiction by choosing a fragment. The company must decide which source has authority.
Access follows the action
Knowing and changing are different permissions. Preparing an answer may require reading a customer record and a current policy. Changing a price, deal status or company detail requires separate authority and perhaps human approval.
Grant only the minimum access needed. An agent handling enquiries does not need the entire contract archive or management correspondence. A narrow boundary reduces the effect of mistakes and makes behaviour easier to understand.
Sensitive information also needs a separate decision: personal data, commercial terms, passwords, keys and closed project material. Technical availability is not a reason to place them in context. Define the purpose first, then choose a safe route.
Context should reveal where an answer came from
A useful AI answer can be inspected. The user sees which document, record or rule supported it, when the source was updated and which part of the question it addresses. Provenance matters more than confidence of tone.
If the sources do not support an answer, the system should say so and hand the question to a person. Filling a gap with plausible prose is particularly dangerous when the answer affects a price, obligation or customer decision.
Verifiability changes the interface. A single answer field becomes a view with source, date, confidence and an action to clarify. This is a working product with visible logic rather than a chat for its own sake.
Maintenance matters more than the first upload
A knowledge base begins ageing on its first day. Products, people, rules and agreements change. A project must answer not only how documents enter the system but how knowledge will continue to live.
There needs to be a clear route for publishing a new version, removing an obsolete one and verifying that the system uses the current source. Change history helps explain why AI answered differently yesterday and today.
The owner of knowledge remains a person or organisational role. A model may notice a conflict and propose an update, but approval of a rule cannot be hidden inside an automated process.
A pilot tests a decision rather than memory
Evaluate the loop on real questions with known answers. The set should include simple cases, ambiguous wording, obsolete information and questions the system must refuse. Useful measures include accuracy, handoff rate and review time.
If answers arrive faster but an employee must reread every document, the system has not created trust. If AI responds confidently without a source, the polished interface conceals risk. A pilot reveals this within a bounded task.
Within the Method, I treat company data as material for a digital product. The Concept defines the decision the system must support; knowledge structure, access and verification turn scattered files into an operational loop. Only then does AI receive context that people can trust.
AI for a specific process
If you are considering AI implementation, we can start with the process, available data, automation boundaries and a way to assess the result.