Clean Business Data Before AI: The Four Registers

Editorial update: 14 November 2026. Unverified historical outcomes have been removed; this version presents practical planning guidance.

Before connecting a business workflow to AI, establish which records it should trust. A customer name, product code or payment status can appear in more than one place. Differences need investigation rather than a confident guess.

The Four Registers are a proposed preparation checklist. This guide does not report a verified trader case, customer-count discrepancy or measured cleanup result.

Money register

Identify the accounting system and reporting period. Agree which reports or exports the workflow may use and who reconciles them. Keep balances, transaction dates and status definitions consistent.

Do not treat an AI-generated explanation as a replacement for the underlying ledger. When a total cannot be reconciled, flag it for the responsible finance reviewer.

Customer register

Choose a maintained customer identifier and record where names, contacts and account ownership are managed. Review potential duplicates before merging them; similar names can belong to different customers.

Restrict access to the fields the task needs. Record who can correct a customer entry and how that correction reaches dependent systems.

Product register

Check identifiers, units, descriptions and applicable price sources. Make effective dates and exceptions visible. A product can have several valid commercial prices, so the workflow needs rules for selecting the right one.

Flag unknown products rather than silently assigning the nearest match when a mistake could affect an order or invoice.

Knowledge register

Collect the policies and operating instructions relevant to the task. Include the document owner, version and review date. Separate approved policy from working notes and informal suggestions.

Record what to do when two sources conflict. The system should request a decision from the responsible person rather than invent a new rule.

Try the new-joiner check

Give an authorised colleague a representative task and the same information the workflow will receive. Can they find the source, understand the fields and explain what needs approval?

The exercise can expose unclear instructions. It is not a scientific test of data readiness, and a colleague's success does not prove that an AI system will perform correctly.

Keep a cleanup register

For each issue record the affected record, problem, proposed correction, reviewer and status. Preserve the original information and record approved changes. Do not bulk-delete or merge data solely on a model's suggestion.

Start with the data required for one pilot. Expand when ownership and correction processes are working. There is no universal cleanup duration; the scope and condition of the records determine the effort.

Explore Bizgenix custom AI solutions to discuss a workflow and the sources it would need.