Before you buy an AI tool: the data foundations finance teams skip

Every finance leader is being sold AI. Agents that reconcile, assistants that draft the board pack, tools that read invoices. The demos are impressive. Then the pilot starts, and within weeks the team is back to spreadsheets. The model was rarely the problem. The data was.

AI does not fix a disorganised finance function; it amplifies it. A process that depends on one person remembering which supplier uses which ledger code will confuse an AI agent in the same way it confuses a new hire, only faster and at scale. The firms getting real value from AI in finance did unglamorous work first.

Why AI projects stall in finance

Language models are good at reading, matching and drafting. They are poor at guessing intent that was never written down. In finance, that intent lives in places a machine cannot see: an inconsistent chart of accounts, duplicate supplier records, approvals given over chat, and documents that exist only as attachments in someone's inbox. When an AI system meets these gaps it either stops and asks for help on every item, or worse, makes a confident wrong choice.

The five foundations

1. One chart of accounts, used the same way everywhere

If two entities or two people post the same kind of cost to different accounts, no system can learn the pattern. Rationalise the chart of accounts, write a one-page posting guide with examples, and align subsidiaries to it. This single step often removes a third of the exceptions an AI tool would otherwise raise.

2. Clean master data

Duplicate vendors, customers with three spellings and missing tax identifiers are the most common reason automated matching fails. Deduplicate the masters, make one person accountable for each, and put a simple approval step on new records.

3. Documents captured digitally at source

Invoices, contracts and bank statements should arrive in one place, in a readable format, linked to the transaction they support. If documents are scattered across email and shared drives, the first AI project should be the intake process itself.

4. Processes written down as they really run

Map the close, payables and receivables as they actually happen, including the workarounds. The map becomes the specification for automation: which steps an agent prepares, where a person reviews, and what evidence the reviewer needs.

5. Controls designed for machines as well as people

Decide now which actions any automated system may take, which it may only prepare, and where approval thresholds sit. Access should be role-based and every action logged. Designing this before deployment is far cheaper than retrofitting it after an auditor asks.

A 90-day plan

• Days 1 to 30: diagnose. Map two core processes, measure exception rates and the time spent on manual matching, and list data quality issues by cost.
• Days 31 to 60: fix the foundations. Clean the masters, tighten the chart of accounts, centralise document intake and write the control rules.
• Days 61 to 90: pilot. Automate one high-volume process with clear data, run it alongside the existing process for a full cycle, and compare accuracy and time.

At the end of 90 days a finance team has a measured result, a cleaner ledger and a pattern it can repeat, whether or not the first pilot is expanded.

Questions to ask a vendor

• What data quality does your tool assume, and what happens when it isn't there?
• Which actions does it take on its own, and which does it prepare for approval?
• What does the audit trail show for each action, and can our auditors rely on it?
• Where is our data processed and stored, and is it used to train your models?

The bottom line

The return on AI in finance is decided before the first tool is switched on. Teams that invest a quarter in clean data, documented processes and clear controls get AI that works on day one. Teams that skip it get an expensive demo. The work is not exciting, but it is the difference between automation that compounds and automation that is quietly switched off.

This article is general information, not advice.