
Here's the lowdown
AI can repeat and amplify bad patterns when vendor records, account mappings or other finance data are incomplete or inconsistent.
Chris Millet of Baker Tilly suggested a 90-day cleanup cycle that starts with the most valuable datasets and gives each one a named owner.
Automation should follow preset rules, with people approving important changes and bank or remittance details kept outside the AI's control.
AI promises to make finance work faster. What happens when the records are wrong? The mistakes can move faster too.
That is the catch behind accounting automation. The software learns from the vendor names, account codes, payment details and old records already inside the business.
If those records are messy, the AI can learn the mess.
Chris Millet, a managing director in Baker Tilly's digital advisory practice for NetSuite, explained the problem at the Finance and Accounting Technology Expo in New York. His warning was simple: better tools do not automatically fix a weak data foundation.
One vendor can look like three
Vendor master records show how a small data problem becomes an operating problem.
A company might have the same supplier entered under several names. A person can often recognize that the records refer to one business. An automated system may treat them as separate vendors.
That split can damage spend analysis because purchases appear to be spread across several suppliers. It can also contribute to payment errors.
General ledger mapping creates another risk. If similar expenses are coded to different accounts, an AI tool can learn those inconsistent choices and keep applying them. The automation looks busy, but the reports become less useful.
Outdated records and disconnected systems add a third layer. One application may hold a current address while another keeps an old one. A workflow that pulls from both systems can carry the conflict into every later step.
The result is not just one bad answer. The system can repeat a bad pattern at scale.
Cleanup needs a cycle
Millet suggested treating data hygiene more like an inventory count than a one-time project.
His example was a 90-day data hygiene cycle. The method starts by profiling the available data and deciding which datasets matter most. A vendor file that controls payments deserves attention before a low-risk archive.
Next, the company assigns an owner to each important dataset. That person might be a controller, processor or vendor manager who understands how the records are used.
The team then reviews one dataset for duplicate, missing, outdated or incorrect information. It fixes the records and adds controls that make the next cleanup easier.
Those data hygiene controls can include required fields, consistent entry rules, ongoing monitoring and an audit trail that preserves the reason for a change.
Then the team moves to the next dataset. When it reaches the end, the cycle begins again.
A clean vendor file does not stay clean forever. New suppliers arrive, employees type names differently and systems fall out of sync. Data hygiene is maintenance, not graduation.
AI can help, but it needs boundaries
AI can also assist with the cleanup. A finance team can ask a model to identify missing fields, possible duplicates or records that look outdated. It can rank issues by likely impact and help people focus on the most valuable corrections first.
But finding a problem is different from having permission to fix it.
A low-risk typo may be safe for automatic correction under an agreed rule. A more important change should go to a person for approval. Some changes, including bank and remittance details, should remain outside the AI's authority.
That boundary protects more than data quality. Payment information can be a fraud target, and an apparently small automated edit can redirect real money.
The same logic applies across accounts payable, bank reconciliations and other finance systems. Teams need to decide what the AI may suggest, what it may change and what it must never touch before the automation begins.
The real AI control is routine
Buying an AI tool is a visible event. Maintaining the records behind it is quieter work.
Yet that routine can determine whether automation saves time or spreads confusion. Clean data gives the system a stable pattern to follow, while approval rules keep sensitive changes in human hands.
The practical lesson is not to wait for a broken report. Start with the datasets that affect money and management decisions, clean them in order, and schedule the next review before the current one ends.
AI can speed up the work. It cannot decide that the underlying facts deserve care.

