Home  /  Insights  /  Data fitness audit

How to run a 500-record data fitness audit before any artificial intelligence pilot

Most failed automation pilots fail on data, not on the model. This is the two-to-four-hour audit that tells you whether your enterprise resource planning data can carry an agent, what it costs to fix if it cannot, and when to stop.

By Lalit Kumar · Published 20 September 2026 · 8 minute read

Why a data readiness assessment comes before any AI pilot

Every agentic artificial intelligence (AI) pilot reads transaction records: invoices, purchase orders, vendor and customer master data, general ledger codes. If those records are incomplete or inconsistent, the agent produces confident wrong answers, the pilot misses its accuracy gate, and the budget is gone. In the SEAS implementation system, the data fitness audit is the first activity of Phase 0, before vendor selection and before any budget approval, because skipping it is the single most expensive shortcut in the program: a failed implementation on bad data costs between $100,000 and $500,000 to unwind and restart.

The audit takes two to four hours with a finance analyst and someone who can export from the enterprise resource planning (ERP) system. It needs no software beyond a spreadsheet.

The ten fields that decide whether an AI agent can work

Pull a random sample of 500 transaction records from the workflow you intend to automate, typically accounts payable invoices or purchase orders. Score each record on ten fields:

FieldWhat "pass" meansWhy an agent needs it
Primary keyPresent and unique for every recordWithout it, nothing can be matched or de-duplicated
Vendor namePresent, one spelling per vendorThree spellings of one supplier become three suppliers to an agent
Customer namePresent, one spelling per customerReceivables and churn analysis break on duplicates
Invoice datePresent, valid, in a consistent formatCycle time and days sales outstanding are date arithmetic
Purchase order datePresent, valid, not after the invoice dateLead-time and late-delivery analysis depend on it
Invoice amountPresent, numeric, non-zeroEvery savings figure is amount multiplied by volume
Purchase order amountPresent, numericThe three-way match compares it to the invoice
Purchase order referencePresent on the invoice and resolvable to a real purchase orderThe match cannot run without it
Vendor identifierPresent and consistent with the vendor nameMaster-data integrity for every downstream agent
General ledger codePresent and validSavings must land in a cost centre to reach the profit and loss statement

Each field carries its own pass target in the scorecard. Count the records that pass on each field, divide by 500, and average the ten percentages. That average is the data fitness score.

Reading the score: the four bands

ScoreDecisionRemediation cost and time
Above 80%Proceed to the pilotNone required before launch
70% to 80%Remediate first, then proceed$30,000 to $75,000, four to eight weeks
40% to 70%Significant remediation before any pilot$75,000 to $150,000
Below 40%Stop. The workflow is not a candidate this yearA data programme, not an AI programme

The bands are deliberately blunt. A score of 62 percent looks close to passing; in practice it means more than one record in three has a defect an agent will act on. The related readiness checks in the SEAS execution manual set the bar at data completeness above 85 percent and accuracy above 90 percent for the fields the pilot actually uses.

The remediation order, if the score falls between 40 and 80 percent

Fix fields in the order that unblocks the most downstream work, not in the order they appear in the export:

  1. Primary keys, weeks 1 to 2. Nothing else can be reconciled until every record has a unique identifier.
  2. Dates, weeks 2 to 3. Standardize formats, remove impossible values, close the invoice-before-order cases.
  3. Names, weeks 3 to 4. One canonical spelling per vendor and customer, mapped to identifiers.
  4. Purchase order references, weeks 4 to 6. The slowest fix, because it usually means changing how invoices are received.
  5. Amounts and general ledger codes, ongoing. Validation rules at entry rather than cleanup after.

The re-audit rule

After remediation, draw a fresh random sample of 500 records and score it again. Do not re-score the cleaned sample; you would be measuring the cleanup, not the system. The re-audit decision is stricter on the low end: 70 percent or better, proceed; 60 to 69 percent, conditional, with two more weeks of remediation and one more audit; below 60 percent, do not proceed. The audit closes with a sign-off sheet naming the auditor and the chief financial officer or chief operating officer who reviewed it, so the go decision has an owner.

What the audit connects to

The score feeds three later decisions. It sets the remediation line in the implementation budget. It determines which workflow becomes the pilot: when two candidates have similar savings, the one with the higher data score goes first. And it is the first evidence a board sees that the programme is being run on measurement rather than enthusiasm. In SEAS the audit is workbook one of nine, with the remediation cost calculator and the sign-off sheet built in, and the free sample data guide includes two datasets to practise the scoring on before touching your own.

Frequently asked questions

How long does an ERP data quality audit take?
Two to four hours for a 500-record sample, with one finance analyst and one person who can export from the ERP system. The re-audit after remediation takes the same time on a fresh sample.
Why 500 records and not the whole ledger?
A random sample of 500 gives a stable estimate of defect rates for each field without the export and cleaning effort of a full extract. The purpose is a decision, not a cleanup; the cleanup comes later and only if the decision says so.
What data fitness score do I need to start an AI pilot?
Above 80 percent on the ten-field average to proceed directly; 70 to 80 percent to proceed after four to eight weeks of remediation; 40 to 70 percent means significant remediation first; below 40 percent means stop.
What does data remediation cost for a mid-market company?
In the SEAS scorecard bands, $30,000 to $75,000 for a score between 70 and 80 percent, and $75,000 to $150,000 for a score between 40 and 70 percent. Skipping remediation and failing the pilot typically costs $100,000 to $500,000.
Can the same audit be used for supply chain or receivables data?
Yes. The ten fields cover the invoice, purchase order, vendor, customer and general ledger elements that supply chain, receivables and payables agents all read. Draw the sample from the workflow you intend to pilot.

The scorecard, the remediation cost calculator and the sign-off sheet are workbook one of the nine live Excel workbooks in SEAS, with sample data to practise on.

See what ships in SEAS   Get the free FLOAT diagnostic book