Why enterprise AI projects fail: the data was never checked

Every AI proposal assumes your data is fine. Nobody has checked.

That one unchecked assumption is where AI projects quietly stop. The vendor demo ran on clean sample data. The business case took the CRM at its word. The board approved a number that rested on records nobody had compared. Six months later the model is live, the outputs look plausible and nobody trusts them, and the project goes quiet: no announcement, just a contract that isn’t renewed.

This article sets out where AI projects actually break, why the usual data quality checks miss it, and the test to run before you sign.

The assumption under every proposal

Read any AI proposal closely and you will find the same sentence, stated or implied: the solution will use your existing customer, finance and operational data. Everything after that sentence depends on it. The forecast accuracy, the churn uplift, the hours saved: all of it assumes the data is complete, consistent and joined up.

Nobody in the room is placed to challenge it. The vendor is paid when you sign. The consultancy is paid to build. Your own team has backed the project before it reached the board. Each of them has a reason to take the data on trust, and none of them is paid to check it.

The wider picture suggests the assumption is fragile: 8% of businesses say their data is fully ready to support AI at scale (Dun & Bradstreet AI Momentum survey of 10,000 businesses, November 2026). And the stakes for the person signing are personal: half of CEOs believe their job is on the line if AI doesn’t pay off (BCG AI Radar, January 2026).

Every system looks clean on its own

Here is why the assumption survives. Ask the head of sales whether the CRM is in good shape and the answer is yes: the fields are filled, the pipeline report runs. Ask finance about the ledger and it balances. Ask HR about the employee list and it is current. Each answer is honest, and each is about one system.

AI doesn’t read one system. A churn model needs the customer in the CRM to be the same customer in billing. A revenue forecast needs deals marked won to have produced invoices. An AI agent working accounts needs the account owner to still work for you. The truth is in the gaps between systems, and those gaps are invisible to anyone looking at one system at a time.

The gaps look like this:

• Customers finance bills that sales never logged. Revenue exists that the CRM can’t see, so any model trained on the CRM learns from part of the customer base.
• Deals marked won with no matching invoice. The pipeline says one thing, the ledger another. A forecast trained on the pipeline inherits the difference.
• Accounts owned by people who have left. The CRM still routes work, alerts and renewals to them.
• The same customer recorded three ways. “Example Foods”, “Example Foods Ltd” and a billing record in capitals: one customer to a person, three to a model.

What it looks like in real output

To make this concrete, we ran our engine on a generated test dataset for a fictional B2B software company, Example Co: 72,837 records across seven exports from CRM, finance and HR. It is real engine output on generated test data, not a client, and the full memo is published as our sample Truth Score Read.

Inside each system, the data looked respectable. Completeness scored 97 out of 100. Accuracy and consistency both scored 88. A dashboard built on any single export would have been green.

Compared with each other, the systems told a different story:

• 823 customers in finance with no account in the CRM.
• Only 13% of deals reconciled to an invoice.
• 3,128 records where the city disagreed between systems.

Connectivity, the pillar that measures whether your systems agree about the same customer, deal or employee, scored 44. That one number changes the verdict. The weighted average of the six pillars was 75, but the Truth Score™ (a score out of 100 for how far your data can be trusted) is capped when a critical pillar is weak, so it came out at 59: Band C, significant work needed. The verdict on the memo was “Don’t sign yet”.

A churn model, a revenue forecast or an AI agent built on that data would have run. It would have produced confident output. And it would have been wrong in exactly the places that matter, because it could only see part of the customer base and couldn’t tie deals to revenue.

Why the usual checks miss it

Traditional data quality work isn’t wasted, but it was built for reporting, not for AI, and it has three blind spots.

It measures one system at a time. Fill rates, duplicate counts and format checks run inside a database. They can’t see that a customer in billing has no twin in the CRM. That is a cross-system question, and it needs the systems compared.

It averages. A field that is 94% complete sounds fine. If the missing 6% are the customers who left, a churn model is blind to exactly the cases it exists to predict.

It asks the people who run the systems. Questionnaires and maturity models record what people believe about their data. Belief and records can disagree in either direction, and the disagreement is often the most useful finding.

What a failed project actually costs

The licence fee is the smallest part. The larger costs are the months of internal time, the decisions made on outputs nobody should have trusted, and the credibility the sponsor spends getting the next AI proposal approved. For an investor-backed company the clock matters too: a project that stalls for a year consumes a year of the hold.

All of this is set in motion at the moment of signing, which is why the check belongs before it.

The test to run before you sign

The test is simple to describe. Take exports from two or more of the systems the proposal depends on, match the customers, people, deals and invoices between them, and see what doesn’t line up. Then read the proposal against what you found: which of the AI ideas in it the data can carry now, which after fixes, and which not at all.

That is what a Truth Score Read does. Our engine matches the customers, people, deals and invoices across the exports you send. Every red flag is checked by hand against sample records before sign-off. Within 5 working days of complete exports you get a signed memo: your Truth Score, a Leverability read on the AI in question (which AI ideas you can use now, after fixes, or not at all), the cross-system red flags, a scope line saying what we didn’t check, and a verdict: sign, fix first, or walk away.

The answer can be “sign”. If the data can carry the proposal, the memo says so in writing, and you go into the board meeting with evidence rather than an assumption. We sell no AI, no software and no fixes, and our fee is fixed before we start, so a “sign” pays us exactly what a “walk away” does.

The method is published on our approach page, including how the score cap works. Clarity before capital.