A $30,000 AI Deployment Failed in Two Weeks. The Problem Was 12,000 Duplicate Records.

A health services provider spent $30,000 on an AI deployment that failed within two weeks. The system was supposed to give patients accurate balance information. Instead, it gave inconsistent answers. One patient would get three different balances depending on when they asked.

The root cause was not the AI. It was 12,000 duplicate patient records sitting in the source system.

If you connect a model to duplicate patient records, you get inconsistent balance answers. The AI worked exactly as designed. It pulled from the data it was given. The data was never cleaned or deduped, so the system pulled from a mess.

Why the balance lookup broke

When a patient texted for their balance, the AI queried the database. If that patient had three records under slightly different name spellings or contact entries, the system might return the balance from any one of them. The query worked. The patient records were duplicated.

Twelve thousand duplicates in a patient database is not unusual. It happens when intake processes change over time, when front-desk staff enter records manually, when system migrations carry over legacy data without reconciliation. Most organizations know they have duplicates. They live with them because the cost of cleaning feels abstract until something breaks.

The project burned $30,000 in fourteen days.

This was not a prevention problem

I wrote recently about preventing duplicate volunteer records using SSO and a unified CRM. That post was about stopping duplicates before they start.

This case is different. The duplicates already existed. They launched the AI before fixing duplicate patient records. The failure was not in the choice to use AI. It was in the sequencing.

Who owned the patient record

Governance stops sounding optional when patients get three different balances. Before launch, assign an owner for the patient schema, define the merge process, and set a minimum data-quality threshold.

Most small and midsized nonprofits and health services organizations do not have a data steward. They do not have a documented data dictionary. They do not have a pre-deployment data audit checklist. Most teams ignore this work until it costs them money.

The $30,000 loss in this case was the AI vendor fee plus internal staff time. But the bigger loss was trust. Patients who received inconsistent balances stopped believing the system. Staff who championed the project lost internal capital. The next AI proposal will face more skepticism.

The check they skipped

Before connecting an AI to a patient database, someone should have run a deduplication audit. The query is not complicated: group records by normalized name and phone number, flag clusters with more than one entry, review and merge. Tools like OpenRefine, or even a careful SQL script, can surface the problem in an afternoon.

The harder part is the process around it. Who has authority to merge records? What happens when two records have conflicting balances? What audit trail is required?

None of this is model work. It is record ownership and merge policy. These questions should be answered before the vendor contract is signed.

What to ask in the first meeting

If you advise nonprofits or health providers on AI, use this case in the first meeting. The risk is simple: the patient records are too dirty to use in production.

Before you hand this to the implementer, count the duplicates, decide whether cleanup happens before go-live, and write into scope what happens if the duplicate rate is high.

Skipping cleanup is what slows the project down.

Start with the duplicate count

This failed for a familiar reason: nobody checked whether the patient records were usable. The difference is that AI failures are faster and more visible. A dashboard with bad data can sit unnoticed for months. An AI giving patients wrong balances gets flagged immediately.

If your organization is considering an AI deployment, ask the data quality question first. Count the duplicates. Identify the owner. Build the audit checklist. Count the duplicates before you sign the contract.