The Algorithm Doesn't Charge You. It Just Suspends You.

On July 14, the Small Business Administration announced that its relationship with Palantir Technologies had entered a permanent, expanded phase, converting a $300,000 pilot contract signed in January into an ongoing deployment running against the agency's entire pandemic-era loan archive. Administrator Kelly Loeffler's line for the occasion: "No amount of fraud is acceptable, whether it is $10,000 or $10 million." Somewhere past that sentence sit more than 150,000 small business owners who are already suspended, none of them convicted of anything, all of them locked out of the SBA on the word of a dataset.

Published August 16, 2026 • Filed under: Suspicion As A Service, The Fraud Prevention Pilot That Never Ends

Rows of illuminated data center server racks, standing in for the loan-file archive Palantir software now runs against

Nobody at LOLSBA thinks a government agency should sit on evidence of $200 billion in fraud and do nothing with it. That estimate belongs to the SBA's own inspector general, drawn against the roughly $1.2 trillion in Paycheck Protection Program and COVID Economic Injury Disaster Loan money the agency pushed out the door in 2020 and 2021. Somebody has to sort through that pile. The question the July 14 announcement never really answers is what happens to the people caught in the sorting before anyone proves they belong there.

Here is what the agency confirmed. The Palantir contract began in January 2026 as a $300,000 fraud prevention pilot and training bootcamp. Seven months later, it is no longer a pilot. The SBA's release describes an arrangement that runs indefinitely, with Palantir software analyzing loan datasets at scale, flagging anomalies, tracing what the agency calls coordinated schemes, and feeding leads to investigators and prosecutors. The pitch is efficiency. The output, so far, is a suspension list.

The Five-State Ledger

The SBA's own numbers, published the same day, break down like this across the states where the dragnet has run longest.

StateBorrowers suspendedSuspected fraud tied to them
California112,000$8.6 billion
Ohio27,000$1.1 billion
Wisconsin7,800$375 million
Minnesota6,900$400 million
Maine1,500$93 million
Five-state total150,000+$10 billion+

Suspension is not a courtroom word. It is an administrative one, and the SBA's release is careful to say the figures represent suspected fraud, not established criminal liability. That caveat is buried under the headline number. A suspended borrower cannot get a new SBA loan. A suspended borrower cannot get into the 8(a) Business Development Program. A suspended borrower's file sits flagged while, separately, the agency has already referred more than 560,000 pandemic-era borrowers worth $22 billion to the Treasury Offset Program for collection, a package the SBA itself calls the largest referral in agency history. None of the referral requires a trial either.

The pilot program worked exactly as advertised: it found more names to add to a list. What it has not done, seven months in and now made permanent, is tell any of those 150,000 people how to get off it.

What "New" Actually Means Here

Reporting on the contract's history complicates the announcement's framing as a fresh breakthrough. Palantir's involvement traces back to a Minnesota fraud investigation that first drew national attention, and independent coverage of the expansion has noted that a meaningful share of the fraud cases the administration has since publicized were already under DOJ investigation before Palantir software ever touched them, flagged during the prior administration and left sitting. What changed in July was not that fraud got discovered. What changed was that the agency put a vendor's name and an AI narrative on top of casework that, in a lot of these files, was already moving.

That distinction matters because it is doing real marketing work. A software platform that finds coordinated fraud rings sounds like a technical leap. A software platform that re-indexes a backlog of leads investigators already had, and packages the result as a "new phase," sounds like what it mostly is: a procurement decision dressed up as a breakthrough.

The Fair Reading

Here is the version that gives the agency its due, because one exists. Running pattern-matching software against 1.2 trillion dollars of loan records by hand was never realistic, and the alternative to automated screening is not some idealized careful process, it is the same rubber-stamp underwriting that let the fraud happen in the first place. A tool that flags anomalous filing patterns across five states in months rather than years is doing something a human review team structurally could not. Suspension, unlike prosecution, is reversible in theory, and the SBA has consistently framed it as a hold rather than a verdict. If the software's hit rate is genuinely high, most of the 150,000 people on that list are not going to be exonerated later. Some of this dragnet is going to be right.

But "in theory reversible" and "in practice navigable" are different claims, and the agency has published no data on how many suspended borrowers have successfully appealed, how long an appeal takes, or what evidence clears a name once software has attached it to a file. The release that announced a permanent AI partnership did not come with a released appeals success rate, a published average resolution time, or a commitment to review already-suspended files under whatever improved detection the new phase supposedly brings. It came with a quote about zero tolerance and a five-state scoreboard. A machine that generates suspicion at scale and an agency that cannot say how fast it clears people at scale is not a fraud solution. It is a queue with a press release attached.

What Would Fix It

If the SBA's algorithm flagged an actual working business you run, send us the story. The rest of the receipts are filed here.

SHARE ON X SHARE ON FACEBOOK SHARE ON LINKEDIN