So here is the new act in the same play. A White House task force run out of the Vance side of the building took a long look at the Small Business Administration's pandemic loan portfolio and produced a sentence that should have ended several careers and instead became a press release. The sentence is this. The prior administration's SBA approved, then internally flagged, then continued to honor, 562,000 pandemic loans worth a combined $22.2 billion. The agency's word for this in the public statement is "protected." Read that word twice. The federal agency that exists to verify small business eligibility says it "protected" 562,000 loans that its own systems had labeled as likely fraud. The agency did not catch the fraud. The agency saw the fraud, then waved it through. That is not a control failure. That is a policy choice.
The follow-up sentence is the part where the cyberpunk lights kick on. The SBA's plan to find the rest of the fraud is to hand the dataset to Palantir, the AI surveillance vendor, and let the analytics platform do the job the agency has refused to staff itself for since 2020. So now the loop has a fourth turn. Step one, the SBA approves the loan. Step two, the SBA's fraud system flags the loan. Step three, the SBA decides to pay the loan anyway. Step four, the SBA hires an AI vendor to identify the loans the SBA already flagged and chose to pay. The agency could have skipped to step four in 2021. Instead it spent five years and $22.2 billion writing the dataset that justifies the contract.
What "Protected" Actually Means In Government English
"Protected," in the press release, is doing a remarkable amount of work. The implication, in the kind of language a press intern uses to keep a senior administrator out of a Senate hearing, is that the prior SBA leadership instructed loan officers and lender partners to keep disbursing on accounts that had been flagged by internal anti-fraud screens. The task force's language is more direct than the SBA's. It uses the verb "shielded" in some materials. It says the disbursements continued after the flags. It says the lender partners were notified that the flags should not be a basis for stopping payment. It says, in places, that the SBA's own inspector general has the receipts.
Translate that into plain English and you get a sentence the SBA will never print on its homepage. The agency designed an automated fraud-detection layer, the layer correctly identified hundreds of thousands of likely fraudulent applications, and the agency overrode the layer in real time. That is not an accident. That is the agency choosing to launder its own fraud-detection signal through a press cycle three years later instead of acting on the signal in 2020 when it first lit up.
Now Enter Palantir, Stage Right, In a Hoodie
Palantir Technologies has spent the last decade building tools that ingest unstructured government data, normalize it, and surface anomalies. It is good at this. It is also expensive at this. The federal contracting record shows Palantir has been hired to do similar work for the IRS, parts of HHS, and various intelligence community customers. The pitch to the SBA is straightforward. The pandemic portfolio is too big and too messy for the SBA's existing analytics shop to chew through, so the agency outsources the chewing. Palantir gets paid by the contract year. The SBA gets a quarterly report it can hand to Treasury for offset. Treasury gets a list. The list lands. Some of it sticks. Most of it does not, because most of these debtors are insolvent shell entities and the recovery rate on shell-entity restitution orders is, statistically speaking, an embarrassment.
The agency that approved 562,000 fraudulent loans is now hiring an AI surveillance company to identify 562,000 fraudulent loans. If you described this as a business model, an investor would ask why the agency does not just stop approving fraudulent loans in the first place. There is no answer to that question that does not implicate the agency.
The Time-Travel Logic Of "AI-Powered Fraud Detection"
The phrase "AI-powered fraud detection" sounds like it points forward. In SBA usage, it points backward. The pandemic loan window has been closed for years. The applications are not still arriving. The fraud has already happened. The money is already gone. The phrase, in this context, means "we are going to use machine learning to comb through transactions that finished clearing during the Trump and Biden administrations and identify, with high confidence, which ones we should have rejected at the time." That is not fraud detection. That is forensic accounting with a marketing budget.
The reason this matters is that "AI-powered" has a specific function in government press copy. It signals technical seriousness. It implies that the next round will be different. It does not, in any of the SBA's recent contracting language, commit to a real-time check at the application stage of any future federally-administered emergency loan program. The agency is buying retrospective surveillance, not prospective controls. If a second pandemic-style emergency lending window opens tomorrow, the SBA will run it on the same intake stack it ran in 2020, because the contract it just signed does not touch the intake stack. It touches the rear-view mirror.
The 562,000 Number, In Context
To understand the scale, line up some other numbers. The total number of PPP loans issued was around 11.5 million. The total number of EIDL grants and advances was several million more. A figure of 562,000 "protected" suspect loans is, depending on how you slice the denominator, somewhere between three and five percent of the entire pandemic disbursement universe. Three to five percent does not sound dramatic until you remember that those are the loans the agency's own internal flags caught and the agency chose to keep paying. The total pool of fraudulent or improper pandemic disbursements, per multiple inspector general estimates, is in the hundreds of billions. The 562,000 is the slice the agency saw, identified, and decided to honor. The rest is the slice the agency did not even see.
- Loans approved that the agency's own anti-fraud system flagged: 562,000.
- Total dollar value of those flagged-and-paid loans: $22.2 billion.
- Independent IG estimates of total likely PPP/EIDL fraud across the program: north of $200 billion.
- Number of borrowers, of those 562,000, who have so far been criminally charged: a fraction of one percent.
- Years between the original disbursement and the current "AI-driven" recovery push: between four and six.
The recovery rate on this kind of restitution is grim and well-documented. Once shell entities are dissolved, once stolen funds have been pushed through real-estate purchases, vehicles, and offshore wires, the government's collection cents on the dollar drop into the low single digits. Palantir's contract will identify the loans. Treasury will refer the borrowers. The cases will land. The money will not come back, in the aggregate, in any volume that meaningfully offsets the original loss.
"Protected" Is The Confession
The reason this story is genuinely worth a page on this site is that the word "protected" is the closest thing the SBA has come to a public confession since the inspector general's first big report. Up until now, the agency's preferred posture has been that fraud was inevitable, the program prioritized speed by congressional design, and the agency has been catching cases steadily ever since. The Vance task force's framing breaks that posture. It says, in writing, that there was a flagging system. The flagging system worked. The flagging system was overridden. The override was a leadership decision. The agency that runs the program is the agency that called the override. None of those facts are compatible with "fraud was inevitable."
This is the part where, in a normal accountability cycle, the agency leadership that authorized the override would be named, the override directive would be published, and the lender partners who were instructed to ignore the flags would be issued letters. None of that has happened. The agency has gone immediately to "AI-powered cleanup" because the AI-powered cleanup is the version of the story where nobody at the SBA has to answer a question about why a flag became a payment.
What The Palantir Contract Actually Buys
Strip out the procurement language and the Palantir engagement appears to deliver four things, in roughly this order.
- An ingestion pipeline that pulls SBA's pandemic loan records, including the historical anti-fraud flag data, into a Palantir Foundry environment.
- A scoring layer that re-ranks borrowers by likelihood of fraud, using updated entity-resolution and external dataset cross-checks (state corporation filings, IRS data where available, bankruptcy records, criminal court records).
- A case-routing layer that hands the highest-confidence fraud cases to the SBA Inspector General, the DOJ, or Treasury for offset.
- A dashboard that lets agency leadership and Congress watch the recovery numbers tick up over time, in a visualization the agency can put in front of any committee that wants one.
The dashboard is the product. Everything upstream of the dashboard is what justifies the dashboard. The dashboard is what the agency will point at, in two years, when somebody asks why the SBA never built this in the first place. The answer will be that the agency now has a partnership with Palantir, the partnership has identified X billion in additional fraud, the agency has referred Y thousand cases for collection, and the agency takes fraud very seriously. That answer will be technically accurate. It will also not contain a single verb that touches the original failure.
Watch The Phrase "Real-Time"
If, six months from now, you see SBA materials describing a "real-time AI fraud screen" deployed at the application stage of a future emergency program, that is the moment to take the agency seriously. If the materials describe Palantir scoring of historical pandemic loans, ongoing case referrals, and a dashboard, the agency is still doing what it has always done. It is closing the door retroactively, then issuing a press release that calls the retroactive closing a forward-looking control. Those are not the same thing. They will never be the same thing. The phrase "real-time" is the difference, and so far it has not appeared anywhere in the SBA's public-facing language about this contract.
The Receipts, Filed Under L For Loop
So here is the closing receipt for the cyberpunk file. The agency designed a fraud-detection layer. The layer worked. The agency overrode the layer. The agency then waited until a politically aligned task force named the override out loud. The agency's response is not to publish the override directive. The agency's response is to hire an AI surveillance vendor to comb through the loans the agency overrode the layer to pay. The agency will then tell Congress and the press that the fraud is being handled, because the AI is now handling it. The AI is handling the dataset that the agency itself produced by ignoring its own fraud signal. The AI cannot un-pay the loans. The AI cannot retroactively staff the 2020 review process. The AI can only sort the wreckage. The agency will call this accountability.
It is not accountability. It is the next page of the same script. The SBA's pandemic loop now has four turns instead of three, and the fourth turn is the one where Palantir gets paid. The dashboard will glow. The numbers will tick up. The agency that approved the fraud will hold a quarterly briefing about the AI that found the fraud the agency approved. The receipts will continue to print. We will continue to file them. The loop, as ever, will close where it always closes, which is exactly nowhere near the desks of the people who decided to "protect" 562,000 loans the agency knew were rotten the moment they cleared.