By Chuck Gallagher — Business Ethics Keynote Speaker and Trainer
TL;DR: I’m an AI ethics speaker and author, and here’s my read: the Air Force’s compressed deadline to purge Anthropic products from contractor systems doesn’t remove risk so much as move it — onto companies that had no seat at the table and no way to verify what they’re being asked to certify.
Tonight, a compliance officer at a defense contractor is staring at an online form.
The form asks a simple question. Does your company use Anthropic products or services anywhere in its systems? Report by August 1. Have it all out by September 1.
She doesn’t know the answer. That’s no knock on her. Nobody knows.
Why Can’t a Contractor Just Answer the Question?
On July 9, the Air Force Research Laboratory sent industry a memo formally notifying contractors of a Department-wide requirement for the removal of all products and services provided by Anthropic. Find every use of Anthropic anywhere in your systems. Report by August 1. Have it gone by September 1, almost a month ahead of the Department of War’s own September 29 deadline. Breaking Defense obtained the memo. AFRL called the earlier date a buffer for administrative processing and contract modifications.
The scope language is where this gets hard. The memo reaches products and services provided by Anthropic and its subsidiaries and affiliated entities. That’s not a list you pull out of a drawer. It’s an archaeology dig. Vendors embed other vendors, and a tool somebody bought back in 2024 may reach a model through a middleware partner who resells through somebody else. The company sitting at the end of that chain has no contractual right to look inside. None. The deadline doesn’t care.
Ban first. Map later. That order’s backwards. As an AI ethics speaker and author, let me say plainly what I’ve watched play out in one industry after another: ban before you map and you don’t eliminate risk. You relocate it. What was supply-chain risk in Washington becomes operational failure risk, and it lands in the lap of a contractor who never had a seat at the table.
How Did a Contract Dispute Turn Into a Purge?
On February 24, Defense Secretary Pete Hegseth demanded that Anthropic drop its usage restrictions and let the military use Claude for all lawful purposes. The company held two lines. No fully autonomous weapons without a human in the loop. No mass surveillance of Americans. It refused. On February 27, President Trump directed federal agencies to stop using the company’s technology, and Hegseth moved to designate Anthropic a supply chain risk to national security — a label reserved for foreign adversaries and never once hung on an American company.
Then the courts split clean in two. On March 26, U.S. District Judge Rita Lin found Anthropic likely to succeed on claims that the designation was First Amendment retaliation for the company’s safety advocacy, violated Fifth Amendment due process, and was arbitrary and capricious under the Administrative Procedure Act. She blocked the presidential directive, the Hegseth directive, and the designation under one statute. On April 8, a D.C. Circuit panel declined to pause a parallel designation under a second statute, saying the equities favored the government. Two statutes. Two courts. Two answers.
So contractors are being told to strip out a vendor whose ban one federal judge has already called likely unlawful, on a clock that keeps sliding, under a scope nobody can fully inventory. Here’s the part that ought to stop you. Breaking Defense also reported that the Pentagon reissued its removal memo on April 2, nearly word for word, with one thing missing — the phrase calling Anthropic an unacceptable supply chain risk. The stated reason quietly left the building. The order stayed.
What Does an Impossible Deadline Actually Produce?
Here’s where I stop commenting on defense policy and start talking about what I actually teach. White-collar crime doesn’t begin with bad people. Never has. It begins with three ingredients. Pressure. Opportunity. Rationalization.
Now look at what this memo hands a contractor. Pressure — a hard date, a contract modification riding on it, and a program manager asking about it every single morning. Opportunity — a self-report form nobody can audit, covering dependencies nobody can see. Rationalization — we’re substantially compliant, the prime would’ve told us, that’s a partner’s stack and not ours.
That’s the fraud triangle, assembled by memo. A self-report to the government isn’t a survey. It’s a representation. Under the False Claims Act, we did our best won’t get you far. We said yes because saying no would have cost us the contract — well, that’s a case. And nobody in that building woke up intending to commit fraud. That’s not how any of this works. Fraud is what happens when decent people get handed a deadline they can’t meet and a form they can’t verify.
Is the Pentagon Wrong to Want Out?
Let me be clear, because this is where most commentary goes soft. The Department is under no obligation to buy from a vendor who won’t accept its terms. Emil Michael, who ran the negotiations, argues the uses Anthropic feared are already barred by law and standing Pentagon policy, and that the company took a contract dispute and turned it into a public campaign. Judge Lin didn’t order the government to keep using Claude. She questioned how it went about the exit.
That distinction is the whole ballgame. You can be dead right about the destination and downright wrong about the road. Anger isn’t a plan. And the folks who pay for a decision made in anger are almost never the folks who made it.
What Should a Contractor Do Right Now?
Don’t guess on the form. As an AI ethics speaker and author, the counsel I give boards applies here without much translation: build the inventory you should have had before the memo ever showed up. Keep a software bill of materials and a model inventory naming every AI dependency you’ve got, direct and inherited. Write down your approved use cases. Keep a tested swap-out path for every model you rely on, so that losing a vendor is an inconvenience instead of an outage. That discipline — the kind formalized in AI management system standards such as ISO/IEC 42001 — turns a political shock into an ordinary Tuesday. It isn’t rocket science, but it is work, and most organizations keep putting it off.
And when you can’t answer, say so. In writing. With the date you asked and the name of whoever didn’t answer you. A documented unknown is defensible. A confident no you can’t prove is a liability with a statute attached to it.
Every choice has a consequence. The choice here got made in Washington back in February. The consequence lands in August, on a compliance officer who never met anyone in this fight.
Frequently Asked Questions
What did the Air Force tell contractors to do about Anthropic?
Find it all and rip it out. A July 9, 2026 memo from the Air Force Research Laboratory notified contractors of a Department of War requirement to remove all products and services provided by Anthropic. Identify every use of Anthropic in your systems, report by August 1, finish the removal by September 1. That last date runs almost a month ahead of the Department-wide September 29 deadline. AFRL called it a buffer for administrative processing and contract modifications.
Why is removing an AI vendor from contractor systems so difficult?
Because software vendors embed other vendors. A contractor may be running a scheduling, coding, or analytics tool that quietly calls an AI model through a reseller or a middleware partner, with no contractual right to open the hood and look. The memo’s scope reaches Anthropic’s subsidiaries and affiliated entities, which widens the search without making it one bit easier to complete. Seven weeks. Finding every dependency is an inventory problem most organizations have never solved even on a normal schedule.
Is the ban on Anthropic still in effect?
The legal picture is split, and that’s the honest answer. On March 26, 2026, a federal judge in California blocked the presidential directive, the Hegseth directive, and the supply chain designation issued under one statute, finding the government’s action likely retaliatory and procedurally defective. On April 8, the D.C. Circuit declined to pause a parallel designation issued under a second statute. Removal deadlines are moving forward on that second track while the appeals grind on. Confirm current status with your counsel. Don’t hang your company on any single ruling.
What is the legal risk of self-reporting AI use to the government?
A self-report is a representation. The government reads it that way. So will a prosecutor. If a company certifies that it has removed a prohibited product and that certification turns out to be wrong, exposure under the False Claims Act doesn’t evaporate because the deadline was unreasonable. As an AI ethics speaker and author, I tell clients the same thing every time: a documented unknown, with dates and names attached to it, is far safer than a confident answer you can’t substantiate.
How can a company protect itself if an AI vendor gets banned overnight?
Keep a software bill of materials and a model inventory that lists every AI dependency you have, including the ones you inherited through partners. Document your approved use cases. Maintain a tested swap-out path for each model, so that losing a vendor becomes a scheduling problem rather than a mission failure. Management system standards such as ISO/IEC 42001 exist to make all of this routine. Here’s the thing, though. The companies handling this deadline calmly are the ones that built the inventory before they needed it.
Before You Go
Here’s the part nobody puts in the memo. Every organization eventually gets handed a decision it didn’t make and can’t appeal, and the only thing left to decide is what it does under pressure. That moment isn’t a compliance problem. It’s a character problem wearing a compliance costume, and boards tend to find out which one they’ve got on the day the deadline lands. Trust me on that one. I speak to leadership teams, boards, and industry audiences about exactly this — how ordinary people under pressure make choices they never planned to make, and how you build the discipline that holds when the clock is unreasonable. If your organization is carrying an AI dependency you haven’t mapped, or a deadline you’re not sure you can meet honestly, that’s the conversation to have now instead of in August. Bring me in at ChuckGallagher.com.
Five Questions for Reflection
1. If your CEO asked you today whether your company uses Anthropic anywhere in its systems, how long would it take to give an answer you would be willing to certify in writing?
2. When a deadline you did not set turns out to be impossible, where does the pressure in your organization actually land — and does the person who set it ever hear about it?
3. Pressure, opportunity, and rationalization are the three ingredients of white-collar crime. Which of the three does your current compliance process quietly supply?
4. Is your organization’s answer to an unanswerable question a documented “we don’t know yet,” or a confident guess that protects the contract?
5. The Department was free to walk away from a vendor. The question was how. Where in your own leadership do you confuse being right about the destination with being right about the road?
