A Tier-1 prime sends you a questionnaire. It asks whether your company, or any of your suppliers, uses Anthropic products, and sometimes whether you use AI at all. It wants an answer by a date, and the last line is a certification of accuracy that cites 18 U.S.C. 1001, the False Statements Act, above a signature block with your name on it.

Manufacturers I work with in New England have been receiving exactly these requests. Most of the advice they hear is about vendors: which AI tool is allowed, which one to switch to. That is the smaller question. The bigger risk sits on the signature line. A company that cannot say where AI runs inside its own walls cannot sign that certification honestly, whichever vendor it picks.

This is operational guidance from an operator, not legal advice. Have counsel review any certification before you sign it.

What is actually in force as of October 29, 2026?

Two separate authorities sit behind these requests, and the courts have treated them differently. One naming note first: the Department now calls itself the Department of War, and the FAR and DFARS still say Department of Defense. Same department.

10 U.S.C. 3252 is a defense-specific authority, carried into contracts through DFARS 252.239-7017 and 252.239-7018. The Department designated Anthropic a supply chain risk on March 5, 2026. On March 26 the Northern District of California preliminarily enjoined that designation, and on August 27 Judge Rita Lin ruled it unlawful and set it aside.

The Federal Acquisition Supply Chain Security Act (FASCSA), 41 U.S.C. 4713, is carried into contracts through FAR 52.204-28, 52.204-29 and 52.204-30. On September 25 the D.C. Circuit, in a 2-1 decision, upheld the Department's FASCSA exclusion of Anthropic. Anthropic petitioned for panel rehearing or rehearing en banc on October 13. No ruling on that petition had been reported as of October 29.

So two courts have reached opposite results under two different statutes, and the FASCSA order is the one reaching contract performance today. Rehearing could change that. The discipline below holds either way.

What does the restriction actually reach?

Three limits matter more than the headlines suggest.

Department of War contracts. FASCSA can apply government-wide, but as invoked against Anthropic it applies to the Department's contracts. Civilian federal work, commercial aerospace work and export orders sit outside it.

Use in contract performance. FAR 52.204-30 says contractors "shall not provide or use as part of the performance of the contract any covered article." That is a narrower question than "does anyone in your company use this tool." Holland & Knight reads unrelated commercial or back-office use as not necessarily prohibited on its own, while advising that you be able to document the separation.

Contracts that carry the clause. FAR 52.204-30 flows down to subcontracts, and under paragraph (b)(4) an order issued after award applies to a contract "only if added by an amendment to the solicitation or modification to the contract." Whether your purchase order carries the clause is a fact you can check this afternoon.

Why are the questionnaires so different?

Each prime writes its own. Mayer Brown documented the pattern in July: some requests ask whether any deliverable is "powered by or integrated with" Anthropic technology, some ask about sub-tier vendors and future adoption plans, and some are "arguably broader than the contractor's actual clause-based obligations, particularly where the underlying contract does not even contain the relevant FAR or DFARS supply-chain clause." Many end with a certification that expressly invokes 18 U.S.C. 1001.

A prime asking for more than your contract requires is managing its own exposure and pushing the uncertainty down a tier. That is a reasonable thing for a prime to do. The trouble starts when a supplier answers the broad question with a broad yes or no that it has not verified.

Why is the signature the real risk?

18 U.S.C. 1001 is a criminal statute. It reaches anyone who, in a matter within the jurisdiction of the executive, legislative or judicial branch of the federal government, "knowingly and willfully" makes "any materially false, fictitious, or fraudulent statement or representation." The penalty is a fine, imprisonment of up to five years, or both.

Keep it separate from the False Claims Act, 31 U.S.C. 3729. The False Claims Act is civil. It imposes treble damages and per-claim penalties for false claims for payment, and it is the statute behind the SPRS score cases covered elsewhere on this site. A false statement that is material to getting paid can raise exposure under both. They are different laws with different standards, and the certifications on these questionnaires name 1001.

Whether a particular statement to a particular prime falls under 1001 is a question for your counsel. A prime that prints the statute on the form is telling you how it intends to treat your answer. Treat the answer the same way.

Now picture an owner who signs "we do not use AI in the performance of this contract" because nobody in the front office uses it. Meanwhile, on the floor:

  • A manufacturing engineer pastes a failing machine macro into a free chatbot on his personal phone to find the bug.
  • A planner uses a consumer AI app to clean up the wording on traveler notes.
  • A quality technician asks an assistant to summarize a forty-page customer spec sheet before first article.

None of that appears on an IT asset list. Any of it could be use in contract performance, depending on the job and the tool. This is shadow AI: the gap between what the company believes it does and what it actually does. The certification asks about what it actually does.

The four-step answer

1. Inventory where AI is actually in use

Start on the floor, not in the IT ticket queue. Ask every role the same three questions: engineering, quality, planning, purchasing and the front office. Which AI tools do you use, on which device, and for what work? Make it safe to answer. The goal is an accurate map, and people report honestly when honesty carries no penalty.

Sort what you find into the three categories Mayer Brown recommends: direct use in performing a federal contract, indirect use in business systems that support that work, and purely commercial use. Record the company behind each tool, including tools that run another company's model underneath. Date the inventory.

Know which standard you are signing to. FAR 52.204-29 asks for a "reasonable inquiry," defined as an inquiry "designed to uncover any information in the entity's possession" about covered articles, and the definition expressly excludes the need for an internal or third-party audit. Some prime certifications ask for more than that, such as verified absence after an audit. Those are different promises.

2. Read which clauses are actually in each contract

First identify the source of each request: a contracting officer, a program office, or a prime acting on its own. Then pull the flowdowns and your purchase orders, job by job. Look for FAR 52.204-28, -29 and -30, and DFARS 252.239-7017 and -7018. Mark which jobs are Department of War work and which are civilian, commercial or export. A shop with one defense program and ten commercial customers has one program the order reaches, not eleven.

Write the result down as a table: contract, customer, clauses present, AI use found against that job. That table is the backbone of every answer you send.

3. Answer exactly what is asked and required, no more and no less

Answer the question on the form, scoped to the contracts it covers, and state that scope inside the answer. When a request reaches beyond the clause or the contract, for example company-wide use when the clause covers contract performance, say so in writing and ask for the scope to be confirmed. Holland & Knight recommends raising that kind of discrepancy with the contracting officer.

Under-answering is a risk. Over-answering is a risk too. Certifying "no AI anywhere in the company" after checking only the systems on one program is a statement you cannot support. The accurate answer is the narrow, documented one.

If a request forces you to pull a tool out of a defense job, record the technical impact, the added labor and the schedule effect. Mayer Brown notes that record may support an equitable adjustment.

4. A named human signs off

One person owns these responses. That person receives every request, keeps the inventory and the clause table current, and signs only what the evidence supports. Holland & Knight recommends the same structure: a single point of contact to receive, assess and coordinate every Anthropic-related certification response.

This is the Human-in-the-Loop pillar of Standard Work 2.0™ applied to compliance. AI can draft, sort and summarize. A verified human owns every affirmation that leaves the building. The discipline that keeps a person accountable for a first article inspection keeps a person accountable for a federal certification.

A sample answer structure

The outline below organizes a response. It is a structure, not language to sign. Your counsel writes the words.

  1. Scope. The request, who sent it, the date received, and the contracts or purchase orders the answer covers.
  2. Clauses reviewed. Which FAR and DFARS supply-chain clauses appear in each covered contract, and which do not.
  3. Method. What the inquiry covered (roles interviewed, systems and devices reviewed, sub-tier suppliers contacted) and the date it was completed.
  4. Findings by category. Use in performance of the covered contracts, use in supporting business systems, and commercial use outside the covered work, each stated plainly.
  5. Sub-tier status. What your suppliers reported, and which have not yet responded.
  6. Assumptions and limits. Anything you could not verify, stated as unverified.
  7. Owner and signatory. The named person, with title, who signs and who will update the answer if the facts or the law change.

Then put a date on the calendar to re-run the inventory. An answer is accurate on the day you sign it. Keeping it accurate is a process.

Governance that holds whichever way the courts rule

Swapping one AI vendor for another feels like action, and it leaves the actual problem in place. If rehearing changes the FASCSA result, the questionnaires will change. If a different vendor draws a designation next year, they will change again. A company that knows where AI runs, which clauses bind which jobs, and who signs for it can answer the next questionnaire in a day, whatever it asks.

That is the asset worth building: an operating discipline rather than a vendor list. It also earns standing with the prime. The supplier who returns a scoped, dated, documented answer, signed by a named owner, becomes the supplier the prime trusts with the next question.

If a certification request is on your desk this week, run step one before anyone signs, and bring your counsel in on the wording.

Sources