Can AI draft government proposals from past performance without compliance risk?

The solicitation is out, the past performance volume is the slowest part of every bid, and someone has a demo of a tool that writes it in an afternoon. The CFO's worry is not whether the draft reads well. It is what the company signs when it submits it.

The risk is misrepresentation, and the rules do not ask who typed

Here is how an AI draft goes wrong in a past performance volume, and every version reads smoothly. It blends two task orders into one citation with the larger one's dollar value. It rounds a CPARS rating up a notch. It writes a teammate's or a predecessor company's contract in the first person. It fills a gap in the record with a plausible sentence, because that is what a language model does with a gap.

None of those is an AI problem under the rules. 18 U.S.C. 1001 makes it an offense to knowingly and willfully make "any materially false, fictitious, or fraudulent statement or representation" in a matter within federal jurisdiction, with up to five years in prison. The False Claims Act reaches anyone who knowingly makes or uses "a false record or statement material to a false or fraudulent claim," at treble damages plus a civil penalty on each false claim, adjusted for inflation, and it defines knowingly to include deliberate ignorance and reckless disregard of the truth, with "no proof of specific intent to defraud" required. A proposal volume nobody checked against the records is hard to distinguish from the reckless-disregard case the statute describes, and the statute names no "the model wrote it" defense. That is our reading of the text, not legal advice.

Before any of that, there is the protest. In Watermark Risk Management International (March 17, 2026), GAO restated its standard: an offeror's "material misrepresentation in its proposal can provide a basis for disqualification of the proposal" where the agency relied upon it and it "likely had a significant impact on the evaluation." GAO rejected that ground because the solicitation did not ask for named staff and the agency gave them no credit. A past performance citation is written to earn credit.

The government scores relevance and candor, not prose

What the evaluator is told to look for is in FAR 15.305(a)(2), and it is not fluency. The solicitation "shall describe the approach for evaluating past performance" and shall let offerors identify similar contracts and report problems encountered and the corrective actions taken. The evaluator considers the currency and relevance of the information, its source, the context of the data and general trends in performance. Predecessor companies, key personnel and subcontractors performing major or critical aspects count when relevant. An offeror with no relevant record "may not be evaluated favorably or unfavorably on past performance."

Three consequences for a draft. It has to be organized by relevance to this requirement, not by which contract was biggest. It has to say what went wrong and what was done about it, because the solicitation asks; a model that writes only the flattering version has left out what the solicitation invites and the evaluator is told to weigh. And it has to be dated and sourced, because currency and source are evaluation factors.

Logged

Today, 2:14 pm

Asked by

Proposal manager

Question

Draft the past performance citation for the DoD logistics support task order, against Section L.4 of the solicitation

Sources

Costpoint, Manage Contracts, contract record and Performance Assessment tab (final CPARS, status Accepted, received 7 weeks ago)
Costpoint project record: period of performance, customer, funded value as of last month's close
Proposal library: the original technical volume and the last two monthly status reports

Answer

Draft of four paragraphs; each figure carries its record reference. Two statements flagged for review: no record found for "zero late deliverables," and the CPARS rates Quality as Satisfactory, not Very Good as the prior proposal said.

Reviewed

Director of Contracts, today, 3:40 pm. Both flagged statements corrected from the CPARS.

Approved as revised and kept with the proposal file

Illustrative entry, not a record from any client. The roles, the task order, the times and the two flagged statements are examples of what a draft log holds: which records the draft read, what it could not support, and who corrected it before submission.

The records are restricted, so the draft should point at them, not restate them

A past performance volume is built from three kinds of record. The CPARS evaluation is a government document. Under FAR 42.1503 it is rated on a five-point scale from exceptional to unsatisfactory, it may be used in source selection for three years after performance (six for construction and architect-engineer work), and while it can be used it "shall not be released to other than Government personnel and the contractor whose performance is being evaluated" and is to be marked "Source Selection Information." Your copy is restricted. A teammate's copy is theirs.

The contract record is yours and lives in Costpoint. The Costpoint 8.2 help describes a Performance Assessment tab on the Manage Contracts screen that documents assessments "such as those from the Contractor Performance Assessment Report System (CPARS)," with the assessed period of performance, the assessor, a status (Pending, Received, In Dispute, Accepted, or Accepted with reservation) and a score. The customer and the assessed period of performance sit on the same tab. That is the record a draft can cite.

The third kind is the proposal itself. FAR 3.104-1 defines "contractor bid or proposal information" to include cost or pricing data and "indirect costs and direct labor rates" submitted with a proposal and not made public. The pricing volume falls within the FAR's own definition of contractor bid or proposal information, and the statement of work may itself be CUI. So the data-handling question comes before the AI question: OpenAI's own help page says consumer ChatGPT (Free, Plus and Pro) may use your content for training by default, while its Enterprise and API tiers do not, and CUI is a separate test. We drew those lines in Is it safe to put CUI or contract documents into ChatGPT or Copilot? The client's data stays the client's. Nothing leaves their control.

There is no AI rule yet, so the solicitation is the rule

As of November 2, 2026, we found no FAR provision that prohibits AI in proposal preparation or requires an offeror to disclose it. GSA has proposed GSAR 552.239-7001, which would make Multiple Award Schedule contractors disclose the AI systems they use in contract performance within 30 days after award; it says nothing about preparing the offer. David Timm of Burr & Forman, writing in Washington Technology in July 2026, wrote that "only a handful" of solicitations that warn offerors generative AI may be used to evaluate bids, all of them DoD.

So the rule is Section L of the solicitation in front of you. If it asks for disclosure, disclose. If it restricts AI, follow the restriction. If it is silent, the content rules above apply in full, and the honest position is the one Bradley's August 2026 guide gives: no misleading statements about the extent of human involvement.

Deltek ships the in-system draft; the cross-system check is the work

Deltek's own tools cover a good part of this already. Deltek announced AI-Powered Proposal Outlines for GovWin IQ in June 2025. The Costpoint 2026.3 release notes list a CPARS Creation Agent that extracts up to 10 assessment factors from an uploaded CPARS PDF into the record above, and a Proposal Generation Agent that builds an outline from a solicitation, both opt-in for cloud customers. Deltek's November 2025 press release says its Deltek Proposals product "is designed to help reduce proposal creation time by as much as" 60 to 70 percent, "based on internal projections and AI automation benchmarks": a projection, not a result. We sorted the options in Deltek Dela, Copilot or a custom assistant: which does a contractor actually need?

None of the release notes above describes the check in the log: take a claim from a proposal library in SharePoint, find the contract and the CPARS it rests on in Costpoint, compare the two, and write down what it could not find. That is the part a protest would test. GovCon Cloud Moderate tenants have an extra step: as of October 2026 Deltek's undated Dela FAQ says Dela is not available in GCCM while Deltek's August 2025 Costpoint Dela/GenAI overview says it runs there on Azure GCC High, so ask your account team before planning around it.

What a defensible draft looks like

Every sentence that states a fact about a contract carries a record reference: the Costpoint contract or project record, the CPARS, the deliverable. The draft flags what it could not support instead of filling the gap; the flags are the most valuable thing it produces. A person who can read Section M compares each citation to the evaluation approach, problems and corrective actions included. The pricing volume comes from rates you already trust, which is why rate monitoring comes before proposals in What should a government contractor automate with AI first? And the log of what was drafted, from what, and who corrected it stays with the proposal file, the same habit we described for accounting data in Can AI read our Costpoint data without creating a DCAA problem?

AI drafts and flags; people approve and sign. Run it that way and the model's speed is a gain and its fluency is harmless, because nothing it writes reaches the government until a person has matched it to a record.

Common questions

Do we have to tell the agency that AI helped write the proposal?

As of November 2, 2026, no government-wide FAR rule prohibits AI in proposal preparation or requires its disclosure. GSA's proposed GSAR 552.239-7001 would require Multiple Award Schedule contractors to disclose AI systems used in contract performance within 30 days after award, which is a different question, and comment on its June 2026 revision closed August 3, 2026. Individual solicitations can and sometimes do address AI, so Section L of each one controls: if it asks for disclosure, disclose; if it restricts AI, follow the restriction.

Can we put our CPARS reports into ChatGPT to write the past performance volume?

Not into a consumer tier. FAR 42.1503(d) says a completed evaluation shall not be released to other than government personnel and the contractor being evaluated while it can be used in source selection, and tells agencies to mark it Source Selection Information. OpenAI's help page says Free, Plus and Pro ChatGPT may use your content for training by default; Enterprise and the API do not. Keep the evaluations in an environment whose signed terms rule out training and give you retention control, and if the solicitation or its attachments carry CUI, the tool has to meet the CUI rules as well.

What happens if the AI gets a past performance detail wrong and we submit it?

The same thing that happens when a person gets it wrong. GAO's standard, restated in Watermark Risk Management International on March 17, 2026, is that an offeror's material misrepresentation in its proposal can provide a basis for disqualification of the proposal where the agency relied upon it and it likely had a significant impact on the evaluation. The False Claims Act defines knowingly to include reckless disregard of the truth and requires no proof of specific intent to defraud, and 18 U.S.C. 1001 covers materially false statements in any matter within federal jurisdiction. A draft no one checked against the contract record and the CPARS is hard to distinguish from the reckless-disregard case the statute describes, so the check is the control. That is our reading, not legal advice.

Can the draft use a subcontractor's or a predecessor company's past performance?

FAR 15.305(a)(2)(iii) says the evaluation should take into account past performance of predecessor companies, key personnel with relevant experience and subcontractors performing major or critical aspects of the requirement, when relevant. The record is still theirs: a subcontractor's CPARS is released only to the government and to that subcontractor, so get the record and written consent from the teammate before the draft cites it, and label whose work it was. The draft must never present a teammate's or a predecessor's contract as the offeror's own.