# Speed Was Your AI Advantage. It Isn't Anymore. > Federal agencies now run AI on proposals too, and a court says its output belongs in the record. Why drafting speed stopped being an edge, and what still is. _By Oswald B. · Published 2026-10-19_ For two years, AI gave proposal teams one clear edge: speed. That edge is closing from both sides. More than half of A/E firms now draft with AI, and federal agencies have started running AI on the proposals they receive. When both sides use it, speed stops separating you. Specific, consistent evidence about your people and projects still does. This piece is about what changed in 2026, what it means for a firm that pursues public work, and where the advantage moved. Every source below is federal, and none of it comes from an A/E selection. We'll say where the evidence stops. ## What Did the Court Decide in TRAX v. United States? _TRAX International Corp. v. United States_, No. 26-796, is the first court ruling we know of on an agency's use of AI in a proposal evaluation. The U.S. Court of Federal Claims publicly reissued the opinion on September 22, 2026. The facts, as reported by [Washington Technology](https://www.washingtontechnology.com/contracts/2026/09/trax-loses-white-sands-protest-judge-hits-government-over-ai-disclosures/416334/) and a [Bradley Arant analysis](https://natlawreview.com/article/government-use-ai-evaluate-proposals-comes-under-scrutiny): - The Army competed a mission support contract at White Sands Missile Range worth about $450 million. - A procurement analyst tested an AI platform called FAST TRACK on the proposals. - The Army first said evaluators saw only one AI output, on screen, and judged it not usable. - After the judge ordered more declarations, the analyst said AI evaluations had been sent to the contracting officer and two members of the evaluation board. - TRAX alleged the tool invented weaknesses in its proposal. It lost the award to a team priced about $29 million higher, per [Defense One](https://www.defenseone.com/policy/2026/08/ai-army-contract-lawsuit/415261/). Judge Carolyn Lerner denied the protest. TRAX could not tie a specific evaluation error to the AI. But she ruled that the AI reports belong in the administrative record, the file a court reviews in a bid protest: > "Whether the SSEB [source selection evaluation board] ultimately found that output 'useable' or 'reliable' has no bearing on whether it was considered." She also called the AI evaluations "relevant to the process" and issued a sealed order asking the government to explain "the errors and misrepresentations in its briefing." In short, an agency can use AI on your proposal. If its people saw the output, that output is part of the record. ## Are Federal Agencies Using AI to Evaluate Proposals? Some are, and they've said so in writing. Here's what's on the public record as of October 2026. | Source | What it shows | What it doesn't show | | ---------------------------------------------------------------------------------------------------------------------------------------------------------------------- | ----------------------------------------------------------------------------------------------------------------------------------------------------------------------- | --------------------------------------- | | _TRAX v. United States_ (Army, Court of Federal Claims) | An AI tool reviewed competing proposals, and its output reached the contracting officer and evaluators | That the AI changed the award | | GSA's CALI tool ([GSA AI use case inventory](https://www.gsa.gov/technology/government-it-initiatives/artificial-intelligence/2025-gsa-ai-use-case-inventory)) | A machine-learning tool that checks proposals against the solicitation for format, forms, reps and certs, and requirements | Use on A/E or SF330 selections | | SOCOM IDIQ solicitation H9240026RE001 ([Burr & Forman](https://www.burr.com/government-contracting/should-the-government-use-ai-tools-to-evaluate-contract-proposals)) | The agency "may employ AI as a tool to assist in the analysis and review of offeror proposals," including "highlighting areas of potential strength, weakness, or risk" | How often the disclosure clause appears | None of these is an A/E selection under the Brooks Act. TRAX was mission support. CALI is being trained on Multiple Award Schedule data. We found no public evidence that a state DOT or a city scores SOQs with AI, and this piece doesn't claim one does. What the record does show is direction. The tools exist, agencies are writing them into solicitations, and a court has said their output is reviewable. A firm that pursues federal work should assume the first reader of a future submittal may be software. ## Why Doesn't AI Drafting Speed Give You an Edge Anymore? AI drafting speed no longer sets a firm apart, because most of its competitors have it too. When every firm on the shortlist drafts in minutes, the deadline and the page limit stay the same, and the evaluator still picks one firm. Speed was a real advantage when few firms had it. A team that could turn a project description around in ten minutes instead of an hour could chase more pursuits with the same people. That window is closing. A 2026 industry study of about 700 North American A/E firms found AI use jumped from 38% to 53% in a single year. Over the same stretch, industry hit rates held near their long-running level. We covered the gap in [53% of A/E Firms Use AI for Proposals. Win Rates Didn't Move.](/blog/ai-adoption-aec-proposals-win-rates) An edge that every competitor can buy for a monthly fee gets competed away. When your rivals also draft in ten minutes, the shortlist doesn't reward the faster draft. It resets the baseline. Speed still matters inside your firm. It frees hours for go/no-go decisions and for pursuits you'd otherwise skip. We made that case in [AI Didn't Double Your Proposal Capacity](/blog/ai-proposal-capacity-myth). But speed is now an operating cost, like having a scanner. It no longer separates you on the shortlist. ## What Does an AI Proposal Screen Check First? GSA describes CALI's checks in four areas. They are a useful preview of what any machine screen is likely to look at first: 1. **Format compliance.** Page limits, font sizes, section order, file type. 2. **Forms.** The required forms are present and filled in. 3. **Reps and certs.** The representations and certifications are complete. 4. **Requirements.** Each requirement in the solicitation has an answer somewhere in the proposal. A screen like this checks whether you answered what was asked, in the form it was asked. Whether your approach is any good is a question for a person, later. The screen is close to pass/fail, and it runs before anyone reads your project approach. Human evaluators already do this check, and many firms already build a [compliance matrix](/blog/compliance-matrix-soq-sf330) to pass it. What changes with software is consistency. A tired evaluator might let a missing form slide. Software applies the same check to every proposal. A requirement answered in the wrong section may simply read as unanswered. ## What Still Sets a Proposal Apart When Both Sides Use AI? If speed is table stakes and compliance is pass/fail, the edge moves to evidence, which AI can't make up. Evidence that holds up has three qualities. **It's specific.** A named engineer, a named project, a dollar value, a completion date, a client who will take a reference call. "Extensive experience with bridge rehabilitation" is what every AI-drafted proposal says. A line like "Led the deck replacement on the Route 9 bridge, completed in 2024, $14 million construction cost" is something only one firm can say. We made the human-evaluator version of this point in [Evaluators Can Tell When AI Wrote Your Proposal](/blog/ai-proposals-sound-the-same). A machine reader lands in the same place: generic text gives it nothing to match against the requirement. **It's consistent.** A machine is good at comparing one part of a document to another. If the org chart lists someone as project manager and their resume lists them as deputy, a person might miss it at 11 p.m. Software built to compare the two is less likely to. The same goes for project dates, contract values and license numbers that differ between the project sheet and the resume. **It's retrievable.** If your evidence lives in last year's submittal, a former employee's folder and three versions of a resume, your team spends the week finding it. Finding it is where the hours go. This is also where the kind of AI your firm uses matters. AI that drafts from your firm's own records can't promote a junior engineer to design lead, because the record says otherwise. A general chat tool can. That kind of mismatch is exactly what a careful evaluator, or a compliance tool, looks for. RFPM.ai is built on that idea: it drafts and searches from the firm's own staff and project records, and a person still picks who and what goes in. ## How Does TRAX Change Debriefings and Protests? TRAX shows the risk runs both ways. An agency's AI can misread your proposal. TRAX alleged exactly that, saying one weakness rested on "made-up references" to its proposal. The ruling means that output, if evaluators saw it, can now be reviewed. That gives you new questions to ask after a loss. The Bradley Arant analysis suggests asking in a debriefing: - Did any AI tool produce a report on our proposal? - When was it generated? - Who on the evaluation team saw it? You won't always get an answer, and the ruling doesn't require an agency to hand over every AI document. It covers what decision-makers considered. But after TRAX, the question is fair to ask. Write down whatever answer you get. For A/E work, there's a catch. The debriefing is where you'd ask, and the FAR overhaul's proposed Part 36 rule would change A/E debriefings from "will be held" to "may be held." We covered that change in our [FAR Part 36 proposed rule breakdown](/blog/far-part-36-proposed-rule), and what a federal A/E debriefing covers today in [our debriefing guide](/blog/federal-ae-debriefing-far-36607). The comment period on that rule closed October 19, 2026. ## What Should A/E Firms Do About AI Evaluation This Quarter? You can do all of this with the tools you have. The work is making the evidence your firm already has specific, consistent and easy to find. 1. **Run a consistency pass before every federal submittal.** Check names, titles, dates, contract values and license numbers across resumes, project sheets and the org chart. Do it the way a machine would: one field at a time. 2. **Answer every requirement where the solicitation asks for it.** Build the compliance matrix first. Put each answer in the section the solicitation names, with its heading. 3. **Replace adjectives with facts.** For each key person and each featured project, list the three facts a reviewer could check. If a sentence would fit any firm's proposal, cut it or make it yours. 4. **Fix the source, not the submittal.** When you correct a resume for a pursuit, correct the master record too. Otherwise the same error comes back on the next pursuit. 5. **Add the AI questions to your debriefing list.** Ask them on federal pursuits, and write down the answers. 6. **Keep the scope in mind.** Today's evidence is federal. If you mostly pursue state DOT and municipal work, the consistency pass still pays off, because human evaluators catch the same errors. You just aren't facing a documented machine screen there yet. ## Frequently Asked Questions ### Do federal agencies use AI to evaluate proposals? Some do. In _TRAX v. United States_, the Army ran an AI tool on competing proposals and sent its output to the contracting officer and two evaluators. GSA lists a proposal compliance tool called CALI in its AI use case inventory. At least one SOCOM solicitation says the agency may use AI to review proposals. ### What did the court rule in TRAX International v. United States? The Court of Federal Claims ruled that AI evaluations belong in the administrative record when agency decision-makers considered them, whether or not they relied on them. TRAX still lost its protest, because it could not tie a specific evaluation error to the AI tool. The opinion was publicly reissued September 22, 2026. ### Do state DOTs use AI to score SOQs? We found no public evidence of it as of October 2026. The documented cases are federal: the Army, GSA and SOCOM. None of them involved an A/E selection. That could change, so watch solicitations for disclosure language about AI-assisted review. ### Does using AI to write a proposal hurt your chances? Not by itself. The risk is what AI tends to produce: generic text that sounds like every other firm's, and confident details that don't match your records. Specific facts, consistent across every document in the submittal, are what hold up whether a person or software reads it first. ### What can I ask in a debriefing about AI evaluation? Ask whether any AI tool produced a report on your proposal, when it was generated, and who on the evaluation team saw it. After _TRAX_, those reports can belong in the record if evaluators considered them. An agency may not answer fully, but the question is a fair one. --- Canonical HTML: https://rfpm.ai/blog/ai-proposal-speed-advantage · Site overview: https://rfpm.ai/llms.txt · Sitemap: https://rfpm.ai/sitemap.xml