Allison Ritz is an experienced leader in the technology sector, with a strong background supporting some of the world’s largest government contracting organizations. A certified APMP member, Allison is a recognized thought leader and frequent speaker at GovCon events, sharing insights on the power of strategic connections. Her passions lie in driving success by helping organizations realize the value of meaningful relationships. With a firm belief in the importance of planning, process, and thoughtful execution, Allison excels at making SaaS solutions personal, fostering long-lasting, profitable client relationships.
- On-Demand Webinar
When the First Reader Is a Machine: Writing Proposals That Survive AI Evaluation
This session covers how to structure proposals for AI evaluation tools that check compliance, consistency, and traceability before human reviewers score them.
- November 12, 2026 , 11.00 am
- ET
Register for this Webinar
In this webinar we will discuss
AI tools are reading your proposals before humans do, and they don’t infer, skim, or give partial credit.
Evaluation is changing. Before a human on the SSEB reads a word of your volume, there is a growing chance a model has already parsed it, mapped it to Section L and M, and flagged gaps. That reader matches language, structure, and traceability. It doesn’t reward a strong executive summary if your requirement responses bury compliance in prose.
This session covers what machine evaluation actually does well, where it falls short, and what both mean for how you write requirement responses, build compliance matrices, manage terminology, and sequence sections. The goal is deterministic writing choices that hold up under both readers.
What you will learn
- Mirror, don’t paraphrase: Why requirement language from the solicitation should appear verbatim in your response, and how restating it in your own words creates traceability gaps machines will flag as missing.
- Structure as a scoring signal: How headings mapped to Section L, consistent numbering, and one-requirement-per-response blocks make your proposal parseable by both automated tools and human evaluators.
- Readability as a hard requirement: Why long sentences, passive voice, and undefined acronyms degrade model extraction and evaluator comprehension, and how to set and enforce readability thresholds before final submission.
- Cross-volume consistency checks: How conflicting staffing numbers, dates, or terminology between technical and management volumes get caught at scale, and why a terminology and figures crosswalk belongs in your process before final review.
- Clearing the first gate: How writing for machine evaluation is about getting your discriminators through cleanly to the human who actually scores them.
Who should attend
Proposal managers, writers, and volume leads who want to understand how AI evaluation tools assess compliance and consistency, and who need practical, structural techniques to improve how proposals perform before human review begins.
Speakers
Allison Ritz
Director of Product Marketing
VisibleThread
Allison Ritz
Director of Product Marketing
VisibleThread
Webinar Transcript
Hello, everyone, and thank you for being here. If you’re new to our office hour sessions, welcome. If you’re back, thank you for joining us again. Today we’re going to talk about machine readers and AI evaluators. I know this is a hot topic right now, and it’s a question I get asked a lot. What happens to your proposal if part of the evaluation is done with help from an AI tool? Does that change our approach? And what should we know? Going into the process, we don’t know whether AI will be involved or whether it will be human-only evaluators, so we have to create a proposal that can stand up to both. We’ll spend a few minutes on what we actually know, where the conversation around AI evaluation is right now, and what we don’t know yet. We can talk about trends, but I really want to nail down what we’re seeing today and what you can do to prepare. Then we’ll walk through an example. In these sessions, I like to take one requirement and follow it from the solicitation to a response that’s ready for review. I’ve created a mock requirement that we’ll follow through that process. Then we’ll cover five things you need to do, along with a short audit you can run before submission. We’ll hopefully have a few minutes for Q&A at the end. As always, if we don’t get to your question, I’ll make sure to follow up. Drop your questions in the chat as we go, and we’ll address them at the end. A quick plug: this session is part of our office hours series. We meet every other Thursday, it’s free, and there are replays. We cover a lot of topics that are top of mind for you right now. Sometimes we talk about VisibleThread, but often, like today, the sessions are neutral. We won’t get into the platform today; we’ll focus on takeaways and share some resources at the end. On screen are some of our recent sessions available on replay. In the console, you’ll also find resources including an AI guide, a buyer’s guide, and links to our upcoming webinars. We have a lot of free educational content on the website, and that’s a great place to start. Now that the housekeeping is done, let’s jump in. Chapter Understanding AI Assisted Evaluation What we know about AI-assisted evaluation Here’s the crux of today’s session: your proposal may have two readers. On some evaluations, one of those readers could be a machine, and you won’t know which tool it is or how it’s been set up. That’s an added challenge, because we don’t know how smart or comprehensive that AI tool might be. These tools can work very differently. Some are rules-based, looking for specific words, numbers and sections. Some use semantic search, so they can find a passage that means the same thing even if you use different words. And some use GenAI to interpret, summarize and suggest strengths and weaknesses. Those are very different approaches, but all of them can miss things, misread things, or put things in the wrong bucket. Even the smarter tools aren’t immune; they just fail in slightly different ways. So I’m not going to give you hard-and-fast rules for writing to one specific kind of machine. The practical approach is to write responses that are clear, explicit and traceable for all of them. Whether the reader is an AI tool or a human evaluator, we want to reduce ambiguity and make sure whoever is scoring you can find all of the information. Let’s separate what’s confirmed from the hype and the fear. I’ll point out three things. Chapter Current Trends And Legal Precedents First, solicitations are starting to disclose that AI may assist evaluation. In a US Special Operations Command solicitation this year, offerors were told up front that AI may help summarize proposals, identify compliance, and highlight strengths, weaknesses and risks, but that final judgments would stay with government evaluators. Be careful how you describe this internally. A solicitation saying AI may be used tells us the agency gave itself permission. It doesn’t tell us whether or how it will actually be used in the evaluation. So prepare for automated analysis, but don’t assume every agency does it the same way, or that all of them are doing it at all. Second, and this is where we have more documented use: AI-generated evaluations have shown up in a bid protest. In Tracks International v. United States, an Army analyst used a tool called Fast Track to generate evaluations of all three proposals as a test. The court ordered those outputs into the administrative record, but it denied the protest. The case showed no challenged evaluation error attributable to AI; the AI use didn’t undermine the evaluators’ findings. That’s an important distinction. Evidence that AI was involved is not the same as evidence that AI improperly shaped the award. Third, people still make the call. AI is not making the award decision. The source selection decision has to reflect independent human judgment. As always, compliance alone won’t win. Your approach, evidence and discriminators still decide it, and the final award decision still sits with human evaluators. So what can a tool check, and what can’t it verify? It varies depending on the tool, how it’s configured, the quality of what you feed it, and what it’s been asked to look at, just like any other application of this technology. Chapter Capabilities And Limitations Of Tools Tools can often help with locating the likely response to a requirement; comparing terms, numbers and dates; checking format; flagging conflicts between volumes; and summarizing a long volume quickly. We’re seeing a lot of summary and overview use. The part I think gets missed is what no tool can independently verify. A generative tool can potentially spot an exception or a dependency in your text. What it can’t do on its own is check your claims against the real world. It can find that you claimed a sixty-day transition, but it can’t verify that you have the people and maturity to deliver it, that the commitment is achievable as written, or that the result you cite really happened. It can’t judge whether your evidence is credible and relevant, and it doesn’t decide who wins. Here’s the principle for the rest of the session: make the evidence of compliance explicit enough for any reader, human or machine, to verify. Chapter Three Levels Of Response Quality Three levels of response quality Let’s frame the discussion with three levels of response quality, using the requirement we’ll work through today. Level one is findable. It matches the requirement word for word, and any tool will find it, but it doesn’t show how you’ll actually do the work. At level two, an evaluator can examine how the commitment gets delivered. Notice there’s no evidence in this statement yet. It explains the approach; substantiating that approach is what we add at level three. Level three is the competitive level. Here the response is backed by credible evidence, and that’s the kind of statement that may support a strength under Section M. Evaluators still decide whether it does; they have the final say. One important takeaway from today: machine-readable doesn’t mean compliant, and compliant doesn’t mean competitive. AI tools entering the evaluation process doesn’t reduce the work we do to make our proposals compliant, compelling and defensible. Mirroring the language gets us to level one. Everything else we cover today is about getting to levels two and three. Chapter Mirroring Requirements And Providing Proof 1. Mirror the requirement, then prove it Let’s start on the five items I mentioned at the top. First: mirror the requirement, then prove it. We’ll follow one requirement through the rest of the session. It’s fictional, but it should look familiar. Section L asks the offeror to describe its approach to transitioning service desk operations within sixty calendar days of award, including knowledge transfer, staffing ramp-up, and the transition plan. Section M says the government will evaluate how well the approach minimizes disruption to end users. The PWS requires full operational capacity within sixty calendar days. On the right of the slide are the seven things any reader, human or not, needs to be able to locate to confirm we’ve satisfied this requirement. Luckily, the solicitation already gives you the words for every one of them. That language matters. If a tool matches exact words, those words are how you get found. If it matches meaning, using them still removes any doubt, so the tool doesn’t have to infer meaning from different terminology. Either way, there’s no downside to mirroring. Here’s the same first paragraph written two ways. On the left is a paraphrased version. The heading is “Getting up and running,” and the text says the team will ramp up over the first two months, bring new staff on board, learn the environment, and users will see no disruption. A general reader might think that’s fine. But a strict reader could flag that the heading doesn’t trace to L.4.3, and two months isn’t the same as sixty calendar days. In a transition, that difference can matter. There’s no mention of knowledge transfer, no CDRL, no full operational capacity, and the words “service desk” never appear. A smart semantic tool might connect some of that, and an evaluator might too. But you’re asking every reader, human or machine, to do the translation for you. On the right is the mirrored version. The heading carries the reference. The first sentence restates both the transition and full operational capacity with the exact time frame and PWS reference. The second names all three required elements, and the third uses Section M’s language and starts the argument. That’s far more complete, regardless of who or what is assessing the response. What does mirroring look like in practice? First, mirror exactly: the key nouns and verbs of the obligation, numbers and time frames, named deliverables, CDRLs, Section L, M and PWS references, and the customer’s own word for the thing. If the PWS says “service desk,” it’s “service desk,” and we’ll follow that through later. Then add what turns compliance into a score: how you’ll actually do it, evidence a reviewer can check, and what it means for the customer and why it’s lower risk. Avoid keyword stuffing. That means pasting a whole Section L paragraph back with “we will comply” and nothing behind it, or repeating requirement terms over and over. In anticipation of AI evaluators, we’re seeing more of this, much like the advice people give about stuffing resumes with keywords or structuring things in an overly simple way. That’s not the answer. Human evaluators will notice, and if they’re using a generative tool to summarize the section, it won’t find any substance either. The rule to give writers: requirement language in the first sentence, then your own language after that. Mirroring helps readers find the answer, but evidence is what makes it compelling and helps evaluators score it. Chapter Structuring For Proposal Traceability 2. Structure for traceability Our second point is structure. What helps a tool parse your proposal is the same thing that helps an evaluator find answers. Here’s a response pattern that can be standardized and is worth considering: a heading that mirrors the requirement and carries the references; an answer-first sentence that restates the obligation as a commitment with exact numbers or figures; the approach, in the order the solicitation lists the elements; evidence a reviewer can check, stated in the text (a comparable contract, its scope, the result, and who can confirm it); and a benefit line that closes in Section M’s words. Real solicitations are messy. I can show perfect examples here, but you need to be able to use this in the heat of battle, so keep three notes in mind. Follow the prescribed organization first. Usually that’s Section L, but Section L and the PWS often don’t line up neatly. When they don’t, use cross-references so every requirement has a place a reader can find it. Be careful when grouping related items. You don’t need a separate block for every line if the solicitation lets you address requirements together; it depends on your instructions. What matters is that each one is visibly answered. When three requirements get blended into one flowing narrative, readers can lose track, and that’s absolutely an area where AI evaluators can miss things. Use real heading styles in Word. A bold line of body text may look like a heading to you, but to a document parser it may just be part of a paragraph. Five or ten minutes of template work can make sure nothing gets lost in the shuffle. Next, the compliance matrix. Most teams have one, especially on larger pursuits. The question is whether it’s detailed enough to check against. The rows on the slide are excerpts so they fit, but in a working matrix you want to quote the full requirement text. The moment someone summarizes the requirement in the matrix, writers start answering the summary, and you get a game of telephone where one person’s summary becomes another’s and important context is lost. Use one row per discrete, evaluable obligation, not one row per paragraph, and not only the “will,” “shall” and “must” statements. Our L.4.3 example alone is several rows. Track section and page, update it at every review gate, and mark gaps as gaps. In our last row, the commitment is there but there’s no evidence behind it yet. That’s a useful matrix: it tells you where to spend time before red team and shows you the gaps. The other piece is file hygiene. If your final PDF is a scan, or a key claim lives inside an image, a tool may not read it at all. Keep claims in text, use real tables, and caption graphics with the requirement they support so the traceability is there. Here’s a quick test: open your PDF and try to select the text on a few key pages, including inside tables and graphics. If you can’t select it, it’s safe to assume a tool can’t read it either. If text that satisfies a requirement is baked into an image, you may want to reconsider. Chapter Readability As A Risk Control 3. Readability as risk control Our next point is readability, which we’re treating as a risk control measure. These are internal quality targets that reduce ambiguity, not a scoring standard any agency has published. I’m not presenting this as what agencies score against. It’s a risk-reduction tool you can use within your organization. There are four readability signals, each with a starting target. To be clear, there’s no federal rule and no published AI evaluation standard right now. Hopefully, as we move forward, we’ll get more structure around where AI sits in the process and what it evaluates. Nothing currently says a sentence has to be under 20 words or structured a certain way. But following a consistent standard limits your risk when AI is part of the evaluation. Sentence length. Aim for an average under 20 words, an industry best practice, and review anything over 30. Long sentences tend to bundle several commitments or requirements together, and any reader, person or tool, can walk away with one and miss the other. Passive voice. Ideally under 10%. “Knowledge transfer will be performed” doesn’t say who’s responsible or who does the work, and that’s often exactly what evaluators are looking for. Limit passive voice where appropriate. Acronyms. The hard-and-fast rule: define them at first use in each volume, since volumes are often read by different people and can be evaluated separately by different tools. Keep undefined acronyms out of headings, which can also cause problems for parsing tools. Grade level. Set a range per volume. An executive summary and a technical approach won’t read the same and aren’t held to the same standard, but every writer should be working toward a shared grade-level and accessibility target. What you’re really managing is the cognitive load it takes to absorb your information. The numbers are variable. These are examples, and you don’t have to take mine. You just need to agree on a set at kickoff so everyone works to the same baseline through the review stages. What do these targets catch in a real sentence? I won’t read the whole “before” example, but it’s one 44-word sentence in passive voice with undefined acronyms. We’ve all seen real versions of it. The “after” reads: “Our transition management team will run knowledge transfer with the incumbent during the sixty-day transition. We’ll take over the service desk one support tier at a time as we reach the staffing levels in our staffing ramp plan.” Two sentences, active voice, clear about who does what. The second version also mirrors much better: “service desk” is back, using the customer’s language. Readability and traceability fixes often turn out to be the same fix. Being more direct for readability and accessibility frequently improves traceability too. Chapter Ensuring Consistency Across Volumes 4. Consistency across volumes Our fourth point is consistency across volumes. When a tool is set up to compare volumes, it can do this quickly, and it’s something manual reviews don’t always do as thoroughly as they should. Just as in our everyday work, AI can do certain things very quickly that take us a long time. So expect that some evaluations may include a more comprehensive review of traceability, language and requirements identification, because that work is now fast. The things tools do well will be the things that get expanded first. I’ve called out four conflicts that are probably sitting in a proposal right now. The figures are just examples, but you’ll likely recognize the pattern. On staffing, we have 42 FTEs in the technical volume, 45 in management, and 44 in price. On transition, technical says sixty calendar days, which matches the PWS; management says ninety days; and price just says sixty days. The program lead is a program manager, a program director, or a project manager, depending on the volume. And the function is a service desk, a help desk, or an IT service center. Each volume may have been reviewed, but not against the others. To an evaluator, that raises real questions. Which number is real? If it’s 45, have we understood the price? Is ninety days a schedule risk? At best, it reads like a communication issue. At worst, it becomes a weakness or a clarification request. An AI evaluator can catch these very quickly. A human evaluator, depending on time and attention, might let it slide or weigh it less, but it will pop immediately if AI is being used to summarize and pull out inconsistencies. The fix is a one-page terminology and figures crosswalk. Every number or term that appears in more than one volume gets a row with a locked value, its source of truth, and an owner. Staffing comes from the current staffing model. The transition period comes from the PWS. The program lead title comes from the key personnel list, and the function name comes from the PWS. Add a row for banned variants, so writers know, in this case, that “help desk” is out. In VisibleThread land, we talk about this a lot as watchwords: using watchwords as an automated check in your process, before anything goes out, to confirm the language is correct and sourced from the right places. How do you put that into practice? Start at kickoff with what you can lock from the solicitation, in terms of both language and traceability to the source. When something changes, update it immediately, whether that’s an amendment, a language change, or a change in your own response. Before reviews, search every volume for each value and for the banned variants. Then search again before final, because price is usually the last thing to move and late-stage changes happen. This matters because inconsistency is something an AI evaluator will surface far more easily than a manual read-through. Chapter Making Discriminators Verifiable 5. Make discriminators findable and verifiable The last point is the reason all of this matters: getting your best material in front of the person who scores it, with evidence they can verify. AI evaluators aren’t making the final decision, but they are doing initial passes: matching language, identifying requirements, spotting gaps, and writing summaries. All of that ends up with a human evaluator. So we want to make sure, first, that the AI can find exactly what it needs to review, and second, that it doesn’t highlight inconsistencies or lose the narrative of our approach in a summary, in that game of AI telephone. How do we get discriminators through to the person scoring them? Four things. First, name it in Section M’s words so nobody has to work out where it earns credit. Nobody should have to infer or guess where it belongs. That’s how you get full credit. Second, put it in the text, early in the block. You can repeat it in a graphic or callout box, but that shouldn’t be the only place it lives. Third, back it with evidence that’s relevant to this scope, with enough context for the reviewer to check it. Fourth, say what it means for the customer in one direct sentence. Look at the difference between a claim and a claim with evidence. Every reader has seen the sentence on the left: “Our proven transition approach ensures minimal disruption and a seamless experience.” A summarizing tool would find nothing in it to point to. On the right, and again the figures are hypothetical, the strength is named in Section M’s language, followed by a comparable contract, its size, a measured result, and where that result is documented. Everything needed to assess it is right there. One caution on numbers before we move on. A number on its own doesn’t make a claim stronger; more data isn’t automatically better. A number without context, like “78%,” invites the question: 78% of what, on what size contract, compared to what? It can raise more questions than it answers. That’s why the example on the right gives 78% against a benchmark: a 70% contract target and a 71% baseline before transition. Now the evaluator can see that you beat the target and actually improved during a period when performance usually dips. Well-chosen numbers with relevant context build confidence. If you use numbers, make sure they have a purpose, are verifiable, and add to the quality of what you say you’ll achieve. One more point on this. A lot of proposal text is being drafted with AI help now, and evaluators are seeing more generic, same-sounding content. That’s why specific, verifiable evidence is the most important part, and what will separate you competitively. If you generate content with AI, it’s critical that it’s grounded in your own information, and that the grounding is traceable and verifiable. After submission, and choosing the right tools Chapter Post Submission Strategies And Tools A short, practical word on what happens after you submit. Before you submit, keep an exact record: the final files you submitted and the approved sources behind your key claims. If a debrief points to a weakness that doesn’t match what you wrote, you need to be able to say exactly what you said and where. A mismatch suggests your content may have been paraphrased or parsed incorrectly. Keep a record of the claims you make and where they are so you can justify them if needed. Ask specific questions: which findings, and where did the evaluators look? If a finding says something is missing and it’s on a specific page, that’s worth a follow-up. Whether it’s a best and final offer, a review, or a debrief after a loss, you need to be able to answer the agency’s questions. You also need to do your own traceability on the back end to identify where you had misses, why, and how your structure might have kept those systems from catching things. That helps you improve over time. If something looks misread, think about how Tracks ended. The AI outputs went into the record, but the protest still failed because it didn’t show an error attributable to AI. Showing that AI was involved isn’t enough; there has to be a real miss. So in your debrief, take all your feedback and responses and identify whether there was actually a miss, because you may have something to raise. Finally, read every solicitation for statements about AI, both how the agency may use it and any rules for how you may use it in drafting. Some agencies are starting to disclose this, and it will remain a moving target. They’re beginning to explain how they use AI or, as in the example I gave earlier, at least making you aware that it could be part of the process. Briefly, on tools for checking and assessing your own drafts: use the right technology in the right place. Use deterministic, rules-based checks when you need the same answer every time. That includes finding requirements language and specific risk items, and it’s also very helpful for scoring sentence length and passive voice, finding undefined acronyms, and checking terms against a locked list. I mentioned the watchword list earlier, and the VisibleThread platform also uses Flesch Reading Ease and Flesch-Kincaid to measure readability, grade level and accessibility. You’re not getting a summarized take; you’re getting measurable targets. GenAI interprets, so it’s very useful for drafting mirrored openings, summarizing gaps, and suggesting plainer sentences, with a person checking the output. As always, the human column doesn’t change. Win strategy, whether the evidence holds up, and what you can and can’t claim will always fall to a human. If anyone has more questions about the VisibleThread approach, reach out to me or one of our team; there’s a link in the box below. VisibleThread or not, think about the technology you apply to your QA review before submission and make sure it’s the right tech for the right job. Audit, next steps, and closing Chapter Pre Submission Audit Checklist A couple of quick action slides on what you can do next. Here’s a short audit to run one requirement at a time, early and often. Run it at pink team, run it again at red team while there’s still time to fix what it finds, and of course run it before you submit. The five steps are: 1. Find. Find the requirement and record where it’s answered. 2. Verify. Confirm the response actually meets the obligation. Matching words is just level one; this step asks whether you’ve committed to and explained what was asked of you. 3. Compare. Check dates, staffing commitments and terms against the crosswalk and against your other volumes. 4. Validate. Confirm the key statements have credible evidence the reviewer can access. Can the reviewer see your proof points? 5. Discriminate. Could an evaluator find your discriminator and connect it to Section M, or is there a gap? You’ll get these slides afterward, so you can review the worked example. We’re almost at time, so I won’t read through each box, but look at the Verify box in particular. That’s where a more experienced reviewer asks the harder question: what does full operational capability actually mean, and how will we show we’ve reached it? If we can’t answer that, the commitment is findable but not meaningful. Verification goes deeper than just finding the piece. Then there are five things you can work into your next proposal: 1. Mirror the requirement in the heading and first sentence: the requirement’s words, exact numbers, and references. 2. Build the compliance matrix to page level, with one row per evaluable obligation, quoted in full, with status. This enhances traceability and your ability to verify. 3. Set readability targets at kickoff and check them again before red team, so everyone works toward the same goal. 4. Start the crosswalk at kickoff and search every volume against it before red team and before final, tracing requirements throughout the process. 5. Run the audit early in the process, not just before submission, starting with your highest-weighted requirements. For every major commitment, ask how you’ll prove you’ve met it, and whether that’s easily traceable. That information has to be there for a human evaluator, and if AI evaluators are in the mix, you want them following your rationale and traceability and passing it on to the evaluator. None of these depend on AI. You could adopt them anyway, and they make a person’s evaluation easier too. But they safeguard you in the situation we’re in right now, where we don’t know who’s on the other end, whether AI is involved, how much, or what type of tool it is. The goal is a response that’s universally easy to digest and universally conveys your message. The pre-submission checks will be shared with the deck as well. Where we land is this: get found by the machine, get believed by the human. The basis of a competitive proposal hasn’t changed. You still need a clear approach, credible evidence and real differentiation. What changes is how quickly gaps, conflicts and unsupported claims can surface, because AI evaluators can find them fast. So make your response easy to find, verify and evaluate, whoever reads it and whatever system is in the mix. We’re at time. I do have a couple of questions, and since I want to respect everyone’s time, I’ll follow up with those folks directly. Thank you for being here and for taking the time; I know everyone is busy. If you’d like to continue the conversation or talk about anything else, I’d love to connect with you on LinkedIn. Our next office hour session is in two weeks, and there’s a lot of educational content available to download on the website right now. I’ll share the slides, the checklist and everything we covered today, and you’ll be able to access the recording. Feel free to visit our office hours page for past recordings and additional resources. I love this conversation and would be glad to keep talking about it. I’m really looking forward to seeing how AI evaluation develops, the parameters around it, and when it becomes more standardized. We know it’s not going anywhere; the question is how we work with it. I’ll keep working on this and share information with you as I get it. Thanks again for joining me. Have a great rest of your day and a wonderful weekend. Bye.
Explore our Past Webinars
On-demand recordings so you can learn on your own time