Writing Proposals for a Machine

For 20 years we’ve told proposal teams the same thing. Write for a tired, busy evaluator. Someone with a stack of 14 proposals, a cold cup of coffee, and a deadline of their own.

That evaluator still exists.

But more and more, they aren’t your first reader anymore. A machine might be. And the tired, busy human may or may not double-check what that machine decided about you.

So… does everything we know about writing winning proposals go out the window?

I’d argue the opposite. Here’s why.

Your first reader might be a machine

his stopped being hypothetical this year.

The Department of Defense (DoD) has said in several recent solicitations that it uses AI to help evaluate bids (Nextgov/FCW). And the first fights over it are already in court.

The TRAX case. The Army awarded a $449.4 million White Sands Missile Range mission support contract to the incumbent, Southwest Range Services, at a price roughly $29.4 million higher than TRAX International’s bid (ENR). During the GAO protest, the Army admitted one weakness it assigned TRAX was wrong. GAO still denied the protest on May 14, calling it a harmless error that didn’t change the outcome (Federal News Network).

So TRAX went to the Court of Federal Claims. Its lawsuit calls that weakness a classic AI hallucination… with made-up references to its proposal that no one on the evaluation board caught (Defense One).

The Army’s first answer? A procurement analyst ran the proposals through an AI tool called FAST TRACK as a test, the output was unusable, and nobody relied on it.

Then the story shifted. The judge found the government hadn’t been forthcoming, ordered the AI evaluations into the record, and told the government’s lawyers to explain the errors in their briefing (Washington Technology).

And TRAX still lost. In an opinion made public September 22, the court ruled TRAX couldn’t connect the Army’s AI use to any actual error in its evaluation (Bloomberg Government).

The kicker? The court noted parts of the Army’s evaluation were nearly identical to an AI-generated assessment… and still said more was needed to prove the Army leaned on it (National Law Review).

Read that again. The AI evaluations were real. They reached the contracting officer and evaluation board members. Chunks of the evaluation read like the machine wrote them. And the protester still walked away with nothing.

The government isn’t off the hook, though. In an October 5 order, the court said it needs more information to decide whether sanctions are warranted for the government’s misrepresentations (Bloomberg Government). That part’s still playing out.

TRAX won’t be the last. Salient CRGT has also alleged that a DoD subagency improperly used AI instead of people to evaluate bids (Defense One). GW Law’s Jessica Tillipman put it simply: TRAX “certainly won’t be the last” (Bloomberg Government).

Here’s what that means for your next bid. A protest after the fact is a long shot, robot or no robot. Your real shot is the proposal itself… before anyone, or anything, scores it.

Assume a machine reads it first and a busy human may or may not check the machine’s work.

The fundamentals didn’t get the memo

Here’s where I’m going to push back on the panic.

My LinkedIn feed is already full of “write for the bot” hacks. Somebody, somewhere, is selling an AI-optimized proposal template as we speak.

I get the instinct. But an AI evaluation tool is asked to do the same job your human evaluator always had… find the strengths, find the weaknesses, score against Section M.

So the stuff that wins hasn’t changed. Customer language. Discriminators the other bidders can’t claim. A solution with proof attached. Benefits the government can actually measure.

One of those deserves its own soapbox. Benefits.

Here’s the trap. A proposal comes from your company, so it ends up being about your company. Our approach. Our team. Our tools. Our 25 years of experience. Read your last technical volume and count the “we” and “our” against the “you” and “the Army.” It’s usually not close.

Meanwhile, look at how the FAR defines a strength: something in your proposal that has merit or exceeds the requirement in a way that will be advantageous to the Government. That last part is the whole game. And it’s exactly the phrase an AI tool scoring strengths will be hunting for.

So every claim needs all three pieces.

The feature is what you do. “Our transition team includes three incumbent-capture specialists.”

The benefit is what the government gets out of it. “So the program keeps 90% of its trained workforce on day one, and your mission doesn’t skip a beat during transition.”

The proof is why they should believe you. “We retained 94% of incumbent staff on our last three transitions, including [contract name].”

Stop at the feature and you’ve described yourself. A human might connect the dots for you. A machine will log it as compliant and move on… no strength, no credit.

Skip the proof and you’ve made a claim. Claims without evidence don’t score. Worse, they read exactly like the AI-generated fluff your competitors submitted.

Write to the requirement and you’re compliant. Write to the benefit, with proof, and you’re winning.

What changed is the margin for error. Evaluators are only supposed to score what’s on the page. That’s always been the rule. But humans are human... an evaluator who knows the program might read a little generosity into a vague answer. A machine reads exactly what you wrote. Nothing more.

That brilliant strength your program lead explains in the hallway? If it’s only implied on the page, it doesn’t exist.

That’s the Propose and Persuade peaks of the Summit Win System™ doing exactly what they always did. The bar for execution just got higher.

A proposal that was clear to a tired human is already clear to a machine.

Robots writing to robots

Now for the plot twist.

If a machine is going to read it… why not have a machine write it?

Plenty of firms are already there. Small businesses especially, and I get it. When your BD lead is also your capture manager, your proposal manager, and the person who orders the printer toner, an AI that spits out 50 pages in an hour looks like a lifeline.

So let’s play this out.

Your AI writes your proposal. The other four bidders’ AIs write theirs. The government’s AI reads all five and drafts the evaluation. A human skims the summary on the way to another meeting.

Not one original thought was harmed in the making of this procurement. Somewhere, a contracting officer is on a beach.

Here’s the catch. Those AIs all learned to write from the same internet. Ask all five to describe a transition approach and you get five versions of “a proven, low-risk transition that ensures continuity of operations.” Five bidders are now “uniquely positioned.” Five bidders will “hit the ground running.”

When every proposal sounds the same, the evaluator has exactly one thing left to compare. Price. Congratulations… you automated your way into a race to the bottom.

The robot also wasn’t in the room. It didn’t hear the program manager admit the last transition was a mess. It has no idea who you’re competing against, which makes ghosting them a little tricky.

And it doesn’t have your past performance. It will happily invent some, though. That’s the same hallucination TRAX alleged, pointed the other direction… except a made-up project in your proposal is a false statement to the federal government. That’s a much bigger problem than a lost bid.

Discriminators come from capture. From relationships, intel, and a real point of view on the customer’s problem. That’s the Position peak, and no prompt replaces it.

Use AI. We do. It’s great at first drafts of boilerplate and catching the requirement you missed on page 87. Hand it your strategy and it’ll help you write faster. Hand it nothing and it’ll hand you nothing back… in 50 beautifully formatted pages.

When both sides are robots, the human who actually knows the customer wins.

What does shift

The principles hold. Here’s what actually changes on the page.

Write every section so it stands alone. Your human evaluator may only be assigned Factor 2. And some AI tools pull passages out of a document rather than reading it cover to cover. Either way, “As described in Section 2.3” assumes a reader who flips back. Don’t count on it. Restate the key point where it gets scored.

Make your win themes survive a summary. The human checking the machine may read the AI’s summary instead of your proposal. Picture your 30 pages squeezed into one paragraph. If your discriminator doesn’t make that paragraph, it didn’t make the evaluation.

Get the critical stuff out of your graphics. Your best discriminator, locked inside a gorgeous infographic, may be invisible to a text-based tool. Put the point in the caption and body text too.

Write so a skimming human and a literal machine land on the same strengths.

Same game, stricter referee

AI in source selection is here. Agencies will get better at using it, contractors will keep protesting when it goes sideways, and the rules will lag behind both.

None of that changes what a winning proposal does. It knows the customer. It answers the question asked. It proves why you, with evidence on the page.

The firms panicking about AI evaluators are mostly the firms that were getting away with vague writing before. The machine just stopped giving them the benefit of the doubt.

So write the proposal you should have been writing all along. Then make it impossible to misread.

‍