Ike HolleyAI has commoditized federal proposal drafting. It has not touched the four things that decide federal outcomes: knowing about an opportunity before it is public, reading a program office accurately, positioning a company so a requirement can be written around what it does, and carrying risk on the result. A firm whose offer is writing is exposed. A firm whose offer is judgment is not.
A founder told us in June that he had produced a full Phase I narrative in an afternoon with an AI tool and wanted to know whether he still needed us. It is a fair question, asked honestly, and the honest answer is that he was right about the drafting and wrong about the job.
He was also going to lose, for reasons that had nothing to do with the quality of the prose.
What the tools genuinely do well now
It is worth being straight about this, because the firms in this market that pretend otherwise are going to look foolish quickly.
- First drafts. A competent structural draft of a technical narrative, a commercialization plan, or a past performance write-up, in minutes rather than days.
- Compliance mechanics. Cross-checking a draft against a solicitation’s instructions and evaluation criteria, catching the missing subsection that used to get caught at red team, if it got caught at all.
- Summarization. Reading a long solicitation, a budget document, or a policy release and producing an accurate outline of what is in it.
- Volume work. Reformatting, consistency passes, tailoring a technical description across several topics.
That is real, and it is a permanent change. The cost of producing a competent-looking proposal has fallen close to zero, and it will not go back up. Which means a competent-looking proposal is no longer evidence of anything.
What it cannot do, and why
1. Know about the opportunity before it is public
The most valuable information in federal business development is not published. It is a program manager saying at an industry day that the next iteration is going to need a different sensor. It is a resource sponsor mentioning that a line survived the internal fight. It is a contracting officer explaining which pathway they are actually able to use, which is often not the one everyone assumes.
A model can read everything that has been written down. By the time something is written down and public, the shape of the answer is usually set. The gap between those two facts is where federal deals are won, and nothing that trains on public text can close it.
2. Read a program office
Every program office has a posture: what it is worried about, what it got burned by last time, whether it prefers a prototype it can touch or a study it can defend, how much risk the current leadership will carry. Those things are not in the solicitation. They are in how the requirement is phrased, what the last three awards looked like, and what people say in the hallway after the briefing.
You learn to read that by having sat on the other side of the table, or by spending years next to people who did. It is pattern recognition on a data set that was never public.
3. Position a company before the requirement exists
The highest-leverage federal work happens eighteen months before there is anything to write. It is responding to a request for information in a way that helps the government understand what is possible. It is being one of the companies whose capability the office had in mind when they described the gap. It is lawful, documented participation in market research, which is exactly what that process exists for.
That work is relationship-shaped and calendar-shaped. There is no prompt for it.
4. Carry risk
This is the one people underrate. When we recommend a bid, we have something at stake in the outcome: our fee structure, our reputation with that program office, and the next conversation we have with your investors. A tool has no exposure. It will produce a confident, well-structured, entirely plausible proposal for an opportunity you should never have entered, and it will never tell you not to bid.
The advice that has saved our clients the most money is the advice a language model is structurally incapable of giving: do not do this one.
What this means for evaluators, which is where it gets interesting
Consider the position of a government evaluator in this environment. The volume of competent-looking submissions goes up. The variance in writing quality goes down. Everything reads fluently and says roughly the right things.
What happens to their decision process is predictable: they lean harder on the things AI cannot fake. Specific past performance. Named customers. Technical detail only someone who built the thing would know. Evidence that the offeror understands this office’s particular problem rather than the general problem. Anything checkable.
Which means the arrival of good drafting tools has, slightly counterintuitively, increased the value of the underlying work. When everyone’s prose is fine, prose stops being a discriminator, and proof takes its place.
How we use these tools ourselves
We use them, daily, and we will tell you exactly where.
- Monitoring. Watching solicitation feeds, policy releases, and budget documents, and flagging what changed. A person then decides whether it matters. The machine reads; the judgment is ours.
- Drafting the parts that do not decide anything. Boilerplate, formatting, first passes on sections that are structurally identical every time.
- Compliance checking. Because a machine is better than a tired human at confirming that every instruction in Section L has a corresponding piece of the response.
What we do not do is let a tool make a bid decision, describe a program office, or invent a number. Our one hard rule on published claims applies internally too: a sourced number, or no number. A model will happily produce a plausible statistic. That is the most dangerous thing about it.
The test to apply to any firm you are considering
Ask them what they think AI has changed about their business. If the answer is nothing, they have not been paying attention. If the answer is that they now produce proposals faster, ask what you are paying for, because you can produce proposals faster too.
The answer you want describes work that happens before writing and continues after submission: which offices, which requirement stage, which pathway, which decision to skip. That work has not been automated, and the reason is not that the tools are not good enough yet. It is that the information it depends on was never written down.