This August, in our own pipeline, we watched the automation manufacture an attachment out of two pages of a solicitation, give it a filename that did not exist anywhere in that solicitation, and apply a signature to it. On the same bid, in the same session, it invented three different filenames for documents that already had one. A submission email it drafted told the contracting officer, in the body text, that the attached offer was unexecuted.
A human gate caught every one of these before anything went out. That is the actual reason AI never submits a bid on our platform. Not a compliance footnote. A demonstration, from the people building the tool, of exactly what the automation will do when nobody is checking its work.
What we watched it do
The solicitation named a required attachment and told offerors to fill it out and return it. There was no separate file. The attachment was printed inline, as a labeled section near the back of the same document, and the notice itself carried no separate attachments at all. The job was to fill in that section, in place, in the document that was already there.
Instead, the automation read the word "attachment" as proof a separate file had to exist somewhere. When it could not find one, it built one, copying two pages out of the existing document into a new file and giving that file a name nobody had used. Told this was wrong, it deleted the invented file. Then it built the same extract again, filled it in, and applied a signature to it.
What ended up carrying a signature was not the document the contracting officer had published. It was a copy of two of its pages, generated by the automation to satisfy a sentence it had misread. On the same bid, the same session produced two more invented filenames for documents that already had a naming convention, and a submission email that printed an internal file-naming artifact, the word "unexecuted," directly into the message body a contracting officer would read.
Why this is not a one-off bug
Every one of these mistakes looked plausible in isolation. An invented filename read like a reasonable name. An invented attachment read like a reasonable interpretation of the word "attachment." A submission email that lists an internal file marker next to a signed file looks, at a glance, like ordinary bookkeeping. None of it was caught by the automation itself, because none of it looked wrong from the inside. It took a person who knew the actual solicitation, the actual file that already existed, and the actual naming convention, checking the output against reality, to catch it.
That is the argument for a human gate that a legal citation alone cannot make. The risk is not only that a system might submit something noncompliant on purpose. It is that a fluent, confident-sounding automation can generate a document, a filename, and even a signature that all look correct, on a bid closing in hours, and be wrong about all three.
What "AI-analyzed" actually covers, and what it does not
Reading a 200-page solicitation, pulling out requirements, checking them against your certifications and capacity, and pricing against real award history: AI genuinely helps with all of that, and does it faster than a person reading line by line. None of it is the moment where legal exposure attaches, and none of it is the moment where an invented file can quietly become the file of record. The exposure and the risk both attach at the same point, when a document leaves the building. That is the one step in the entire pipeline we build for a human to do, every time, no exception.
Submitting a federal offer is also a legal act in its own right. FAR 52.204-8 makes this explicit: by submitting an offer, "the offeror verifies" that its representations and certifications are current, accurate, complete, and applicable to that specific solicitation (FAR 52.204-8, Annual Representations and Certifications). That verification is not a checkbox a system ticks on your behalf. It is you, attesting to something, with your name and your business behind it. Our own August record is the concrete reason that attestation cannot be automated away. The FAR citation is the reason it should never have been legal to try.
Why a fast tool and an automatic one are different products
A tool that reads, scores, and prices bids faster than you could by hand saves real time and does not carry new risk, because you still review the output before anything happens. A tool that submits automatically is a different product entirely: it makes a legal representation on your behalf, on a timeline you do not control, based on a model's read of a document that might contain an ambiguity only a human catches. The upside of automatic submission is speed. The downside is a certification you did not actually verify, attached to your business, that you cannot take back once it is in.
Where this shows up in how we built it
Every bid gets prepped all the way to a signed, ready-to-send state. Nothing leaves your account without you clicking submit yourself. You also set hard daily caps and a kill switch on AI usage, so the work happening before that point stays bounded and visible, not a black box running unattended. On BYOK, your AI spend goes straight to your own key, not through us, which keeps the incentive structure honest: we are not paid more for you to run more AI jobs.
The honest tradeoff
This means the pipeline is not fully hands-off, and it never will be by design. That is the actual tradeoff, stated plainly: you get speed on everything up to the decision, and you keep the decision. How the AI analysis itself works, start to finish, is worth reading if you want to see exactly where the automation stops and the review starts.
If you are comparing this against a pure research database like HigherGov, this is one of the real differences: a database does not submit anything for you either, but it also is not running your pipeline day to day the way an operator tool does. See how the whole pipeline runs, end to end.
Related reading
- How AI bid analysis actually works
- A Blank on a Submitted Form Is an Answer
- HigherGov alternative: what to look for if you're a small business