What we're hearing from bid and proposal teams about AI

Where it's helping, where it's stalling, and where the judgment call still has to be human

Written by Elisabeth Rotnicki, Founder & CEO at Vyavos

Before anyone writes a word of a tender response, somebody reads the thing. 40 pages, 60, sometimes 80, plus the annexures that arrive late. Everything downstream inherits whatever that person missed.

That is where the tooling went. On the evidence I have, it is working. It is also pointed at the wrong half of the job for most firms, and when it fails, it does so in a way you will not notice.

I ran proposal processes at KPMG Law, working across partners, pricing, legal and risk, and trained as a data protection lawyer before that. Since then, I have held more than 45 interviews with bid leads, proposal managers and partners in professional-services firms, and in early testing, a smaller group put live opportunities through an AI-assisted system end to end. Between them, the firms span Australia, the US and Europe. One of those systems is one I am building, which is also why I checked it harder than the others.

It reads a tender better than I expected

Not one tester rated the analysis poor or unusable, and the gains they named were not about writing. They were about being able to see a document.

A tester who put a commercial RFP through it found it had gone past the tender itself. "Beyond restating the complications described in the RFP, it went a step further by adding elements like risk exposure if unresolved and urgency drivers." A bid writer will recognise that as thinking rather than typing.

Asked for the biggest gain, she was plain: "Time saved, and help organising and executing on the mundane stuff, leaving us more time to think about the proposal." Neither of us has an hours figure for that. But it describes where the hours moved: to strategy and win themes, away from page 52.

So I took a 44-page public tender and checked every field against the source. It was right on what the tender states: nine appendices under their exact names, the question and submission deadlines down to the time zone, the five things wanted in every biller's CV. It was also right about what the tender does not say. Eight fields came back marked "not specified" - no pre-proposal conference, no contract start date - because the tender specifies neither. A system willing to return an empty field is one you can work with. These are reliable on what a tender states and on what it plainly leaves out.

When I sketched this piece out I would have put the evaluation criteria in that category too. That is the one field it invented.

It is aimed at the wrong half of the job

The first reason has nothing to do with the technology. Asked what actually costs them, more people named the content they already own than named reading the tender. A bids and tenders manager at a mid-size Australian law firm said the time goes on "finding the right credentials to include", because partners "keep stuff on their own folders and forget that other people need to see those credentials". Not everyone finds the reading burdensome: one business group leader described his own intake as "probably no different to how anyone else would do it".

That gap explains a lot of disappointment. The reading is what AI does reliably; the credentials are where the hours go. Buy the first expecting relief from the second and you will conclude the technology underdelivered.

Second, without a firm's own approved material loaded, the output reads generic: more editing, not less. "Where it falls a bit short is in how specifically it reflects our company, since we weren't able to upload detailed information about our organization." More than one tester said so unprompted. For whoever has to make a draft sound like the firm, that is not a saving.

Third, permission. "I would definitely use something like this, if I were allowed to," a managing partner told me. But the sharper finding is what nobody said: not one person described a process inside their firm for approving a tool like this. For whoever owns risk, that gap sits upstream of any product decision.

The failures do not announce themselves

Three failures, on three documents, found by three different people.

Incomplete. The tester who found his deck immediately understandable then audited it against the grant guidelines line by line: nine omissions on one application, three of them capable of automatic disqualification.

Misattributed. Across five live client projects, another tester found her own firm named as the client in one of them.

Invented. On that same tender, which it otherwise read accurately down to the gaps, one field came back fabricated. The evaluation weightings: 70% technical, 30% cost, eight sub-criteria. They sum to 105%. No real scoring model sums to 105. That tender publishes no weightings at all.

Not one of them showed on the face of the output. For a practice leader signing off, that is the difference between a tool that saves time and one that creates new risk.

Fluent prose announces itself as prose and gets read sceptically. A populated field looks like a fact.

What to ask for, and what to fix first

Look at where it held and where it broke. A missing conference date is obvious, and it reported that one. A missing scoring model is not, and that is the one it filled. So the distinction is not whether a system will admit a gap. It is whether you could have spotted the gap yourself, and with the weightings you could not, because there was nothing in the tender to compare the output against.

So the question to put to a supplier is not how good the output looks. It is whether you can trace any field back to a line in the tender. Until you can, the tester who audited the grant wrote the rule himself: "the content needs to be carefully reviewed and compared with the original file before sharing it with the client."

And if your real problem is the one named more often than any other in these conversations, knowledge and credentials scattered across personal folders and out of date, start there rather than with the reading. Getting your own approved material into one place is what makes everything downstream worth automating.

AI has changed the first pass, and nobody I spoke to wanted to go back. What it has not changed is who signs. The suppliers worth your time are the ones that can show you which fields they read and which they wrote.

Elisabeth Rotnicki, Founder & CEO at Vyavos