The draft is done, the deadline is Friday, and some or most of the text came out of a model. What stands between you and submission is the step most teams compress or skip: reading the bid the way the evaluator will. The nine checks below are the ones that catch the failures we actually find in AI-assisted bids. Run them in order; the early ones are the cheapest and the most lethal.
- Search for residue. Before reading a word for meaning, search the document for the mechanical tells: assistant scaffolding (“here’s a”, “I’ve drafted”, “as an AI”), square brackets and placeholder text ([client name], [insert]), markdown artefacts (##, **, stray pipes |), and every client and place name that is not this buyer. The full residue list takes minutes to run and catches the errors that end careers.
- Verify every fact. List each figure, date, named contract, certification and person in the bid. Trace each to a source document. Anything that cannot be traced was possibly invented - remove it or replace it with a claim you can stand behind. Pay particular attention to statements about the buyer’s own organisation and service: one factual error there undermines the whole submission.
- Answer the question as asked. Put each response beside the question in the RFT. Check it answers every part, in the buyer’s own terminology, using every data point, sub-group and appendix the question names. Fluent text that drifts off the brief is non-responsive text.
- Check compliance mechanics. Page limits, word counts, mandatory formats, required declarations, certifications held at bid date, every mandatory document present and in the right place. Generated text is verbose by default; limits are the cheapest exclusions in procurement.
- Read for contradictions. One person reads the complete bid end to end. Team names, reporting frequencies, timelines, prices and commitments must match across every answer, the CVs and the pricing schedule. Models drafted your sections independently; the evaluator will not read them that way.
- Apply the competitor test. For each claim, ask whether a competitor could paste it into their bid unchanged. If yes, it earns nothing - cut it or make it specific to your delivery record.
- Count the evidence. Each scored answer needs verifiable proof: a named contract, a dated outcome, a real figure, a consequence for the client. An answer with none is a mid-band answer regardless of quality of prose.
- Strip the hedges and the padding. “We would hope to”, “subject to”, “endeavour to” read as reservations - commit in plain future tense with an owner. Delete empty intensifiers. If a section exists to fill space rather than earn marks, it is spending your page limit against you.
- Have someone who did not write it score it. The author of a passage cannot see its gaps; the writer of a prompt cannot see what the model quietly dropped. An independent read against the buyer’s criteria and band scale - human, machine, or both - is the only check that resembles what happens next.
The takeaway
An AI-generated draft is a starting position, not a submission. The nine checks are one afternoon’s disciplined work, and they operate on a simple principle: everything in the bid gets read, verified and owned by a human before an evaluator reads it. That is also, not coincidentally, exactly what buyers’ new AI-disclosure clauses ask you to confirm - and why using the tools is fine when the checking is real.
Step 9, done properly, in hours not days. BidReview audits your finished bid against the buyer’s own criteria - compliance, evidence, contradictions and residue flagged with fixes before you submit. Run the free scorecard →
