- Nothing a model writes can fix the two things that actually exclude a bid, which are a missing registration and a missed deadline.
- Three regimes now make you register once and reuse the answers. The UK platform, the EU self declaration, and US federal registration all work this way.
- Fluency is not a scoring criterion. Evaluators award points against a published scheme, and a smooth answer that misses an evidence requirement scores zero.
- Detection of AI text in a bid is unreliable in a documented direction. Detectors flag non native English writing as machine written.
- The reusable asset is not a bid. It is an answer library with dated evidence, and it is what makes the second bid take a third of the time.
- A first public sector bid is a two week job, and most of that time is gathering documents rather than writing prose.
A tender pack arrives as a zip file. Inside are eleven documents, a pricing spreadsheet with locked cells, sixty questions with word limits, and a deadline that falls on a Friday at noon. You sell a service, you are one person or three, and the buyer wants a quality management statement, two case studies from the last three years, and evidence of professional indemnity cover at a level you may not currently hold.
This is the moment people reach for a language model, and it is a reasonable instinct that is usually pointed at the wrong part of the pack. The writing is not what loses these bids.
What actually disqualifies a bid?
Two things, and neither of them is prose quality. Missing a mandatory registration and missing the submission deadline account for most of the responses that never reach an evaluator at all.
The registration point is worth spelling out, because the rules are now explicit in each of the three markets a small seller is most likely to bid into. US federal procurement puts it plainly: the Federal Acquisition Regulation policy on registration states that offerors and quoters are required to be registered in SAM at the time an offer or quotation is submitted. Not by award, not by contract start. At submission. The regulation lists exceptions, and they cover things like classified contracts, deployed military operations and disaster response, none of which describes a small firm bidding for a service contract.
The UK works the same way through a different door. The central digital platform behind Find a Tender, live since 24 February 2025, is where all UK contracting authorities publish procurement information, and suppliers register once with core business details that are reused across bids. What you enter is your name, address, company number, tax registration, financial records, connected persons and exclusion information, then you share it with a share code. The platform also records whether you qualify as a small or medium enterprise, which matters because that field is what buyers filter on.
In the EU, the European Single Procurement Document is the same idea as a form rather than a portal. It is a self declaration of your financial status, abilities and suitability, it counts as preliminary evidence of the conditions required, and since October 2018 it has been electronic only. The stated purpose is that tenderers no longer have to provide full documentary evidence and different forms for every procedure, which is a genuine simplification for a small firm bidding across borders.
| Where you are bidding | What you register with | What you declare once | What it does not do |
|---|---|---|---|
| UK public sector | The central digital platform, via Find a Tender | Company details, financials, connected persons, exclusions | Complete the bid itself, which each authority still sets |
| EU above threshold | The European Single Procurement Document | Financial status, abilities, suitability, exclusion grounds | Replace the full evidence, which is requested later |
| US federal | SAM registration, required at submission | Annual representations and certifications | Cover you for a bid submitted before registration completes |
| Private corporate | Whatever portal the buyer licenses | Usually a supplier questionnaire with insurance and policies | Carry over to the next buyer, who uses a different portal |
That fourth row is the one that catches people who have only ever bid privately. Public registration is reusable by design. Corporate supplier onboarding is not, and the same nine documents get uploaded again for every buyer.
Where does AI genuinely save time on a tender?
In three places, all of them upstream of the writing. Building the compliance matrix, cutting answers to length, and turning your existing material into an answer for a question phrased differently.
The compliance matrix is the highest value use and the least glamorous. Feed the pack in and ask for every requirement, every document requested, every word limit and every submission instruction as a list with the page it came from. What you are buying is completeness under time pressure, which is exactly the failure mode of a tired human reading eleven documents at 11pm. Check the output against the pack, because a missed row is worse than no matrix, but the first pass is genuinely faster.
Cutting to length is the second. A 500 word limit on an answer you have written at 900 words is mechanical editing, and models are good at it in a way that preserves the specifics if you tell them which specifics are load bearing. Say which numbers and names must survive, or they will be the first casualties.
The third is reframing. Buyers ask the same eight things in forty different phrasings, and a model that has your previous answers can map a new question onto existing material quickly. This is the same mechanic that makes an assistant useful for putting together a written quote for a service job, where the underlying facts are stable and only the framing changes.
What does not work is asking for the answer itself with no input. A generated response to "describe your approach to social value" contains nothing about you, and an evaluator reads twelve of them a week. It is not that the text is detectably artificial. It is that it is detectably empty, which is a much older failure.
Can an evaluator tell you used AI?
Reliably, no, and the research on detection points in an uncomfortable direction. Detectors do not measure origin. They measure how statistically ordinary the writing is, and those two things are not the same.
The clearest published result on this is the Stanford study on bias in GPT detectors, which evaluated several widely used detectors on writing by native and non native English writers. The finding was that the detectors consistently misclassify non native English writing samples as AI generated, while native writing samples are accurately identified. The authors also showed that simple prompting can bypass the detectors, and they explicitly caution against use in evaluative settings. A buyer running bids through a detector is therefore most likely to penalise a compliant supplier whose second language is English, which is a discrimination problem rather than a detection capability.
The practical implication is not that you should feel safe. It is that you should stop optimising for the wrong risk. The thing an evaluator notices is not machine writing, it is the absence of anything only you could have written. We looked at the same accuracy question from the seller's side in what an AI content detector can and cannot prove about your own pages, and the conclusion holds here: the score is not evidence.
Disclosure is a separate question from detection, and it is moving. Some buyers now ask whether generative tools were used in preparing a response, and some contracts carry obligations about AI used in delivering the work. Answer those questions accurately. A written internal position on what your business does with these tools makes that answer a lookup rather than an improvisation, which is most of the argument for having a short written AI policy for a small business at all.
Why does a fluent answer score badly?
Because it is not being read as writing. It is being scored against a published marking scheme, usually in bands, and the band descriptors ask for evidence rather than for clarity.
A typical scheme awards the top band for a response that addresses every element of the question with specific, relevant evidence, the middle band for one that addresses the question with limited evidence, and zero for one that fails to address it. Read those words as instructions rather than as bureaucracy. "Every element" means a question with four clauses needs four answers, which is why the compliance matrix matters more than the prose. "Specific evidence" means a named project, a date, a number, an outcome. A beautifully written paragraph containing none of those sits in the middle band, and the middle band loses.
This is the structural reason generated text underperforms at tender. A model produces the average of everything written about a subject, and the average contains no proper nouns. Your advantage as a small supplier is entirely in proper nouns: the actual site, the actual month, the actual person who ran it, the actual number that came out. The tool should be holding those and arranging them, never inventing the frame they sit in.
Which questions are worth answering properly?
The ones carrying the most marks, which is published and almost nobody reads before writing. A tender pack states the weighting, and the weighting frequently does not match the question order or the word limits.
Work it out before you draft anything. If price carries 40 percent, quality 50 and social value 10, then a 500 word social value answer is worth a tenth of the outcome and a pricing error is worth four times that. People reliably invert this, because the social value question feels hardest and the pricing spreadsheet feels mechanical. The spreadsheet is where the contract is won or lost, and it is also the part where an error is fatal rather than merely costly: a locked cell filled in wrongly, a rate entered per day where the form wants per hour, a total that does not match the sum of its parts.
Within the quality section, the weightings usually differ per question too. A question worth 15 percent deserves three times the effort of one worth 5, and the word limits will not tell you which is which. Sort the questions by marks, answer downwards, and accept that the last few will be thinner. An evaluator scoring strictly rewards that allocation.
There is one exception worth naming. Pass or fail questions carry no marks at all and end your bid if you answer them wrongly. They are usually about insurance levels, exclusion grounds, or a mandatory standard, and they look like administration. Treat them as the highest priority items in the pack precisely because they carry no points: a zero weighted question that can disqualify you has an effectively infinite weighting.
What does a buyer do with your answers afterwards?
Puts them in the contract, in most cases, which changes how you should write them. A commitment made in a tender response to win a mark becomes a term you are held to for the contract duration.
This is the most common self inflicted wound in small firm bidding, and generated text makes it easier to commit. A model asked to write a strong answer about response times will happily promise a four hour response window, because that is what strong answers in its training data say. If you win, you have promised a four hour response window to a buyer who will measure it.
So read every draft answer once with a single question in mind: what have I just agreed to do, and can I do it on the worst week of the year. Anything that survives that reading is fine. Anything that does not gets rewritten to what you can actually sustain, and a modest promise you keep scores better over a three year contract than an ambitious one you breach in month two. The same discipline applies to the pricing: a rate that works at full utilisation and fails at realistic utilisation is a bid you should want to lose.
What should a small bidder actually build?
An answer library, and it is the only artefact from this whole process with compounding value. A bid is worth nothing after the deadline. The library is worth more after every bid.
The shape that works is boring. One file per recurring topic, each holding the facts rather than the prose: quality management, health and safety, data protection, social value, environmental measures, staffing and cover, insurance, references, financial standing. Inside each, dated evidence. Not "we take data protection seriously" but the date of your last policy review, the name of the person responsible, the retention period you actually apply, and the one incident you handled and what changed after it.
Dated is the operative word. Buyers ask for evidence from the last three years, and an undated library quietly ages into non compliance. A line saying which month each fact was last checked turns a rewrite into an update.
Then the model becomes genuinely useful, because it is arranging facts rather than inventing them. Give it the library and the question and the word limit, and the draft that comes back is specific because the input was specific. The quality of a generated bid answer is almost entirely a function of what you fed it, which is an unglamorous finding and the most useful one on this page.
How long does a first bid take?
About two weeks of calendar time for a first public sector submission, and the split surprises people. Roughly a quarter of it is writing. The rest is registration, chasing documents you do not have, and getting other people to confirm things.
The items with external dependencies are the ones that decide whether you make the deadline. Insurance certificates at the required level, which may mean buying cover. Accounts filed and current. A referee who has agreed in writing to be contacted. Certification such as an information security standard, which takes months rather than days and cannot be acquired inside a bid window. Bank or accountant confirmation of financial standing.
Read that list against a typical four week tender window and the arithmetic explains most failed first attempts. The prose was never the constraint. So the first decision on any pack is not how to answer question 34, it is whether the mandatory requirements are things you already hold, and that decision takes twenty minutes with the pack and an honest inventory.
A tender pack is confidential and often contains the buyer's own commercial detail. Before pasting it into any assistant, check what that tool does with uploads and whether the account has training turned off. The same question applies to every document you hand a vendor, and our own position on it is on the security page.
The habit that wins the second bid
Ask for the feedback and read it against your own answers. Most public buyers will tell you your score per question and the winning score, and a surprising number of small suppliers never request it.
That feedback is the only externally validated signal you will ever get about your library. A question where you scored in the middle band is a library gap, not a writing failure, and it names itself. Fix the underlying fact, date it, and the next bid in that category starts from a stronger position.
The wider pattern here applies past procurement. Every use of these tools that works in a small business has the same shape: the model handles the volume, the arrangement and the second draft, while the facts, the commitments and the evidence come from the business. A bid is just an unusually strict test of that division, because an evaluator with a marking scheme is the only reader who scores your answer against a published standard and tells you the number afterwards. Reading the contract terms attached to the tender before you price it belongs in the same fortnight, since the terms are what you are actually bidding to accept.
One honest caveat about all of this. Public procurement is slow, the win rates for first time bidders are low, and a fortnight spent on a bid you lose is a fortnight. The reason to do it anyway is that the library you build is reusable and the buyer relationship is long, which is the opposite of most sales channels available to a small firm. If the register and the mandatory documents are beyond you today, the useful work is closing that gap rather than writing a response that cannot be accepted.