Can ChatGPT write a public tender bid? What it does well, and where it stops

What can ChatGPT do in a tender?

Quite a lot, provided you know what to ask and give it the right material. A general-purpose model is good with the text in front of it: it summarises it, organises it, explains it and rewrites it. In a public tender that makes it useful for specific, well-bounded tasks, as long as someone with judgement checks what comes back.

• Explaining something you have not come across before: what a lot is, what a clause means, what an award criterion is really asking for.
• Summarising a section of the specifications you paste in, and turning it into a list of requirements.
• Suggesting a table of contents for the technical proposal from the criteria you give it.
• Writing a first draft of a section, or tightening one you already have.
• Going over a text for typos, repetition and unclear sentences.

None of that is trivial. But every one of those tasks begins after you have found the tender, decided to bid and worked out what has to be prepared. The model works on the part of the problem you hand it, and a tender is a good deal more than that part.

Where does it fall short?

In everything that is not in the conversation. A general-purpose model does not know what your company sells, which projects it has delivered, what experience it can evidence, what team and certifications it has, which contracts it is after or how much risk it is prepared to take on. Without that context, three things happen, and none of them should come as a surprise.

• The technical proposal turns generic or, worse, claims capabilities the company cannot back up.
• The decision to bid becomes guesswork, because the model has no way to weigh what the specifications demand against what the company really has.
• In Spanish public procurement the requirements are spread across the administrative specifications (PCAP), the technical specifications (PPT) and the annexes, and they slip through unless the model has been given every document. It seldom has.

Then there is the process. Finding the tender, deciding whether it is worth pursuing, reading the specifications, preparing the technical, administrative and financial parts, reviewing the whole, following the file through to the award and learning from the outcome are links in a single chain. With a general-purpose model each one is a separate conversation, and holding them together, checking that nothing has been missed and keeping on top of the deadlines is still your job.

Four ways to get the work done

ChatGPT is one option, not the only one. These are the four usual ways companies take on public tenders in Spain, with what each one solves and the point at which it stops.

A general-purpose model, used directly

Good for reading, explaining and drafting from the material you supply. It is quick and always to hand. It stops where the conversation ends: it does not look for opportunities, knows your company only as far as you describe it each time, cannot tell whether you meet the solvency requirements and does not follow the file once the bid is in.

An alerts platform or tender search tool

Built to find opportunities and tell you they have been published, which is useful if you do not want to miss anything. That is usually where it ends: deciding which ones are worth it, reading the specifications and preparing the bid fall to your own team.

A tender consultancy

Brings human judgement and experience of public procurement, and can prepare bids from start to finish. Its capacity is measured in people's hours: every extra tender means more hours, and that puts a ceiling on how many bids you can submit.

A specialised, autonomous system

This is what Carabela is: a system designed to run the whole public procurement process. It brings the company's context, the analysis of the file, the award criteria, the documentation, the strategy, the preparation of the bid, the follow-up and the learning together in one system, and it works through opportunities and documents continuously. The decisions that commit the company stay with the company. And it usually works on a success-based model, with part of the fee tied to the tenders the client wins.

None of the four is better in the abstract. It depends on what you need:

• To understand a concept or polish a paragraph: a general-purpose model will do.
• To keep track of what is published in your sector: a search tool or an alerts platform.
• Expert judgement on one specific file: a consultancy.
• To bid for more tenders without growing your team at the same rate: a system that takes on the whole process.

If you use ChatGPT anyway, do it like this

If you decide to prepare a tender with a general-purpose model, these precautions head off the most expensive mistakes.

1. Give it the complete specifications (PCAP, PPT and annexes), not a summary. The requirements are spread across all three; if you are not sure what each one covers, start with what the PCAP and PPT are.
2. Put the award criteria and their points in front of it, and ask it to structure the response against them, section by section.
3. Supply real company documentation: projects, team, certifications, references. Whatever you leave out, it will be tempted to fill in.
4. Tell it in so many words not to claim anything the company cannot evidence, and check every claim regardless.
5. Check the result against the specifications: unanswered requirements, page limits, mandatory formats and annexes, contradictions with the financial offer and the administrative documentation.
6. Before uploading internal documents, check what terms the tool applies to that information. The specifications are published; your company's data is not.
7. If you need a binding legal interpretation, ask a lawyer. No AI model is a substitute for that advice.

And one rule that has nothing to do with the tool: whether to bid, the final price and the signature on the offer are the company's decisions, taken by people with names.

What changes with a specialised system like Carabela?

The difference lies less in the model than in the context and the process. Carabela works with the leading AI models on the market and picks the right one for each task rather than relying on any single one. What it adds is everything around the model: a context built for each company (what it sells, what it has delivered, what it can evidence, which contracts it wants) and a process that runs from search to learning without anyone having to stitch separate conversations together.

With that context, the search ranks the opportunities that genuinely fit instead of sending hundreds of alerts; the analysis tells you whether you can bid and what is missing; the technical proposal is written against the criteria and from the company's real information, without inventing capabilities; and the final review looks for unanswered requirements, contradictions and grounds for exclusion. Once the bid is in, the system follows the file and analyses the result to improve the next one. No tender is a one-off any more.

The level of autonomy is set for each company: you can approve every opportunity before a bid is prepared, review only the major pricing decisions, or keep a human sign-off before final submission. No one can guarantee that a public administration will award a particular contract, and Carabela does not claim to. What changes is how many good opportunities you can pursue, and how consistently. For the full picture of what AI can already do, see what AI can do in a public tender; for the system itself, AI for public tenders.