What does work?
A first pass at structure. Headings, a close date if it is printed clearly, a list of documents in the pack, a plain-language restatement of a scope you have already opened. Used as a junior reader, with you holding the pack, ChatGPT and Claude save the hour you would have spent making a contents page.
That is not nothing. It is also not a bid/no-bid. Firms that stop there have a fluent memo and an unread indemnity. The memo feels like work. The call has not been made.
Where does it break?
Mandatories. They hide in schedules, in a “must” that is not labelled, in a prequalification on another portal. A model that missed one will still write a confident summary. Evaluators who missed one will still void you. Held today is a human tick, against a named instrument, not a paragraph that sounds right.
Indemnity and liability. Tables, cross-references, definitions that move the cap. A lawyer reads those as a system. A model reads them as nearby sentences. Uncapped consequential loss sitting in a definition you did not paste will not appear in the chat. The fifteen flags exist so that hunt has a beginning. The chat does not replace it.
Consistency. Ask twice, get two lists. A gate that cannot be reconstructed on Thursday is not a gate. The room needs the same clauses named, the same three scores, an owner and a date. If the working changes when you refresh, you do not have working. You have a vibe with a better vocabulary.
Citations. “Somewhere in the conditions” is not a clause. A partner who was not in the room cannot sign a paragraph without a page. If the model cannot point, the call cannot travel.
What about Claude, Copilot, a private GPT?
Same lane. A stronger summary is still a summary. A model on your tenant, with the pack not used for training, is a better confidentiality story. It is not a better gate unless it cites, repeats, and scores against your firm rather than a generic “good bid” prior.
selrob sits in that last job: the pack against your strategy, capacity and risk, clause-named. That is not a chatbot with a better prompt. It is a different layer. The qualification vs response page is the comparison. Do not buy a writer, or a chat window, to do the gate.
Is it safe to upload the pack?
Treat a consumer chatbot as a decision your insurance and your client may both want to hear about. Training, logging, residency: see security and confidentiality. Assume the pack has left the building. Commonwealth tenders marked sensitive, defence volumes, customer data in schedules: if you would not email it to a stranger, do not paste it into a free window.
A private tenant with a contract is a different conversation. Still a conversation. Still not a substitute for the four reads.
So what should you actually do?
Use a model as a junior reader if the pack is allowed to go there. Do not let it own the call. Mandatories, liability, roster, weights: you, or software that cites and repeats. The weekly gate is twenty minutes with the working on one page. A chat log is not that page.
If the honest reason you opened ChatGPT is that nobody will read the pack, the problem is the gate, not the model. Install the read. Then decide whether a model, a qualification layer, or a library is the next purchase. Order matters. The DIY question is usually a qualification question wearing a chatbot.