Short answer
Ask which of your systems it will write to, what happens when it is not confident, what the ongoing run cost is at your volume, who maintains accuracy in month six, what share of volume will be handled unattended and on what basis, whether your data trains a model, what happens to the system if you stop paying, what they have declined to build, and what the assessment concludes if the answer is that you should not proceed.
Why the usual questions do not discriminate
Asking which model a vendor uses, or whether they have experience in your sector, produces answers that every vendor can give. The questions below are useful because a vendor who has not actually delivered this work cannot answer them fluently, and the hesitation is informative.
The nine
- Which of our systems will it write to? A vendor selling a chatbot answers this with training data rather than integrations.
- What happens when it is not confident? The answer should describe a threshold, an escalation path and a queue — not reassurance about accuracy.
- What are the ongoing run costs at our projected volume? Model usage, hosting and monitoring. Vendors who omit this are quoting a build, not a system.
- Who maintains accuracy in month six, and what does that cost? Systems drift as your business changes. Unmaintained ones quietly stop earning.
- What share of volume do you expect to handle unattended, and what is that based on? A number offered before anyone has looked at your actual data is a guess wearing a suit.
- Is our data used to train a model? This should be answerable in writing, in the contract, without hedging.
- What happens if we stop paying? Where does the configuration live, can you export it, and does the workflow survive?
- What have you declined to build? A vendor who has never turned work down has either been lucky or has no boundaries.
- What does your assessment conclude if we should not proceed? If there is no version of the analysis that says no, it is not an analysis.
The answers that should concern you
A firm price for a workflow nobody has mapped. A deflection or accuracy figure quoted before anyone has seen your data. A proprietary schema or technique described as secret. Reluctance to put data-processing terms in writing. And any proposal whose business case only works at very high automation rates — because a first pass that handles half the volume is a good outcome, and a plan that requires more than that is a plan that fails.
If the business case only works at ninety per cent automation, it does not work.
Last updated 2026-08-18 · Koderead
Related
Apply this to your own business
Eleven questions, scored in your browser, with the arithmetic shown.
Find my AI opportunities