Before paying for AI, ask what operating problem will change, what information the system will use, what authority it will receive, who remains responsible, and how the business will know whether the implementation created value.

Most bad AI purchases aren't obviously bad at the sales call. They pass because nobody asked what happens after the invoice: who owns the account, what the provider still controls, and what it costs to walk away if the workflow doesn't work.

Decision criteria

  • The proposal names a workflow, not just a model or agent platform.
  • The provider can say when AI is unnecessary.
  • Data, permissions, account ownership, and offboarding are clear.
  • Commercial, sensitive, or unusual cases stay with people.

A proposal that skips straight to model names is telling you something: nobody has mapped the actual workflow yet. Ask for the trigger, the input, and the exception path before the demo, not after the contract, because a vendor who can't answer those questions in plain language usually can't operate the system after it ships either.

Practical steps

  • Ask the provider to describe the trigger, input, current delay, expected output, owner, exception path, and measurement.
  • Challenge whether a CRM rule, template, checklist, form, or existing software feature would be enough.
  • Confirm retention, vendor terms, logs, permissions, and who can disable the integration.
  • Define success, unacceptable errors, and a stop rule before expansion.

If the answers to those questions aren't clear before you sign, they won't get clearer after. A provider who can't say what "unacceptable" looks like for their own system is asking you to define their stop rule for them, after the money has already moved.

Scope boundaries

A prominent honest point: your business might not need AI. AI is useful when language or information varies enough that fixed rules cannot produce a dependable result by themselves. It is not useful when the real problem is an unclear process, stale data, poor ownership, or missing follow-up discipline.

Ask what "not useful yet" actually costs. A provider confident enough to say AI won't help a given problem, before taking payment for it, is telling you more about how they'll handle the next disagreement than any feature list will.

Questions to ask before you start

  • Can we explain the first version in plain language?
  • What information may the system use?
  • What is it never allowed to decide?
  • How do we know whether it helped?

Get these answers in writing before you sign, not verbally on the sales call. A vendor willing to put "what it's never allowed to decide" in the contract is a different kind of vendor than one who'd rather keep that vague.

A responsible first version

Buy the smallest version that answers one real question first. If it holds up after a few weeks of actual use, the case for expanding it is evidence, not a sales pitch. If it doesn't, you've learned that for the price of one workflow instead of the whole platform.

FAQs

What is a red flag in an AI proposal?

A proposal that leads with model names, vague agents, or transformation language without defining workflow, access, review, and measurement.

Should AI talk directly to customers?

Only when the use case is narrow, disclosed when appropriate, tested, and backed by an easy human escalation path.

Who should own the system?

The business should retain access to important accounts, data, integrations, documentation, and offboarding instructions.