Seven questions to ask anyone selling you AI
Everyone has an AI pitch now. These are the questions that separate the people who will return hours from the people who will return invoices.
If you run a service firm, you've probably been pitched AI more than once this year. A software rep, a consultant, the IT provider, someone's nephew. The pitches sound the same, and that is the problem: when every offer sounds the same, you end up choosing on price or on how much you liked the person.
The gap between firms that get something from AI and firms that don't is real. In McKinsey's latest survey, 88% of organizations use AI somewhere, but only 37% see any effect on profit. Most of the difference isn't the technology. It's whether anyone redesigned the work and measured it.
So before you sign anything, ask these seven questions. Ask us too.
1. What is my baseline?
If they can't tell you how many hours your team spends on the work today, they can't tell you how many they gave back. A serious provider times the work before building anything and asks you to sign the number. A vague one talks about "efficiency gains".
Worry if you hear: "You'll feel the difference."
2. Which three workflows first, and why those?
A good answer names your work: bid intake, quoting, field reports to invoice, owner reports, survey write-ups. It ranks them by hours and by how rule-based they are. A weak answer is a product menu that looks the same for a strata manager and a shipyard.
Worry if you hear: "Our platform does all of that out of the box."
3. What does it touch, and who approves what goes out?
For a property manager, an owner report with the wrong figure is a relationship problem. For an oil and gas services firm, an invoice that doesn't match the field ticket is an audit problem. Ask exactly which systems it reads from, which it writes to, and which outputs wait for a person before they reach a client. Quotes, invoices and client-facing reports should always wait.
Worry if you hear: "It runs on its own, so you don't have to think about it."
4. Where does my data live, and is it used to train anything?
You want a plain answer: which accounts, under whose admin, in which tenant, and whether anything you put in is used to train a model. Ask for the usage policy in writing. If there isn't one, you are the one writing it, after something goes wrong.
Worry if you hear: "It's all in the cloud, it's secure."
5. Who owns it if we stop?
Your accounts, your data, the automations and the documentation should all be yours. If the answer is that everything lives in their platform and stops working when you cancel, you are not buying a system. You are renting one, with your own processes as the deposit.
Worry if you hear: "You'd just export a CSV."
6. Who runs it in month seven?
AI systems are not a one-time install. Models change, your price book changes, a supplier changes their form, someone renames a folder. Something that worked in March quietly gets worse by September. Ask who watches it, who fixes it, and how you'll know.
Worry if you hear: "Once it's set up, it just runs."
7. What happens if it doesn't work?
Ask for the answer in writing, tied to a number you can check. "We'll make it right" is not a term. A measured target with a consequence is.
Worry if you hear: "It always works."
How we answer them
We wrote these knowing you'd ask us. Briefly: we time your team's week and you sign the baseline (1). We pick the first builds from that study, not from a menu (2). Nothing reaches a client without a person approving it (3). Everything runs in your own accounts under your admin, with a written usage policy, and nothing is used to train models (4). You own all of it (5). We run it every month and report the hours in one page (6). And if your ledger isn't at ten hours a week by the end of month three, we keep working free until it is (7).
If someone else answers all seven better than we do, hire them.