Choosing an AI and automation partner has become hard for one simple reason: everyone can run a demo. Today's tools let anyone build something that looks like it works in a single evening. What cannot be built in an evening is a process still running in eighteen months, once the CRM interface has changed, once the model has been updated, and once the person who built it has left.
So we gathered the ten questions that, in our view, separate the two. We publish them together with our own answers, article by article. You can ask them of any provider, including us.
1. The starting method
"If you walk into a company that doesn't yet know what to automate, what is your method?" A good answer talks about process mapping, interviews with the teams, volumes, time spent, costs and risks. A poor answer starts with a tool name. Our answer.
2. A real case, end to end
"Show me a process you automated from start to finish." Starting situation, breakdown, tools, automated share, human share, result, maintenance. If the answer is a client logo with no description of the mechanism, the question is still unanswered. Our answer.
3. The technology
"What do you actually use?" Models, programming interfaces, databases, queues, permissions, logging, monitoring. The tool matters less than the ability to turn it into an architecture that stands up. Our answer.
4. Where the human sits
"Where do you put the human?" There is a considerable gap between "the AI decides and acts" and "the AI analyses, proposes, a human validates, then the action goes out". On a sensitive process, that gap is the whole subject. Our answer.
5. Errors
"What happens when the AI gets it wrong?" Expect precise words: thresholds, deterministic rules, fallback, logs, alerts, undo. If you are told the model is smart enough not to be wrong, the conversation can stop there. Our answer.
6. Return on investment
"How do you calculate it?" Time saved multiplied by a loaded hourly cost, plus errors avoided, plus additional revenue, minus the cost of the solution. Automating a twenty hour per week task and a thirty minute task are not the same decision. Our answer.
7. Maintenance
"Who looks after it once it runs?" Who steps in when a programming interface changes, who monitors, who fixes, what it costs, and whether you can take back control yourself. It is the most forgotten question, and the most expensive one. Our answer.
8. Ownership
"Who owns the workflows, the code, the instructions, the databases, the accounts and the data?" An automation living inside the provider's account is not an asset, it is a dependency. Our answer.
9. The ninety day plan
"If you had to automate 20 % of our operations in three months, how would you do it?" Do not feed them the answer. Let them build the plan in front of you: that is where experience shows, or does not. Our answer.
10. What they refuse
"What would you refuse to automate?" This is our favourite question. A provider with no limits to offer has probably never had to carry an automation in production. Our answer.
A provider who refuses nothing protects nothing.
One extra question, if you hear the word orchestrator
The word is fashionable and covers almost anything. Ask what sits behind it: which components, where the data is stored, how agents pass information along, how permissions and errors are handled. We answer that in a dedicated article, and we explain exactly what our offer covers on the AI Orchestrator page.
And after the meeting
Score every provider on the same ten criteria, cold, the next day. We publish the grid we use, with its reading thresholds. It has one deliberate flaw: it does not reward the demo. That is exactly why it is useful.
