At ReignOverTech, Tristan Waite leads our AI work, and every agent we build starts from one rule I would put on the door: the agent asks a person instead of guessing, and anything risky waits for approval.
It is the least exciting thing you can say about an AI agent. It is also why the agents we build can run unattended.
Guessing is the expensive failure
A question costs a person a moment. A wrong guess costs the mistake plus the trust in everything the agent does afterwards, because nobody can tell which of its other actions were guesses too.
So when an agent we build cannot tell what a request means, it does not pick the likeliest reading and carry on. It routes the request to the right person and waits.
On the equipment billing ledger we built for a Texas construction contractor, that rule decided the result. Across 72 real call-off messages from crews, the system booked 62 correctly, got none wrong, and handed the other 10 to a person. The number that matters there is the zero, and how we checked that ledger against the books is its own story.
The actions are a list we agree up front
The second limit is what an agent may do at all. Before an agent touches a client's inbox, CRM, calendar or documents, the actions it may take are written down and agreed with the client. Reading and filing might be on the list. Sending money or deleting records usually is not.
An agent with an open-ended brief is a liability waiting for an edge case. One with a list can be checked: anyone can look at what it did and see whether it stayed inside the lines.
Risky means a person says yes
Some actions belong on the list but should never happen on their own. For those, the agent prepares the work and a person approves it before it goes out.
Some limits sit above any single action. In INTERMEW, outreach is controlled centrally: opt-outs are honored before any message is written, pacing is enforced, and no client setting or config change can switch sending on by itself.
Watching after it ships
An agent that ran well in testing can drift once it meets real traffic. Every agent we deploy ships with monitoring, so the client can see what the agent did and catch it if it starts to wander.
That is also where the limits earn their keep. A log of what an agent did is only useful if there is a short, written list to compare it to.
When AI is the wrong tool
The last limit is the one we apply before building anything. Plenty of problems need a script, not a model. If a task is fully rules-based, a plain program does it more cheaply and more predictably, and we say so.
We put AI on jobs with a name: reading the requests that arrive by chat and email and filing each one where it belongs, reading documents that used to take an afternoon, predicting the one number a business plans around. In each of those, the limits above are what make it safe to let the agent run unattended.
If you have a job you would like off a person's plate, tell us what it is. We will tell you whether an agent can do it reliably, and what limits it would run inside. Our AI agents and machine learning page has the rest.