The demo is never the hard part. An agent that answers questions over a handful of documents takes an afternoon.

What decides whether it survives contact with real users is far less glamorous: when it produces something odd, how fast can anyone tell what went wrong, and how much can it have cost by then?

The number that matters

Without citations With citations
First step ask where it came from; nobody knows open it, see the document and the passage
Then search drives, ask around, compare versions check whether that document is right
Typical time half a day to a week about a minute
What you learn the answer was probably wrong whether the source, the retrieval or the reading was wrong

The last row is worth more than the time saved. A wrong answer has three possible causes — the document itself is wrong, the right document was never retrieved, or the document was fine and the model misread it. From the outside they look identical and the fixes are completely different.

Three boundaries that make that possible

It reads only what the asker can read. Permissions are inherited from the systems you already run rather than redeclared for the AI, so an agent cannot surface something its user could not have opened. That is a stronger guarantee than filtering after retrieval, because it does not depend on anyone remembering to tag a document correctly.

It says so when it finds nothing. The dangerous output is not an obvious mistake — it is a confident, well-formatted answer describing something that does not exist. Nobody fact-checks a reply that reads like it came from the manual.

Committing actions wait for a person. Submitting, paying, deleting, sending outward: these are gated, not because the model is unreliable, but because the cost is asymmetric.

Retrieving the wrong figure Submitting the wrong order
How it is fixed look again phone call, reversal process
Cost minutes hours, sometimes money

The gate is also what lets the rest expand. A team that knows nothing is committed without a human will let an agent do considerably more of the work that leads up to the commitment.

Tools are granted per agent

Reading a service history is one capability; creating the case is another; closing it is a third. They are usually discussed as one feature and they should not be.

Granted per agent, a reader and a writer are two different agents with two different tool lists, and the one that talks to customers does not need the one that can close tickets. The practical test: for any action the system can take, you should be able to name the agent that can take it without looking it up.

What this adds up to

Not a system that cannot be wrong. A system where being wrong is visible, bounded, and diagnosable — which is the standard a deployment can actually be held to.