There are two ways an automated desk can fail you, and they are opposites.
The first is obvious: it guesses. It needs an account number, a commercial position, a confirmation you hold a particular certification, and rather than stopping it produces something plausible. That is the failure everyone worries about, and it is the easier one to design against.
The second is quieter and more common. It stops constantly, on everything, without distinguishing between a question that needs your judgement and a question it could have answered from the record. A review queue of forty items where six actually needed you is not oversight. It is a queue nobody reads carefully by item thirty.
An automation that asks you about everything has not given you control. It has given you a second inbox.
What genuinely requires a person
The category is narrower than most tools assume. Three kinds of thing.
- Facts only you hold. An account number, an internal reference, whether a certification is current, what a specific customer was told on a call last month. No amount of context recovers these, because they were never written down.
- Commercial judgement. What margin to take. Whether to accept a risk. Whether to bid at all on a job where you are unlikely to be competitive. These are positions, not answers, and they belong to whoever owns the number.
- Anything that commits the company. A quote, a proposal, a submission. Not because the draft is likely wrong, but because a commitment should have a person's name on it regardless.
Everything else, and it is most things, is a lookup. What did the supplier say about lead time. What does the specification require for coating. What did we already tell this buyer about scope. If the record contains it, asking you is not caution, it is a failure to read.
Ask the question, not the task
When a person is needed, the shape of the request matters as much as the fact of it. The bad version routes you the whole item: here is a deal, deal with it. That hands back the context-reconstruction problem you adopted the system to remove.
The useful version isolates the one thing. A specific question, with the drafted reply sitting beside it, so that one answer resolves both. You supply the fact. The reply completes and goes to the gate. You are not asked to write anything.
Sample data. One question, the draft beside it, one answer resolves both.
And say why it is waiting
Every card that is not on the automatic track should state its reason in plain words. Waiting on you for a fact. Held because the draft scored below the bar. Held because it commits the company and needs a click.
This sounds like a small interface courtesy. It is the difference between a system people trust and one they route around. An item stuck for an unexplained reason gets handled outside the system by someone in a hurry, and once that starts, the record is no longer complete, and everything downstream of the record degrades.
The honest version of hands off
We do not claim the desk runs without you. It runs the correspondence, which is most of the volume and almost none of the judgement, and it brings you the decisions.
The measure worth holding a vendor to is not how few things reach you. It is whether the things that reach you are worth your attention, and whether you can tell at a glance why each one is there.