Human always sends
Agents draft and fill forms. A person owns every message that leaves the building.
The problem
Client-servicing teams spend most of their day reading email, working out what the client wants, looking things up, and writing a reply. That is exactly the work language models are good at, and exactly the work where one confident wrong answer reaches a client with the bank’s name on it.
Full automation fails the first time it sends something wrong. No automation leaves the obvious gains on the table. The useful question is where to put the person so they add judgment without redoing the work.
How it works
- Client email
- Classify requestwith confidence
- Confidence gate
- Scope + lookups
- Draft reply + form
- Person approvesor edits, or rejects
- Sent + logged
The gate sits at the boundary, not at every step
Inside the system, agents classify, extract, look things up and draft without asking permission. The one hard stop is the point where something leaves: an email, a status change the client can see, a booking. Reviewing one finished draft is fast; approving five intermediate steps is not.
The reviewer gets a finished artifact
The specialist sees the original email, the classification and its confidence, what was looked up, and a ready-to-send draft with the request form already filled. Their job becomes checking and deciding, which is far faster than reading, researching and writing from scratch.
Every decision is labeled data
Approve, edit and reject are all captured with the draft they applied to. Edits show where drafts are close but wrong; rejections with a reason become test cases. The eval set grows from real decisions instead of from a one-off labeling exercise.
Authority checks are code, not prompts
Whether the sender may ask for something is answered by a lookup against the system of record. The model never decides who is allowed to do what.
See it run
Pick a scenario and step through it. When the run reaches a decision, you make the call.
A corporate client asks to add a user to their payments portal. The agents do the work; a person still decides what gets sent.
Where it breaks
Rubber-stamping
If the drafts are almost always right, reviewers stop reading them. Track edit and reject rates per reviewer, and seed known-bad drafts occasionally to check the gate is still a gate.
The gate becomes the bottleneck
If review takes as long as doing the work, nobody wins. The draft has to arrive complete, with the evidence beside it, or specialists will go back to their inbox.
Treating "human in the loop" as the safety case
A person reviewing output does not fix a system that can call dangerous tools. Least-privilege tools and input screening still have to hold on their own.
In practice
I use this as the operating principle for agentic client-servicing work: agents classify, extract and draft, and a person sends every message.
The principle also gives risk and compliance partners one well-defined control to evaluate, instead of a promise that the model will behave.