Point an agent at a job, not just a client

"Review every VAT return still in preparation" and "write to every client who has one" are different questions, and they need different answers. Agents can now be targeted three ways.

An agent used to run against a saved client filter and nothing else. That is the right shape for plenty of work and the wrong shape for a surprising amount of it, because it fixes what one unit of work is — and that decides what the agent can sensibly say.

Three ways to target

• Clients matching a client filter — one answer per client, the original behaviour and still the default
• Jobs matching a task filter — one answer per job
• Clients that have a job matching a task filter — one answer per client, chosen by their work

Every agent your firm already has behaves exactly as it did; this adds options rather than changing the existing one.

Why the last two are different questions

"Review every VAT return still in preparation" is a question about jobs — one answer per return. "Write to every client who has a VAT return still in preparation" is a question about clients — one email, not one per return.

The same saved filter produces very different runs depending which you pick. One filter in testing read as three jobs in the first mode and one client in the second. The form tells you which it counted before you save, because that number is also what a run costs.

Looking at one job

In per-job mode the agent is handed a single task and asked about its state: its dates, who is on it, its subtasks and how many are done, the time booked to it. The client's wider context is still available, so it can place the job — but the instruction is blunt about the distinction:

You are looking at this ONE task, not at the client in general. If your answer would be the same for any other job on this client, you have answered the wrong question.

Clients, chosen by their jobs

In the third mode the client is the unit, but the jobs that selected them are carried onto the result and named in the instruction — so the agent knows why it is looking, and the results screen shows you the work that put each client on the list. Without that the agent drifts into answering about the client at large and never mentions the job that prompted the run.