Every automation pitch aimed at recruiters promises the same thing: more output. On a specialist desk, more output is rarely the constraint. Judgement is. So the useful question is not what can be automated, but what should be.
The test
Automate the work where being wrong is cheap and recoverable. Keep the work where being wrong costs you a relationship.
That single line resolves most cases.
Safe to automate
Extraction. Pulling names, companies, dates, credentials, and stated requirements out of documents you have already received. Getting this wrong is visible immediately and costs nothing.
Retrieval. Surfacing what you already know about a person or an account at the moment you need it. This is the highest-value automation on any specialist desk and the least glamorous.
Formatting and repurposing. Turning one conversation into a candidate summary, a client update, and a market note. The thinking already happened; only the shaping is repetitive.
Ordering. Deciding what is worth your attention first, based on what is still recent. A ranked list is a suggestion, not a decision.
Never fully automate
Sending. A draft that goes out without a human read is how a specialist desk loses a decade-long relationship in one afternoon.
Judgement about people. Whether someone is right for a role, whether a client is serious, whether now is the moment. Nothing observable in the data reliably answers these.
Anything unattributed. If a system tells you something about a person and cannot say where it came from, it is not automation, it is invention. Holi never asserts a fact without a visible source line, for exactly this reason.
The honest middle: drafting
Drafting sits between the two lists and is where most of the value is. A first draft written from your own captured notes, in your own patterns, saves the blank-page tax without removing the decision.
The design rule we hold: the draft is always editable, always attributed, and never sends itself. Your rewrite is the most valuable feedback there is, because it shows the gap between what a machine inferred and what you actually meant.
A sensible order to adopt in
- Capture, so nothing you read is read twice
- Retrieval, so the desk remembers on your behalf
- Repurposing, so one conversation produces several artefacts
- Drafting, once the first three are feeding it real material
Skipping to step four is why most automation projects on specialist desks stall. A draft written from nothing is a template, and recruiters can already write templates faster than they can fix bad ones.
What you should feel after a month
Not that you are sending more. That you are opening fewer tabs, retyping less, and starting conversations that are already half informed. If automation makes you louder rather than sharper, it is pointed at the wrong work.