Why we made our AI copilot ask before it does anything

The AI inside Recruitly does not send emails. It does not text candidates. It does not move a job through the pipeline, change a salary field, or book an interview. Not on its own. It writes all of those things, shows them to you, and waits. You read what it produced, change what needs changing, and press confirm. Only then does anything actually happen.

People assume that is us being cautious. A bit timid with the AI. It is the opposite. Letting the copilot draft everything and commit nothing was the most deliberate decision we made when we built it, and we would make it again tomorrow.

It is the kind of decision that sounds obvious once it is said and is anything but obvious while you are building the thing, when every AI demo you watch has the copilot firing off actions by itself and looking clever doing it.

An AI that acts on its own is an AI you check twice

The pitch for an autonomous copilot is speed. It reads the call, updates the record, sends the follow-up, and you never lift a finger. That is the demo. It looks like magic for about a minute.

Then you think about the first time it is wrong. The AI heard the candidate say their notice was four weeks. It was fourteen. It has already updated the field and already told the client the candidate is available next month. Nobody saw it happen. The mistake is now sitting in your CRM and in your client's inbox, and the first you hear of it is when the client calls to ask why the dates do not add up.

A sensible recruiter does the only thing they can with an AI that acts on its own. They check everything it did. Every email before it goes, except it already went. Every field before it saves, except it already saved. An AI you cannot trust to be right every single time, acting without you, does not save you work. It moves the work from doing to auditing. And auditing someone else's mistakes after the fact is slower than just doing the job yourself.

Draft everything, commit nothing

So we inverted it. The copilot does all the work right up to the last step and then stops. It drafts the email to the candidate. It writes the call summary and the notes. It proposes the salary update, the new notice period, the next action, the pipeline move. It assembles the shortlist. Then it hands all of that to you as a draft and waits for a yes.

You read it. Most of the time it is right and you confirm in two seconds. Sometimes it is close and you fix one line. Occasionally it has the wrong end of the stick entirely and you bin it. The point is that the recruiter is the one pressing the button, every time, on anything that touches a record or leaves the building.

This is exactly how the call summaries work in the phone system we are building. The call ends, the AI has already written the summary and the field updates, and you read it and confirm. The AI did the typing. You did the deciding. That split runs through everything the copilot touches.

The confirm step is where your judgement goes

There is a temptation to read "the recruiter confirms everything" as friction. One more click on every action. It is not friction. It is the entire point.

A copilot is right most of the time. Call it nine times in ten. The trouble is always the tenth. It misheard a number, it matched the wrong candidate to the role, it took a throwaway comment on a call as a firm commitment. Nine-in-ten is excellent for a draft and completely unacceptable for something that sends itself to a client. The confirm step is the half-second where a human who knows the candidate, knows the client, and was actually on the call catches the one in ten before it does any damage.

That half-second is cheap. The wrong CV landing in your best client's inbox is not.

Reversible by default

The second rule sits underneath the first. Anything the AI does, you can undo.

We learned this building the Data Agent, which runs across hundreds of thousands of candidate records and refreshes them. An agent working at that scale will get some records wrong. So every change it makes is tracked and reversible. You can see exactly what it changed and roll any of it back. We carried the same rule straight into the copilot. A confirmed action is logged, attributed to whoever confirmed it, and undoable. The confirm step stops most mistakes before they happen. Reversibility cleans up the few that slip through.

Between the two, there is no AI action anywhere in Recruitly that a recruiter cannot see, did not approve, and cannot reverse. That is the floor we hold the whole system to, and we do not make exceptions to it because a feature would demo better without them.

Yes, this is slower. On purpose.

I will be honest about the trade. A copilot that drafts and waits is slower than one that fires on its own. There is a click that a fully autonomous system would not have. For a single action, in that single moment, the autonomous version wins.

It wins right up until it sends the wrong thing. Then the autonomous version is dramatically slower, because now you are apologising to a client, correcting a record, and rebuilding trust in a tool your whole team has quietly stopped believing in. Speed you cannot trust is not speed. A team that has started double-checking the AI in secret, because it burned them once, has lost every minute the AI was ever supposed to save.

The draft-and-confirm version is a fraction slower on every single action and far faster across a quarter, because nobody is auditing it and nobody is cleaning up after it. AI that actually helps you bill has to be AI your recruiters trust enough to lean on every day. And they only lean on it when they are the ones pressing send.

Where we do let it run on its own

There is a boundary, and it is worth being precise about where it sits. Not everything needs a human pressing confirm. Some things should just happen.

The test is two things. Whether the action leaves the building, and whether it is hard to undo. An email to a candidate, an SMS, a CRM field a colleague will rely on tomorrow, a job moving stage. All of those either leave the building or get leaned on by someone else, and all of those wait for you. Background work that stays inside the system and is fully reversible does not need to interrupt you. Flagging two records that look like the same person. Suggesting a candidate might suit an open role. Tidying inconsistent formatting overnight. Those run on their own and surface their results for you to act on when you are ready.

The line is not AI bad, human good. The line is that anything outward-facing, or relied upon by a person, waits for a person. Everything else can hum along in the background where a mistake costs nothing and fixes itself.

The copilot does the work, you make the call

That is the whole philosophy on one line. The copilot does the work, you make the call. It drafts, you decide, and nothing it produces reaches a candidate, a client, or a colleague until you have said yes. It costs a click. That click is the reason recruiters actually let it near their desk instead of switching it off after the second time it embarrassed them.

If you want to see how the copilot drafts and waits across the CRM, sourcing and calls, book a demo and we will walk you through it. Or start a trial and try saying yes to it yourself.