The AI agent that works when nobody writes to it

When someone says "we've got an AI agent", they almost always mean the same thing: something that answers when a customer writes in. Message arrives, reply goes out. But a lot of the work that eats your week doesn't start with anyone writing to you. It starts with a date arriving, a deadline expiring, a cell changing in a spreadsheet. There's a different kind of agent for that — one that isn't waiting to be spoken to. It wakes up at the hour you set, does its job and goes back to sleep. That changes what you can hand off, and it completely changes how you know whether it's working.

Reactive or triggered: the difference in one sentence

A reactive agent waits for a message. Everything it does is tied to a person writing: a customer on WhatsApp, a web form, an email landing. If nobody writes, the agent may as well not exist.

A triggered agent waits for nobody. It has a trigger, and there are two kinds. The clock: every day at seven, every Monday, the first of the month. And the event: an invoice gets paid, an order comes in, someone flips a status to "needs review". When that happens the agent starts up, with no conversation open and no one at the desk.

The distinction sounds technical. It isn't. It's the difference between handing off customer contact and handing off the work nobody sees: the review, the reconciliation, the reminder that slips on a bad day. That second pile is usually the one eating the most hours and the one fewest owners automate — precisely because there's no customer standing there complaining about it.

What an unattended agent is actually good at

Not all background work qualifies. The stuff that works has a pattern: it repeats, you can describe it in one sentence, and the result fits in a list.

• Checking what's coming due. Contracts renewing, warranties expiring, quotes sent ten days ago with no reply, insurance. Anything with a date on it that nobody looks at until it's already late.
• Reconciling two lists. Orders against delivery notes, invoices raised against payments received, hours logged against hours billed. Mechanical, dull, and error-prone when a human does it at seven in the evening.
• Watching a threshold. Stock under X, a new review, a form sitting unanswered for two days, a regular customer who hasn't ordered in three months.
• Getting things ready. A draft quote with the details already filled in, a list of who to chase today, the client folder tidied before you open your laptop.

Notice what these share: none needs fine judgement, and in none of them is a customer waiting for an answer right now. That's exactly why they can happen at four in the morning without consequences.

The new failure mode is silence

Here's the shift that matters, and it's worth understanding before you build anything.

When a reactive agent fails, you find out fast. The customer gets a strange reply, or no reply, and shortly after you get a complaint. The failure makes noise.

When a background agent fails, nothing happens at all. That's the problem: "nothing happened" looks exactly like "nothing needed doing." The agent hasn't run in nine days because a connection dropped, and you're perfectly relaxed, assuming no contracts came up for renewal this week. You didn't notice, because there was nothing to notice.

Which gives you the one rule no unattended agent should break: it has to report even when it did nothing. Zero is a number. "Checked 34 contracts, none expiring this week" is information. Silence isn't, because silence means both things at once.

The morning report

The output of a background agent isn't the work it does — it's what it tells you. If the summary is poor, it doesn't matter how well the thing runs, because you'll stop opening it.

A decent report reads in twenty seconds on your phone over coffee, and carries four things:

1. What it looked at, and when. "Last night at 3:10 — 214 invoices, 34 contracts." That's the proof of life.
2. What it handled on its own. Already done, nothing needed from you.
3. What's waiting for you. Things prepared and queued for approval, with a link to approve in one tap.
4. What it couldn't work out. The odd cases, in plain English: "three invoices with no tax ID, couldn't tell who they belong to."

Plus one line at the top saying whether you need to do anything today. Most days the answer is no — and that's the whole point. You close the report and get on with your day knowing it's covered.

Prepared at four, sent at nine

A night-shift agent runs into a problem the reactive kind never has: at four in the morning there's nobody around to approve anything. And approval stays yours. You approve, it executes — that part doesn't move.

The clean way through is to split preparing from sending. Overnight, the agent does everything invisible from the outside: read, reconcile, sort, draft, stack the queue. In the morning, with the report in front of you, you release whatever needs to go out. The heavy lifting is done before you arrive; what's left is deciding, which is the part you're good at and it isn't.

One practical exception: anything purely internal and reversible — tidying a folder, updating a record, tagging email — can happen overnight without asking. Anything that leaves the building shouldn't. A message landing on a customer's phone at 3:15am shouts "a machine sent this" however well written it is, and burns the exact trust you were trying to build.

When this adds nothing

Three situations where a background agent is wasted effort, and they're worth saying out loud.

When the data it reads isn't updated automatically. If the list your agent checks at three in the morning gets filled in by hand the following afternoon, it's reading the day before yesterday. The agent isn't the problem — the information arrives late, and automating on top of that just gets you a fast, wrong answer.

When the trigger barely fires. If the deadline you want watched comes round twice a year, you'll see it coming unaided. Building and maintaining something for that costs more than it returns.

When you know you won't read the report. This is the worst of the three, because it's invisible. A background agent whose summary nobody opens is worse than nothing: it gives you the feeling that the job is covered when it isn't. If two weeks pass and you haven't opened it once, switch it off — the task was never the issue.

Where to start

Pick one task, the dullest you can think of, and run it in watch-and-tell mode: it reviews, it sends you the report, it touches nothing. Two weeks of that is enough to see whether it gets things right, and enough for you to build the habit of reading the summary. Once you've had ten reports in a row where you think "yes, that's what I'd have done", let it act on the reversible part. Then the next task, one at a time.

If you want to see what this looks like applied to back-office work — renewals, follow-ups and reminders prepared before you open your laptop — have a look at the operations agent.