It is 03:10. A process nobody has looked at in six weeks has just decided that fourteen invoices are duplicates and flagged them as such. There is no one awake to argue, so it keeps going: by 03:40 it has emailed eleven suppliers to tell them their invoice was already paid. At nine in the morning, the first person to find out will not be anyone on your team. It will be the supplier calling.
The thesis in one line: what can run overnight is not decided by how good the agent is. It is decided by whether that night’s mistake can be undone in the morning. Unattended execution is not a reward for accuracy, it is a bet on reversibility — and the bet is almost always placed backwards: what gets scheduled at 2am is whatever is most urgent, not whatever hurts least when it goes wrong.
Automating processes after hours with AI is not a question of trusting the agent
The pitch sells itself: the agent works while you sleep, does not charge a night premium, and by eight you have the work done. It is true, which is exactly what makes it dangerous. The sentence hides the fact that night is not just a cheap gap in the calendar: it is the only stretch of the day where a failure gets eight hours of runway with nobody watching.
An agent that gets it wrong at eleven in the morning and one that gets it wrong at three in the morning get it wrong equally well. What changes is not quality, it is how long the mistake has to compound before someone stops it. At eleven, somebody spots it in five minutes and halts the batch. At three, the mistake compounds: every badly resolved case feeds the next one, and the morning does not find a failure, it finds a state.
What that silence costs is not a hunch. In its eighth annual outage analysis, published on 13 May 2026, the Uptime Institute reports that 57% of respondents put the cost of their most recent major outage above $100,000, and one in five put it above $1 million — the second year running. That is a global infrastructure sample, not a small-business one: it sizes the problem, it does not promise a result of ours. But the same report flags what matters here: the leading driver of human error in 2026 outages is still failure to follow established procedures. At night nobody skips the procedure. What you have at night is a procedure nobody ever wrote.
The three questions that decide what stays on the night shift
You do not need a twelve-axis risk matrix. You need three questions, in this order — and the order matters, because the first one can veto on its own.
- Is it reversible? If tomorrow it can be undone in minutes without apologising to anyone, it is a candidate. If it leaves a state you have to negotiate your way out of — a payment issued, a purchase order sent to a supplier, an accounting period closed — it is not, however often the agent gets it right.
- Does it leave the company? An internal mistake is a mistake. A mistake that has gone out by email, WhatsApp or API to a customer, a supplier or a regulator is no longer a mistake: it is an incident, with its awkward conversation and its reputational cost. The outbound door is the natural boundary of the night.
- Would anyone notice before nine? If the answer is no, you are accepting an eight-hour blind window. Sometimes that is perfectly fine. But it has to be a decision said out loud, not the side effect of having put a cron job at 2am.
All three answer yes or no, and the result is not a score: it is a placement. Anything that answers "yes, no, yes" stays on the night shift. Anything that fails the first drops to preparation: it runs overnight, leaves the proposal ready, and a person clicks the button in daylight. That is a different thing, and almost always enough.
Is it reversible? The question that can veto on its own
Reversibility is the only one of the three that nothing else compensates for. An agent can have excellent accuracy, impeccable supervision and months of clean history: if the action cannot be undone, that accuracy only tells you how often you will have an irreversible problem, not whether you will have one. At enough volume, frequency stops being the interesting question.
It also helps to separate two things that get conflated daily: reversible is not the same as correctable. A journal entry is corrected with a reversing entry, but the trail stays and somebody will have to explain it at review. A sent email cannot be corrected at all — you send another one. The honest test is asking what it costs to get back to the previous state across three columns: time, money and conversations with third parties. If any of them makes you uncomfortable, that action does not run at night.
Where each action sits on the permissions ladder is laid out in the autonomy levels of an agent. What the clock adds is easy to state and hard to accept: night moves every action down one rung. What runs as "act and notify" in daylight runs as "leave it ready" at 3am.
The sort, with real processes
| Process | Overnight? | Why |
|---|---|---|
| Drafting replies to inbound email | Yes | Fully reversible: nothing leaves until a person sends it. The morning starts with the work done and the veto intact. |
| Sending those replies to the customer | No | It leaves the company and cannot be undone. Prepare at night, send by day. |
| Reconciling bank movements and proposing matches | Yes | The proposal lives in a queue, not in the ledger. Reviewing a hundred proposals is far faster than doing a hundred reconciliations. |
| Posting those matches to the ledger | No | It closes a state that later gets audited, and the saving from doing it at 3am is exactly zero. |
| Triaging and routing tickets that arrive overnight | Yes, with a caveat | Reversible and internal — as long as routing does not fire an automatic reply. If it does, that is a different question. |
| Ingesting, normalising and enriching data in your own warehouse | Yes | The canonical case: internal, repeatable and undoable by reloading. |
| Placing supplier orders on automatic replenishment | No | Money committed to a third party while nobody is watching. Prepare the proposal; the signature happens by day. |
| Generating the report somebody reads first thing | Yes | It decides nothing. It is the exact piece that makes the rest of the night auditable. |
The boring parts almost nobody ships: the night report and the handbrake
This is where nearly everyone stops, because the flashy part is done and what is left does not demo well. Two pieces, both cheap, and without them the sort above is worth nothing.
- The night report. A short summary somebody reads before coffee: how many cases were processed, how many are queued, what went off the rails and what got decided by default. Not a dashboard nobody opens: a message that reaches a named person and gets noticed when it does not arrive. Its real job is not informing — it is turning an eight-hour blind window into an eight-hour window with minutes.
- The handbrake. A rule that halts the batch by itself when the rate of errors, exceptions or volume spikes above a written threshold. A misaligned agent does not fail once: it fails the next nine hundred in exactly the same way. The difference between fourteen wrongly flagged invoices and fourteen hundred is literally this rule, and it takes an afternoon to write.
The handbrake has a non-obvious virtue: it forces you to write down what normal looks like. Most teams already running things overnight do not know how many exceptions their process produces on a good day, which is why they cannot recognise a bad one. Writing the threshold is the real work; stopping is just the consequence.
There is a third piece, and it is not technical: who picks up the phone if the handbrake trips at four. If the answer is "the team", there is no answer. How to build a realistic on-call rota in a small company — severities, one name per week, a one-page runbook — is in who answers when an automation goes down. Without that piece, the handbrake only ensures the problem waits for you, switched off, until nine. Which is much better than the alternative, but it is not a plan.
The three mistakes that keep repeating when scheduling the night
- Scheduling whatever is most urgent. Urgency pushes towards things with immediate consequence, which is exactly what handles eight unwitnessed hours worst. The right criterion pushes the other way: night takes whatever hurts least when it goes wrong, however unimpressive that sounds in the meeting.
- Confusing "it has not failed in months" with "it can run alone". A process that has not failed in half a year has not proven it is reversible; it has proven its turn has not come up yet. They are independent criteria and only one of them wins the night.
- Letting the night grow on its own. It starts with one task and three months later there are seven things in the scheduler nobody has revisited, depending on each other in an order written down nowhere. The night needs an inventory with the same discipline as a human shift: what runs, in what order, who approved it, and when it was last reviewed.
What to do this week
- List what already runs after hours. Not what you think runs: what is in the scheduler. Half the surprises live in that first inventory.
- Run the three questions against every entry. Anything that fails the first drops to preparation, no debate.
- Ship the night report for whatever stays up there. A named person, a fixed hour, four numbers.
- Write a stop threshold per process, even if the first one is a guess. A rough threshold that exists beats a perfect one that is pending.
- Put a date on the review. An unattended process with no review date is a process that degrades quietly until somebody finds it through a complaint.
None of this is the agent. It is the governance around the agent, which is where it gets decided whether unattended execution ends up a competitive advantage or an expensive way to discover problems in the morning. When volume outgrows anyone looking case by case, the next step is human supervision of AI at scale: sampling, exception queues and thresholds, instead of reviewing everything or reviewing nothing. And the general pattern for where the person sits inside the flow is in human in the loop.
The line worth bringing to the next argument about what to schedule at 2am: night is not for deciding, night is for preparing. Whatever decides, decides in daylight with somebody watching.