Most automation advice tells you how to build an agent. Almost nobody tells you which tasks deserve one. That's the expensive gap — because the wrong process automated badly costs more than the manual version it replaced.
So this week, a different angle: a decision framework. We call it the 3am Test.
What an autonomous agent actually does
Quick grounding first. An AI agent isn't a chatbot. A chatbot waits for you to ask. An agent watches, decides and acts — on a schedule or a trigger — then reports back or escalates when it hits something it can't handle.
The loop is simple: sense, decide, act, log. It reads an inbox, scans a tender portal, watches a folder. It applies rules and judgement. It does something — drafts, files, flags, replies. And it leaves a trail so a human can audit what happened.
That's it. The magic isn't the model; it's pointing that loop at the right problem.
The 3am Test
Here's the question we ask before building anything: if this task fired at 3am with nobody watching, would you be comfortable?
It sounds flippant. It's actually a sharp filter, because it forces three sub-questions.
1. Is the task repetitive and rule-shaped?
Good agent work has a recognisable shape. Same inputs, similar decisions, predictable outputs. Triaging email by urgency. Checking a procurement portal for new tenders matching your criteria. Turning a meeting note into a draft proposal.
If a competent new hire could learn the task from a one-page SOP, an agent can probably run it. If the task changes shape every time and depends on context no document captures, you're not ready — yet.
2. What happens when it gets it wrong?
This is the one people skip. Every agent will be wrong sometimes. The right question isn't "will it err?" but "how bad is the worst error, and how fast do we catch it?"
Drafting an email for review? Low stakes — a human reads it before it sends. Auto-paying an invoice with no approval gate? High stakes. The 3am Test passes when a mistake is recoverable and visible, not when the agent is perfect.
This is why we build approval steps into anything touching money, contracts or customers. The agent does 90% of the work; a person clicks 'send' on the 10% that matters.
3. Is the volume worth it?
Automating a task you do twice a year is a hobby, not a business case. Automating something you do 40 times a day — or that you should do daily but never get round to — is where the return lives.
The second half of that matters. The best agents don't just speed up work you're already doing; they do the work you keep meaning to do and never have time for.
Use cases that pass the test
Email triage
The classic. An agent reads incoming mail, sorts by intent and urgency, drafts replies for routine queries, and flags the handful that need you. You wake up to a tidy inbox and three things that genuinely need a human. Recoverable errors, high volume, rule-shaped. Clean pass.
Tender and opportunity monitoring
If you bid for work, you know the pain of checking portals manually — and the cost of missing a deadline. An agent watches the sources, filters against your real criteria (value, sector, geography, framework), and alerts you with a summary the moment something fits. We build this into our hi-tech procurement work because in defence and export, a missed tender is a quarter's pipeline gone.
Content engines
A repeatable pipeline: research, draft, format, schedule. The agent produces the first 80% on brand and on cadence; a human edits for judgement and signs off. Our marketing manager app runs exactly this kind of loop — consistent output without a content calendar that quietly dies in month two.
Custom builds
Most real value is bespoke. Pooled knowledge agents that answer staff questions from your own documents. Multi-step agents that chain several jobs together. That's the heart of our AI agent systems work — we map your actual process, find the 3am-safe parts, and build the loop around them.
When the answer is 'not yet'
Sometimes the honest answer is no. The process is too messy, the stakes too high, or the volume too low to justify the build. That's fine. We'd rather tell you a task isn't worth automating than sell you an agent that creates more cleanup than it saves.
Often the right first move is smaller: tidy the forms and workflow underneath the task so it becomes rule-shaped. Get the inputs clean, then automate. An agent on a chaotic process just produces faster chaos.
Start with one
Don't try to automate the whole business at once. Pick a single task that passes the 3am Test — repetitive, recoverable, high-volume — and build one agent that owns it end to end. Prove it. Then move to the next.
That's how operators do it. Not a moonshot; a series of small, auditable wins that compound.
Want a second pair of eyes on which of your processes pass the test? Book a call and we'll walk through your workflows and tell you straight which ones are worth automating — and which aren't yet.