Here's a test we use with clients before we build anything: would this task survive if it ran at 2am with nobody watching?
If the answer is yes, it's probably a good candidate for an AI agent. If the answer makes you wince, you've just learned something important — and saved yourself a bill.
Most advice about AI agents starts with the technology. We'd rather start with the decision: which processes deserve an agent, and which ones you should leave well alone. Get that wrong and you automate a mess faster. Get it right and you free up real hours.
What an agent actually does (briefly)
An AI agent is software that takes a goal, breaks it into steps, uses tools — your inbox, a database, a web search, an API — and works through them with limited supervision. The language model is the brain. The tools are the hands. The guardrails are what stop it doing something daft.
That's the whole trick. A chatbot answers. An agent acts. The difference matters because action has consequences, and consequences are exactly why you need a framework for what to hand over.
The four questions that decide it
Before we build, we run every candidate process through four questions.
1. Is it repetitive and rule-shaped?
Agents thrive on tasks you do the same way, over and over, with decisions that follow patterns. Triaging inbound email, screening tender portals, drafting first-pass content, chasing missing form fields — these have shape. A one-off judgement call that depends on a hallway conversation does not.
2. Is the input messy but the output structured?
This is the sweet spot. Agents are brilliant at turning chaos into order. A hundred unread emails become a prioritised list with draft replies. Forty tender notices become three that match your capability. A pile of meeting notes becomes a clean summary with actions. If your input is unpredictable and your desired output is tidy, you've found a job worth doing.
3. Can you tolerate — and catch — a mistake?
This is the question people skip. An agent that drafts a reply for you to approve is low-risk. An agent that sends irreversible payments at 2am is not. The honest move is to match the agent's autonomy to the cost of being wrong. Early on, keep a human in the loop. Widen the leash once it's earned trust.
4. Does it run often enough to be worth building?
Automating something you do twice a year rarely pays back. Automating something you do forty times a week almost always does. Volume is what turns a clever build into a boring, reliable asset.
Where this plays out in practice
Email triage. Not "AI writes all your emails" — that's the hype version. The real win is an agent that sorts, labels, flags the three messages that actually need you today, and drafts replies to the routine ones. You still hit send. You just stop drowning.
Tender monitoring. If your business lives on contracts, an agent can watch the portals, filter by what you can actually win, and surface the live ones with deadlines attached — before your competitors have finished their morning coffee. We've written plenty about how we approach this in our AI agent systems work.
Content engines. An agent that takes your source material — case studies, specs, past posts — and produces consistent first drafts across channels. The agent does the heavy lifting; your team edits for voice and accuracy. That's the right division of labour, and it's the model behind our marketing manager app.
Forms and workflow. The quiet killer of small-business productivity is the half-finished form, the email attachment that should've been a database entry, the follow-up nobody chased. An agent closes those loops. See our forms and workflow automation for the specifics.
Custom builds. Sometimes the process is unique to you — a pricing workflow, a compliance check, a client onboarding sequence that touches six systems. That's where a bespoke agent earns its place rather than bending an off-the-shelf tool to fit.
The traps worth naming
Two mistakes come up again and again.
The first is automating a broken process. If your tender response is slow because nobody owns it, an agent won't fix the ownership problem — it'll just produce the wrong output faster. Fix the process on paper first, then automate the clean version.
The second is over-trusting too soon. New agents should start supervised. You watch what it does, correct it, tighten the rules, and only then let it run quieter. The teams who skip this step are the ones who end up with horror stories.
Start with one task, not a transformation
You don't need an "AI strategy." You need one annoying, repetitive, high-volume task off your plate — proven, trusted, and running in the background. Then the next one.
We're operators, not theorists, so that's genuinely how we build: smallest useful agent first, real results, then expand. If you want help spotting which of your processes passes the 2am test, book a call or get in touch and we'll walk through your actual workflow — not a demo.