Posted on: 29 10 2026.

Your AI Agent Might Just Be a Very Expensive Timer Switch

COMTRADE 360 · AI, BRIEFLY DEMYSTIFIED

Your AI Agent Might Just Be a Very Expensive Timer Switch

Everyone wants one. Almost nobody can define one. A short, unserious guide to a word that has been stretched until it means nothing — and how to tell whether you actually need the thing, or just a list.

There is a moment in every management meeting in 2026 where someone says, with great confidence, “we should look at AI agents.”

Everyone nods. Nobody asks what one is. The nodding is very important, because the alternative is admitting you have been reading the same three LinkedIn posts as everyone else. So the meeting moves on, a line item appears in next year’s budget, and somewhere a consultant feels a disturbance in the force and begins preparing a slide.

I want to spoil this for you. Most of the things people call AI agents are not agents. Quite a lot of them are, in a technical sense, a timer switch with better marketing. This is good news. Timer switches are cheap.

The word has been stretched until it means nothing

“Agent” now covers everything from a chatbot on a website, to a script that moves files at midnight, to a genuinely sophisticated system that reasons about your supply chain. When a word covers that much ground it has stopped being a description and become a vibe.

So let’s take the vibe apart. There are three different animals hiding under the label, and they cost wildly different amounts.

AUTOMATION
The timer switch

Does one thing, the same way, forever. No judgement, no cleverness, no surprises.

Costs about as much as a decent office chair.

ASSISTANT
The goldfish with a PhD

Brilliant, fast, and it has no idea who you are the moment the conversation ends.

You are the engine. It is the passenger reading the map.

AGENT
The keen junior colleague

You don’t give it a task. You give it a job. Nobody has to remember to ask.

Never sleeps. Occasionally needs supervising.

Three animals, one word. Copilot and ChatGPT are the middle one. Most of what you want is the left one.

What the third one actually does all day

“Keep an eye on our late suppliers and chase them” is a job, not a task. An agent wakes up, works out which orders have slipped, decides whether each one is worth chasing, drafts the email — and then you approve it before it goes anywhere near an actual human being.

Then it writes down what it did, so tomorrow it knows. When a reply lands it reads it, updates the order and tells the buyer what changed. If nobody answers for a week it escalates to a person, rather than nagging into the void forever.

THE AGENTIC LOOP (4-Step Process):

1. LOOKS

What’s changed since last time?

2. DECIDES

Chase now, wait, or leave it alone?

3. DOES IT

Drafts the email, updates the order

4. CHECKS

Did it work? Did anyone reply?

↺ You approve, first — then round again, until it’s done.

The memory bit is not optional, and here is why

We once reviewed a system that had been proudly described as an agent. It did everything on the list except remember. It chased the same supplier eleven times in one morning.

“He has not sent a Christmas card since.”

An agent without memory isn’t an agent. It’s a very fast person with amnesia and access to your email account.

4 / 15

Tasks genuinely needed an agent

We went through fifteen of the things businesses most commonly ask us for — research leads, draft proposals, sort tickets, match invoices, close the month faster, chase suppliers, screen candidates, onboarding admin, answer HR questions, the usual parade.

4 out of 15 actually needed an agent. The other eleven were better served by a rule, a checklist, or an assistant that drafts while a person decides.

Greatest hits: things confidently sold as agents that are not

A man with a hammer sees nails. A vendor with an agent-shaped product sees agent-shaped problems — including, alarmingly, your payroll.

Claimed ‘Agent’ Solution

What It Actually Is

“AI-powered onboarding”

Accounts, laptop, sign-offs, a welcome email. This is a checklist. A magnificent checklist, possibly the best checklist in the building — but a checklist.

“An AI agent for month-end close”

This is a list of things that must happen before other things. Your grandmother had one. She called it a list.

“Intelligent invoice matching”

The matching part is arithmetic. Numbers either agree or they don’t, and a rule does it faster and without ever hallucinating a supplier. The exceptions are the interesting bit — about 8% of the volume and 100% of the pain.

“An agent that answers policy questions”

It looks things up in your own documents. That’s a search box with good manners. And if the policy isn’t written down, no amount of artificial intelligence will find it — because it does not exist.

The most valuable thing we find is usually embarrassing

Here is what we keep discovering, and it never makes anyone’s press release. The expensive problem is usually not that the work is slow. It’s that some of it quietly vanishes. An enquiry nobody picked up. A request that fell between two shifts. A quote never followed up because Dave was on holiday and Dave was, structurally, the process.

6% of customer email threads in one assessment had no traceable reply at all. Nobody knew. Nothing measured it.

That single number was worth more than every hour of typing we could have automated — several times over. The fix was not an agent. It was giving every job an owner and a status.

Nobody has ever won an innovation award for introducing a status column. It remains one of the highest-return things we find.

Fine. When do you actually need one?

You have an agent-shaped problem only when ALL THREE of the following criteria are true:

✔ AUTOMATIC START

The work has to happen without anyone remembering to start it.

If you always press go, you want an assistant — and you should pay assistant prices.

✔ MULTI-DAY RUNTIME

It runs over days, not seconds.

Chasing, following up, watching for something to change. Anything with the word “keep” in the description.

✔ DECISION-BASED

Each case needs a decision, not a rule.

If you can write the logic on a Post-it, write it on a Post-it. A rule will do it better and won’t need a subscription.

All three true? You have an agent-shaped problem. They’re real, they’re valuable, and when they’re the right answer they’re genuinely the right answer.

One question to ask anybody selling you one

“What’s the simplest version that would still help?”

A good supplier answers happily and immediately. A bad one gets a slightly haunted look. It is the fastest way to find out whether you are being diagnosed or sold to.

So — should you build an agent?

Maybe. Honestly, statistically, probably not the first thing you thought of — but quite possibly the second. The trick is finding out before the invoice rather than after, which takes about two weeks and considerably less money than building the wrong thing beautifully.

We’ll look at how your work actually happens, tell you which of your problems is worth spending money on, and say so plainly even when the answer is “this one’s a spreadsheet, keep your budget.” And when it genuinely is an agent, we’ll build you a good one — with a memory, a human on the door, and a very clear list of things it is never allowed to do.

Contact Comtrade 360 to help you build your first agent

Even if it turns out to be a timer switch. Especially then.

Comtrade 360 · Agentic AI Process Transformation Assessment | Examples are illustrative and anonymized