Two years ago the word was chatbot. Then assistant. Now every vendor deck says agentic, and the products underneath have changed less than the vocabulary has.
The word is worth keeping, though, because it does mean something specific. And the definition is testable, which makes it useful when someone is quoting you £400 a month for something.
The three-part test
A system is agentic if it does all three of these. Two out of three is an assistant with good marketing.
- It decides the steps. You give it an outcome, not a sequence. “Chase the outstanding quotes” rather than “send this template to these six people.”
- It uses tools on its own initiative. It can read your inbox, query a database, call an API, and choose which to reach for based on what it found in the last step.
- It acts without you in the loop. Something changes in the world, a message sent, a record updated, a booking moved, without a human pressing send each time.
Most “AI agents” sold to small businesses do the first two and stop dead at the third, because the vendor knows that acting without supervision is where the liability lives. That is a reasonable product decision. It is not agentic.
The ladder, and where products actually sit
| Level | What it does | Who decides the steps | Typical example |
|---|---|---|---|
| Chatbot | Answers questions from a knowledge base | Nobody, it is reactive | Website FAQ widget |
| Assistant | Drafts and suggests, you approve | You | Email reply suggestions |
| Workflow automation | Runs a fixed sequence on a trigger | You, in advance | Form submission creates a task |
| Agent | Chooses the sequence and acts | The system, within limits you set | Triages the inbox and replies to the routine 70% |
Nothing on that ladder is better than the rows above it in the abstract. Workflow automation is the correct answer for a great many small business problems and costs a fraction of an agent. The mistake is paying agent prices for row three.
What agentic genuinely buys you
The jobs where it earns its keep share a shape: high frequency, moderate judgment, cheap to get wrong occasionally.
- Inbox triage with a reply. Not “summarise my email.” Sorting the enquiry from the invoice from the newsletter, answering the ones with a known answer, and flagging the rest with a draft.
- Chasing. Quotes, unsigned documents, overdue invoices. The work is not the message; it is remembering who is at which stage and when the last nudge went out.
- Reconciling two systems that will never integrate. The booking system says one thing, the accounts package says another, and someone spends Thursday morning making them agree.
- First-line scheduling. Filling a cancellation from a waiting list at 7am, when the alternative is an empty slot.
The common thread is that a human currently does these badly, not because they lack skill but because the work is interrupt-driven and nobody is at their desk at 7am.
Where it still breaks
Irreversible actions. Refunds, deletions, anything that touches money leaving your account. Keep a human on those, permanently. Not as a phase one caution to be removed later.
Sparse data. An agent handling your enquiries needs to know what you have said before. If your answers live in three people’s heads and a WhatsApp group, the agent will invent a house style, and it will be wrong in ways you only find out about from a customer.
Systems without an API. This is the quiet killer for UK small businesses. Plenty of trade-specific software has no integration surface at all, and an agent that cannot reach your job management system is limited to the parts of your business that happen in email.
Accountability. When an agent gets it wrong, you are the one answering for it. That is not a technology limitation, and no vendor term of service changes it.
Adoption numbers show the same split. The government’s AI activity in UK businesses survey ↗ has consistently found small firms far behind large ones, and the gap is not enthusiasm. It is that the integration work is the actual project, and always was.
The one question to ask a vendor
Ask to see the run log.
Every real agent produces a trace: it decided to do this, it called that, it got this back, it chose the next step. If a vendor cannot show you that view, one of two things is true. Either there is no agent underneath and you are buying a workflow with a language model bolted to the front, or there is an agent and you have no way to audit it. Both are reasons to walk.
The follow-up question is what happens when it fails. A good answer is specific: it stops, it flags, it hands the item to a named person with everything it had gathered. A bad answer is “it retries,” which mostly means it will make the same mistake more expensively. We covered how that shows up on the bill in what an AI agent actually costs a small business.
What to do this quarter
Pick one task that is boring, frequent, and reversible. Run it agentically for a month, with the run log open and one person responsible for reading it weekly.
Do not start with customer-facing work. Start with something internal where the failure mode is mild embarrassment rather than a lost client. And before you connect anything to your systems, decide what it is allowed to reach and how you revoke that access, which we walked through in giving an AI agent access to your business data safely.
If a month of that shows nothing, the honest read is that the job was workflow automation all along. That is a good outcome. It is a cheaper bill and a system that will still be doing the same thing correctly in three years.
At Forgd we build and run the software underneath this: the integrations, the boring reliability, and the person who reads the log when something looks off. Get in touch if you want a straight answer about whether your problem needs an agent or an afternoon of automation.