AI agent or plain automation: the £20,000 question
« We’d like an AI agent to move orders from the website into the system. » No: you want a connector. It will cost ten times less, never get it wrong, and give you nothing to proofread.
The three-question test
- Are the inputs structured? A form, a database row, a webhook — yes. An email written by a person, a scanned PDF — no.
- Can the rule be written down? « If amount > £5,000, route to the director » can. « If the customer sounds annoyed » cannot.
- Would a mistake be visible? A failed transfer shows up. A poor summary does not — which is exactly why agents need more framing.
Three yeses: automation. A no on either of the first two: agent. A no on the third: agent with mandatory review.
What the mistake costs, both ways
Putting a model where a rule would do means paying per token forever, adding a source of error, and proofreading an output that could have been correct by construction. We regularly see five-figure projects for what twenty n8n nodes do better.
Putting a rule where reading is required costs differently: a library of conditions that grows with every edge case, that nobody dares touch after a year, and a coverage rate stuck around 60%.
The test is not « can AI do this », it is « does this require understanding ».
The common case: both
Most systems we open are mixed, and the assembly is what matters. The automation carries the circuit — trigger, tool calls, writing into the line-of-business system, error handling — and the agent steps in only at the precise point where something must be read or written.
An example: a quotation request arrives by email. The circuit files it, creates the job, notifies. The agent reads the request and finds the comparable job — that is its only task. Then the circuit resumes, pre-fills the template and waits for human approval.
Two benefits: the cost per item collapses, and when something breaks you know which of the two is at fault.
And the tool?
It comes last, and it is reversible. n8n, Make, code: the choice depends on who will maintain the circuit in six months, not on what is fashionable this quarter. A team that can read a visual flow keeps control; a team with developers will prefer versioned code.
What is not reversible is having handed a decision to a machine. That is where to be demanding — not about the logo on the tool.
- The ten systems of a company — the survey grid
- AI for manufacturing — quoting, a textbook mixed case
- AI for logistics — the receiving discrepancy, also mixed
- The survey of your ten systems — free, delivered within 72 hours
Other notes
Your document base, or why the agent answers beside the point
An agent is only worth what it can read. The four defects that produce answers beside the point — duplicates, undated versions, scanned PDFs, implicit context — and the preparation that fixes them.
Responsible AI · 14 August 2026« What if it makes things up? » — getting an agent to cite its sources
A model produces a plausible sentence, not a true one. The four designs that make invention visible, and the single rule that matters: every claim points to its source.
Systems · 10 July 2026Getting cited by ChatGPT and Perplexity: what actually works
Generative engines cite what they can read, structure and verify. Five concrete moves — including serving your pages as Markdown — and what is a waste of time.