The most expensive mistake in our diagnostic intake is the same one every time: a company has called us in to scope an “AI project” and the right answer turns out to be standard automation. The costs differ by roughly an order of magnitude. The timelines differ by months. And nobody in the original meeting wanted to hear that, because the board had already authorised the AI initiative.

This piece is the framing we use to head that off before anyone signs an SoW. If you’re trying to work out which technology fits your problem, work through this before you take a vendor pitch.

What’s actually different

Automation follows rules. You write the rules, the system executes them, and every output is predictable from the inputs. If a form field contains “X”, create a ticket of type Y and assign it to team Z. Run it a million times and it does exactly that, a million times. When it breaks, it breaks loudly and traceably.

AI handles things the rule-writer hasn’t seen. It generalises from examples, which means it works on inputs nobody anticipated — and it fails in ways nobody anticipated either. An AI system reads an invoice in a layout it’s never been trained on and extracts the field anyway. That’s the upside. The downside is that the same system, on a slightly different layout, might extract the wrong field with high confidence, and nobody on the ops team notices until quarter-close.

The core difference is whether your input space is bounded and predictable, or open and variable. Everything else is downstream of that.

Three honest cost comparisons

We’ve delivered both kinds of project. The numbers below are what we typically quote, not aspirational floors.

Invoice intake. A traditional automation that ingests structured PDFs from three known suppliers and pushes the extracted fields into your accounts system: roughly £8–15k, four to six weeks. The same problem with two hundred suppliers, mixed formats, occasional handwriting, and a confidence-scored escalation queue: roughly £25–45k, eight to twelve weeks. Same problem on the surface. Different problem underneath.

Customer query routing. A rule-based router that reads the subject line, checks against a keyword table, and assigns to the right team: £4–10k, three to four weeks. An AI system that reads the body, classifies intent, attaches similar resolved tickets, drafts a candidate response, and escalates the rest with full context: £20–40k for the build, plus ongoing API costs. Different ambitions, different price tags.

Inventory ordering. A reorder-point automation triggered by stock level dropping below a threshold: £6–12k, three to five weeks. A demand-forecasting model that factors in seasonality, promotions, supplier lead times, and historical accuracy: £15–30k, six to ten weeks, plus ongoing model maintenance.

In all three cases, the AI version is more capable. In all three cases, the automation version is enough if your problem is the simple shape. Picking AI when automation is enough is the most common form of overspending in our intake.

When to pick automation

You almost certainly want automation, not AI, if any of the following are true:

The inputs come in a fixed, predictable structure (forms, database rows, API payloads from a known set of upstream systems).
The rules behind the decision can be written down on a page and read aloud without controversy.
Exceptions are rare, well-defined, and have a clear escalation route already.
The cost of a wrong output is high enough that you’d want a human in the loop on every edge case anyway.
You need to defend every decision in audit, and a “the model thought this” answer won’t fly.

If you’re using a phrase like “robotic process automation,” “workflow automation,” or “integration platform,” what you want is almost certainly automation. The marketing slides will sometimes use the word AI to be fashionable. Read past it.

When to pick AI

You probably want AI if:

The inputs are unstructured (free text, images, scanned documents, audio).
The rules can’t be written down without filling a small book, or change too often for rule maintenance to keep up.
The variation is wide enough that rule coverage is patchy, and patchy is worse than statistical-with-confidence.
The work involves judgment that’s currently sitting in an experienced person’s head, and that experience is what you’re trying to scale.

The honest signal is: the work currently requires someone to read something — a document, an email, an image, a transcript — and form a view. Rules-based systems struggle with reading. AI systems are useful precisely because reading is what they do.

The hybrid is usually right

For most operational systems we build, the answer isn’t “use AI” or “use automation” — it’s both, in the right places. A typical pattern:

The automation layer handles the routine cases — anything where a rule reliably gets the right answer. The AI layer handles the variable cases, with a confidence score on every output. Cases below the confidence threshold go to a human queue with everything the system has already worked out attached.

The result is a system that’s cheaper than full AI, more capable than pure automation, and gives the team a clear escalation path that doesn’t depend on them spotting silent failures. Most of the financial services case study on our work page is exactly this shape — the agent handles the 40% of tickets that pattern-match cleanly, and the rest land on a specialist’s desk with a draft attached.

A real example, end to end

A manufacturer we worked with came in describing an “AI project” to handle inbound supplier emails. The brief: read every email, work out what it’s asking, route it.

We ran a diagnostic. Two findings landed:

The first 70% of the inbound traffic was reorder notifications from a fixed set of suppliers, in templates that hadn’t changed in years. That’s automation. We wrote a parser, integrated with their ordering system, total cost £9k.

The remaining 30% was genuinely variable — price changes, lead-time shifts, freight delays, occasional complaints about quality. That’s AI. A classifier reads the email, identifies the category, extracts the relevant fields, and routes with the original attached. £18k.

Total spend: £27k. The version where everything got handled by AI would have been £35–50k, slower to build, and would have applied the AI’s hammer to a nail that didn’t need it. The hybrid was cheaper and more reliable.

The decision framework

If you’re trying to work out which technology you actually need, ask yourself three questions in this order.

First, what does the input look like? If it’s structured and bounded, lean automation. If it’s unstructured or unbounded, lean AI. If it’s mixed, the answer is probably hybrid.

Second, what does failure cost? Rule-based failures are usually traceable and loud. AI failures can be quiet and confident. If quiet-and-confident is unacceptable in your domain (financial reporting, clinical decisions, legal positions), build the audit trail and confidence thresholds into the design from day one — and budget for them.

Third, what does the team need to maintain? Automation is maintainable by anyone who can read the rules. AI is maintainable by someone who can read accuracy and confidence metrics, retrain on drift, and have an opinion about retrieval strategy. If you don’t have that person in-house, plan for either a partner who’ll handle it or a simpler architecture you can run yourself.

The diagnostic shortcut

If you’re staring at this and unsure, the diagnostic is designed to answer exactly this question. Two weeks, fixed price, we look at the actual problem, recommend the simpler technology where it works, and refund the second week if the honest answer is to do something else first.

The cheapest version of an AI project is the one where we tell you it should have been automation.

Talk to us about your process.

AI vs automation: which one your business actually needs