AI automation for business

How AI automation differs from ordinary automation

The difference is not that one is more modern than the other. It is what happens when reality does not match the rule.

Classic automation executes rules written in advance. With valid data and tested rules, it produces consistent results: if an invoice carries its number in the same place as the last one, it will read that number the same way every time. Meet an invoice laid out differently, or one missing the information it needs, and it will either stop or copy the wrong thing — consistently, until somebody updates the rule. That is why these solutions also need data checks and a defined way of handling exceptions.

AI automation handles that case: it recognises what a document is even when nobody has described that layout before, and makes its decision based on similarity rather than a fixed rule. There is a definite price for this — the output no longer follows a fixed rule but the model's judgement, so it is harder to predict when it will be wrong. An AI automation project that has not settled in advance what happens on a mistake is not a project ready to go live.

In practice the good solutions combine both: rules wherever rules suffice, because they are cheaper and predictable, and an AI model only where the rules cannot be written down.

Which processes qualify

A good first candidate meets five conditions at once. Missing one does not rule the project out, but it changes the scope — and that is better known before the quote than after it.

• It recurs often — tens of times a month or more. Integration cost is largely fixed, so an infrequent process rarely justifies it.
• Its rules are reasonably stable. If two people do it differently today and both believe they are right, somebody has to settle which one is. Automation will not settle it — it will entrench one of the versions.
• The input data is available in a form a computer can process, not only in someone's head. This is the condition that most often turns out not to be obvious.
• A mistake is detectable and reversible. For a first project, avoid any process where an error only surfaces at the customer or the regulator.
• It has an owner — a person who can resolve questions and cares about the result.

The data test takes one afternoon and is worth every hour of it: try to export the last hundred instances of the process as a single table. Not describe how you would — actually do it. Whatever you run into is the real scope of the first stage.

Where the human stands in the process

This is the decision with the largest effect on cost and on risk, and it is often taken by accident, late, during roll-out. In practice there are three options: a human approves every output, a human handles the exceptions, or oversight consists of monitoring indicators and analysing them statistically.

We start with the first even when the third is the goal. Moving between them is far easier than retreating after an incident, and a few weeks of human approval produces quality data that cannot be obtained any other way. The full split of decisions — what we settle and what stays with you — is set out in our implementation methodology, and the full set — together with the things we will not build — on what we don't do.

What we do, and what we do not

Automation is often sold as something that always pays for itself. It does not, and it is worth knowing when we advise against a project:

• When the bottleneck is not the work the system would take over. Sometimes cases are not held up because nobody is processing them, but because they are waiting on a decision, or on a form nobody fills in. Automation then speeds up the part that was never the problem.
• When the process has no settled rules. That is a job to do before the project, not during it.
• When a product already does it. Do not build what you can subscribe to — we will say so even when it means there is no engagement.

What it costs

Three things drive the cost: how many systems have to be joined up, the state of the data, and how reliable the results must be, that is, how rarely they may be wrong. Company size barely matters — a small business with one tidy process will pay less than a larger organisation running the same process across four tools.

The line most often left out is upkeep: per-operation fees, quality monitoring, and periodic retesting, because when a provider retires the AI model version the solution uses, it has to be retested. We break all of those components down in our note on what an AI implementation costs.

The team that will live with it

When nobody in the business understands an automation, it quietly loses quality: a document format changes, quality drops, and nobody notices for a quarter. Handover therefore includes explaining how the thing works, where it is unreliable, and how you would know it had broken.

If AI systems are already running in your business, note that the duty to support AI literacy among the people operating them comes from the AI Act itself and has applied since 2 February 2025. What that means in engineering terms is set out alongside our AI Act technical compliance sprint.

How we start

Automation is one stage of a wider implementation and follows the same rules: we pick one process and measure one indicator before we start. Then we run the smallest possible pilot. Before switching over, the system runs for a while alongside the current way of working. The full scope — from process audit through to upkeep — is set out on our AI implementation page. If you would rather work through the first steps yourself before we talk, our guide to how to implement AI in your company covers them.

Frequently asked questions

How does AI automation differ from ordinary automation?

Classic automation executes rules somebody wrote down in advance — if the invoice has this field in this position, copy it there. AI automation copes with cases where the rules are incomplete: the document is laid out differently from the last one, the customer message fits no category, the description is ambiguous. The price of that flexibility is that the output is no longer guaranteed, only probable — which is why the project has to settle up front what happens when the system gets it wrong.

Which processes are best suited to AI automation?

Ones that recur often, follow reasonably stable rules, have input data in a form a computer can process, and fail in ways that can be spotted and undone. Most often they are text-based: working out what enquiries are about and routing them on, reading documents, drafting replies for approval, filling in records in a system. A process run five times a month rarely justifies the cost of an integration, because that cost is largely fixed.

What does AI automation cost?

Cost is driven by the number of systems that have to be joined up, the state of the data, and how reliable the results must be, that is, how rarely they may be wrong — not by the size of the company. The same process in a business with its data in one system, and in a business with that process scattered across mailboxes and scans, are two different projects and two different quotes. We break the components down in our note on what an AI implementation costs; a range quoted before anyone asks about your systems and data is a range pulled from the air.

Does AI automation mean redundancies?

In the projects we run, usually not — and not because that sounds better, but because the first option we recommend requires a person to approve every output. That person is the source of the quality data the system needs. The realistic early effect is shorter handling time and less copying work, rather than fewer roles. We also help plan what work the time freed up will go to. Employment decisions belong to your company. If you intend the system to run with less human involvement, it is worth saying so at the start, because that changes how reliable the results must be, and therefore the cost.

How long does an automation project take?

A single process with one integration and legible data is a matter of weeks. A process touching several systems, with data that needs tidying and with oversight requirements, is a matter of months. The single longest item in the schedule is frequently obtaining access to the systems the solution must read from — so start requesting it in the same week you pick the process.