Automation and agentic AI · One client at a time · London & remote

Most of what slows a business down does not need AI.

Some of it does. The work is telling the two apart, then building both properly.

The first engagement Automation or an agent Enquire
§ 01 · The second channel
What it is, next to the training

Map how the business works. Then build the systems that take the manual work away.

Training changes what your people can do. This changes what they have to do at all.

A business runs on a long tail of small manual steps. The same details keyed into a second system. A file renamed, moved and attached. A reminder somebody has to hold in their head. A weekly figure assembled by hand from three places because nothing adds it up. None of it is difficult, and all of it takes hours.

So I map those steps, agree the map with you, and build the systems that remove them. Some of that is plain automation. Some of it is an agent. Most of the value is in knowing which is which.

§ 02 · Automation or an agent
The question that decides what it costs

Fixed rules and tidy data need a script, not an agent.

An agent earns its place where the work has several steps, real judgement, and enough variation that a fixed process cannot cope. Everything else is over-engineering, and over-engineering is how budgets get spent on things nobody uses.

Plain automation is quicker to build, cheaper to run, and it does the same thing every time. That last part matters more than it sounds. A script that always does the same thing can be checked once. Something that decides for itself has to be watched.

The three things it could be, and how to tell
What the work looks like
What to build
Fixed rules, tidy data, the same steps every time
Plain automation
A few fields pulled out of a document, or moved between two systems
A simple, well-built tool
Several steps, real judgement, and enough variation that a fixed process cannot cope
An agent
Only the last row needs an agent. The first two are quicker, cheaper and steadier without one.
Almost everyone selling AI starts at the bottom row and works backwards. Starting at the top is cheaper for you and slower for me, which is why it is worth saying out loud.
Two ways to count the same list, and both are true
Count the stepsAutomation wins
The great majority of what a business does by hand is deterministic. Same input, same rule, same answer.
It belongs in plain automation, and that is where most of the list ends up.
Most of the steps, and the cheap half to fix
Count the hoursThe judgement work wins
The steps that need judgement are fewer, and they take longer. The exceptions, the awkward cases, the decisions nobody wrote down.
That is where the owner's time goes, which is why it is worth building for even though the list is short.
Fewer steps, more hours
Count the steps and automation is nearly all of it. Count the hours and the judgement work is where the week goes.
A proposal that only counts one of these is selling you something. The order to build in falls out of holding both at once.
§ 03 · The first engagement
Half a day on site, then a record you sign off

Nothing is recommended until you have agreed the record of how things are.

01 · HALF A DAY, ON SITE
The map.
I sit with the people who do the work and map how the business actually runs, rather than how it says it runs. Who does what, which systems they touch, in what order, and where the manual work sits.
02
The written record.
The current state in writing: the people, the systems, the flows step by step, and every point where somebody is doing by hand what a system should be doing. You read it, you amend it, and you sign it off.
03 · ONLY THEN
Recommendations.
What to change, in what order, and what each one takes. Plain automation where the rules are fixed. An agent only where the work needs one. And the parts where the fix is the process itself rather than any software.
04
A phased build.
Built in phases, smallest useful piece first, each phase in use before the next one starts. Everything has an owner, a way to switch it off, and a person signing off the parts that matter.
The sign-off is the point. A recommendation built on a misunderstanding is worse than no recommendation, because it looks like a plan and gets funded like one. So the record of the current state comes first, and you are the one who says it is right.

This is not the free self-assessment, which is twenty questions you answer about yourself, scored on screen, costing nothing. This is chargeable work, and the written record is yours whether or not anything gets built afterwards. Where the work is spread across channels that nothing records, the fuller practice on the method page applies.

§ 04 · What the map finds
The step that looks like waste

A workaround is usually somebody's defence.

What looks like wasted effort in a business is usually a defence somebody built around a system that does not do its job. The duplicate spreadsheet. The second check. The note kept on the side because the field does not exist. Someone got caught out once, and has been protecting the business ever since.

So the work is not to delete the step. It is to find the requirement inside the workaround and satisfy it properly. Delete the step and the requirement is still there, unmet, and the person who built the defence is now working without one.

§ 05 · The two channels
How this sits next to the training

One question decides which you need: does the work need a person at all?

The work needs a person

Your judgement runs it the whole way through. What is slow is how the job gets done, not what the job is. That is the training channel: your own workflows rebuilt on Claude with you, and you taught while it happens.

The work should not need a person

It is repeatable, it crosses systems, and it has been taking hours from people who should be doing something else. That is this channel: mapped, agreed with you in writing, then built.

Most firms need some of both, and the map is what says how much of each. Start with an enquiry, or answer twenty questions on the free self-assessment first.