A decade of drift: where to start when you're asked to modernise an automation estate?

The client came to us with a list of things they'd heard about.

Agentic AI. Copilots. Digital workers. Orchestration. Large language models reading the documents nobody wants to read. Machine learning making the calls people currently make. AI woven through the whole operation, end to end.

They wanted to "modernise" their automation estate.

But what does that actually mean? And where do you start?

Modernisation is one of those words that everybody uses but nobody defines, and in an automation context it can legitimately mean rebuilding a process, moving it to a different platform, replacing it with an integration, wrapping an agent around it, or switching it off entirely. Five different answers, five different budgets, five different outcomes.

Without a definition, "modernise the estate" is a direction of travel, not a plan.

This is the first of a five part series on modernising an automation estate that had been running for over a decade. It covers where we told them to start, why that answer surprised them, and what we found when we checked the one word everybody had taken for granted.

Start at the beginning

Our answer was just common sense.

Nobody opens a book at the last chapter and expects it to make sense. You'd know what happens without knowing why any of it mattered, and you certainly couldn't tell anyone whether it was a good ending.

Look at the shape of the question they started with. Agents, copilots, AI across the operation. Every statement is a solution, and they were being weighed before anyone had settled what problem they were solving. They had opened at the last chapter.

Technology can be the solution, but it shouldn't become the problem. Which is obvious right up until the moment a new capability arrives, and with it the pilots, the POCs, the MVPs and all the other pressures to be first to use the shiny new toy.

So the question isn't "where do we put AI?"

It's "what have we got, what is the value, and how would we do it differently now?"

Answer those three and the technology question answers itself, because you'll know which processes deserve investment and what each one actually needs. Skip them and you'll bolt new capability onto processes that shouldn't exist, and call it modernisation.

A decade-old estate is a record of decisions, not a set of assets

This is the part that reframes the whole exercise.

Every automation in a mature estate encodes the conditions that existed on the day it was built. If a system had no API, you drove the screen. If an integration wasn't funded, you automated around it. If a step needed judgement, you kept a person in the middle of it.

Every one of those was correct at the time. None of them were revisited, because working automations don't ask to be revisited.

Until you know which of those decisions still hold, you can't tell where any of the new capability belongs.

That's why "start at the beginning" isn't a delaying tactic. The estate is the map. Without it you're guessing at where agents help, and guessing expensively.

How we set about answering it

The brief was one line. Turning it into something anyone could act on meant deciding what we'd have to establish first.

Four things:

And from that, a target architecture and a decision on every individual automation:

Five options if you count remediation. Most estate reviews only ever price the last one.

Then we checked one word

We'd been told the automations were running.

We started there, because it was the cheapest thing to verify and the most dangerous thing to assume.

They were running. The completion rate across the estate was 96.4% over more than 11,000 runs. On paper, healthier than most estates I've seen.

They were running. That turned out to be a very different thing from working.

What was underneath that number is the subject of the next article, and it changed the shape of everything that followed. For now the point is narrower: the word "running" had been doing an enormous amount of unexamined work in every conversation about this estate, including the ones about what to do with it next.

The question that actually matters

Not whether an automation is running.

Where the value is, and how that value is best delivered today.

Those are two separate questions and they need separate answers. Value tells you whether something still deserves to exist. Delivery tells you whether a bot is still the right mechanism, or whether an API, an integration, an orchestration layer or an agent would now do it better.

An estate review that only asks the first produces a retirement list. One that only asks the second produces a rebuild plan. You need both, per automation, to produce a roadmap anyone can fund.

This is not a platform replacement

Worth being explicit, because "modernisation" gets heard as "migration", and migration gets heard as eighteen months, a business case and a risk register.

The platform was sound. Current, supported, doing what it should.

The question was never whether the platform is right. It was whether each individual thing running on it is still the right answer - and that's something you decide process by process, and act on process by process.

Nothing was paused for the assessment. No instrumentation added, no downtime taken, no service interrupted. Everything came from data the platform was already producing, while the estate carried on running normally.

In summary

When the question is "where do we start with AI?", the answer is almost never the technology. It's the estate you already have, what it's genuinely worth, and which of the assumptions underneath it are still true.

A well-maintained estate and a modern estate are different things, and the first can hide the absence of the second. Ten years of continuous support is a real achievement. It still leaves you with the architecture you chose a decade ago.