How to Prioritise AI Automation: Knowing What Comes First

You prioritise AI automation by putting each step after the things it depends on, rather than starting with whatever hurts most. That sounds obvious written down, and it is where most plans quietly go wrong. Ask a leadership team what they could automate and the list arrives in minutes. Ask what should come first, and what should wait, and the room goes quiet.

That quiet is the reason we built a free AI roadmap tool and launched it earlier this month. Over the past year I have watched capable teams stall at exactly this point, with plenty of ideas and no confidence about which one to pick up on Monday morning.

This post is about the order. It covers why the order is the hard part, why no two businesses share one, and why the roadmap you write this month should look different by the spring. If you want the shape of a first 90 days, our 30/60/90-day template covers it. What follows is about what goes into that shape for your business in particular.

Everyone has the list, nobody has the order

The list is easy because it is a list of pains. Sales would like the lead research done for them, finance would like invoicing to run itself, and operations would like the quoting backlog gone. Every item is real, and every item has someone in the business who will argue for it.

The order is hard because the items depend on each other in ways the list cannot show. Automating quotes needs the quoting process written down somewhere, the customer and job data kept in one place, and systems that will actually connect. None of that appears on the list, because nobody has ever felt it as a pain. So the natural instinct is to start with the biggest problem, and the biggest problem often turns out to be sitting on top of two smaller jobs that nobody wants to do.

The government's own guidance for its digital teams is blunt about this. Prioritising what you should work on and in what order is an important part of delivering good products and services, it says, and the phrase that matters there is "in what order". A backlog of good ideas is not a plan until somebody has decided which one goes first and why.

The starting points also vary far more than most AI coverage suggests. The ONS's latest look at AI in UK businesses found that among businesses with 10 or more employees, over half of businesses in information and communication (58%) reporting using AI, compared with much lower levels in construction (13%). Those are self-reported figures from a survey that leaves out some sectors entirely, but the gap between them is the point. Two leadership teams reading the same article about AI agents are rarely standing in the same place.

Same 14 questions, two different first steps

The clearest way to show this is to put two businesses through the same questions. The roadmap tool asks 14 of them, about how the business works and where the time goes, and it builds the roadmap from a library of recommendations we wrote from real projects. Nothing in it is generated on the spot. Below are two of the example businesses we use to test it. They are illustrations, not clients, but every line of their results comes straight from the tool.

The first is a construction firm of about 60 people with nothing automated yet. Quotes are priced by hand from emailed drawings, which takes more than 10 hours a week, the business runs on spreadsheets, and nobody owns AI.

The second is a marketing agency of a similar size, which is a good deal further along. AI is part of everyday work there, one agent is already running in production, and there is no AI usage policy and no clear view of where customer data goes.

The construction firm's first move is to write its quoting process down. The tool's advice is to "ask the person who does quoting best to talk through it once while someone records it", and then to turn that into steps, decisions and sources. "Rough is fine. What matters is that it's out of their head." Automating the quoting itself, which is the thing that hurts, sits in the second phase of the roadmap, behind picking one source of truth for customers and jobs and checking that the core systems can be connected at all. The biggest drain is real, and it still is not the first move.

The agency's first move is a one-page AI usage policy, which surprises people because the agency is further ahead. The tool's reasoning is that "your team is almost certainly using AI already, so this turns quiet risk into an agreed way of working". Writing down a process, the construction firm's first step, moves back into the agency's second phase. The business that is further along has to take a step back before it goes forward, because it already has something live without guardrails, and governance needs to come before more agents.

What strikes me most, looking at the two results side by side, is how much of the library they share. Both roadmaps include a usage policy, a single source of truth, measuring the hours before anything changes, and writing the process down. The steps overlap heavily, and what separates the two businesses is almost entirely the order those steps come in.

The "not yet" lists differ too, and they are worth as much as the steps. The same government guidance makes the point that a roadmap also explains what you’re not doing. For the construction firm, that means not hiring an AI specialist before the first win and not signing up to a company-wide platform yet. For the agency, it means not letting AI send anything without a human check, and not training its own model on company knowledge, for the reasons we set out in how systems that remember change behaviour. A good roadmap saves you from the expensive wrong turn as often as it points out the right one.

What sets your order

If the steps are broadly shared, the order comes from your answers. Five things do most of the work.

The first is where you stand today. A business on the first rung of the AI maturity ladder needs different groundwork from one with agents already running, and skipping a rung tends to mean coming back for it later at a higher price.

The second is your foundations, meaning your data, your processes, who owns AI, your governance and your tools. The weakest of the five usually decides the first move, because everything built on top of it inherits the weakness.

The third is where the hours actually go. This is the part people expect to drive the whole roadmap, and it does decide what the roadmap is for. It just rarely decides what happens first.

The fourth is people. Whether there is somebody with the time and the authority to own this changes the order more than almost anything else. A plan with no owner tends to start with naming one.

The fifth is how soon you want to move. A business that wants results this quarter gets a different sequence from one that is exploring, and both are reasonable positions to hold.

Sector, size and systems run through all five. A construction firm on spreadsheets and an agency with half a dozen connected tools will answer the same questions very differently, which is exactly why a generic plan struggles to tell either of them what to do on Monday.

We saw this play out with Designs4You, a UK commercial flooring business. Their RFQ automation took handling time per request from around 20 minutes to under 90 seconds and freed more than 12 hours a week of staff capacity. The case study is explicit about why it moved so quickly. The legacy project management system had been replaced first, and "the reason the team could move this fast on Phase 1 is that the foundation was already in place". The visible win came second, and it was fast because of what came first.

Why your roadmap should change

A roadmap that never changes is usually one nobody is reading. The government guidance builds change into its definition, saying that roadmaps need to be easy to understand, and simple to adjust when priorities change, and it adds that "At GDS roadmaps are iterated on at least a quarterly basis." Those teams run public digital services, not mid-market businesses, but the habit translates well.

Eisenhower put the deeper reason better than anyone. Speaking in 1957, he said Plans are worthless, but planning is everything, and explained that an emergency "is not going to happen the way you are planning". AI adoption is not an emergency, but it shares the same property. The plan will be wrong in places. The planning is what lets you re-plan in an afternoon when it is, because you already understand what depends on what.

In practice, four kinds of event change an AI roadmap. Finishing a step changes it, because the next one becomes visible and is often not quite what you expected. Learning something changes it, such as discovering that the process you wrote down turns out to have three variants. The business changes it, through a new hire, a lost client or a system you are moving off. And the tools change it, because what was hard in the spring is sometimes routine by the autumn.

The answer is to hold the near steps firmly and the later ones loosely. The first quarter of a roadmap should be specific enough to act on this week. By the second half of the year it should read more like a direction of travel, because nobody knows yet what the first two quarters will teach you. The 30/60/90-day template makes the same point from the other side, by putting decision points into the plan where you can continue, pause or change direction.

What a good next step looks like

The most useful thing a roadmap gives you is a clear next step that you can start on with confidence. The tool holds every first move to the same test. "It costs little, it needs nothing from us, and it makes every later step easier."

Both example businesses get a first move like that. Recording someone talking through how they price a quote takes an hour. Writing a one-page usage policy takes an afternoon. Neither needs a budget or a supplier, and both make everything after them faster and safer.

That is also a fair test to apply to any roadmap you are shown, including one from a supplier. If the first step is a purchase, it is worth asking what has to be true before that purchase can work, because the answer is usually your real first step.

If you would like to see where you stand, the AI roadmap tool takes about 4 minutes. You see your position, your foundations and your first move straight away, with no email needed, and the full 12-month plan comes to you as a private link if you want it. Take it again in a quarter and see what has moved. If you are earlier on and want a feel for the first workflow before any of this, where to start with AI automation is a gentler place to begin.

Practical takeaways

1. Separate the list from the order. Write down everything you could automate, then ask what each item depends on before deciding what comes first.
2. Expect the biggest pain to come second. It often sits on top of groundwork, such as a documented process or one source of truth, that has to go in first.
3. Fix the weakest foundation first. Data, process, ownership, governance or tools, whichever is weakest will shape everything built on it.
4. Keep a "not yet" list. It saves you from expensive wrong turns, and it stops good ideas being argued over every month.
5. Revisit it every quarter. Hold the next 3 months firmly, treat the rest as direction, and expect it to change.

Frequently asked questions

How do you prioritise AI automation?

Put each step after the things it depends on. Starting with whatever hurts most is the natural instinct, and it is usually the wrong one. List the work you could automate, then ask what each item needs to be in place first, such as a documented process, data in one place or systems that connect. The biggest time drain usually decides what the roadmap is for, but your weakest foundation usually decides what happens first. A good first step costs little, needs nothing from outside the business and makes every later step easier.

Why is every company's AI roadmap different?

Because the order comes from the business rather than from the technology. Two companies can share most of the same steps and still need them in a different sequence, depending on where they stand today, how strong their data, processes, ownership, governance and tools are, where their hours go, who can own the work and how soon they want to move. Sector, size and existing systems shape all of those answers, which is why a generic plan rarely tells anyone what to do first.

How often should you update an AI roadmap?

At least once a quarter, and whenever something significant changes. Finishing a step, learning that a process is messier than it looked, a change in the business or a change in the tools can all move the order. The government's digital teams iterate their roadmaps at least quarterly, and that rhythm suits most mid-market businesses well. Update the next 3 months in detail and treat the rest as direction.

Is an AI roadmap worth having if it will change?

Yes, because most of the value is in the thinking rather than the document. Working out what depends on what is the hard part, and once you have done it you can re-plan quickly when something changes. A roadmap also records what you have decided not to do yet, which stops the same ideas being argued over every month. A plan that changes because you learned something is doing its job.

How far ahead should an AI roadmap plan?

About 12 months, with the detail fading as it goes. The first quarter should be specific enough to act on this week, with named steps and an owner. The middle of the year should name the likely steps and what they depend on. The far end should read as a direction of travel, because the first two quarters will teach you things that reshape it. A 30, 60 and 90 day view is a good way to run the first stretch.