Anna Totterdell
Projects Director
If you have read any of the other articles on this site, you will have noticed a pattern. We keep saying the same thing in different ways: AI does not work when it is bolted onto broken operations. Your data needs to be structured. Your systems need to be connected. Your processes need to be mapped before any intelligence layer can sit on top of them.
That is all true. But it raises an obvious question: so what does the actual work look like?
This article is the answer. Not theory. Not a framework for the sake of having a framework. A practical, step-by-step model based on how we actually approach operational AI in mid-market businesses - and the order in which things need to happen for the results to stick.
There are five steps. They are not optional, and they are not interchangeable. Skip one and the ones that follow will fail.
Step 1: Map
Before you touch a single system, you need to know how your business actually runs.
Not how the process documents say it runs. Not how the org chart implies it runs. How it actually runs - today, on the ground, including the workarounds, the manual steps, the emails that trigger actions, and the spreadsheets that fill the gaps.
This is harder than it sounds. In most mid-market businesses, the real processes are invisible. They live in the habits of experienced employees, in undocumented handoffs between departments, and in shadow systems that nobody officially acknowledges.
Mapping means sitting down with the people who do the work and walking through every step of a process from trigger to completion. Where does the request come in? Who touches it? What data gets entered, where, and by whom? Where do things get stuck? Where do errors occur? What happens when something goes wrong?
The output is not a pretty diagram. It is an honest, detailed picture of how the operation works - including the parts that nobody is proud of. That honesty is the foundation for everything that follows.
Step 2: Structure
Once you can see the processes clearly, the next job is to structure the data that flows through them.
This is where most AI projects should start, but almost none do. They jump straight to tools and models, and then wonder why the outputs are unreliable.
Structuring means answering basic questions that most businesses have never formally addressed. What is the single source of truth for your customer data? How are products categorised, and is that categorisation consistent across systems? When someone enters an order, what fields are mandatory, what formats are enforced, and what happens when someone deviates?
In practice, this is the first half of data and systems integration: normalising fields across systems, deduplicating records, agreeing on naming conventions, and establishing data governance rules that prevent the mess from recurring. It is painstaking, cross-departmental work. It requires decisions that have been deferred for years.
It is also the single highest-value activity in the entire process. Every step that follows depends on this one being done properly.
Step 3: Integrate
With mapped processes and structured data, you can now connect your systems.
Most mid-market businesses run on a patchwork of platforms - an ERP, a CRM, a finance system, and a collection of point solutions for specific functions. These systems typically do not talk to each other, or they talk badly - through manual exports, CSV uploads, or integrations that broke six months ago and nobody fixed.
Integration means building reliable, automated data flows between these systems so that information moves without human intervention. When a customer record is updated in one system, it is reflected in the others. When an order is placed, every downstream system receives the same structured data. When a payment is recorded, the relevant reports update in real time.
This is not about replacing systems. It is about making the ones you already have work together. The technology for this is mature and well-understood. The challenge is not technical - it is operational. It requires clarity about what data moves where, what the rules are, and who owns the process.
Step 4: Automate
With connected systems and structured data, you can now automate the manual handoffs and repetitive tasks that consume your team's time.
This is the step most people want to jump to first. Resist that instinct. Automation built on unmapped processes and unstructured data just makes bad processes run faster. You end up automating the workaround instead of fixing the underlying problem.
But when the foundations are right, business automation is transformative. Not in the grand, press-release sense. In the practical sense: the reconciliation that took two days now takes twenty minutes. The status report that required someone to pull data from four systems is generated automatically every morning. The approval chain that relied on someone remembering to send an email is now triggered, tracked, and escalated without intervention.
The key principle is this: automate the workflow, not the task. A task is a single action. A workflow is the entire sequence - from trigger to completion - including the rules, exceptions, and handoffs. Automating tasks gives you small time savings. Automating workflows gives you operational change.
Step 5: Apply AI
This is where AI enters. Not at the beginning. At the end - once the processes are mapped, the data is structured, the systems are connected, and the workflows are automated.
At this point, AI has something to work with. It has clean, consistent, accessible data. It has structured workflows with clear inputs and outputs. It has a governed environment where its outputs can be validated and acted upon.
And that means it can do things that are genuinely useful. Not impressive in a demo. Useful in an operation.
AI that categorises incoming requests based on content and routes them to the right team - because the workflow exists for it to route into. AI that flags exceptions in your supply chain data - because the data is normalised enough for it to spot the anomaly. AI that summarises a customer's full history before a call - because that history is connected across systems, not scattered across four platforms and a dozen email threads. AI that recommends pricing adjustments - because it has access to consistent margin data it can actually trust.
This is operational AI enablement. It is not a product you buy. It is a layer that sits on top of everything you have already built. And it only works when the four steps before it have been done properly.
Why the order matters
Every failed AI project I have seen has the same root cause: they started at step five.
Someone bought a tool. Someone built a proof of concept. Someone connected a language model to a dataset. And it produced impressive outputs in a controlled environment - and completely fell apart when exposed to real operational data.
That is not a technology failure. That is a sequencing failure. The technology works. It just needs the right foundations underneath it.
The businesses that are getting real, measurable value from AI right now are not the ones with the most sophisticated models or the biggest budgets. They are the ones that did steps one through four properly - often without thinking of it as an "AI project" at all.
They mapped their processes. They cleaned their data. They connected their systems. They automated their workflows. And then they asked: where can intelligence make this even better?
That question, asked in that order, is where the real value lives.
What this looks like in practice
In a typical mid-market engagement, this model plays out over weeks, not months. Not because the work is shallow, but because it is focused.
You do not map every process in the business. You map the ones that matter most - the critical paths where time, cost, and risk are concentrated. You do not structure every data field. You normalise the ones that feed the workflows you are fixing. You do not integrate every system. You connect the ones that are involved in the processes you have prioritised.
The result is not a transformation programme. It is a series of targeted, measurable improvements that compound over time. Each one builds the foundation for the next. And by the time AI enters the picture, it has a solid, structured, connected operation to work with.
That is what actually works.
