Every operational business wants AI right now. Predictive analytics, natural language queries, automated decisions. The ambition is right. The sequencing is usually wrong.

AI does not fix a broken process. It amplifies it. Feed a poorly structured workflow into a machine learning model and you get faster, more confident, more expensive versions of the same mistakes you were already making.

The problem is almost never the technology. It is what comes before it.

Data cannot flow from a process that does not exist yet

Most non-IT businesses run on Excel. Not because their teams are unsophisticated — often the opposite. The people closest to the work built spreadsheets that solved real problems, and those spreadsheets became the system of record.

The issue is that Excel-based workflows do not produce usable data at scale. The structure changes between users. Columns get renamed. Rows get merged. Formulas break. By the time you try to connect that to an AI model, you are not feeding it data — you are feeding it noise.

Before AI, you need a workflow. Before a workflow, you need the data to be structured, consistent, and accessible.

Fix the process, then digitise it, then add intelligence

The right sequence is not glamorous but it is reliable.

Start by mapping the actual workflow — not the ideal version, but what people are really doing today. Identify where the data lives, who enters it, and what decisions it drives.

Then digitise it. Move the process into an application where data entry is controlled, outputs are consistent, and every transaction creates a record that can be queried.

Only once data is flowing cleanly does AI become a viable next step — and at that point, it becomes a powerful one.

In practice, that sequencing looks like this. Start with the processes still running on spreadsheets — those are almost always the highest-friction points and the easiest wins to identify. Make them simple and repeatable first. Then move to standardisation: get teams aligned on shared data structures for things like cost breakdowns and sales reporting, so that a cost line means the same thing in procurement as it does in finance. Audit the systems in use and remove the redundant ones — consolidate where it makes sense, and make sure data flows between what remains without manual intervention. Only once the process is clean, the structures are consistent, and the tooling is rationalised does it make sense to introduce AI. At that point, you are giving it something worth working with.

What this looks like in practice

Consider a representative manufacturing business where pricing, procurement, and sales data have evolved over time across a patchwork of systems that were never designed to work together. Quote data lives in one place, approval workflows run through another system like Salesforce, bills of materials and invoice history sit in SAP, and the rest of the ERP data is scattered across exports and shared drives.

Each system is doing its job in isolation — the problem is that no one has a coherent view across all of them.

In that kind of environment, annual planning cycles often become slow and manual. Teams spend weeks pulling numbers together from different platforms, reconciling conflicting versions, chasing approvals across systems, and rebuilding planning documents from scratch. Highly paid leaders end up absorbed in administrative data work instead of focusing on commercial or operational decisions.

That is not a technology problem. It is a process problem wearing a technology costume.

The first step is not to layer AI on top of the chaos. It is to convert the spreadsheet-driven workflow into a structured application with controlled inputs, consistent data models, and a single source of truth.

Once that system is in place, the business can start using analytics and AI in a way that actually works — forecasting raw material movements, identifying cost anomalies, and letting non-technical users query business data in plain English.

The value comes not from adding intelligence first, but from giving the process a reliable foundation.

The question to ask before any AI project

Before committing to an AI initiative, ask one question: if you removed the AI entirely, would the underlying process still work?

If the answer is no — if the data is inconsistent, the workflow is manual and undocumented, or the people closest to the work do not trust the outputs — then the AI project is not ready. The process work is.

Get the foundation right. The intelligence layer will follow, and when it does, it will actually work.

Continue the Discussion

If your organization is trying to introduce AI into spreadsheet-heavy or fragmented workflows, I can help you design a practical process-first roadmap. Book a CTO consultation.

You can also connect with me on LinkedIn for a deeper discussion.