Clear thinking on AI,systems and operations.

The pilot works. The rollout does not.

A pilot is usually run by a small, motivated group with clean examples and someone watching every output. That is the right way to test an idea. It also hides the questions that decide whether the idea survives normal work: who owns the step, what happens when the output is wrong, and who checks the result.

Three things to define before scaling

• The workflow: where the AI step starts and ends, what goes in and what comes out.
• Ownership: one named person responsible for the step and its results, not just for the tool.
• Review cadence: how often outputs are sampled, who checks them and what triggers a change.

None of this is technical. It is ordinary operating discipline, and it tends to be missing in exactly the places where teams hope AI will fill the gap.

A steadier order

Map the current process first, even roughly. Agree who owns it. Then add the AI step and decide how it will be reviewed. A pilot built this way is slower to start, but much more likely to become part of everyday work.

When an ERP feels slow or awkward, the first instinct is often to look for a newer platform. Sometimes that is the right call. More often, the friction comes from how the system was set up and how it connects to the rest of the business.

Where the friction usually sits

• Process design that does not match how work actually flows, so people work around the system.
• Weak master-data rules: duplicate customers, inconsistent product codes, units that do not agree.
• Poor handoffs between systems, filled in with exports, re-keying and spreadsheets.

A new platform inherits all three problems unless they are fixed along the way. The migration takes months, and the same complaints can return a year later.

Test the current core first

Before replacing the stack, ask whether better structure, data governance and integration could make the current digital core reliable. If the answer is yes, that is usually the faster path. If the answer is no, you will at least know what the next system must do differently.

By digital core we mean the few systems and data definitions that the rest of the business depends on: usually ERP, CRM, inventory and finance, and the rules that tie them together.

What a stronger core looks like

• Shared definitions: an order, a customer and a margin mean the same thing in every report.
• Clean system interaction: data is entered once and flows to where it is needed.
• Clear reporting ownership: each key number has someone who maintains it and can explain it.

Dashboards and automation built on a weak core tend to multiply the confusion: more numbers, less agreement. Built on a stronger core, they become useful quickly, because information moves with less delay and less manual correction.

Start small

Strengthening the core rarely needs a big programme. Pick one recurring report that people argue about, trace each number back to its source, and fix the definitions and handoffs along the way.

The same six steps work whether the answer turns out to be a small integration or a larger custom system.

1. Name the problem and the outcomeDescribe it in business terms, such as slow quotes, stock surprises or a long month-end, and agree what better looks like.
2. Review the workflow, data and toolsFollow the work from start to finish. Note where it waits, where it is re-keyed and where it depends on one person.
3. Keep what worksExisting systems that do their job stay. Change is aimed at the gaps.
4. Choose the lightest approach that fitsThat may be an integration between systems, a small custom application or an AI assistant. Often it is a combination.
5. Pilot, review, then expandStart with the priority problem, check the results against the outcome you agreed, and widen the scope only after that.
6. Keep supporting the teamPeople need help as the solution changes how they work, and the solution needs adjusting as the business changes.

The order matters more than the technology. Starting from a tool tends to produce a solution looking for a problem.

Use these questions on one workflow at a time, not on the whole business. A no is not a reason to stop. It shows what to fix first.

• OwnershipIs there one person responsible for this workflow and its results?
• InputsAre the documents and data going in reasonably consistent and easy to find?
• RulesCould someone write down how a good result is produced, including the common exceptions?
• EscalationIs it agreed when a person must step in, and who that person is?
• OutcomeIs there a measure, such as turnaround time or rework, that you can compare before and after?
• Access and privacyIs it clear which data the AI may use and who can see its outputs?

If most answers are yes, the workflow is a good candidate for an AI step. If several are no, start with the workflow itself. AI added to an inconsistent process usually amplifies the inconsistency.

Look back over a month of normal work and mark each sign you recognise.

• Rekeyed dataThe same order, invoice or item is typed into more than one system.
• Duplicate approvalsThe same decision is signed off in email, on paper and again in the system.
• Spreadsheet bridgesA spreadsheet is the only place where two systems’ data meet.
• Manual reportsA regular report is rebuilt by hand every week or month.
• Status chasingPeople ask each other where an order or job is, because the system cannot tell them.
• Single-person knowledgeOnly one person knows how to make the numbers line up.

What to do with the list

Rank the signs by how often they happen and how much rework they cause. The top one or two are usually the best place to start an integration or a small custom tool. Fixing them also makes later AI work easier, because the data becomes more reliable.

In many plants the information already exists. Planning has a schedule, production has counts, quality has inspection results and the warehouse has stock levels. The difficulty is that each lives in its own report and is updated at its own pace.

One operating rhythm

Performance tends to improve when these signals are reviewed together, at a regular rhythm, by the people who can act on them. A short daily or shift review of plan against actual, open quality issues and material shortages often does more than a new module.

Where to focus

• Exception handling: make deviations visible early, each with a clear next step.
• KPI ownership: each measure has an owner who explains it and acts on it.
• Faster escalation: agree when an issue moves from the line to a supervisor or manager.

Software helps when it supports this rhythm, for example by bringing plan, production and quality data onto one screen. Adding complexity before the rhythm exists rarely does.

Distribution businesses make many small promises every day: this item is in stock, this order ships today, this customer gets this price. Each promise depends on information that sits in different places.

Different versions of the truth

When the ERP, the warehouse and the sales team each work from their own numbers, small gaps turn into missed shipments, back orders and long email threads. Service suffers even when everyone is working hard.

What usually helps most

• Order-flow visibility: anyone can see where an order is, from entry to dispatch.
• Cleaner stock signals: available stock reflects allocations, receipts and returns promptly.
• Clear recovery ownership: when something goes wrong, one person owns putting it right with the customer.

These changes usually create more value than another layer of coordination, such as an extra meeting or one more spreadsheet to reconcile the others.

Engineering firms, installers, technical service companies and other project-led businesses depend on specialist judgement. That judgement is often spread thin: a few experienced people interpret drawings, exceptions and handovers for everyone else.

Where the strain shows

• Documents are scattered across email, shared drives and personal folders.
• Approvals happen informally, so it is hard to see later what was agreed and why.
• Handovers between sales, design and delivery lose detail along the way.

What reduces the risk

Stronger knowledge flow, better structured approvals and clearer delivery context. In practice that can mean one place for project documents, simple approval steps that leave a record, and a handover template that travels with the job. A company knowledge assistant can then help people find answers in their own material, with sources they can check.

The aim is not to replace specialists. It is to give them better context and make their expertise easier to share.

New AI products appear every week, and their descriptions blur together. Two questions cut through most of the noise. Does the tool only advise, or can it act? Is it standalone, or connected to your systems, data and permissions?

Put connection on the vertical axis and action on the horizontal one, and you get four places a tool can sit.

How to use the map

Place each tool you use or are considering. Standalone tools suit individual work, as long as the rules for company data are clear. The upper right is where AI starts to change how the business runs, and also where it needs the most care: defined permissions, a record of what was done and a person accountable for the outcome.

Moving a tool up or to the right should be a deliberate decision, not an upgrade accepted by default.

Four stages describe the shift. They overlap, and most businesses use more than one at the same time.

1. Generative AI

   Creates text, images, analysis, summaries and prototypes. It improves knowledge work, but often stops at recommendation or output.

2. Agentic AI

   Plans, calls tools, uses systems and completes governed multi-step tasks. It turns AI from an assistant into an operating workflow participant.

3. Physical AI

   Connects AI models to sensors, robots, machines and physical work environments so AI can help inspect, move, handle, guide or coordinate real work.

4. Human + AI work

   People remain responsible for judgement, safety, exception handling and improvement while AI systems reduce repetitive coordination and execution load.

What this means for a business

There is no need to move through these stages in order, or all at once. The practical question is which work could move one step along, and what controls that step needs: who approves, what is recorded and how errors are caught.

At every stage, people stay responsible for the result. Good AI work takes away repetitive load so that attention goes where judgement is needed.