A Better AI Strategy Starts with the Work, Not the Tool

Most organisations do not have an AI strategy problem. They have a sequencing problem.

They choose a tool first, then look for a problem it might solve. The result is often a growing list of licences, pilots and demonstrations, without enough meaningful change in the work itself.

A better AI strategy starts with what needs to improve. It identifies a real business problem, decides whether now is the right time to act, and builds the conditions for a successful rollout.

Key takeaways

• Start with a business problem, not an AI product.
• Manage AI as a portfolio of investments, not a collection of pilots.
• Design for employee adoption as carefully as technical performance.
• Measure business value, quality, cost and risk.
• Scale only the use cases that produce repeatable results.

1. Start with the work that needs to improve

“Use AI to improve the business” is not a strategy. It is an ambition, and one that makes prioritisation almost impossible.

Start with a more useful question: where is the business losing time, money, quality or customer confidence?

For a customer-service team, the issue may be slow resolution times because agents search across disconnected knowledge bases. For finance, it may be a manual invoice-reconciliation process. For marketing, it may be the gap between collecting campaign data and turning it into a timely decision.

These are problems that can be measured. They may benefit from AI, but they may not.

That distinction matters. Sometimes the best answer is a better workflow, cleaner data, clearer approvals or a system integration. Choosing a non-AI solution when it is the best fit is a sign of strong strategy.

A useful problem statement has three parts:

1. The outcome that needs to improve.
2. The operational cause holding it back.
3. The measure that will show whether it improved.

For example:

Reduce average customer-response time from 18 hours to under six hours by helping support staff find approved answers faster, while maintaining customer satisfaction.

This creates a concrete test. It also stops teams from measuring AI adoption instead of business impact. A hundred employees using a new tool is not necessarily progress. Faster, safer and more consistent work might be.

2. Treat AI opportunities as an investment portfolio

Not every promising idea should be funded immediately. Nor should every AI initiative be expected to create the same kind of return.

A healthy AI strategy balances three types of investment:

An organisation that only pursues quick wins may save time but never build a meaningful advantage. One that only pursues ambitious transformations may spend heavily without delivering enough early value to sustain momentum.

For every potential use case, assess four dimensions:

• Value: What material outcome could this improve?
• Feasibility: Are the data, workflow and people ready enough?
• Risk: What happens if the output is wrong, biased or exposed?
• Repeatability: Could the capability be reused elsewhere?

This provides a more reliable way to choose work than simply backing the loudest idea in the room.

3. Decide whether the organisation is ready to act

AI projects often fail for ordinary organisational reasons. The data is incomplete. No team owns the workflow. The people expected to use the output were not involved in the design. A pilot runs indefinitely because nobody agreed on what success would look like.

This does not mean waiting for perfect readiness. Perfect readiness is usually an excuse for inaction.

Instead, separate opportunities that need foundational work from those ready for a focused test. A good early use case is usually:

• Narrow enough to evaluate in weeks, not years.
• Valuable enough that a successful result matters.
• Low-risk enough to trial responsibly.
• Owned by people who understand the existing process.
• Measurable against a clear baseline.

The goal is to choose the next sensible move, rather than attempting to predict every AI decision the organisation will make over the next five years.

4. Build a roadmap, not a collection of pilots

A pilot proves that something might work. A roadmap explains how it becomes useful.

Without one, organisations accumulate disconnected experiments. One team buys a writing tool, another trials a chatbot, and another tests forecasting software. Each might be reasonable in isolation, but together they can create duplication, security concerns, inconsistent standards and unnecessary cost.

A practical roadmap should answer five questions:

For each use case, define a simple sequence:

1. Establish the baseline.
2. Design the smallest viable test.
3. Run the pilot with real users and clear controls.
4. Measure the operational and commercial result.
5. Decide whether to stop, refine, expand or integrate it into normal work.

The decision rule should be defined before the pilot begins.

For example, a proposal-writing assistant might progress to wider rollout only if it reduces first-draft time by 30 per cent, keeps factual-error rates below an agreed threshold and is adopted by a defined share of the team. If it misses those standards, the organisation has still learned something useful without committing to a poor investment.

5. Design for adoption, not just technical performance

An AI tool that works in a demonstration but is ignored in daily work has not succeeded.

People adopt technology when it fits naturally into their work, helps them perform better and gives them confidence that they remain accountable for the result. This makes adoption a design requirement from the beginning, not a communications exercise at the end.

Include the people closest to the work when defining the problem and testing the solution. They will see exceptions, risks and friction that are easy to miss from a leadership meeting.

Then make the new way of working explicit:

• Which tasks will change?
• Which decisions remain human-led?
• When should a person override or reject an AI output?
• What training and examples will help people use it well?
• How will feedback be captured and acted on?

The best AI implementations do not merely automate a flawed process faster. They redesign the work around what people do best: judgement, relationships, context and accountability.

6. Build the right foundations

Technology matters, but it should follow the business case and operating model, not lead them.

Data: is the information fit for purpose?

AI is only as useful as the information it can access and the rules that govern that access.

Before selecting a solution, ask:

• Where does the relevant data live?
• Is it accurate, current and sufficiently complete?
• Who is permitted to access it?
• Does it contain personal, commercially sensitive or regulated information?
• Can people understand where an output came from?

More data is not automatically better. For many use cases, a smaller, well-maintained and clearly governed source of truth is more valuable than a large, messy collection of documents.

Models: what capability is actually required?

Not every problem needs a custom model. Often, a well-designed workflow using an existing model is more reliable, cheaper and easier to maintain.

The right question is not simply, “Which model is best?” It is, “What level of capability does this problem require, and what level of risk can we accept?”

Drafting internal meeting summaries is very different from making eligibility decisions, providing legal advice or approving payments. Higher-risk decisions require stronger controls, human review, testing and accountability.

Choose the simplest approach that can deliver the required outcome safely.

Infrastructure: can the solution work beyond a demonstration?

Infrastructure includes the systems, integrations, access controls, monitoring and support arrangements that make an AI solution dependable in daily operations.

Before scaling, clarify:

• How will it fit into existing workflows?
• Who will support it when it produces a poor output?
• How will quality, cost and usage be monitored?
• What is the fallback if the system is unavailable or wrong?
• How will it be updated as policies, products and information change?

7. Manage AI as a living capability

AI is not a one-off technology purchase. Models change, costs change, employee behaviour changes, and the quality of outputs can drift over time.

That makes ongoing evaluation essential.

Set up a lightweight review rhythm for each active use case. Look at the value delivered, operating cost, user feedback, quality issues and emerging risks. This does not need a large governance committee for every tool. It does need a clear owner and a regular decision point.

The most useful questions are straightforward:

• Is this still solving the problem we set out to solve?
• Is the output accurate and useful enough?
• Are people using it in the intended way?
• Is the value greater than the financial and operational cost?
• Have new risks appeared as usage has grown?

Treating AI as a living capability lets organisations improve what is working, retire what is not, and avoid carrying yesterday’s experiments into tomorrow’s operating model.

Better decisions are the goal

The strongest AI strategies are not defined by how many tools an organisation buys. They are defined by whether the organisation makes better decisions and improves important work.

Start with a problem worth solving. Choose a balanced set of opportunities. Build adoption into the design. Give every pilot a clear decision point. Then invest in the data, technology and operating model required to make success repeatable.

AI can be transformative. The transformation begins with being clear about what needs to change.