The Board Wants an AI Strategy. Start With a Use-Case Register.

Every quarter, another board member asks: what is our AI strategy? The usual response is a slide deck with a vision, a maturity model and a roadmap that shows proofs of concept turning into production by next year. The deck gets approved. The proofs of concept run. Most of them stop there.

My experience is that the problem is not the strategy. It is that the strategy has no inventory underneath it. Nobody has written down what the business actually does, process by process, and scored each one for whether AI would make a measurable difference. Without that register, the AI strategy is a set of guesses dressed in a framework.

This piece is for CIOs, CTOs and heads of data who have been asked for an AI strategy and want to build it on something solid.

The short version

Before writing an AI strategy, build a use-case register. List every business process, score each one for AI potential, and let that register drive prioritisation. A strategy without an inventory is a deck that dies after the board meeting.

• A use-case register lists business processes, not AI technologies.
• Score each process on data readiness, business impact and implementation risk.
• The register replaces opinion-based prioritisation with evidence.
• Start with the processes where you already have clean, labelled data.
• Update the register quarterly as you learn what works.

This article sits inside the ai and enterprise solutions architecture cluster, where the wider argument is set out in full.

What a use-case register actually is

A use-case register is a structured list of business processes with enough detail to evaluate whether AI adds value. It is not a list of AI ideas, and it is not a technology roadmap.

Each entry answers five questions:

1. What is the process? A plain description of what happens, who does it, and how often. 2. Where is the data? Which systems hold the data this process uses or generates, and in what condition. 3. What would improvement look like? Faster cycle time, fewer errors, lower cost per transaction, higher accuracy — stated as a measurable outcome. 4. What is the current performance? The baseline number for that outcome today. 5. What is the constraint? What stops the process from performing better right now — data quality, system access, manual steps, regulatory requirements, skill gaps.

Notice that none of these questions mention AI. That is deliberate. The register describes the business reality. AI is one possible intervention, and it competes with simpler options like process redesign, automation without AI, or fixing the data.

Why most AI strategies skip this step

Building a use-case register is slow, unglamorous work. It requires talking to operations teams, auditing data sources, and documenting baseline performance numbers that nobody has measured before.

A strategy deck, by contrast, takes two weeks and a good designer. It references industry trends, names a few promising use cases picked from vendor presentations, and shows a phased timeline. It looks decisive. The board approves it. Then someone has to figure out where the data actually lives, and the timeline slips.

I have built both. The register takes longer but survives contact with reality. The deck without a register survives contact with the board but nothing after.

Should your organisation do this now?

• Yes, if you have approved an AI strategy but no proof of concept has reached production.
• Not yet, if you have fewer than ten business processes that use structured data.
• Instead, first: digitise the processes where AI might apply.
• Measure it by: number of processes with documented baseline performance.

How to score the register

Every entry in the register gets three scores, each on a one-to-five scale:

Data readiness (1–5). Is the data available, clean, labelled and accessible? A one means the data does not exist in digital form. A five means structured, labelled data sitting in a system with an API.

Business impact (1–5). How much does improving this process matter to the business? A one means minor efficiency. A five means direct revenue impact or regulatory requirement.

Implementation risk (1–5, inverted — lower is better). How hard is it to build and deploy? A one means a well-understood problem with available tools. A five means novel research, cross-system integration and regulatory approval needed.

Multiply data readiness by business impact, then divide by implementation risk. The result is a priority score. Sort by it. The top of the list is where you start.

This scoring is deliberately simple. A more complex model adds precision but reduces adoption. The point is to replace gut-feel prioritisation with a structured conversation.

The data readiness check that saves you months

The single biggest predictor of whether an AI use case succeeds is data readiness. Not algorithm sophistication, not compute power, not team size — data readiness.

Before committing to any use case, I run five checks:

1. Does the data exist? If it is in people's heads or on paper, AI cannot use it. 2. Is it accessible? Data locked in a legacy system without an API is data you cannot use at scale. 3. Is it clean enough? Missing fields, inconsistent formats and duplicates need to be fixed before training, not during. 4. Is it labelled? Supervised learning needs labelled examples. If nobody has labelled the data, add three to six months for that step. 5. Is there enough of it? A process that happens ten times a month does not generate enough data for most AI approaches.

Any use case that fails on more than two of these checks is not ready for AI. It is ready for a data improvement project, which should run first.

What the register reveals

The first version of a use-case register typically surfaces three things that a strategy deck never would:

The best opportunities are boring. The process with the highest score is rarely the one the board mentioned. It is usually something operational — invoice matching, delivery scheduling, warranty classification — where good data already exists and the business impact of a ten per cent improvement is large because the volume is large.

The exciting ideas are not ready. The use case the CEO saw at a conference usually scores low on data readiness. The data does not exist in the right form, or it sits across three systems that have never been connected.

Quick wins exist. Some processes can be improved with simple rules or classical automation, not AI at all. The register shows these because it describes the constraint, and sometimes the constraint is a manual step that can be eliminated without a model.

Should your organisation do this now?

• Yes, if your AI pilots are chosen based on executive interest rather than data.
• Not yet, if you have not assigned an owner for the register.
• Instead, first: name a person responsible for maintaining it.
• Measure it by: percentage of AI budget spent on processes with a data readiness score of four or above.

Keep the register alive

A register that is built once and shelved is a document. A register that is updated quarterly is a governance tool.

After each quarter, update three things:

1. Scores. Data readiness changes as systems are upgraded and data quality improves. Business impact changes as strategy shifts. 2. Status. Mark which use cases are in exploration, pilot, production or retired. 3. Results. For pilots and production use cases, record the actual performance against the baseline. This is where the strategy meets reality.

The quarterly review also catches drift. A use case that was high priority six months ago may have been overtaken by a regulation, a market change, or the discovery that the data is worse than expected.

Before you build one

Checklist:

• Named owner for the register with authority to request process data from any team.
• Agreement on the scoring dimensions and scale.
• Baseline performance data for the top twenty processes by volume.
• Data readiness assessment for each process in the register.
• Quarterly review scheduled with the leadership team.
• Clear rule: no AI budget approved for processes not in the register.

Questions to ask:

• Your operations team: which five processes consume the most time or cost per transaction?
• Your data team: for each of those processes, where is the data and in what condition?
• Your finance team: what is the current cost per transaction for high-volume processes?
• Your compliance team: which processes have regulatory constraints on automation?
• Your board: are you willing to fund data improvement for processes where AI readiness is low?

How to measure it

• Register coverage. Percentage of business processes with a scored entry. Baseline: first version. Owner: enterprise architecture. Review: quarterly. Leading.
• Data readiness improvement. Average data readiness score across the register. Baseline: first assessment. Owner: data governance. Review: quarterly. Leading.
• Priority alignment. Percentage of AI spend on top-quartile scored processes. Baseline: current budget split. Owner: CTO office. Review: quarterly. Leading.
• Pilot-to-production rate. Share of AI pilots that reach production. Baseline: last twelve months. Owner: AI programme lead. Review: semi-annual. Lagging.
• Business outcome delivery. Actual performance improvement versus baseline for production use cases. Baseline: register baseline. Owner: process owner. Review: quarterly. Lagging.

Mistakes that cost the most

Starting with the technology

A team picks a tool and looks for problems it can solve.

• Why it happens: vendor demos are persuasive.
• Prevention: require the register entry before the vendor conversation.
• Early warning: the pilot brief names a product but not a baseline metric.

Scoring without data

The register exists, but data readiness is estimated rather than checked.

• Why it happens: the data audit is slow and nobody owns it.
• Prevention: make data readiness a prerequisite, not an estimate.
• Early warning: pilots delayed by data issues that the score did not predict.

No baseline measurement

You cannot prove improvement without a before number.

• Why it happens: nobody measured the process before the AI project started.
• Prevention: require baseline measurement before approving any pilot.
• Early warning: the pilot report says "significant improvement" without a number.

What to do next

Pick your ten highest-volume business processes and write a one-page register entry for each. Score them. Share the result with your leadership team before your next AI strategy conversation. Then read the related piece on why agentic AI projects get cancelled and the enterprise AI architecture topic page.

Last reviewed: 14 October 2026.

Views are my own and do not represent my employer.