← All news and guides
AI adoption

AI Use-Case Prioritisation Matrix (Template): Score Value, Risk and Data Readiness

A copy-paste template to rank AI ideas: gates, a 100-point score for value, risk and data readiness, a one-day readiness check and an ROI baseline.

Mono Team ·

Rows of identical blank wooden tiles laid out in a grid on a pale oak table, seen from above in soft daylight, a few tiles moved to the top edge.

Short answer: score every AI idea on the same six criteria, apply a few hard gates first, and pilot only the top one or two. The template below scores value size, volume, data readiness, risk, integration effort and adoption on a 1 to 5 scale, weights them, and gives a total out of 100. Risk and data readiness carry 40% of the weight between them, because a use case that looks good in a demo can still fail on messy data or on mistakes it cannot afford. Before you build anything, record a baseline of how the work is done today, so you can show later whether the AI helped.

This template is for a founder, COO or engineering lead with a long list of AI ideas and a small budget. It covers the matrix, gates, a worked example, a one-day readiness check and an ROI baseline. The example scores are made up to show the method. The weights and cut-offs are our template choices, not an industry standard, so change them.

What is an AI use-case prioritisation matrix?

It is a scored list of candidate AI projects. Each row is one use case, such as summarising support tickets. Each column is a criterion. You score each cell, add up the weighted total and sort the list. The point is to replace "whoever shouts loudest" with a comparison the team can argue with and revise.

What do you need before you start scoring?

  • A list of candidates. Come from the work, not from the tools. Ask each team which tasks are repetitive, high-volume or slow. Our AI adoption mapping step does this by listing every task, how long it takes and which ones AI or a better tool could take over. Aim for 10 to 25 rows.
  • A named owner per row. One person in the business who wants the outcome and will use the result. If nobody will own it, it does not make the list.
  • Two or three scorers. Include someone from the team that does the work and someone who knows the data. Score alone first, then compare.
  • About an hour per ten rows. Scores are rough on purpose. If a row needs a day of research to score, it is not ready.

Which gates should stop a use case before it is scored?

Some ideas should not be traded off against value. Check these gates first. A row that fails a gate is either fixed or parked. It does not go through to the scoring.

GateQuestionIf the answer is no
OwnerQuestion: Does a named person own the outcome and the budget line?If the answer is no: Park it
MeasurableQuestion: Can we measure how the work is done today, in minutes, cost, errors or cycle time?If the answer is no: Define the measure first, or park it
Permitted dataQuestion: Are we allowed to use this data for this purpose, under our contracts, notices and consent?If the answer is no: Fix the permission first. If the data is personal data, involve counsel and read our guide to DPDP-compliant AI architecture
Human in the loopQuestion: If the output affects a person's job, pay, credit, health or access to a service, will a trained human review it before it takes effect?If the answer is no: Add the review step to the design, or park it
FallbackQuestion: If the AI is wrong or unavailable, can the work still be done by hand?If the answer is no: Design the fallback first

The gates are a design stance, not a legal test. Whether a use is lawful is a question for your counsel.

How do you score each AI use case?

A grid of identical small blank wooden tiles being arranged in columns on a pale oak table, seen from above in soft daylight, a few tiles set apart from the rest.

Score each criterion from 1 to 5. Use the same scale direction for all six, so that a higher number is always better. That means risk is scored as 5 for low risk and integration effort is scored as 5 for low effort.

CriterionWeightWhat a 5 looks likeWhat a 1 looks like
Value sizeWeight: 25What a 5 looks like: Large saving or revenue effect that you can state in hours or moneyWhat a 1 looks like: Nice to have, nobody can say what it is worth
Volume and frequencyWeight: 15What a 5 looks like: Done many times a day or week by many peopleWhat a 1 looks like: Rare or one-off
Data readinessWeight: 20What a 5 looks like: Data is in one place, clean, permitted and currentWhat a 1 looks like: Scattered, messy, unclear permission, or does not exist
Risk (5 is low)Weight: 20What a 5 looks like: A wrong answer is cheap, visible and easily correctedWhat a 1 looks like: A wrong answer harms a person, breaks a rule or is hard to detect
Integration effort (5 is low)Weight: 10What a 5 looks like: Works with a tool you already use, setup in daysWhat a 1 looks like: Needs deep changes to core systems
Adoption and ownershipWeight: 10What a 5 looks like: Users asked for it and the owner is engagedWhat a 1 looks like: Users are sceptical and the owner is unclear

The total out of 100 is each score multiplied by its weight, added up, then divided by 5. In a spreadsheet: `=(25V + 15F + 20D + 20R + 10E + 10A) / 5`. If every score is 5, the total is 100. If every score is 1, it is 20.

For your own sheet, use these columns: use case, owner, gates passed (Y or N), the six scores, total out of 100, and route (buy, configure or build).

What is a worked example of the scoring?

These five rows are invented to show how the arithmetic and the discussion work. Your numbers will differ.

Use caseValueVolumeDataRiskEffortAdoptionTotal /100
Draft summaries of support tickets (agent reviews before sending)Value: 3Volume: 5Data: 4Risk: 4Effort: 4Adoption: 4Total /100: 78
Extract fields from supplier invoices into the finance systemValue: 4Volume: 5Data: 3Risk: 4Effort: 3Adoption: 4Total /100: 77
Pull key clauses out of customer contracts for legal reviewValue: 4Volume: 3Data: 3Risk: 3Effort: 3Adoption: 3Total /100: 65
Screen incoming CVs and rank candidatesValue: 4Volume: 4Data: 2Risk: 1Effort: 3Adoption: 3Total /100: 56
Staff Q&A assistant over the shared driveValue: 3Volume: 3Data: 2Risk: 2Effort: 3Adoption: 3Total /100: 52

How to read it:

  • Ticket summaries and invoice extraction go to pilot. They score within a point of each other. Pick the one with the clearer owner and the easier baseline.
  • Contract clause extraction is a maybe. A missed clause matters, so it needs mandatory lawyer review.
  • CV screening is scored here only to show the arithmetic. As described, it fails the human-in-the-loop gate, so in a real round it would be parked before scoring until a trained reviewer checks every shortlist decision. A good total never overrides a failed gate.
  • The shared-drive assistant is held back by data. Its score is low because the data score is low. See the next section.

How do you score data readiness?

Data readiness is easy to overrate. A demo works on a hand-picked folder. Production works on everything the system can reach. Score the weakest point, not the average.

ScoreWhat it means
5What it means: One source of truth. Fields are consistent. You know who owns it, you are permitted to use it for this purpose, and it is current.
4What it means: Mostly clean, in one or two systems. Small gaps you can fix in the pilot.
3What it means: Usable but spread across several systems or formats. Needs a clean-up pass or a connector before it is reliable.
2What it means: Messy, partly out of date, or access rights are unclear. The project would start with data work.
1What it means: Data does not exist in usable form, or you are not sure you may use it.

Permissions are part of readiness. Microsoft says Microsoft 365 Copilot "only surfaces organizational data to which individual users have at least view permissions". If a shared drive is over-shared today, an assistant built on it can show a user documents they were never meant to find. That is why the staff Q&A example above scores 2 on data. Fix the permissions first. Our post on permission-aware RAG covers this.

If the data is personal data, add it to your data map before you build. The core DPDP duties start on 13 May 2027.

How do you score risk?

Score risk by what happens if the AI is wrong, not by how likely that is. Ask how bad and how visible the mistake would be.

ScoreTypical situation
5Typical situation: Internal draft that a person reads and edits. Mistakes are caught fast and cost little.
4Typical situation: Internal output that feeds a process with a downstream check.
3Typical situation: Output reaches customers, but a person approves it first.
2Typical situation: Output feeds decisions on money, contracts or compliance, with review that could be rushed.
1Typical situation: Output affects a person's job, pay, credit, health or legal position, or acts with no human check.

Also check where the data goes. Read each vendor's terms on training, retention and processing location when you buy, because they change. OpenAI says it does not train on business data by default, and Anthropic says the same for its commercial products. Microsoft says Copilot prompts and responses are not used to train its foundation models. Do not assume processing in India. OpenAI's data residency article lists India for storage at rest but not for model inference, and Microsoft says customers outside the EU may have queries processed in the US, EU or other regions.

How do you run a one-day AI readiness check for the whole company?

The matrix scores ideas. A readiness check scores the organisation. Run it in a morning with the owners of data, operations and IT. Score each dimension 0 (not in place), 1 (partly) or 2 (in place).

DimensionCheckScore 0 / 1 / 2
DataCheck: We know where our main business data lives, who owns it, and who can access itScore 0 / 1 / 2:
ProcessCheck: The processes we want to improve are written down or can be shown by a team memberScore 0 / 1 / 2:
PeopleCheck: There is a named sponsor and a team willing to try a change for four to six weeksScore 0 / 1 / 2:
Tools and accessCheck: We have approved tools and single sign-on, and staff are not using unapproved AI tools without us knowingScore 0 / 1 / 2:
GovernanceCheck: We have a short AI use policy, a list of AI tools in use, and a way to review a new use caseScore 0 / 1 / 2:
  • 0 to 4: fix foundations first. Start with a data map, an access review and an AI use policy. Our shadow AI audit is a quick way to see what staff already use.
  • 5 to 7: pilot low-risk use cases (risk score of 4 or 5) and work on the gaps in parallel.
  • 8 to 10: pilot the top rows, with measurement in place from week one.

These bands are our template choices. They are a prompt for discussion, not a benchmark against other companies.

How do you baseline ROI before you build?

A spirit level resting on a plain wooden workbench beside a steel tape measure, soft side light from a window, the background blurred.

An AI project cannot show a return if nobody measured the starting point. Do this before the pilot, for each row that passes your shortlist.

  • Sample real cases. Time 20 to 30 recent items end to end. Do not rely on a survey of how long people think it takes.
  • Record volume per week or month, and how it varies.
  • Record quality: the error or rework rate, and what an error costs.
  • Record cost per unit: minutes multiplied by a loaded hourly cost from finance.
  • Record cycle time: how long a customer or colleague waits.

Then work out the saving with the review step included, because someone must still check the AI output:

`hours saved per month = volume x (baseline minutes - new minutes including review) / 60`

For the invented invoice example: 2,000 invoices a month, 6 minutes each today and 2 with AI extraction and a human check. That is 2,000 x 4 / 60, about 133 hours a month. Multiply by your loaded hourly cost, then subtract the cost side.

Cost lineWhat to include
LicencesWhat to include: Seats for the tool, at the vendor's current list price. For example, Microsoft lists Microsoft 365 Copilot at $30.00 per user per month paid yearly, and Anthropic lists the Claude Team Standard seat at $20 per seat per month billed annually (US list prices, ex-tax, as of 20 November 2026). Copilot also needs a qualifying Microsoft 365 plan.
UsageWhat to include: Token or credit charges. Some Copilot features use usage-based billing on top of the seat price. See our post on LLM bill shock for how usage grows.
Build and runWhat to include: Engineering time to integrate, test, monitor and update the tool.
Review timeWhat to include: Minutes a human spends checking output. It is easy to leave out of a business case.

Why do licences and satisfaction not prove ROI?

Two UK government studies of Microsoft 365 Copilot show why you need your own baseline. Our guide to fixing low Copilot adoption covers them, and the rollout side, in more detail.

  • UK Department for Business and Trade pilot. It ran 1,000 licences from October to December 2024. Mean daily active use was 21% (30% if you exclude weekends and bank holidays), and mean weekly active use was 64%. Satisfaction was 72%. The evaluation found no evidence that time savings had led to improved productivity.
  • UK cross-government experiment. It covered 20,000 government employees in 12 organisations from 30 September to 31 December 2024. Users reported saving an average of 26 minutes a day. That figure is self-reported. Satisfaction averaged 7.7 out of 10.

Together they suggest that people like the tools but use them less often than the licence count implies, and that self-reported time saved is not measured output. So track usage by user and task, and measure outcomes such as cycle time, error rate and cost per unit.

Vendor ROI figures need the same care. Microsoft's quoted return for Copilot comes from a Microsoft-commissioned Forrester study of a composite organisation. It is a modelled projection, not evidence about your business.

What do you do with the scores?

PositionWhat it meansAction
High total, gates passedWhat it means: Valuable and doableAction: Pilot now. Set the baseline, a four to six week window and a stop rule.
High value, low data scoreWhat it means: Worth doing, not readyAction: Fund the data work first: clean-up, permissions, a connector. Re-score in a month.
High value, low risk scoreWhat it means: Needs controlsAction: Design the review step and fallback first, then pilot with a small group.
Low on value or most criteriaWhat it means: Small or not nowAction: Drop it from this round. Re-score next quarter.

Write the stop rule before each pilot starts. For example: "if we do not see the saving after six weeks of use by the target team, we stop or change scope." Pilots without one become permanent costs.

Route each row too. Many are solved by configuring a tool you already have. Some need a workflow tool. A few need a custom build. If the data must stay on your own servers, see Own Your AI and our guide to private LLM versus API cost in rupees.

What mistakes should you avoid when prioritising AI projects?

  • Scoring on the demo. Score on production data and real users, not on a hand-picked sample.
  • Letting value dominate. A high-value use case with a 1 on risk or data is a trap.
  • Skipping the baseline. Without it, any result is an anecdote.
  • Scoring once. Re-score each quarter, because models, prices, data and policies change.
  • Treating the total as the decision. The matrix feeds a conversation with the owner. It does not replace it.

This is an engineering and operations view, not legal advice.

Sources: Microsoft, Microsoft 365 Copilot privacy page (learn.microsoft.com/en-us/copilot/microsoft-365/microsoft-365-copilot-privacy, updated 4 November 2026) and Microsoft 365 Copilot enterprise pricing page (microsoft.com/en-us/copilot/solutions/enterprise), checked 20 November 2026. Anthropic, Claude pricing (claude.com/pricing) and data-use article (privacy.claude.com/en/articles/7996868, dated 18 August 2026). OpenAI, Enterprise privacy (openai.com/enterprise-privacy, updated 8 January 2026) and data residency help article (help.openai.com/en/articles/9903489). UK Department for Business and Trade, Microsoft 365 Copilot pilot evaluation report (gov.uk). UK Government Digital Service, Microsoft 365 Copilot experiment cross-government findings report, 2 June 2025 (gov.uk). Digital Personal Data Protection Rules, 2025 (core duties from 13 May 2027). Last reviewed 20 November 2026.

FAQ

What is an AI use-case prioritisation matrix?

It is a scored list of candidate AI projects. Each use case is rated on the same criteria, such as value, volume, data readiness, risk, integration effort and adoption. The weighted total ranks the list so a team can choose one or two pilots on evidence instead of opinion. Hard gates such as a named owner and a human review step are checked before scoring.

Which criteria should I use to score AI use cases?

A practical set is value size, volume and frequency, data readiness, risk, integration effort and adoption. Score each from 1 to 5, with higher always better, so risk of 5 means low risk. Weight data readiness and risk heavily, for example 20% each, because a use case that demos well can still fail on poor data or on errors it cannot afford.

How do I measure an AI readiness score for my company?

Rate five dimensions as 0, 1 or 2: data, process, people, tools and access, and governance. A total of 0 to 4 suggests fixing foundations first, 5 to 7 suggests low-risk pilots, and 8 to 10 suggests going ahead with pilots across the top-ranked use cases. These bands are a template choice for discussion and not an industry benchmark.

How do I baseline ROI before building an AI tool?

Time 20 to 30 recent real cases end to end, record volume, error or rework rate, cost per unit and cycle time, then compare them with the same measures during a pilot. Calculate saved hours as volume times the minutes saved per item after allowing for human review, and subtract licence, usage, build and run costs.

Why do Copilot licences often not deliver the expected ROI?

Licences are not usage and satisfaction is not output. The UK Department for Business and Trade pilot of 1,000 licences found mean daily active use of 21% and no evidence that time savings had improved productivity. A wider UK government study relied on self-reported savings of 26 minutes a day. Measure your own usage and outcomes.

Turn your AI idea list into a ranked pilot plan.