How to choose the right use case for your first AI project

The best first use case is rarely the largest or most spectacular. It is important enough to create value, sufficiently bounded to be evaluable, and safe enough for the organization to learn quickly.

A practical prioritization weighs utility and feasibility against risk. The model below can be used in a workshop to compare ideas on the same basis.

Start with the problem

Formulate the problem in business terms. "We want an agent" says nothing about what will get better. "Case officers spend two hours per week compiling the same information from four systems" is something that can be investigated and measured.

Describe users, frequency, current state, consequence, and desired outcome. Then identify several possible solutions. Generative technology could be one, traditional automation or improved information management another.

Digg recommends that organizations start from the need and consider alternatives before acquiring generative technology.[1]

AI in the management team: 10 questions that take you from idea to business value

Six criteria for a good first use case

1. Business benefit

How large is the possible improvement in time, quality, cost, revenue, or service? Score highest when the benefit is clear, relevant to a prioritized goal, and has a named owner.

2. Data availability

Is data available in sufficient quantity and quality? Is it permitted to use, does it have an owner, and can it be connected? A document-intensive case may be attractive but unsuitable as a first project if the documents are contradictory and lack an authorization structure.

3. Risk

What happens if the solution answers incorrectly, leaks information, or acts wrongly? Assess the impact on individuals, finances, security, rights, and trust. Low risk gives higher priority for a first test, all else being equal.

4. Frequency of use

A task that occurs often provides more test opportunities and greater potential benefit. Very rare events may be important but difficult to validate during a short pilot.

5. Measurability

Can you establish a baseline and see an effect within the pilot's timeframe? Clear metrics make the decision less dependent on isolated positive reactions.

6. Ability to change

Is there a sponsor, problem owner, test users, and the opportunity to change the workflow? A good model creates no value if no one has the mandate to implement it.

The best first use case is important enough to create value and sufficiently bounded to be evaluable.

Benefit versus feasibility

Place the ideas in a matrix with expected benefit on one axis and feasibility on the other. At the same time, mark risk with a separate color or symbol.

High benefit, high feasibility

Strong candidate for a pilot.

High benefit, low feasibility

Conduct a feasibility study or solve data issues first.

Low benefit, high feasibility

Can work for learning but risks lacking a sponsor.

Low benefit, low feasibility

De-prioritize.

The matrix is not an automatic decision. A high-risk pilot may require more control even if the benefit is great. A smaller pilot can be strategically valuable if it builds common infrastructure for multiple future needs.

What data exists?

Inventory both technical access and the right to use the data. Ask:

• Where is the information and who owns it?

• Is it structured, current, and representative?

• Does it contain personal data or sensitive information?

• Do permissions carry over to the index and model?

• Is there a ground truth or are there domain experts for evaluation?

• Can the data continue to be updated after the pilot?

Avoid use cases where the pilot's main task becomes repairing the entire organization's data management, unless that is the intended project.

What happens to the company's data when you use AI?

Risk and human control

Preferably, choose a first use case where the result is used as support and can be reviewed before it has consequences. An internal assistant that finds sources is often easier to bound than a system that makes or executes decisions about individuals.

Define what the model should do in case of uncertainty, what answers are forbidden, and when a human must approve. The NIST playbook can be used as support to link risks to governance, measurement, and actions.[2]

AI Act, NIS2, DORA & GDPR – what requirements affect the company's AI solution?

Can the value be measured within a pilot?

Choose a timeframe in which enough real tasks occur. Measure before and during the pilot.

Examples:

• time to find the right document

• share of suggestions accepted after review

• forecast error compared to baseline

• reduced manual registration

• share of answers with correct source

• number of errors and need for escalation

Also calculate control time, training, and new work steps. The users' experience is important but should complement, not replace, objective measures.

From idea to pilot: what is PoC, prototype, and pilot?

Examples of high and low priority use cases

High priority

A department has well-maintained manuals, clear permissions, and recurring questions. An assistant can be tested with known answers and search time can be measured.

How to connect a secure chat to your organization's own data

Lower priority

The organization wants an agent to automatically make complex decisions in multiple systems, but the process lacks an owner and the data is incomplete.

AI assistant or AI agent – what is the difference?

Feasibility study first

A forecast could free up significant capital, but stock shortages and campaign data are not registered. Start with data and a baseline.

Prioritization template

Score each criterion 1–5:

• business benefit

• data availability and quality

• low consequence in case of error

• frequency of use

• measurability within 8–12 weeks

• sponsor and ability to change

Write a comment and owner for each score. An automatic total sum is not sufficient; a legal showstopper can outweigh high benefit.

Select the two or three strongest ideas for a brief feasibility study. Only then do you decide which one will become the pilot.

Sources

1. Digg, Make conscious decisions about acquiring generative AI:
   https://www.digg.se/ai-for-offentlig-forvaltning/riktlinjer-for-generativ-ai/ta-medvetna-beslut-om-att-skaffa-generativ-ai

2. NIST, AI RMF Playbook:
   https://www.nist.gov/itl/ai-risk-management-framework/nist-ai-rmf-playbook