How to pick your first AI use case

What makes a good first AI project?

A good first AI project is one where the company sees results quickly, errors can be corrected, and lessons carry over to the next project. The goal is not to solve the biggest problem. It is for the organisation to learn how to work with AI.

Good first projects share a few traits:

• Frequent. The task happens every day or every week, not a few times a month.
• Clear inputs and outputs. It can be described simply: “an invoice arrives, the fields go into the ERP”.
• Recoverable errors. A wrong suggestion can be caught by a person reviewing it.
• An owner. A manager is accountable for the outcome and can make decisions.
• Measurable. You know, or can easily measure, how long the task takes today and how often it goes wrong.

Picking the wrong first project costs more than budget. A failed pilot leads teams to conclude that “AI does not work here”. Keep the ambition small and the learning large.

How do you list and score use cases?

Start the list from the work, not the technology. Ask each department two questions: “Which repetitive task takes most of your time?” and “Which decision would you make better with better data?” These two questions usually produce 15 to 30 candidates.

Then score every candidate on the same criteria. The matrix below defines each on a 1 to 3 scale. The point is not a precise calculation; it is a shared language.

Criterion	1 (weak)	2 (moderate)	3 (strong)
Business value	Unclear benefit, or it affects one person	Noticeably speeds up a team	Direct effect on revenue, cost or customer experience
Data readiness	No data, or scattered and compiled by hand	Data exists but needs cleaning and joining	Data in one place, regular and accessible
Process repeatability	Rare and different every time	Weekly, with a recognisable pattern	Daily and standard
Cost of error (reverse scale)	Errors reach customers, health or legal matters	Errors cause internal delays	Errors are easily caught by human review
Compliance risk (reverse scale)	Significant decisions about people, or high-risk use	Personal data is processed and needs controls	No or very limited personal data
Ownership	No owner, “everyone’s job”	An owner with limited time	An owner who takes the pilot on

Four examples show how to read the matrix. We do not give scores; the result differs in every company.

• Demand forecasting. Business value is usually high. The deciding criterion is data readiness: is sales history regular and long enough, and are promotions and stock-outs marked?
• Visual quality inspection. Repeatability on a production line is high. The deciding criterion is cost of error: if a missed defect reaches the customer, the system should only alert a person at first.
• Credit risk scoring. High value, but also high compliance risk. It makes decisions about people, so it needs careful review under GDPR, Türkiye’s KVKK and the EU AI Act. It is rarely a good first project.
• Invoice processing. Repeatability and data readiness are usually high, and the cost of error is low. For many companies it is a strong first candidate.

How do you check data readiness quickly?

Five questions show where you stand before you start any data project:

1. Where? Which system holds the data: ERP, CRM, email, spreadsheets, paper?
2. How much? Are there records for the period the use case needs?
3. How clean? Of 50 randomly chosen records, how many are complete and consistent?
4. Accessible? Can it be exported or pulled through an API, or does someone copy it by hand?
5. Whose? Who has to approve its use, and does it contain personal data?

The answers are also a quick picture of your digital maturity. If most of them are “we don’t know”, your first project may be collecting the data, not AI.

How do you apply a risk and compliance filter?

After scoring, run the top candidates through a filter. This section is not legal advice; review it with your legal counsel.

• Data protection. Does the use case process personal data? If so, the legal basis, the privacy notice and the country where data is processed need to be clear under GDPR, and under KVKK if you operate in Türkiye.
• EU AI Act. Uses such as creditworthiness assessment, recruitment or assessment in education can fall into the high-risk category. Staying away from these areas in a first project is sensible.
• Risk management frameworks. The NIST AI Risk Management Framework organises risk into govern, map, measure and manage functions. The OECD AI Principles stress transparency, accountability and human oversight. You do not need the whole framework for a first project, but you do need answers to “who oversees this, and how do we roll it back?”

How do you define and measure pilot success?

Write down what success means before the pilot starts. Otherwise every result gets called “promising”.

A good pilot definition contains four things:

• Baseline. How the task performs today: time per transaction, error rate, waiting time.
• Target. The value at which you will say “continue”.
• Scope and duration. Which team, which data, until which date.
• User response. Does the team actually use the tool, and where do they give up?

Google’s People + AI Guidebook recommends designing up front for how much users will trust the system and how they will deal with its errors. In the pilot, measure not only model accuracy but how people check and correct the output. That is why basic skills such as prompt engineering belong in the pilot.

Which signs say “not AI yet, fix the process and data first”?

Sometimes the right decision is to wait. These signs point that way:

• Every team runs the process differently, and nobody can describe the right way.
• The data lives in several spreadsheets, in different formats, maintained by hand.
• The real problem is a missing approval step or integration.
• The task has no owner, or nobody wants to measure the outcome.
• A simple rule or automation would do the same job.

In that case, simplifying the process and putting the data in one place prepares the ground for the next AI project. For SMEs in particular, this step often brings a quick, visible gain.

Where should you start?

A half-day workshop with department heads is enough to produce the candidate list. Score it with the matrix above, run the top three through the data and risk filter, and pick one.

At Corrplus we start with a free discovery call, then assess the candidates with you during our two-week AI discovery sprint. See our AI consulting, digital transformation strategy and corporate AI training services. For sector-specific use cases, see manufacturing and ecommerce and retail.