AI project feasibility: what to check before choosing the technology.
Before investing in an AI project, it pays to answer a few concrete questions: is the data there, and can it be used? What does it cost to prepare it? Who will run the service after release? This page sets them out in order. Answering them takes no technical skills: it takes knowing your own business.
These are the same points where many businesses stop: according to Istat, the Italian national statistics institute, in 2025 businesses that considered AI without adopting it named as obstacles a lack of skills (58.6%), unavailable or poor-quality data (45.2%), data protection (43.2%) and costs (43.0%).
What this page assesses, and what it does not
These checks tell you whether the project is sustainable for the business. Whether the system's answers are correct and useful is a separate, technical piece of work, to be assessed with tests on the real case: passing the former does not guarantee the latter. For both, the intended service must be clear.
1. Can the business provide and prepare the data?
The availability of data and the people who can prepare it is an initial investment to estimate.
- Where is the required data, and who can authorise access and use?
- Is it sufficiently complete, current and representative of the intended work?
- Who can extract, correct and organise it, stating what each item of data contains?
- Do you need examples already classified by hand, or correct answers to compare the system against? Who can prepare and verify them, knowing how the business works?
- What people, time and costs will this work require? How will the data be kept up to date?
Classifying examples by hand (in the jargon, labelling them) is not necessary for every project: it depends on the task and approach. When it is needed, the business must be able to do it reliably. Even when it is not, preparing and maintaining data can account for a substantial part of the investment.
2. Can the data leave the business?
The first check concerns the information that must enter the AI workflow:
- Can the data be sent to an external AI service?
- Does it include personal or confidential data, or data subject to contractual or regulatory restrictions?
- Where is the data processed and stored, for how long, and can the provider use it for other purposes?
- Can the data be minimised or anonymised? If names are replaced with codes (pseudonymisation), what could still lead back to a person?
Replacing names with codes reduces risk, but data that can still be linked to a person remains personal data. If some data cannot be sent to an external service, would the project still work without it? Could the request be reframed, or the sensitive information kept in a separate, controlled stage? If the model needs that data to produce a useful result, the environment and workflow must meet the relevant constraints.
3. Performance, costs and ongoing operation
Once the data constraints are clear, estimate the practical requirements:
- Response time: how long can a user wait for each request?
- Concurrent users: how many people will use it at the same time, on normal days and at peak times?
- Continuity: what availability and growth are required?
- Total cost: what does each request cost at the expected volume? The cost per request cannot be derived from the price per token: it has to be measured with trials on real requests (why). What will data preparation, integration, human review and infrastructure management add?
- Operation after release: who will oversee the service, collect reports and detect any decline in result quality? What data, people and budget will be available to respond?
These factors must be considered together. A solution may be too slow, costly or capacity-limited for the workload, or incompatible with the project's data. Work continues after release: monitoring and subsequent intervention bring costs and responsibilities that must be planned from the start.
4. External service or your own server?
- Pay-as-you-go external service: you pay according to use and there is nothing to manage in-house, but the data leaves the business.
- Dedicated cloud resources: more control over data and access, with a fixed cost even when the service is little used.
- Server on your premises: the greatest control over the data's path, at the price of hardware to buy, update and monitor.
5. Does the service you integrate hold up across all your cases?
You do not need to build a product based on generative AI to face this problem: it also applies when you integrate a ready-made service into your own applications, with or without AI.
An external service should be assessed on the business's real cases, not on those of the demonstration. An example: a service that turns addresses and place names into coordinates (geocoding) may be very accurate in large cities and less so in smaller towns. If your users are there, that is the quality they will get.
- What level of quality do you want to guarantee your users, and for which cases?
- Does the service you rely on guarantee it for all those cases, or only for the most common ones?
- Where it falls short, can you make up for it with your own means?
If the answer to the last two questions is no, there is a problem. It needs to be understood straight away, before integrating the service and not after release.
In short: the project holds up if…
- It is clear which service you want to offer, and to whom.
- The necessary data exists, can be used, and there is someone who can prepare it and keep it up to date.
- The expected response times and number of users are compatible with the chosen solution.
- The total cost, including operation after release, is sustainable.
- The external service you rely on delivers the required quality in the cases that matter, or there is a way to make up for it where it falls short.
- Someone, inside or outside the business, is responsible for the service over time.
What to ask whoever proposes the project to you
- Which of our data have you tried it on, and with what results?
- What does it cost per month with our number of users, and what happens if that number doubles? How did you estimate it: from trials on our cases or from the price per token?
- Where does our data end up, and who can see it?
- Who steps in if the answers get worse, and what does that cost?
- What happens if the model provider changes its prices or withdraws the service?
- In which cases does the service perform worst, and what happens there?
- What is the simplest version that would already bring us a benefit?
I can carry out this assessment with you, before any commitment to development.
Supporting sources
A selection of primary sources supporting the criteria above:
- NIST — AI Risk Management Framework, Core (opens in a new tab): testing and measuring quality before release, data assessment and production monitoring.
- OpenAI — evaluation best practices (opens in a new tab): success criteria, test cases, metrics and repeated checks to assess answer quality.
- Microsoft — evaluating answers in RAG systems (opens in a new tab): a concrete example of assessing relevance, grounding and completeness.
- Google Cloud — data labelling (opens in a new tab): the role of labelled data in projects that need it.
- EDPB — summary of its pseudonymisation guidelines (opens in a new tab): pseudonymised data remains personal when it can be linked to a person.
- Cayo and Talbot — positional error in geocoding residential addresses (opens in a new tab): 2003 study of 3,000 addresses in New York State: the error grows as population density falls.
- Chow et al. — geographic disparities among geocoding solutions (opens in a new tab): 2016 comparison of eight solutions on 1,100 addresses in Texas: significant differences between urban and rural areas, though not always in the same direction.
- Istat — Imprese e ICT, 2025 (in Italian) (opens in a new tab): AI use in Italian businesses with at least 10 employees (15.7% among SMEs, 53.1% among large businesses) and the obstacles named by those that considered it without adopting it.