# 7 questions before starting an AI project

> Most AI projects struggle not because of the model itself but because of questions left unanswered at the start. This guide covers seven questions to ask before writing code and why each one matters.

**In short:**

- Define the decision or process first, then choose the method.
- Data quality and representativeness matter more than the model itself.
- Write success down as a measurable criterion up front.
- A plan for production, monitoring and retraining is part of the project.

## 1. Which decision or process are we improving?

AI is a tool, not a goal. Instead of "using AI", define a concrete decision or process, such as "improving return forecasts" or "routing support requests to the right team". Once the goal is clear, whether AI is really needed becomes visible too.

Problems with clear, rarely changing rules are often solved more reliably and more cheaply with classic software. AI makes sense when rules are too complex to write or keep changing.

## 2. How will we measure success?

"Better" is not a measure. Before you start, write down which metric, above which value, will make the project a success. The choice of metric depends on the cost of the problem: is a false alarm more expensive, or a missed event?

- In classification, evaluate accuracy, precision and recall together.
- In forecasting, watch the worst-case error as well as the average error.
- Use the result of the current method (rules or a manual process) as the baseline to compare against.

## 3. Is our data sufficient and representative?

A model learns the world of the data it sees. If training data does not reflect real use (one period, one customer group, one device), the model fails in real life. Missing, mislabelled or skewed data makes even the most advanced model useless.

How much data is enough varies with the problem, so building a quick prototype on a small dataset first shows whether collecting more data would add value.

## 4. What are the permissions and privacy requirements for using the data?

Who owns the data, why it was collected and whether it may be used for model training should be settled up front. If it contains personal data, compliance with the relevant regulation and data minimisation (using only the necessary fields) should be part of the design.

## 5. Who will use the model's output, and how?

The model's result enters a screen, a workflow or another system. Can the user see how confident the model is? Is there a way to correct a wrong result? For high-risk decisions, the model's suggestion and human approval should be designed together.

## 6. How will we move the model to production?

A model that works in a notebook is not a working product. It must be served as a service, versioned, tested and meet latency and cost goals.

1. Record the versions of the model and the data.
2. Release the service with automated tests.
3. Leave a way to roll back to the previous version on failure.

## 7. How will we notice if it gets worse over time?

The real world changes; customer behaviour, products and data formats drift (model drift). Without monitoring, a model quietly degrades. Keep watching the distribution of input data and model performance, get an alert at a set threshold and plan the retraining process from the start.

## After you answer these questions

Once the answers are clear, building the smallest working example and measuring is the fastest way to confirm your direction without a big investment. VeriSet A.Ş. runs this process through the steps research, prototype, build and scale. For more, see our artificial intelligence area page.

---
https://veriset.org/en/guides/before-starting-an-ai-project/ · destek@veriset.org