← Thinking
2026-04-15Methodology

5 Questions We Ask Before Taking an AI Project

Why We Screen

After years of AI implementation work, we've learned a key lesson: projects fail not because the technology doesn't work, but because the problem itself shouldn't be solved with AI.

Some problems can be solved with a three-line rule engine in seconds, but forcing a large model onto them makes them ten times slower and a hundred times more expensive. Some problems lack sufficient data for the model to learn anything meaningful. Some problems change three times mid-project because the client themselves haven't figured out what they want.

So we made a rule for ourselves: before taking on any AI project, we run it through 5 questions first. If it doesn't pass, we don't take it.

This isn't about being picky — it's about being responsible to both sides.

The Five Questions

Question 1: How far can traditional methods get with this problem?

Purpose: Confirm whether AI is truly necessary.

Many clients come to us saying "we want to use AI" as their first sentence. But we first ask: How is this problem currently being solved? With rules, with manual work, with traditional software — how far can those approaches get?

If traditional methods can achieve 80% and AI can achieve 90%, is that 10% improvement worth hundreds of thousands of dollars? If traditional methods can only achieve 40% and AI can achieve 85%, then the investment makes sense.

We don't do nice-to-haves. We do must-haves.

For example: a client wanted to use AI for data validation. When we asked, we found their data format was very regular, and regex could handle 99% of cases. So why use a large model? We recommended a rule engine instead — saved money and got better results.

Question 2: Do you have enough clean data?

Purpose: Confirm whether AI can actually learn.

AI isn't magic — it learns patterns from data. Without data, it's like asking a student to take an exam without reading any books — no matter how smart they are, they won't do well.

We ask:

  • How much historical data do you have?
  • Is the data format consistent?
  • Is there human-labeled "correct answer" data?
  • How's the data quality? Any dirty data, missing values?

If the data is insufficient or too dirty, we recommend data cleaning first rather than jumping straight to modeling.

Some projects require two months just for data cleaning. This isn't wasted time — it's laying the foundation. A weak foundation means the building will collapse no matter how tall you make it.

Question 3: What are the boundaries of this scenario?

Purpose: Confirm what happens when AI gets it wrong.

AI isn't 100% accurate. We need to ask: if AI makes a mistake, what's the consequence?

  • If it's a recommendation system and recommends the wrong song, the user skips it — minor consequence.
  • If it's a medical diagnosis system and misdiagnoses — major consequence.
  • If it's a financial risk control system and misses a fraud case — major consequence.

The heavier the consequence, the more we need human oversight mechanisms. This means higher costs and more complex workflows.

We design different approaches based on risk level:

  • Low risk: AI outputs directly, humans spot-check
  • Medium risk: AI outputs + human review
  • High risk: AI assists + human decision-making

If a client wants a "fully automated, zero human involvement" solution for a high-risk scenario, we tell them directly: it can't be done, or it can be done but the risk is something they can't bear.

Question 4: Can your team handle it?

Purpose: Confirm whether it can be used after delivery.

AI projects don't end when you deliver a model. Models need continuous iteration, data needs continuous collection, systems need continuous maintenance.

We ask:

  • Do you have technical staff who can maintain this system?
  • Is your business team willing to change their workflows to work with AI?
  • Does your management have the patience to wait for AI to gradually optimize?

Some projects succeed technically but the business team won't use them. Because using AI means changing habits, learning new things, and trusting a "black box."

If the team can't handle it, we recommend starting with a small pilot to let the team gradually adapt, rather than going all-in from the start.

Question 5: Are you willing to pay for "imperfection"?

Purpose: Confirm whether expectations are realistic.

AI isn't omnipotent. It will have errors, delays, and limitations.

We ask:

  • Can you accept AI's accuracy being 85% instead of 100%?
  • Can you accept AI needing 3 months of optimization to reach its best performance?
  • Can you accept AI performing poorly in certain edge cases?

If a client is追求 100% perfection, then AI isn't the answer. Traditional rule engines, manual review, or simple if-else statements might be more appropriate.

AI is suited for scenarios where "there's enough room for error, enough data, and the problem is complex enough."

After Screening

Projects that pass these 5 questions are the ones we take on. After taking them on, we do three things:

  1. Research first, then develop. Spend a week deeply understanding the problem before writing any code.
  2. Build MVP first, then scale. Spend two weeks building a working prototype so the client can see results before deciding whether to continue investing.
  3. Build trust first, then collaborate. We explain the technical approach, risk points, and expected outcomes clearly so the client knows what we're doing and why.

We believe good AI projects aren't taken on — they're screened into existence.

Final Thoughts

These 5 questions aren't set in stone. As we complete more projects, these questions continue to evolve.

But the core logic remains: AI is a tool, not a goal. What we need to do is find the problems that truly need this tool, and then solve them in the most appropriate way.

If you have an AI project idea, feel free to reach out. We may not take it on, but we'll definitely tell you: whether this problem is worth solving with AI, and whether it can be solved well.