How to prepare an AI project before asking for a quote: the 11 questions

Come to a software company with a vague idea and you get vague quotes. Three vendors, three numbers far apart, and no serious way to compare them: each one imagined a different project from the same two lines of description.

In AI projects the problem is even sharper. AI is easy to try and hard to put into production, and the reason is almost never the model: it's what happens before it. This article explains what to clarify before the first call, with the eleven questions we use at Gorilli during discovery.

Why so many AI projects stall before production

The most cited research tells the same story. According to estimates reported by RAND Corporation (2024), more than 80% of AI projects fail: twice the rate of IT projects that don't involve AI. MIT NANDA's report The GenAI Divide (2025) found that 95% of generative AI pilots in companies delivered no measurable impact on the P&L.

RAND identifies five recurring causes:

• A poorly defined problem. Leaders don't clarify what to solve or pick the wrong metric, so the team optimizes for the wrong goal.
• Not enough data. Good-quality data is missing, hard to access or not understood by domain experts.
• Technology before the problem. Teams chase the newest tool instead of business value.
• Neglected infrastructure. Data pipelines, monitoring and deployment come too late.
• Expectations beyond today's limits. Some problems can't be solved with AI yet, however much effort goes in.

The good news is that the first three causes can be tackled before writing a single line of code. All it takes is writing down the right answers.

What changes compared to a traditional software project

An AI project is planned differently from a management system or a website. Knowing the differences avoids the most common surprises on timing, cost and results.

In practice: set the error margin before you see the demo, bring real data as early as possible (a small but real sample beats a perfect dataset that arrives in three months), and ask for an estimate of monthly usage costs, not just the development cost.

The 11 questions to answer before the first call

You don't need a perfect document. Short answers, written with the people who live the problem every day, are enough. "I don't know" is a valid answer: it shows where more digging is needed.

The starting point

1. The problem. What isn't working today, who suffers from it, and what does it cost in time, money or lost customers? "We want a chatbot" isn't a problem; "we answer customers too slowly" is.

2. The goal. What changes if the project succeeds? You need a measurable result within six months of launch, for example halving the average response time.

3. The users. Who will use the software: internal teams, customers, partners? How many are there, and how technical are they? A tool for twelve non-technical operators is a different project from an assistant for three thousand customers.

The scope

4. What it must do. List the features and mark the ones essential for the first release. The essentials define the MVP; the rest comes later.

5. Where AI comes in. What do you expect the AI to do: answer, search documents, classify, extract data, generate text, suggest decisions? And above all: how much error is acceptable, and who checks the output? An AI that suggests a reply to an operator, who approves it before it's sent, has very different requirements from one that replies on its own.

The context

6. The data. What data exists, where is it, and who is responsible for it? Is there personal or sensitive data? This is often the question that moves timing and cost the most.

7. Integrations. Which systems must the software talk to: ERP, CRM, e-commerce, email, WhatsApp, company login?

8. Constraints. Timing, budget and requirements the project must meet. If there's a deadline, explain why: a trade fair, a season, a grant. An indicative budget helps propose the right solution, not inflate the quote. Technical or regulatory requirements count too, such as EU hosting, GDPR or accessibility.

The ground rules

9. What the project must not do. What stays out, at least in the first version? Saying it explicitly stops each vendor from imagining a different scope.

10. How you measure success. Which numbers will you look at three and six months after launch? Without a baseline number, you won't be able to tell whether the project worked.

11. Who decides. Who is the day-to-day contact, who decides, and who approves the budget? It helps understand decision timelines and who to involve in calls.

From idea to production, step by step

A good AI project reduces uncertainty one step at a time. Each phase answers one question and decides whether the next one is worth the investment.

1. Discovery: what are we solving, with which data? Problem, users, data and constraints written down. You leave with a scope and an estimate. The eleven questions are the work of this phase.
2. Proof of concept: does it work on our data? The most uncertain part is tested on real cases. You leave with a yes, a no or a "yes, but".
3. MVP: do users actually use it? The smallest version, used by a small group in their daily work.
4. Pilot and production: does it hold up at scale? Rollout to more users, quality and cost monitoring, full integrations.

Three mistakes to avoid

• Starting from the tool. Technology is chosen after understanding the problem, not before.
• Skipping the test on real data. Demos on sample data almost always work.
• Having no baseline metric. Without today's number, tomorrow's result can't be judged.

How to compare quotes

With the same answers in hand, quotes become comparable. Before choosing, check:

• Do they all start from the same document?
• Do they include data preparation and quality testing?
• Do they estimate monthly model usage costs?
• Who owns the code, data and instructions at the end of the project?
• What happens after launch, in terms of monitoring and maintenance?

For the criteria to evaluate the vendor itself, we wrote a dedicated guide: how to choose an AI software house in Italy.

Frequently asked questions

How long does it take to answer the 11 questions?

About thirty minutes, ideally together with the people who live the problem every day. Short sentences are worth more than a perfect document.

Do I need to know which technology to use?

No. The questions are about the problem, data, users and constraints. Choosing models and architecture is the vendor's job, and it should come afterwards.

How much does an AI project cost?

It depends mostly on data, integrations and the acceptable error margin, which are questions 5, 6 and 7. That's why an indicative budget helps: it lets the vendor propose a fitting solution instead of a generic one. Remember to ask for monthly usage costs too, not just development.

What if I can't answer a question?

Write "I don't know". It's useful information: it shows where discovery needs to dig deeper.

Can I use the answers with any vendor?

Yes, and that's the best way to use them: the same document for every vendor, so quotes start from the same basis.

Want the version to fill in?

The eleven questions, with practical examples and space for your answers, are in the free guide What you need to know before building an AI project, in PDF and Word. If you're already clear on your idea, you can book a free 30-minute AI Check-up.