Why AI Projects Fail: The Checklist Before You Start

You have a job in mind for AI. Maybe it's the calls nobody answers after 6pm, or the quotes that sit in someone's inbox for three days. A vendor has shown you a demo and it looked good. Before anyone starts building, a handful of decisions settle whether the tool ends up in daily use or in a browser tab nobody opens.

This article is for the weeks before the build. If your project has already stalled, read how to rescue a failing AI project instead. It covers whether a stuck project can be saved and the order to fix it in.

Why AI projects fail

The most useful answer I know of comes from RAND. In a 2024 report, The Root Causes of Failure for Artificial Intelligence Projects and How They Can Succeed, its researchers interviewed 65 data scientists and engineers, each with at least five years of experience building AI and machine-learning models. Fifty of them worked in industry and 15 in academia. The five root causes the report names come from the 50 industry interviews:

• The people asking for the project misunderstand or miscommunicate the problem it should solve.
• The organization lacks the data to train a model that works.
• The team chases the newest technology instead of the problem its users have.
• The organization has no infrastructure to manage its data and run the finished model.
• The problem is too hard for AI to solve.

Read the scope before you apply it to a ten-person business. RAND says it wrote the report so that "U.S. Department of Defense leadership and others" can avoid these failures. Its industry interviewees came from companies of many sizes, not small businesses in particular. It also left out projects that only used a pretrained language model without training or customizing one. If the tool you're considering runs on an off-the-shelf model like the ones behind ChatGPT, it sits outside the study's scope.

Still, every one of the five can be checked before anyone writes code. The sections below turn them into questions you can answer this week.

What to decide before anything else

Decide which job the AI will do, and what that job costs you today. "Use AI in the business" isn't a project. "Answer the calls we miss after 6pm and book the job into the calendar" is. RAND's first and third causes both start here, with a problem nobody pinned down or a tool picked before the problem.

Write these down on one page:

• The task, in one sentence a new hire would understand.
• How often it happens in a normal week.
• What it costs now, in hours or lost jobs. A rough number is fine.
• What a mistake would cost. A wrong reply about opening hours is cheap to fix. A wrong invoice isn't.

If you can't fill in that page yet, the project isn't ready, and no vendor can fill it in for you. To choose between several candidate jobs, where to start your first AI project has a four-part test.

Is your data ready for AI?

Every AI tool answers from something: your price list, your calendar, past quotes, the shared inbox. Pull twenty real examples of the job from last month, messy ones included. For each one, check whether the right answer was written down somewhere a tool could read. If half of them depended on someone remembering, fix that before the build. Your data is the hard part walks through finding the one source that holds the right answer.

Then check whether the tool can reach that source. Does your booking system or CRM let other software read from it and write to it? Who on the team can export the data today? For a small business, that's what RAND's infrastructure point comes down to. A tool that can't book into your calendar will only ever take messages.

If the data includes customer or patient details, decide now who is allowed to see it and where it may be stored. Changing that after the build is expensive.

Who owns the AI tool after launch?

Name one person in the business who owns the tool once it runs. It doesn't need to be the person who builds it. The owner decides when it's ready, reads what it got wrong each week, and switches off the old way of doing the job once the new one works. Without an owner, the old process keeps running beside the new tool, and the tool becomes optional.

Decide where a person checks the output before a customer sees it, and who that person is on a Saturday. Start with more review than you think you need. You can pull it back once the weekly review shows the tool getting the job right.

Settle the questions about the builder before you sign. Who fixes the tool when it breaks? Who pays the monthly running cost? If you part ways, do you keep the code, the prompts and the data? Questions to ask before you sign with an AI vendor has the full list.

How to scope a first AI build

Keep the first build to one job on one channel. After-hours calls only, say, and leave email and web chat for later. The narrow version ships sooner, and running it teaches you what the wider one needs.

Write down what the tool won't do and hands to a person instead: complaints, refunds, safety questions, anything it isn't sure about. A short out-of-scope list stops the project from growing every time someone in a meeting has an idea.

Ask the builder outright whether current AI handles your task well. That's RAND's fifth cause. Reading a web inquiry and offering open slots is well within what current AI does. Judging from a blurry photo whether a roof needs replacing is a much harder one. You want to hear which side of that line your job sits on before you see a quote.

Get two numbers in writing: the price of the narrow build and the monthly running cost at your real volume. How much custom AI costs explains what drives both.

What does "live" mean for your project?

Agree in writing what "live" means before the build starts. A useful definition names the job, a time period and the number you'll check. For example: "For four weeks the tool answers every after-hours call and books jobs into the calendar. We compare booked jobs with the same four weeks last year."

Measure that number today, so you have something to compare against, and agree the go/no-go line in advance. Without it, a pilot can't pass or fail, and it stays "in testing" for months. Once the build is done, the AI pilot to production checklist covers the seven gates between a working demo and daily use.

The AI project checklist

Run through this before you sign anything:

• The job is written down in one sentence, with how often it happens and what it costs now.
• You know where the right answer lives, and twenty real examples check out against it.
• The tool can read from and write to the systems the job needs.
• You know who may see the data and where it will be stored.
• One named person owns the tool after launch, and you know who reviews its output.
• The first build covers one job on one channel, with a written list of what goes to a person.
• The builder has told you whether current AI handles this task well.
• You have the build price and the monthly running cost at your volume.
• "Live" is defined in writing, with one number, today's value and a go/no-go line.

If you can tick every line, you're ready to talk to a builder. If the first two are still blank, nobody should start building yet.

Our free AI-readiness audit is a short guided interview that scores your business on these questions and points to a good first job. When you're ready to build, the AI Lead Responder and AI Quote Automation pages show what a narrow first build looks like. And if a project you started earlier is stuck, an AI project rescue starts with a written diagnosis of why.

Frequently asked questions

Why do AI projects fail?

In the 50 industry interviews behind RAND's 2024 report, five root causes came up: a misunderstood problem, missing data, chasing technology instead of the problem, missing infrastructure, and problems too hard for AI. A small business can check for each of them before the build starts.

What should an AI project checklist include?

The job and what it costs today, where the data for it lives and whether the tool can reach it, who owns the tool after launch, a narrow first scope with a list of what goes to a person, the build price and running cost, and a written definition of "live" with one number to measure.

How do you know if your business is ready for an AI project?

You can name one job, say how often it happens and what it costs, and point to where the right answer for it is written down. If you can't do that yet, sort it out before you spend money on a build.