Guide · 8 min read

Should you hire an AI consultant or build in-house? What the failure-rate studies show

Should you hire an AI consultant or build AI in-house? What research on AI project failure says, where each option fits, and a simple way to decide.

Reviewed by Caelen Yung, Founder, True Meridian AI · Published October 13, 2026 · Updated October 17, 2026

Two colleagues discussing a project at a table with a laptop in a bright office — Photo by Alena Darmel on Pexels

On this page

What the studies say about failure rates
Why in-house AI projects stall
What outside help does well, and where it goes wrong
When building in-house is the right call
A simple way to decide
The cost question
Start with an audit
Sources

Sooner or later most owners ask the same question: do we hire someone for this AI work, train a person we already have, or bring in outside help? The honest answer depends on the business. The research does lean one way on how often projects reach real results, though, and it is worth knowing what it says and how far it can be trusted.

A disclosure before we start. True Meridian AI sells outside help, so read our reasoning with that in mind. Every study below is public and linked at the end, so you can check what we say about it.

What the studies say about failure rates

AI projects fail often, whoever builds them

A 2024 report from the RAND Corporation is based on interviews with 65 data scientists and engineers, each with at least five years of experience building AI and machine learning systems. It says some estimates put the failure rate of AI projects above 80 percent, about twice the rate for corporate IT projects that do not involve AI. RAND is quoting those estimates, not measuring them itself. What its interviews add is a list of causes. Leaders and technical teams misunderstand what problem the AI is meant to solve, which was the most common cause. The data is not sufficient. Teams chase the technology instead of a real user problem. The infrastructure to manage data and deploy models is missing. And the problem is too hard for current AI.

In July 2024, Gartner forecast that at least 30 percent of generative AI projects would be abandoned after the proof-of-concept stage by the end of 2025. It named poor data quality, inadequate risk controls, escalating costs, and unclear business value. That was a forecast, and we did not check how it turned out.

A BCG survey of 1,000 executives in 59 countries, published in October 2024, found that 74 percent of companies had yet to show tangible value from AI. Only 4 percent consistently generated significant value, and another 22 percent were beginning to see substantial gains. BCG also reported that roughly 70 percent of the challenges were about people and processes, about 20 percent were about technology, and about 10 percent were about the AI algorithms themselves.

One widely cited study compares building with buying

The comparison closest to your question comes from the MIT NANDA initiative. Its 2025 report, The GenAI Divide, found that about 95 percent of the generative AI pilots it examined delivered little or no measurable effect on profit and loss. As Fortune reported it, AI tools bought from specialized vendors and built through partnerships succeeded about 67 percent of the time, while internal builds succeeded only one-third as often. The research drew on 150 interviews with leaders, a survey of 350 employees, and an analysis of 300 public AI deployments. We could not open the original report, so these numbers come from Fortune's coverage of it.

Large IT projects overrun even without AI

AI is not the only thing that goes wrong. A 2012 study by McKinsey and the University of Oxford looked at more than 5,400 large IT projects, each with an initial price above $15 million. On average they ran 45 percent over budget and 7 percent over schedule, and they delivered 56 percent less value than predicted. Those projects are far larger than anything a small business builds, and the data is more than ten years old. Treat it as a reminder that technology projects are hard, not as a forecast for yours.

What the studies cannot tell you

None of them compares small-business in-house AI work with small-business agency work, head to head.
Each study defines success differently, and some of the numbers are forecasts or estimates, not measurements.
Companies that buy outside tools may pick easier problems, so part of the gap between buying and building could come from what each group attempted.
Several of the studies cover large companies or large projects. The McKinsey study only includes projects above $15 million.

One Canadian number gives some context. Statistics Canada reports that 12.2 percent of businesses used AI to produce goods or deliver services in the second quarter of 2025, up from 6.1 percent a year earlier. In our reading, that means most small businesses have little in-house experience to build on and few peers to copy.

The research supports a modest conclusion. In at least one major study, outside partnerships reached production more often than internal builds. The most common causes of failure, an unclear problem, weak data, and poor process, can hit anyone, whoever does the building.

Why in-house AI projects stall

This section is our interpretation of the causes above, applied to a small business. It is not a finding from any one study.

Nobody has it as their job. The person asked to build the automation also runs operations, so the project happens in the gaps. A part-time project is where a mismatch about the problem goes unchecked, which RAND found to be the most common cause of failure.
One person holds all the knowledge. If they leave or get busy, nobody else can maintain what they built.
Nobody has done it before. An outside team has seen similar problems in other businesses. A first-time builder learns on your time.
People and process are the hard part. BCG puts about 70 percent of the challenges there, and an in-house builder has to change habits among colleagues they work with every day.
Maintenance gets forgotten. Software tools change, connections break, and automations need someone watching them.

What outside help does well, and where it goes wrong

A good outside team brings experience from other businesses, can start on a defined scope within weeks, and gives you a fixed price for a fixed result. It also leaves your own people free to run the business.

Outside help also has real failure modes:

Lock-in. The work lives in accounts and tools you cannot access or take elsewhere.
A black box. Nobody on your side understands how it works, so you cannot fix it or judge it.
Misaligned incentives. A provider paid to build has a reason to recommend building.
Shallow understanding. The team learns your business for a week and misses what your staff know.
A weak handover. The project ends and the knowledge leaves with the team.

You can guard against most of these. Own the accounts and the workflows. Ask for written documentation. Agree a fixed scope and price. Ask what the provider would recommend against. Speak to a past client. The AI audit guide ends with a list of questions to ask any provider.

When building in-house is the right call

AI is part of what you sell, so it is your product and not a tool for running the business.
There is enough volume and change that the system needs constant adjustment.
The data cannot leave your systems under any contract.
You already have a person with the skills and the time.
You have a long horizon and a named owner who will keep it running.

Many businesses end up with a hybrid: an outside team builds the first version and trains your people to run it. If you go that way, ask any provider whether they will do it and what the handover includes.

A simple way to decide

Question	Leans in-house	Leans outside help
Is this what you sell, or how you run the business?	It is what you sell	It supports operations
How many hours of AI work a year?	Enough to keep a skilled person busy	A few projects a year
Do you have someone with the skills and spare time?	Yes	No
How soon do you need a result?	You can wait several months	Within weeks
How sensitive is the data?	It cannot leave your systems	It can be handled under a contract
Who will maintain it?	A named person on your team	You want support after launch

If most of your answers land on the right, outside help is probably the better fit. If most land on the left, build in-house, and still consider a short outside review before you commit.

The cost question

A hire costs a full salary whether or not there is enough AI work to keep them busy. The number that matters is the cost per hour of useful work: the annual cost divided by the hours of AI work you can actually give them.

An illustration with invented numbers. Suppose a skilled hire costs $110,000 a year including benefits and software, and your business has about 400 hours of AI work a year. That is $275 for every hour of work. With 1,600 hours of work, the same hire comes to about $69 an hour. The same person is expensive at low volume and efficient at high volume. Your own figures will differ, so replace the assumptions before you rely on the result.

Outside help is priced per project, so you pay for the work and not for idle time. Compare quotes from two or three providers against the same written scope, and include maintenance in the comparison for both paths.

Start with an audit

You do not have to decide before you know what to build. An AI audit reviews how your business runs, ranks where AI would help and where it would not, and ends with a written report you can take to a hire, a provider, or your own team. Read the AI audit guide to run a basic version yourself, or book an AI audit and we will do it with you.