Guide

Why AI projects fail, and how to prevent it

A study of 33 European organisations found 7 reasons AI projects fail. Each one has a fix you can put in place before you buy, at launch or after it.

7 reasons AI projects fail

  1. 01Adopted on hype instead of fit

    What goes wrong

    The tool gets chosen because peers or vendors are using it. "We need an AI strategy" gets mistaken for having one.

    What to do

    Name the business objective before you choose a tool. Then ask whether you'd do this if no competitor had announced it.

    What the study found

    Organisations in the study often followed trends without a thorough evaluation. Firms under 500 people lacked the expertise for due diligence. In enterprises of over 10,000, department managers felt pressure to be seen innovating.

  2. 02Whole-life cost is undercounted

    What goes wrong

    The business case counts licences and build effort. Maintenance, retraining, governance and expert time are left out.

    What to do

    Before approval, build a whole-life cost model with lines for maintenance, retraining, governance, compliance and expert time.

    What the study found

    Cost models across the study left out maintenance, model retraining, continuous optimisation, governance overhead and compliance updates. The time process experts spent in workshops went uncounted too.

  3. 03Savings don't arrive on their own

    What goes wrong

    The hours and headcount savings get assumed at approval, and nobody checks them afterwards.

    What to do

    Measure the process as it runs today first. Then book a post-launch review and budget time to tune.

    What the study found

    Several organisations saw the savings they expected only after they used post-launch data to remove bottlenecks, rebalance work or tune the thresholds at which the AI decides.

  4. 04The cost of doing nothing is left out

    What goes wrong

    The business case compares the project with zero. It doesn't count what standing still costs.

    What to do

    Add a cost-of-inaction line to every business case.

    What the study found

    Turnover, customer perception and missed innovation were rarely priced in. A product manager in the study noted that people left doing the same repetitive tasks may eventually change jobs.

  5. 05Freed time has no plan

    What goes wrong

    Hours are released and nobody decides where they go, so they drift into other work.

    What to do

    Before launch, decide where the released hours go and who agrees it.

    What the study found

    The strongest organisations moved the time to growth work such as customer conversations, and redesigned roles as well as cutting cost. Others asked their staff what they'd do with the time and got good ideas.

  6. 06No owner after launch

    What goes wrong

    The project ends at go-live. So nobody monitors, retrains or improves the system.

    What to do

    Name an accountable owner and give them the time. Set the review cadence and the handover.

    What the study found

    The best organisations planned scaling, retraining and retirement from day one, and treated automation as a capability that keeps changing.

  7. 07No in-house expertise, and hiring takes time

    What goes wrong

    The team doesn't know enough yet to judge or run AI, so recruiting looks like the answer. We've found a new hire needs months to learn your business before they can help.

    What to do

    Bring in expertise that starts from your process and leaves the knowledge behind. Pair it with a named internal owner.

    What the study found

    Smaller firms in the study lacked in-house expertise, so popularity filled the gap.

7 questions to ask before you buy

Ask them before you approve any new AI project.

Fit

Which business objective does this serve, and would we do it if no competitor had announced it?

A strong answer is one objective, in one sentence, that the business would pursue whatever competitors do.

Pain point 1
Evidence

Have we measured the baseline process, and is this a proof of concept or a scalable investment?

A strong answer has volume, handling time, errors and rework measured, and a case that says which of the two it is.

Pain point 3
Whole-life cost

Does our cost model include maintenance, retraining, governance, training and expert time?

A strong answer prices every line before approval. Licences and the build are two of them.

Pain point 2
Cost of inaction

What do turnover, customer perception and missed innovation cost if we do nothing?

A strong answer puts a cost-of-inaction line in the case, set against the cost of the project.

Pain point 4
Value beyond savings

What will people do with the time released, and what could this enable that we can't do today?

A strong answer is an agreed use for the released hours, signed off by a named person.

Pain point 5
Ownership

Who monitors, retrains and improves it after launch, and do they have the time to?

A strong answer names one owner with protected time and a review cadence.

Pain point 6
Expertise

Do we have the skills in-house to judge and run this, and if not, how do we get them in time?

A strong answer names who has the skills today, or brings in outside expertise that works with a named internal owner and leaves the know-how behind.

Pain point 7

Answer all 7 for one AI project and get a score out of 100 and your 3 weakest areas. It takes about 2 minutes.

Take the 7-question check

What to do before, at and after launch

Before you buy

  • State the business objective in one sentence and apply the "no competitor announced it" test.

    Owner
    Executive sponsor
    When
    Before any vendor demo
  • Measure the baseline process: volume, handling time, error and rework rates.

    Owner
    Process owner
    When
    Before approval
  • Decide whether this is a proof of concept or a scalable investment, and say so in the case.

    Owner
    Sponsor with finance
    When
    Before approval
  • Cost the whole life: licences, build, maintenance, retraining, governance, compliance and expert time.

    Owner
    Finance and technology lead
    When
    Before approval
  • Price the cost of inaction: turnover, customer perception, missed innovation.

    Owner
    Sponsor
    When
    Before approval
  • Agree where released hours will go and who signs that off.

    Owner
    Business unit lead
    When
    Before approval
  • Assess in-house skills honestly. If there's a gap, plan how and how fast it closes.

    Owner
    Technology lead
    When
    Before approval

At launch

  • Name one accountable owner for the result and protect their time.

    Owner
    Executive sponsor
    When
    At go-live
  • Set a balanced scorecard: processing time, error rates, throughput, rework and employee experience. For generative AI, add how often it makes things up and how often a person overrides it.

    Owner
    Process owner
    When
    At go-live

After launch

  • Review post-launch data, remove bottlenecks and tune thresholds.

    Owner
    Process owner
    When
    First review at an agreed date
  • Check what time was recovered after checking and rework, where it went, and whether a customer outcome improved.

    Owner
    Owner with sponsor
    When
    Each review
  • Plan retraining, scaling and retirement as a standing cycle.

    Owner
    Process owner
    When
    Ongoing

What the research found

Researchers at LUT University followed European organisations as they adopted RPA, low-code platforms and generative and agentic AI.

  • 33European organisations, from 11 staff to 30,000
  • 3.5+ yearsspent with those organisations
  • 70challenges catalogued along the way

Findings on strategy

  • Organisations adopted on hype instead of fit.
  • They judged results almost entirely on direct cost savings.

What we take from it

  • Adopt it for fit. Start from a business objective and one piece of work, and leave AI out where it doesn't fit.
  • Fund its whole life. Count running, retraining, governance and your team's time next to the build price.
  • Measure what it enables. Measure the process before the build, then check where the released hours went.

Damian Kedziora, Dominik Siemon and Joanna Kedziora, Identifying and Overcoming Challenges in Intelligent Process Automation, California Management Review, volume 68, issue 4, 6 May 2026. DOI 10.1177/00081256261434509.

How we prevent each one

We help with the business case, build the system and run it after launch.

01

Adopted on hype instead of fit

We start from the objective and the process, and we'll tell you when AI is the wrong answer for it.

02

Whole-life cost is undercounted

Our estimates put the running cost and the hours your team gives the project next to the build price.

03

Savings don't arrive on their own

We agree the starting numbers before the build and report against them every month after launch.

04

The cost of doing nothing is left out

We price what the process costs today, so doing nothing has a number next to it.

05

Freed time has no plan

We agree with the business unit lead where the released hours go before the build starts.

06

No owner after launch

We stay on after launch with a monthly scorecard and tuning, and hand over to a named owner.

07

No in-house expertise, and hiring takes time

Our engineers work from your process with your people, and leave the documentation and the know-how with your owner.

See it in practice

8 builds, each measured after launch in hours, euros or days.

Read the case studies

Score your next AI project before you fund it

7 questions, about 2 minutes. You get a score and your 3 weakest areas, each with what to do next.

Ask your AI assistant about us

Open our summary in the assistant you use, or copy the prompt.

Read the prompt