Skip to content
SolvraONE

Knowledge

Why AI projects fail, and what makes them last.

AI illustration: a row of graphite blocks with a line of amber light running through without a break and ending in a checkmark
AI illustration
Updated (first published )

In short · 6 min read

  • Gartner names rising costs, unclear business value and inadequate risk controls as reasons for cancelled agent projects.
  • Add too large a scope, technical limits, no rules for errors and no care after launch.
  • If you treat the agent as an ongoing service – measure, review, improve – you spot these risks earlier.

The short answer

An AI agent is not a program you set up once and then leave running. It works with real customers, real data and real systems, and all of that keeps changing. Our experience from operation: the hard part is often not the language model, but two gaps:

  • Before: nobody defined what the agent should do.
  • After: nobody checks whether it does it.

What the market says

The research firm Gartner predicts that over 40% of agentic AI projects will be cancelled by the end of 2027. That is a forecast, not a measurement. Gartner names three reasons:

  • rising costs,
  • unclear business value,
  • inadequate risk controls.

Many projects are early experiments driven by hype, it says, and underestimate the cost and difficulty of deploying agents at scale.

Our experience from operation fits: what works in a demo still has to prove itself day to day with real cases.

Why agents fail differently from other software

A wrong answer can also cause harm, for example when a customer relies on a wrong deadline. An agent can also act itself: it rebooks, writes an email, closes a ticket. So the difficulties lie not only in the model but also in how the parts work together:

  • the interfaces,
  • the rules,
  • the error handling,
  • the data.

Such errors often only show up when real cases arrive that no test covered. The page AI agent vs chatbot vs copilot explains the difference between these forms.

Seven common reasons

Reason How you notice it What helps
Technology first You look for a tool before the problem is described. Start with one process that costs time today and can be measured.
Scope too large The agent is supposed to take over “customer service”. One task, one channel, clear limits; expand only once there is evidence.
Nobody is responsible IT built it, the business team is meant to use it, nobody decides. One person owns the rules, the knowledge and the results.
No plan for errors It is open what happens when the agent is unsure. Define the handover to people before the first customer writes.
Data not ready Knowledge is outdated, scattered or contradictory. Clean up and maintain the knowledge base before the agent uses it.
Law at the end Data protection and AI labelling come just before launch. Plan them in from the start, see AI agent and GDPR.
No operation after launch The project ends with go-live; nobody looks after that. Measure, review and improve continuously – as a service, not a project.

All seven reasons can surface during the build or only in operation. The last one can go unnoticed for a long time: answers slowly get worse when prices, processes or products change and the agent’s knowledge does not. That is why ongoing care matters so much.

One-off project or operation

This is how we tell the two paths apart here: a one-off project is planned up to launch, operation carries on after it.

One-off project Operation
Goal the agent is built the agent does work you can prove
End at launch none – it runs as long as it is useful
Knowledge as of launch kept up to date continuously
Errors customers notice them are analysed before they pile up
Review once before launch regularly, by a second party
Numbers estimate in the offer measured values every week
Responsibility spreads out after launch clearly with one owner

What operation looks like

Operation does not mean someone sits next to the agent all the time. It means there is a fixed routine:

  1. The AI does the work: the agent handles the cases it has rules for and hands over the rest.
  2. A second party checks: samples and unusual cases are checked against knowledge and rules – by a second AI, a person or both.
  3. A person approves: new rules, principles and bigger changes only go live after approval; the AI may read in knowledge from help pages and terms automatically once the source is fixed.
  4. Numbers every week: how many cases were handed over, how many resolved, where were there complaints? That needs a fixed definition of “resolved”. The numbers point to the next improvement.

With FRED, AI agent in customer service for a mobility provider in Germany we currently analyse conversations and the tickets created. These numbers do not measure whether a case was actually resolved. What is checked there and where the limits are is shown in the customer service agent proof.

Before you start: a short checklist

  1. Which single process costs the most time today and follows clear rules?
  2. How will you measure in a month whether the agent helps?
  3. Who decides what the agent may do, and who maintains its knowledge?
  4. What happens when the agent is stuck – and who takes over then?
  5. Who looks at the numbers every week after launch?

If you have no answer to one of these questions, that is no reason to give up. It is the point where you should start.

Your next step

Planning an AI agent, or have one that does not work as expected? We talk it through in a first call of 30 minutes, based on your process. The website check, which is free, looks at your website instead, not your agent project. Get in touch

What could AI take off your plate?

In a first call (30 minutes) we look at your work and tell you openly what is worth it and what is not.

Request the website check

Frequently asked questions.

Why do so many AI agent projects fail?

The research firm Gartner names rising costs, unclear business value and inadequate risk controls. From our experience, too large a scope, no rules for errors and no care after launch add to that.

Is it the technology?

Partly, but not only. According to Gartner, many projects underestimate the cost and difficulty of deploying agents at scale. In our experience the hurdles often lie in interfaces, data quality, rules and ongoing care.

What is the difference between a one-off project and operation?

A one-off project is planned up to the agent's launch. Launch alone does not secure lasting value: care, review and responsibility have to be settled after it too. In operation the agent is measured, reviewed and improved continuously, with a clear owner and weekly numbers.

How small should a first AI agent be?

Small enough that after a short time you can say for sure whether it helps: one process, one channel, clear limits. Expand only when the numbers support it.

Who in the company should be responsible for an AI agent?

A person from the business team who knows the process and may set rules, supported by the technical side. Without that ownership the agent's knowledge goes out of date quickly.

How can I tell early that a project is going wrong?

When nobody can say how success is measured, when there is no plan for uncertain cases, or when no numbers come in after launch. These are warning signs you should check, ideally before the first customer complaints.

How is your website doing?

The check is free and without obligation. You get the report by email.

Request the website checkor ask for a first call

Sources.

  1. Gartner, press release: forecast on cancelled agentic AI projectsThird-party report · As of
  2. European Commission AI Act Service Desk, Article 50 (transparency obligations)Vendor / primary · As of
All 3 sources
  1. European Commission AI Act Service Desk, Article 14 (human oversight for high-risk AI)Vendor / primary · As of