Agents don't fail, empires do: why 40% of agentic AI projects get cancelled

The pilot that stalled

Walk into almost any leadership meeting about AI right now and you'll hear a version of the same conversation. The pilot showed promise, but the rollout has stalled. IT points to how complicated the integration is, and the business points back at IT. Someone mentions change management, someone else mentions data quality, and the meeting ends with a decision to run another proof of concept while the technology matures.

Gartner expects a lot of these projects to end there. In June 2025 it predicted that over 40% of agentic AI projects will be cancelled by the end of 2027, because of escalating costs, unclear business value or inadequate risk controls (Gartner, 2025). The same research found most current projects are early experiments or proofs of concept, many of them driven by hype. Many are also an exploration of 'what might be possible' rather than 'why this is being deployed'.

I'd argue that "unclear business value" is where most of the story rests, and it's rarely a technology problem. In my experience, value goes unclear when nobody has agreed what the whole workflow is for, and each team involved is measuring success in its own way, or from its own functional view.

Where organisations stood on agentic AI investment

Gartner poll of 3,412 webinar attendees, January 2025

Conservative investment

42%

Wait and see, or unsure

31%

Significant investment

19%

No investment

8%

0%25%50%75%100%

What the research says about why agents fail

A team led by researchers at UC Berkeley looked at why systems of several AI agents working together so often result in disappointment (Cemri et al., 2025). They annotated more than 1,600 runs across seven popular frameworks and grouped the failures into three families: problems with how the system was designed, misalignment between the agents, and weak or inconsistent checking of whether the task was actually completed as expected.

Consider this with your commercial team in mind. Unclear roles, people working to different goals, and nobody confirming the job is finished properly before it moves on. Those are the same problems that trip up people handing work over from marketing to sales and account management. The agents have inherited the way the 'human' organisation worked.

Why every team builds its own empire

Leaders build empires, and I say that without any judgement. It's a rational response to how organisations have rewarded people for decades. Deep expertise in your own function is how you show your value, earn authority and justify the size of your team, or protect it. So you draw a boundary around it, and over time the function develops its own goals, its own measures and sometimes its own micro-culture. It's been the accepted career path for a long time.

Multiply that across a leadership team and you get the structure most businesses have today: a set of well-run departments, each making sense on its own terms and each slightly out of step with the next. Harvard's Amy Edmondson and her co-authors found that people naturally focus on the relationships above and below them in the hierarchy rather than across it, and that redesigning the org chart to fix this tends to be costly, confusing and slow (Harvard Business Review, 2019).

In a £5m to £30m business this usually looks quite ordinary. Imagine a sales rep with an agent reading CRM activity, a marketing exec with one reading campaign data, and a BD lead with one working through their own contacts. Each agent does exactly what it was set up to do, and none of them has seen what the other two know about the same customer.

What agents expose

For years, relationships and informal conversations have covered the gaps between teams, whether intentionally or by coincidence. A good account manager knows who to call in finance to get a contract agreed, and a marketer who gets on well with the sales team hears which leads were a waste of time. Agents don't have any of that. They work to the task and the goal they've been given, and they need clarity and instructions at every handover.

So they show you, very quickly, every point where ownership is unclear, where two teams' goals pull against each other, and where one team's definition of done is another team's problem. That's when the workflow stalls. The only thing worse is when an agent is trained to follow those broken handovers, because then it repeats the problem faster than any person could, and potentially at scale.

Organising around the customer

The big consultancies are converging on a similar diagnosis. McKinsey argues that structure will move towards small, outcome-focused teams delivering end-to-end results across the whole value chain, and estimates that a human team of two to five people can already supervise 50 to 100 specialised agents running an end-to-end process (McKinsey, 2025). It also found that 89% of organisations are still structured for an industrial-age operating model, linear and functional.

What the diagnosis often leaves out is what the new structure should organise around. A flatter team only works better if everyone in it is pointing at the same thing. My answer, and a belief I've held for a long time, is the customer. That means treating the customer as the single source of truth that every team, every process and every agent workflow takes its direction from. There is usually a trade-off in making that shift, so check whether it's product knowledge, deep functional skills or something else.

Each function serves its purpose. Put together, those purposes produce an organisation that's optimised for its own internal logic and is regularly surprised when customers don't behave the way the plan expected.

What Haier did

The best-known example of rebuilding a business around the customer comes from a white goods manufacturer in Qingdao. Haier's former CEO, Zhang Ruimin, saw departments becoming self-centred and the customer signal being filtered beyond recognition by the time it reached anyone who could act on it. Haier removed more than 12,000 middle management roles and reorganised into around 4,000 microenterprises, each responsible for creating value for its own users (McKinsey Quarterly, 2021).

The model, called RenDanHeYi, is built on zero distance to the customer. Teams deal with users directly, without layers of approval in between, and resources follow the teams that create value for real customers (Gary Fox, 2025). By 2020, Haier Smart Home's revenue had passed $32 billion, with growth of more than 18% a year since 2015, and the model carried over to GE Appliances after Haier bought it in 2016 (McKinsey Quarterly, 2021).

You don't need to take out a layer of management to borrow the principle. For a business of 50 to 200 people, it might mean one person owning the customer view across sales, marketing and BD, a shared record everyone works from, and measures that don't pull the three teams in opposite directions. It's the same thinking I cover in how to lead a team that works with AI agents, applied to the whole workflow rather than one team.

Five questions to ask before your next agent deployment

Most businesses go straight to designing the workflow. These questions are worth working through before that work begins.

1. Who are your customer champions? Every business has people with a real, lived understanding of what customers need, built through years of direct contact. They aren't always the most senior people in the room. They're often the ones who get pulled into escalations, or whose opinion the sales team seeks before a big pitch. Identify them so they can be the customer voice in your business, in the same way you'd likely name your AI champions.
2. Are they in the room when you design agent workflows? Most agent projects are led by technology teams or advisers, and the people with the deepest customer knowledge are brought in late in the process. A workflow designed without them will work well by internal measures and keep missing what the customer needs at the point of contact.
3. What is each agent ultimately working towards? Every workflow has a goal built into it, and in most businesses it's an internal one, such as closing the support ticket or hitting the SLA. At each handover, ask whether the work is aligned to delivering what the customer needs, or to your own measure of completion.
4. Where are the empire boundaries in the workflow? Map the workflow against your org chart. Every time a task crosses from one team to another, ask who owns the handover, what each side measures, and where their definitions of success conflict. Answer these before the technology is built.
5. Where will the time saved go? If the capacity agents free up is used to cut customer-facing roles, you shrink the human judgement your agents rely on. If it's used to deepen customer relationships and build knowledge, you're building an advantage that's hard for a competitor to copy. This often plays out when automation is built into customer service operations to create capacity for human judgement at the all-important 'conversion' stage, where it creates real strategic advantage.

Where to start

If you only do one thing, start with question 4. Take one workflow that's already using AI, or about to, and walk it across the org chart with the people who own each step. The places where the conversation gets uncomfortable or unclear are usually the places where your agents will stall. If you'd like help surfacing these gaps, it's the kind of work we do together.

Sources

• Gartner, Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027 (June 2025)
• Cemri et al., Why Do Multi-Agent LLM Systems Fail? (UC Berkeley, 2025)
• Edmondson, Jang and Casciaro, Cross-Silo Leadership (Harvard Business Review, May 2019)
• McKinsey, The agentic organization: Contours of the next paradigm for the AI era (September 2025)
• McKinsey Quarterly, Shattering the status quo: A conversation with Haier's Zhang Ruimin (July 2021)
• Gary Fox, RenDanHeYi: How the Haier ecosystem model works (2025)

Questions people also ask

Keep reading

← All insights