Track · AI in Product4 min read
Gartner Expects 40% of Agentic AI Projects to Be Cancelled by 2027
Gartner expects over 40% of agentic AI projects cancelled by 2027. A practitioner framework separates assistants, copilots and agents by behaviour — and maps each failure mode.

Context
Gartner predicts over 40% of agentic AI projects will be cancelled by end of 2027
MIT Project NANDA (2025) found 95% of studied organisations saw no measurable generative AI return
UC Berkeley catalogued 14 failure modes in multi-agent AI systems, mostly from design and inter-agent misalignment
Gartner estimates only ~130 of thousands of vendors claiming agentic AI offer the real thing
Gartner forecasts up to 40% of enterprise apps with task-specific agents by end of 2026, up from under 5% in 2025
Gartner predicts more than 40% of agentic AI projects will be cancelled by the end of 2027, citing rising costs and unclear business value. MIT's Project NANDA reported in 2025 that 95% of the organisations it studied saw no measurable return on generative AI spend — a figure with critics, but a signal product leaders cannot ignore.
A practitioner essay circulating among product teams argues the root problem is vocabulary. Vendors use "assistant", "copilot" and "agent" interchangeably. Buyers who cannot tell them apart purchase the wrong tool for the job and then wonder why ROI never appears.
What separates the three tool types?
The author's mental model — explicitly not an industry standard, because none exists — sorts tools by behaviour, not brand name.
- Assistant: accesses live company data, runs unprompted, explains its recommendations, works on any LLM. It advises; it never acts. Its failure mode is confident advice built on bad data, plus recommendations with no named owner.
- Copilot: answers questions from company data and executes simple actions, but needs a prompt every time. It fails when it invents answers with no source record to cite, when role-based permissions are missing, and when per-query costs scale without a learning curve.
- Agent: reasons step by step, executes complex action sequences unprompted, and — if advanced — improves itself over time. It cannot reliably pursue two outcomes at once. Autonomy means blast radius: an agent that is wrong books forty deliveries. Irreversible actions need approval gates, and constraints that live only in people's heads get optimised straight through.
Above all three sits orchestration: software that triggers tasks in order, coordinates AI tools across an end-to-end process, and enforces predictable behaviour through a rules engine. The author's caution: once you orchestrate, you own a platform with monitoring and on-call, not a feature.
Why did the agent fail on the wedding day?
The essay carries its argument through a scenario: a turnaround at a heritage palace hotel in Udaipur before a film star's three-day wedding, with hundreds of VIP guests and the CEO attending. An assistant pinpoints front-desk wait times and kitchen delivery as the top satisfaction levers in under a minute. A copilot clears a reception queue by answering staff questions instantly and executing small follow-ups, but refuses multi-step supplier negotiation. An agent rebuilds kitchen procurement, confirming a full delivery schedule in minutes and replanning when a constraint appears.
Then it breaks. On the wedding morning the agent sends trucks to wrong storage areas, double-books supplier slots and spoils a batch of food. The diagnosis: it was given two outcomes — planning deliveries and receiving them. Overloaded, it loops, then hallucinates. The fix is to split the work across three single-outcome agents wired together through orchestration: one plans, one receives, one updates inventory and alerts the kitchen. Prep finishes hours early.
What does the research say about failure?
The scenario's failure modes map to published findings. UC Berkeley researchers studying multi-agent systems catalogued 14 distinct failure modes across three categories: specification and design problems, inter-agent misalignment, and weak task verification. Their conclusion: most failures stem from system design and agents misaligning with each other, not from the underlying model. The taxonomy and data are open source and usable as a design checklist.
Gartner calls vendor relabelling "agent washing": rebranding chatbots, RPA and assistants as agentic AI. Its estimate is that only about 130 of the thousands of vendors claiming agentic capabilities offer the real thing. Its August 2025 forecast describes a staged market: assistants embedded in most enterprise apps first, then up to 40% of enterprise apps featuring task-specific agents by the end of 2026, up from under 5% in 2025. "Task-specific" is the one-outcome rule showing up in the market.
Is the 95% failure figure solid?
Not entirely. Wharton's Kevin Werbach questioned the MIT report's transparency: its own explicit finding was that only 5% of custom enterprise tools reached production — a narrower claim than "95% of all AI pilots fail" — and the underlying data was never released. Critics also noted the report's diagnosis fits the kind of system NANDA itself builds. Even sceptics agree most pilots stall before production.
How should teams choose?
The author's decision rules:
- Do not know what to fix yet: start with an assistant.
- People drowning in questions and small admin: use a copilot.
- Multi-step process with a clear goal and clear rules: deploy an agent, with real permissions and real constraints.
- Several tools in one end-to-end flow: design the orchestration first.
Practical governance follows: define tools by behaviour, record vendor labels in brackets, pick one reference vocabulary organisation-wide, and force business cases to state what a proposed "agent" does when nobody is watching and what it may touch.
Gartner's advice is blunt — pursue agentic AI only where it delivers clear value. As enterprises move from embedded assistants to narrow task-specific agents through 2026, the teams that write down their definitions and gate irreversible actions will be the ones whose pilots reach production.
via towardsdatascience.com (Source)
More from Rebecca Stone
Show full bio
Correspondent covering regulation, policy and their impact on product management at Roadmap File.
72 articles