40% of agentic AI projects will be cancelled by 2027: the cost and governance controls that keep you in the 60%
Summary. Over 40% of agentic AI projects will be cancelled by the end of 2027 on escalating cost, unclear business value and inadequate risk controls, according to a Gartner prediction published on 25 June 2025. A separate Gartner forecast on 26 May 2026 expects 40% of enterprises to demote or decommission autonomous agents by 2027 after governance failures surface in production. The money side is just as blunt: EY puts the jump from a simple 2023 chatbot call at about $0.04 to a 2026 orchestrated agent interaction at roughly $1.20, near 30 times higher. An MIT study of enterprise deployments found 95% of generative AI pilots returned nothing measurable despite $30 to $40 billion in spending. The projects that survive are not the ones with the best demo. They are the ones that pick a use case with real return, cap the cost of every agent run, and govern each agent by how much it is allowed to do. This guide sets out those controls for CTOs and heads of AI, with the numbers behind each one.
Agentic AI is still growing fast. Gartner forecast on 20 July 2026 that the worldwide AI platforms and models market will grow 63% in 2026. The cancellations and the growth are the same story told twice: demand is real, but most teams are buying the hype before they build the controls. The gap between those two facts is where budgets die.
Why agentic AI projects get cancelled
Gartner names three causes in its 25 June 2025 prediction: escalating costs, unclear business value, and inadequate risk controls. Each one is avoidable, and each one is common.
The first cause is overreach dressed as ambition. "Most agentic AI projects right now are early stage experiments or proof of concepts that are mostly driven by hype and are often misapplied," said Anushree Verma, Senior Director Analyst at Gartner. In a January 2025 Gartner poll of 3,412 webinar attendees, 19% had made significant investments in agentic AI, 42% had invested conservatively, 8% had invested nothing, and the remaining 31% were waiting or unsure. Most spend was going into experiments, not production systems with owners and budgets.
The second cause is a market full of relabelled software. Gartner calls it "agent washing", the rebranding of AI assistants, robotic process automation and chatbots as agents without real agentic capability. Gartner estimates only about 130 of the thousands of self-described agentic AI vendors are building something real. Buy one of the other several thousand and the project fails on procurement grounds before an engineer writes a line of orchestration.
The third cause shows up after launch. Gartner's 26 May 2026 forecast is that 40% of enterprises will demote or decommission autonomous agents by 2027 "due to governance gaps identified only after production incidents occur". The demo worked. The pilot worked. Then an agent took an action nobody had scoped, and the safest available response was to switch it off.
The independent evidence is harsher still. The MIT report "The GenAI Divide: State of AI in Business 2025", based on 52 executive interviews, surveys of 153 leaders and analysis of 300 public deployments, found that 95% of generative AI pilots delivered no measurable profit-and-loss impact. Its authors traced the failures to brittle workflows and tools that could not retain feedback or adapt to context, and found that deployments built with external vendors succeeded about twice as often as internal builds.
The cost problem: why agent bills explode
A single chatbot reply is cheap. An agent that plans, calls tools, reads the results, and tries again is not. EY's analysis of enterprise token cost puts a simple 2023 linear workflow at about $0.04 per interaction and a 2026 orchestrated system, with tools, reasoning and iterative loops, at roughly $1.20 per interaction, close to 30 times higher. The per-token price kept falling over that period. The bill still rose, because an agent spends far more tokens per task.
The mechanism is the loop. A reasoning agent can consume 5 to 30 times the tokens of a single model call on the same task, as retries, tool calls and context re-reads stack up, an inference-cost pattern documented across 2026 agent deployments. Multiply that by every user, every day, and a workflow that looked cheap in a demo becomes the line item finance asks about.
| Workflow type | Approx cost per interaction | Why it costs that |
|---|---|---|
| Simple chatbot reply, 2023 | $0.04 | One model call, fixed input and output |
| Orchestrated agent, 2026 | $1.20 (about 30x) | Planning, tool calls, reasoning and retry loops |
| Autonomous multi-step agent at scale | Highest, 5 to 30x token use per task | Loops and context re-reads repeat on every step |
The fix is not a cheaper model. It is treating tokens as a metered utility and engineering the agent to spend fewer of them: route simple steps to a small model and reserve the frontier model for the hard step, cache repeated context and tool results, cap the number of reasoning loops and tool calls per task, and set a hard token budget per session with a circuit breaker that stops the run when it is breached. Cost control is a design decision made before launch, not a dashboard bolted on after the invoice.
The governance problem: binary governance is the root cause
The reason so many agents get switched off is that most teams govern them with one setting for everything. "Enterprises are treating AI agent governance as binary, either locked down or fully trusted, and that is the root cause of failure," said Shiva Varma, Senior Director Analyst at Gartner, in the 26 May 2026 release. Lock every agent down and you slow delivery and push teams into shadow builds. Trust every agent and you hand write access to systems that were never tested for it.
Gartner's answer is proportional governance: classify each agent by how much it can do, and match the controls to that level. The four levels below are drawn from Gartner's autonomy model, with the governance weight rising as the agent's reach grows.
| Autonomy level | What the agent does | Governance the level needs |
|---|---|---|
| Level 1, Observe | Read-only retrieval and summaries, output to the requester only | Scoped data access, authentication, usage logging, basic testing |
| Level 2, Advise | Drafts and recommendations a human executes manually | Level 1 controls plus accuracy and hallucination testing, user training on over-reliance |
| Level 3, Act with approval | Writes data or sends messages, but only after explicit sign-off | Approval workflows with audit trails, security testing, agent-specific incident response |
| Level 4, Act autonomously | Executes within guardrails, humans review exceptions and logs | Continuous monitoring, enforced guardrails, rapid rollback, circuit breakers, named owner |
The classification is the point. An Observe agent that summarises documents does not need the same paperwork as an agent that can move money, and forcing it to carry that weight is how you get the shadow project that nobody governs at all. An autonomous agent that can act on production systems needs far more than a demo and a hopeful launch. Get the mapping wrong in either direction and you land in one of Gartner's two failure modes: over-restriction that drives delivery underground, or under-restriction that raises operational, security and compliance risk until an incident forces a shutdown.
The value problem: pick the use case that pays
Cost control and governance keep an agent alive. They do not make it worth running. Gartner's guidance is to pursue agentic AI only where it delivers clear value or return, and to be honest that many use cases sold as agentic today do not need an agent at all. A routine, deterministic workflow is cheaper and more reliable as automation. A simple lookup is a retrieval assistant. The agent earns its cost only when a task genuinely needs planning and judgement across steps.
The other half of the value problem is where the agent runs. Bolting an agent onto a legacy system is often the expensive path, because it disrupts existing workflows and forces costly changes around code that was never designed for an autonomous caller. Gartner's own advice is that rethinking the workflow around the agent from the start is frequently the better route. The MIT finding that external-vendor builds succeeded about twice as often as internal ones points the same way: the teams that win treat this as a product to be designed, not a feature to be sprinkled on.
| Failure driver (Gartner, EY, MIT) | What it looks like in production | The control that keeps you in the 60% |
|---|---|---|
| Escalating token cost | Agent bill rises 30x versus a chatbot as loops stack up | Model routing, caching, loop and tool caps, per-session token budget |
| Unclear business value | Pilot with no owner, no P&L target, no baseline | Pick use cases with measured return; kill the ones without it |
| Inadequate risk controls | Autonomous agent acts outside its intended scope | Proportional governance by autonomy level, guardrails, rollback |
| Agent washing | Relabelled RPA or chatbot sold as an agent | Vendor due diligence; only about 130 vendors are the real thing |
| Binary governance | One control setting for every agent | Classify each agent, match controls to its reach |
| No evaluation discipline | Quality drifts, nobody notices until an incident | Continuous eval, hallucination testing, monitoring, circuit breakers |
A control plan to stay in the 60%
The surviving projects share a short list of habits. None of them is exotic. All of them are decided before the first agent reaches a user.
Start with one use case that has a measured baseline and a clear return, and refuse to scale until it clears that bar. Classify every agent by autonomy level on day one, and write the governance for that level, not for the whole programme. Put a token budget and a circuit breaker on every agent so a runaway loop stops itself. Route cheap steps to cheap models. Build evaluation and monitoring before launch, not after the first incident, and give every autonomous agent a named human owner with the authority to roll it back. Keep a human approval step wherever an agent can write to a system that matters. Treat vendor claims as claims: ask what the agent actually does, not what the label says.
Done together, these controls convert Gartner's three cancellation causes into three managed line items. Cost becomes a budget. Value becomes a target. Risk becomes a level with matching controls. That is the difference between the 40% that get switched off and the 60% that keep running.
India-specific considerations
Indian enterprises face the same cost and governance maths, plus a compliance layer that is tightening. The Digital Personal Data Protection Act, 2023 (DPDP) governs how agents may access and process personal data, which matters the moment a Level 3 or Level 4 agent can read customer records or act on them. Design for data minimisation and scoped access from the start rather than retrofitting it after an audit.
India's approach to AI itself is principle-based rather than a single hard statute. The India AI Governance Guidelines, published in 2026, set out a techno-legal framework built on guiding principles and new institutions, including an AI governance group and an expert committee, with a stated preference for voluntary compliance and self-certification over prescriptive mandates. Financial services carry more weight: in June 2026 the Reserve Bank of India moved toward requiring banks and regulated entities to establish governance frameworks for AI and machine-learning models, which raises the Level 3 and Level 4 bar for any fintech deploying agents against core systems. On the cost side, Indian teams should price agent runs in rupees against the same 30x escalation EY documents, and set token budgets in the same design pass, because a workflow that costs a few paise per call at demo scale can still run to lakhs per month once every user is on it.
How eCorpIT can help
eCorpIT (eCorp Information Technologies Private Limited) is a Gurugram technology consultancy, founded in 2021, that designs and ships enterprise AI systems for teams that need them to survive contact with production. We are CMMI Level 5 assessed and MSME registered, with senior-led, multi-disciplinary teams and partnerships with AWS, Microsoft and Google. For agentic AI, that means scoping the use case against a real return, classifying each agent by autonomy level and building the matching governance, and engineering the cost controls, token budgets, circuit breakers and evaluation harnesses that keep an agent in the surviving 60%. We design systems aligned with DPDP and RBI model-governance requirements rather than claiming blanket certification. If you are planning or rescuing an agentic AI programme, talk to our team about a control plan before the budget review, not after it. You can also read our approach to an enterprise GTM and RevOps AI agent service and to buying versus building enterprise AI services.
FAQ
Why does Gartner expect 40% of agentic AI projects to be cancelled by 2027?
Gartner's 25 June 2025 prediction names three causes: escalating costs, unclear business value, and inadequate risk controls. Most current projects are early experiments driven by hype rather than production systems with owners, budgets and measured return, which stalls them before they reach real deployment or justify their spend.
How much more does an agentic AI interaction cost than a chatbot reply?
EY's enterprise token-cost analysis puts a simple 2023 chatbot interaction at about $0.04 and a 2026 orchestrated agent interaction, with tools, reasoning and retry loops, at roughly $1.20, close to 30 times higher. Token prices fell over that period, but agents spend far more tokens per task, so the bill rose.
What is agent washing?
Agent washing is Gartner's term for rebranding existing software, such as AI assistants, robotic process automation and chatbots, as agents without genuine agentic capability. Gartner estimates only about 130 of the thousands of self-described agentic AI vendors are building real agents, so vendor due diligence is a core control against a failed procurement.
What is proportional agent governance?
Proportional governance classifies each agent by its autonomy level, from read-only Observe agents to fully autonomous ones, and matches the controls to that level. Gartner's Shiva Varma calls binary governance, treating every agent as either locked down or fully trusted, the root cause of failure, because it either slows delivery or under-controls risk.
Which cost controls keep agent bills under control?
Route simple steps to small models and reserve frontier models for hard steps, cache repeated context and tool results, cap reasoning loops and tool calls per task, and set a hard token budget per session with a circuit breaker that halts a run when it is breached. These are design decisions made before launch, not dashboards added later.
Why do so many enterprise AI pilots fail to show a return?
An MIT study, "The GenAI Divide: State of AI in Business 2025", found 95% of generative AI pilots delivered no measurable profit-and-loss impact despite $30 to $40 billion in spending. It traced failures to brittle workflows and tools that could not retain feedback or adapt, and found external-vendor builds succeeded about twice as often as internal ones.
Does agentic AI compliance work differently in India?
The same cost and governance controls apply, plus DPDP data-protection duties whenever an agent handles personal data. India's 2026 AI Governance Guidelines favour principle-based, voluntary compliance, while the Reserve Bank of India moved in June 2026 toward mandatory AI and machine-learning model governance for banks and regulated financial entities, raising the bar for fintech agents on core systems.
When should a workflow not use an agent at all?
When the task is routine and deterministic, automation is cheaper and more reliable, and a simple lookup only needs a retrieval assistant. Gartner notes many use cases sold as agentic do not need an agent. An agent earns its higher cost only when a task genuinely requires planning and judgement across multiple steps.
References
- Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027 (25 June 2025)
- Gartner Says Applying Uniform Governance Across AI Agents Will Lead to Enterprise AI Agent Failure (26 May 2026)
- EY: Agentic AI enterprise token cost
- The Hill: Most firms see no profit boost from generative AI, MIT report
- Gartner Forecasts Worldwide AI Platforms and Models Market to Grow 63% in 2026 (20 July 2026)
- Forbes: Why 40% Of Agentic AI Projects May Be Canceled By 2027 (7 July 2026)
- Spheron: Agentic AI inference cost, why agents burn 5 to 30x tokens (2026)
- Lexology: India Publishes AI Governance Guidelines
- AIRiskAware: India AI Policy, DPDP, SEBI and RBI (2026)
- Splunk: The hidden cost of agentic AI, why most projects die before production
Last updated: 26 July 2026.
Top comments (0)