Gartner: Over 40% of Agentic AI Projects Will Be Canceled by 2027

Gartner expects more than 40% of agentic AI projects to be canceled by the end of 2027, and the analyst firm points at rising costs, unclear business value and weak risk controls. The deeper problem, practitioners argue, is that these projects get approved on the strength of a demo, with little view of what the work costs today or who owns the outcome.

The risk sits in the business the agent is dropped into, not the model. A customer asks for a refund; the agent finds the order in the CRM, checks the returns policy and drafts a reply in under a minute. Then it stalls, because refund approval lives in a finance system it was never connected to and the customer's address doesn't match between two databases. Gaps like that never surface in a demo, yet they decide what the agent costs to run and whether anyone trusts it.

The readiness numbers back this up. Deloitte's 2026 research found only 5% of organizations rate their business processes as highly prepared for AI agents — a figure more telling than any model benchmark, because it suggests the AI is ready well before most processes are. Routine cases are seldom the problem. Exceptions are, and demos are built to avoid exactly that part.

What does the work after automation actually cost?

The easiest way to overstate an agent's value is to time how fast it produces an output. Consider a request that takes an employee 20 minutes. The agent gets the first pass down to two — an 18-minute saving on paper, right up until someone spends 15 minutes checking and fixing errors. Three minutes saved is a thin return once the agent's running cost goes on the other side of the ledger.

McKinsey's analysis of a banking customer-service agent shows the same pattern at scale: tokens made up 20–25% of variable operating costs, while human oversight ran 70–75%. Most of the money sat outside the model. Before anyone writes code, the business owner needs a baseline on five points:

• Volume: how often the process runs and how many people touch it.
• Completion time: the whole task, checking and corrections included.
• Accuracy: how often it goes wrong or needs redoing.
• Cost per task: what a successfully completed task costs today.
• Business impact: what a real improvement would change for customers, staff or the business.

Where the agent sits decides whether anyone uses it

Adoption usually gets filed under change management. In practice, it's settled earlier — by placement. A global professional services company built agents for support and compliance, but its support agent lived in a separate application people had to open, and they mostly didn't bother. Once the same workflow moved into the existing support platform, it appeared inside the ticket associates were already working on. A compliance screening agent went the same way, built into Salesforce so it ran whenever a case was created.

Nothing about the agents changed; their position in the workflow did. Placement is an architecture decision as much as an adoption one. The output has to feed the next step without someone copying it across by hand.

Autonomy changes the security equation

An agent that can only read information carries one kind of risk; one that can change records or email customers carries another entirely. Alongside excessive permissions and data leakage, prompt injection deserves the closest attention — instructions hidden in a document or email that the agent picks up mid-task.

Gartner's May 2026 guidance on agent governance breaks autonomy into four levels:

• Observe: read-only retrieval and pattern-spotting.
• Advise: it recommends an action; an employee decides.
• Act with approval: it prepares an action a person must authorize.
• Act autonomously: it carries out approved actions within set limits, with monitoring and escalation for exceptions.

At every level, the agent needs its own identity, task-limited permissions and a record of each action, plus someone with the authority to pause it and pull access the moment something looks wrong. Each control carries a running cost, so the right autonomy level is an economic decision as much as a security one.

Pick a process that can justify the agent

Start with a process whose cost or delays are already obvious, not with the technology. A global expert-network company was matching survey requests to experts across three systems that didn't talk to each other, checking profiles, past participation and compliance by hand. An agent combining semantic search with participation history and compliance rules cut mismatch tickets by 80%, raised mapping accuracy 40% and gave the team back 22% of its bandwidth.

Specificity also answers whether you need an agent at all. Fixed-rule work is usually cheaper with conventional automation; finding-and-summarizing work suits an AI assistant. Agents earn their extra cost and risk only when the job means reading context, choosing between permitted actions and chaining steps toward an outcome.

How the eight-week pilot runs

The whole pilot hangs on one commercial question: can this agent improve the process enough to justify running it at scale? The structure:

• Weeks 1–2. Document the process as it runs today — volume, completion time, error rates, rework, cost — and get the business owner and technical team to agree on the one outcome to improve. Decision point: keep going once the baseline, owner and success measures are clear.
• Week 3. Set the agent's boundaries: identity, permissions, data access, tools and degree of autonomy. Test prompt injection, unauthorized access, sensitive-data exposure and tool misuse; put logging, incident response and a kill switch in place. Decision point: keep going once controls are actually tested.
• Weeks 4–6. Give it realistic work — incomplete records, conflicts, odd requests, system failures. Work out what a successfully completed task really costs once model usage, integrations, monitoring and review are counted. Decision point: move to live testing when thresholds are hit and failure modes are understood.
• Weeks 7–8. Put it in front of a small group of users on a controlled slice of real work, with the old process still available. Compare against the week-one baseline on productivity, accuracy, adoption, human interventions, total operating cost and open security risks.

The final call is to scale, fix specific gaps and retest, or stop. A pilot that shows the agent is too expensive or unreliable has still done its work — it has prevented a much bigger investment built on guesswork.

Task throughput alone says little; a system can churn through thousands of requests while employees verify most of them. Cost per successfully completed task, read alongside accuracy and exception rates, is the measure that matters — and the same evidence becomes the case for granting more autonomy. Enterprises that pull ahead will be the ones designing for the work after automation from week one and funding agents on the tasks they actually complete.