More Than 40% of Agentic AI Projects Face Cancellation by 2027

By Daniel Okafor

Gartner expects more than 40% of agentic AI projects to be canceled by the end of 2027, citing rising costs, unclear business value and weak risk controls. Okafor links the forecast to demo-led approvals that omit current costs and ownership.

Where do agentic AI projects fail?

Most risk sits in inherited business processes, not the model, Okafor argues. He calls the remaining human effort “the work after automation.”

In Okafor's refund example, a customer asks for a refund; the agent retrieves a CRM order, checks the returns policy and drafts a reply in under a minute. It then stalls because the refund approval sits in a disconnected finance system and customer addresses conflict across two databases.

Exceptions, not routine cases, drive operating costs and determine trust. Deloitte's 2026 research found only 5% of organizations rate their processes as highly prepared for AI agents. Okafor calls that gap more telling than a model benchmark.

What does a completed task cost?

A task that takes an employee 20 minutes can take two for the agent's first pass, apparently saving 18 minutes. But a person then spends 15 minutes checking and fixing errors, leaving three minutes of savings before model expense.

McKinsey's analysis of a banking customer-service agent put tokens at 20–25% of variable operating costs and human oversight at 70–75%. Most spending sat outside the model.

Before code, Okafor requires a baseline:

• Volume: Process frequency and people involved.
• Completion time: Total task time, including checks and corrections.
• Accuracy: Failure and rework rates.
• Cost per task: Current cost of a successful completion.
• Business impact: What better performance changes for customers, staff or the business.

Where should the agent sit?

Adoption is usually filed under change management, but Okafor places it earlier, in workflow design. A global professional services company built support and compliance agents. Its support agent pulled from several internal systems but sat in a separate app associates mostly avoided.

Moving support into the existing platform put the agent inside the ticket associates already used. A compliance screening agent was built into Salesforce to run whenever a case opened. The agents did not change; their position did.

Placement is an architecture decision as much as an adoption decision. It supplies context and feeds the next step without manual copying.

How much autonomy can it have?\n

Permissions set the control burden. Read-only agents differ from agents that change records or email customers. Okafor flags excessive permissions and sensitive-data leakage, but gives prompt injection the closest attention: hidden document or email instructions can redirect legitimate work.

Gartner's May 2026 governance guidance divides autonomy into four levels:

• Observe: Retrieve information and spot patterns with read-only access.
• Advise: Recommend an action; an employee decides.
• Act with approval: Prepare or start an action that a person authorizes.
• Act autonomously: Execute approved actions within set limits, with monitoring and escalation for exceptions.

Every level requires an identity, task-limited permissions and action records. Organizations need data handling and retention rules, clarity about what reaches outside model providers, and authority to pause the agent or revoke access.

Wider permissions, monitoring and approvals add cost. Okafor treats autonomy as an economic and security choice, tied to worthwhile outcomes at acceptable risk.

Which work justifies an agent?

Start with a costly or delayed process, not technology hunting for a problem, Okafor advises.

Okafor reports a global expert-network company matching survey requests to experts across three disconnected systems, with manual checks of profiles, participation history and compliance. Its agent combined semantic search, participation history and compliance rules.

In production, mismatch tickets fell 80%, mapping accuracy rose 40%, and the team recovered 22% of its bandwidth.

Fixed-rule work is usually cheaper with conventional automation. Finding and summarizing information often suits an AI assistant. Agents justify added cost and risk when they read context, choose permitted actions and chain several steps.

What must an eight-week pilot prove?\n

Okafor's commercial test is: “can this agent improve the process enough to justify running it at scale?” Eight weeks tests the business case, technology and the organization's management ability.

• Weeks 1–2: Document volume, completion time, errors, rework and cost. Map systems, stage owners and exceptions. The business owner and technical team must agree one outcome. Continue only when the problem, baseline, owner and success measures are clear.
• Week 3: Set identity, permissions, data access, tools and autonomy. Test prompt injection, unauthorized access, sensitive-data exposure and tool misuse. Add logging, incident response, suspension and access revocation. Continue only after controls are tested and everyone understands the agent's limits.
• Weeks 4–6: Use incomplete records, conflicting information, odd requests and system failures. Track completions, accuracy and human corrections. Calculate the cost of a successfully completed task, including model usage, integrations, monitoring and review. Move to live testing only after agreed thresholds hold and failures are understood.
• Weeks 7–8: Give the agent to a small employee group and a controlled slice of live work, with the old process as fallback. Observe usage, effort saved and trust in the output.

Compare results with week one across productivity, accuracy, adoption, interventions, total operating cost and unresolved security risks. The decision is to scale, fix specific gaps and retest, or stop. Okafor says a pilot that proves the agent too expensive or unreliable still prevents a larger guesswork-based investment.

What should leadership decide?\n

Task volume says little if employees still verify, correct or finish most requests. Cost per successful task, read with accuracy and exception rates, better measures end-to-end progress. Routine work that proves reliable can justify more autonomy; ambiguous cases should keep human approval, with permissions reviewed as evidence grows.

Scaling requires ownership, access controls, exception handling and budget for keeping the agent current. Leaders should leave week eight knowing what it can handle, where people must step in and what it costs to run. Okafor's forward test is simple: fund agents on completed tasks, not raw speed.