Why most AI pilots never go live, and the fixes that work

How many AI pilots actually go live

Fewer than most teams expect. In Deloitte's State of AI in the Enterprise survey (opens in a new tab) of 3,235 business and IT leaders, published in January 2026, only 25% of respondents had taken 40% or more of their AI pilots live.

Other research points the same way. A 2025 survey by S&P Global Market Intelligence of more than 1,000 companies in North America and Europe found that the average organization dropped 46% of its AI proofs of concept before they went live, and 42% of companies abandoned most of their AI initiatives, up from 17% a year earlier (CIO Dive (opens in a new tab)). Looking ahead, Gartner predicts (opens in a new tab) that over 40% of AI agent projects will be canceled by the end of 2027 because of escalating costs, unclear business value or inadequate risk controls.

Notice what is missing from these reasons: the AI model itself. Cost, value and risk are planning questions, and a team can answer most of them before the pilot starts. Below are seven common causes, each with its fix.

1. No measurable goal

Many pilots start with a broad question, such as what AI could do for the sales team, instead of a target. The pilot ends with a demo that impresses some people and worries others, and nobody can make a decision. Unclear business value is one of the three reasons Gartner gives for the cancellations it expects.

The fix: write the success criteria before the build starts. Pick one or two measures you can compare with today, such as minutes per task, response time, error rate or cost per case. Record the baseline, set the target that would justify going live and ask the budget owner to sign both. Our article on measuring AI pilot ROI covers baselines and targets in more detail.

2. A pilot on a toy workflow

A pilot that answers ten hand-picked questions or reads clean sample invoices will look good. It says little about incomplete records, unusual requests and upset customers, which is what real work looks like. Some teams pick a task so small that success would not matter. Others pick the most complex process in the company. The first proves nothing useful, and the second takes too long to show any result.

The fix: choose a workflow that repeats often, has a clear cost today and follows rules you can write down. Test it on real historical cases, including the difficult ones, not a curated sample. The result should be large enough to justify the next step if the pilot succeeds.

3. No access to real systems and data

Pilots often run on exported spreadsheets because connecting to the CRM, help desk or ERP needs approvals nobody requested. When the time comes to connect for real, the project meets missing APIs, unclear permissions and data in the wrong format. In Anthropic's 2026 State of AI Agents report (opens in a new tab), a survey of more than 500 technical leaders, 46% named integration with existing systems as a primary obstacle and 42% named data access and quality.

The fix: connect the pilot to the real systems from the first weeks, with read-only access where possible and a test environment for anything the agent writes. List every system, field and permission the workflow needs during planning, and request access early. If the agent will work in your CRM, use our checklist of what to check before connecting an AI agent to your CRM.

4. Security and legal review arrive too late

In many companies, the security, privacy and legal teams first see the agent when the business wants to launch it. They ask fair questions: what data it sees, where that data is processed, what it can change and what happens if someone manipulates it. If the answers call for a redesign, the launch slips or stops. S&P Global's respondents named data privacy and security risks among the top obstacles, next to cost.

The fix: involve security and legal in the first week and give them a short design note: data used, systems touched, actions allowed and who approves what. Build the controls into the pilot itself: limited access, approval steps for high-impact actions, logging and testing for prompt injection, which means tricking an AI with hidden instructions. The review at launch then confirms decisions already made.

5. Nobody owns the agent after launch

A pilot has a project team. A live agent needs an owner who reviews quality, approves changes to instructions and permissions, watches costs and handles complaints. Without an owner, an agent can slowly get worse, because nobody reviews its answers or updates its instructions. Deloitte found that only 21% of companies planning to deploy AI agents have a mature governance model for them.

The fix: name the business owner before the pilot starts. Agree on what they review each month, such as quality samples, error reports and AI usage costs, and who is allowed to change the agent. Put the operating guide and the hand-off rules in writing.

6. Costs that nobody modeled

Pilot costs are small and easy to predict. Live costs depend on volume, the length of each conversation, the number of AI calls per task and the staff time spent on reviews. Gartner research reported by Campus Technology (opens in a new tab) predicts that through 2028, at least 50% of generative AI projects will overrun their budgets because of poor architectural choices and a lack of operational know-how. In the Anthropic survey, 43% named implementation costs as an obstacle.

The fix: measure AI usage costs per task during the pilot, then model them at full volume, including monitoring, maintenance and review time. Set budget alerts. Use the smallest model that meets the quality target, and limit how many steps the agent can take on one task.

7. Low trust from staff

If the people who work next to the agent don't trust it, they check every output by hand or route work around it, and the savings disappear. Trust is not a given: in a global study by the University of Melbourne and KPMG (opens in a new tab) of more than 48,000 people, only 46% were willing to trust AI systems, and only 47% of employees said they had received AI training. In the Anthropic survey, small and mid-sized businesses were more likely than large ones to struggle with employee resistance and training needs (51%).

The fix: involve the staff who do the work today when you define the workflow and the test cases. Show them what the agent does and how a person can take over. Start with the agent preparing work and staff approving it, then reduce reviews where the numbers show it is safe. Make it easy to flag a wrong answer, and act on every flag.

From AI pilot to production: what changes

A pilot proves the agent can do the task. Going live proves it can do the task every day, at full volume, without someone watching each step. Four things change:

• Variety. Cases that were rare in testing appear every day.
• Exposure. Customers, and possibly attackers, write the input instead of your team.
• Cost. AI usage costs grow with volume and conversation length.
• Accountability. Mistakes now affect customers, revenue and compliance, so someone must own them.

A pilot designed with these four changes in mind needs hardening before launch, not a rebuild.

How we structure a pilot so it can go live

Each pilot we run is stage 2 of our four-stage process, after a 1–2 week audit that ranks workflows by payback and risk. It follows the same rules:

1. 1One workflow. We pilot the highest-ranked workflow from the audit, not five at once.
2. 2Success criteria agreed upfront. The baseline, the target and the way we measure them are written down before we build.
3. 3Real systems. The agent works in your CRM, help desk or ERP, on your own accounts, with the smallest access that works.
4. 4Security built in. Limited permissions, approval steps for high-impact actions, logging and attack testing before the agent talks to a customer.
5. 5A go/no-go decision. After 4–6 weeks you receive a results report against the agreed criteria and decide whether to continue, change the scope or stop.

Stopping is a valid result. A pilot that shows a workflow is not worth automating saves the cost of a full build. If the answer is go, our AI automation services cover the secure launch and the monitoring that follows.