AI Agents: Why Most Pilots Never Reach Production

Only 25% of leaders surveyed by Deloitte moved 40% or more of their AI pilots into production, and Gartner expects over 40% of agentic AI projects to be cancelled. Here is why, and how to avoid it.

Many companies have already tested AI agents. Few manage to take them to production. In Deloitte's State of AI in the Enterprise survey (3,235 leaders in 24 countries, August to September 2025), only 25% of respondents had moved 40% or more of their AI pilots into production, and just 21% of companies planning to use agents said they had a mature governance model to control them. Gartner predicted in June 2025 that over 40% of agentic AI projects will be cancelled by the end of 2027, because of rising costs, unclear value or inadequate risk controls.

For an SME, that is not a reason to hold back. It is a reason to pick the first project carefully. This article explains why pilots stall and how to avoid it, with a 90-day plan you can start applying today.

What an AI agent is (and what it is not)

A chatbot answers. An AI agent acts: it queries systems, books meetings, updates records and decides the next step to reach a goal, always within limits you define. That ability to carry out tasks is what creates value, and it is also what demands the most care when you implement it.

Beware of agent washing: Gartner estimates that only around 130 of the thousands of vendors calling themselves "agentic" actually deliver that capability. Many simply rebrand chatbots and automations that already existed.

Why so many AI pilots never reach production

1. No measurable goal. The pilot was approved to "try AI", not to improve a specific number (missed calls, response time, bookings). Without a baseline, you cannot prove value.
2. Systems that are not connected. An agent that cannot read the calendar, the CRM or the ERP is a demo, not a solution. Integration is almost always the slowest part.
3. No governance. Who decides what the agent can do on its own? When does it hand over to a person? Who reviews the conversations? This is where most companies planning to use agents are still not ready: according to Deloitte, only 21% have a mature governance model.
4. The team is left out. The people on the front line have to trust the tool. The Starbucks case shows how competent technology fails when it ignores the real working context.
5. Vendor chosen on the demo. A controlled demo does not show how the system behaves with your data, your customers and everyday noise. Always ask for a test with real cases.

Checklist: is your pilot ready for production?

Before you scale, answer these six questions. If any is left blank, the pilot is not ready yet.

• Metric: which number will improve, and what is its current value?
• Integrations: which systems must it read and write, and are they already accessible?
• Limits: what must the agent never do on its own (pricing, refunds, sensitive information)?
• Human handover: when and how does the conversation pass to a person, with full context?
• Oversight: who reviews the logs, and how often?
• Cost and stop criteria: what does it cost per contact or task in real conditions, and at what results do you stop or adjust?

Why voice customer service is a good first project

For many Portuguese SMEs, the phone is still the main channel. Voice customer service has the traits of a good first agent: a repetitive process, a small scope and obvious metrics such as missed calls, response time and bookings made.

Best for: clinics, real estate agencies, law firms, workshops and service companies with many out-of-hours calls.

See how it works in our AI voice agent and, for the broader picture, in AI Agents in Customer Service.

90-day plan: from pilot to production

Days 1 to 15: baseline and scope. Measure the current situation, choose a single process and define what the agent can and cannot do.

Days 16 to 45: parallel pilot. Run the agent alongside human service, compare results and fix the most frequent failures.

Days 46 to 75: limited production. Open it to a real subset of contacts (for example, out of hours), with weekly oversight and human handover always available.

Days 76 to 90: review and scale. Compare with the baseline, calculate the return (see how in AI Return on Investment for SMEs) and decide whether to scale, adjust or stop.

Frequently asked questions

How long does it take to move an AI agent from pilot to production?

For a well-scoped case such as phone customer service, we typically talk about 3 to 6 weeks, depending on the integrations needed. What slows things down most is not the technology but the lack of a defined goal and baseline from the start.

Do I need an autonomous agent, or is an automated workflow enough?

If the process follows fixed rules, an automated workflow is cheaper and more predictable. Keep agents for tasks with variation and natural language. We explain the difference in Automation vs. Artificial Intelligence.

How do I avoid agent washing when choosing a vendor?

Ask for a demo with your own data and systems, not a prepared scenario. Ask what the agent does on its own, how it escalates to a person and which logs are available for audit.

Should customers know they are talking to AI?

We always recommend transparency: identify the assistant as AI and offer a transfer to a person. It is good practice and follows the direction of regulation. See also GDPR and Artificial Intelligence.

Conclusion: start small, measure, then scale

In our reading, the Deloitte and Gartner figures show that the obstacle is rarely the technology: it is method, governance and clarity of value. SMEs have a real advantage, because they can pick a process, measure and decide quickly, without the long cycles of large organisations.

At TecLab, we help Portuguese SMEs move from pilots to agents in production, from designing the use case to measuring results.