← Back to blog
AI Agents

From pilot to production: the three factors that decide whether your AI agent ever pays for itself

By Zarioh Digital Solutions4 min read
Share

Research among more than two thousand decision-makers in twenty countries shows what agents that do reach production do differently from the rest. Not better models, but clean data at the right moment, a narrow task, and an escalation path that existed before launch.

Almost every organisation today runs at least one AI agent, or a pilot of one. An assistant that checks invoices, a bot that triages support tickets, a system that processes orders without anyone needing to watch it. The enthusiasm is real and the budgets keep growing. Yet research from recent months keeps pointing to the same uncomfortable pattern: most of those pilots never reach production. They run for a few weeks, produce a demo for leadership, and then quietly disappear from the agenda.

What 'production' actually means

The gap between a pilot and production is bigger than most organisations assume. A pilot runs on a selected dataset, with a handful of test users, and someone from the project team stands ready to step in the moment something goes wrong. Production means the agent runs on real data, with real customers or employees, and exceptions get handled without anyone watching over its shoulder. That last part turns out to be the bottleneck. As long as an agent only works under the supervision of the person who built it, a successful pilot says little about what happens once that person goes on holiday.

What the research shows

Gartner forecasts that more than forty percent of current agentic AI projects will be scrapped before 2027, before delivering anything structural. Other 2026 research confirms the pattern from a different angle: nearly every organisation now deploys AI agents, but a far smaller share actually runs them at scale in day-to-day processes. The exact figures vary by study and methodology, but the direction is always the same.

In May 2026, Salesforce surveyed more than two thousand decision-makers on agentic AI across twenty countries, asking specifically what separated agents that moved past the test phase from those that did not. Clean data at the moment the agent acts, and a narrowly defined task, tied as the top factor, each cited by thirty-six percent of respondents. An escalation path that already existed before the agent went live followed close behind at thirty-five percent. None of the three factors is about the AI model itself.

Factor one: clean data at the right moment

Most organisations hear 'clean data' and picture a fully tidied data landscape, then back away because that project never finishes. That is not what successful deployments actually do. They make sure the data the agent needs at the exact moment it acts is correct and accessible, not that every piece of data in the organisation eventually lands in one pristine system. An agent that checks invoices needs current supplier data at the moment of checking, not a perfectly structured record of the entire purchasing history.

Factor two: a narrow task instead of a broad promise

The agents that make it tend to do surprisingly little. They check one type of form, answer one category of question, or handle one step in a process. The ones that stall usually get handed a broad mandate instead, such as 'take over customer service' or 'automate procurement'. A narrow task is easier to test, easier to explain to the people who work alongside it, and easier to roll back if something goes wrong. A broad promise sounds better in a sales pitch, but in practice it is almost never what actually gets delivered.

Factor three: the escalation path existed before launch

Every agent sooner or later runs into a situation it was not built for. An invoice that falls just outside the known pattern, a customer question that does not fit the categories, an order with a deviation nobody anticipated. What happens next decides whether users keep trusting the agent. Organisations that defined in advance who an exception gets routed to, and how, keep that agent standing. Organisations that improvise this afterwards see users quietly work around the agent within a few weeks, and at that point the project is effectively over, even if it is technically still switched on.

Plan in months, not weeks

A well set-up agent does not deliver measurable return within a few weeks. The same research points to an average of around eight months before organisations describe visible, measurable gains. Organisations that start with a limited set of governance agreements, such as who may adjust the agent and how errors get logged, almost always expand that set after going live. That is not a failure. It is exactly how it should go: an agent that is actually running shows you where the rules still need sharpening.

For a business owner or operations manager considering an AI agent, the practical lesson is simple: start smaller than you'd like to, decide in advance who steps in when something goes wrong, and plan in months rather than weeks. That is less exciting than the broad promise in a sales deck, but it is the difference between an agent still running a year from now and a pilot nobody remembers. Want help scoping an AI agent, getting the right data in order, or setting up an escalation path before you go live? Zarioh helps you work out what an agent should concretely do in your organisation, and what it takes to keep it running.

Z

Zarioh Digital Solutions

IT specialists from Utrecht, the Netherlands. We help businesses with Microsoft 365, AI agents, hosting and telephony, and share what we learn in practice. Follow us on LinkedIn

Related articles

← Back to all articles
Share