How to scale an AI pilot from one team to the whole company

To scale an AI pilot beyond one team, write down what that team did by hand, fix five things, then add one team at a time. The steps, in order.

By Published 3 min read

To scale an AI pilot beyond its first team, write down everything that team did by hand to make it work, close those gaps in the system itself, then add one team at a time against targets agreed in advance. The pilot did not fail when the second team arrived. It stopped being carried by people who had learned to work around it.

McKinsey’s 2025 survey of 1,993 participants found that nearly two-thirds of respondents said their organizations had not yet begun scaling AI across the enterprise (McKinsey, The state of AI in 2025). The pattern is common, and it has causes you can fix.

Key takeaways

  • A pilot works for one team because that team absorbs the gaps.
  • Scaling means moving that absorbed work into the system, then adding teams one at a time.
  • Each team gets its own targets and a named contact before it starts.

Why does the second team struggle?

The first team chose the use case, knows the data and accepts rough edges. It cleans an input before sending it, knows which answers to double-check and tells the builders when something breaks. None of this is written down, so none of it transfers.

A second team brings different data, different habits and a different tolerance for error. The gaps the first team absorbed now reach users who did not agree to absorb them.

The five steps

1. List the workarounds. Sit with the first team for an hour. Ask what they do before, during and after each use. Every manual fix, every “I always check that” and every rule only one person knows goes on the list.

2. Move each workaround into the system or into a written rule. A cleaning step becomes code. A double-check becomes an approval gate or a visible confidence signal. A private rule becomes documentation. What cannot be moved is a risk to name.

3. Replace personal access with proper access. Pilots often run on one person’s credentials and one shared folder. A system for several teams needs its own service accounts, access by role and a way to add and remove users. Settle this with your security team before the second team arrives, not after.

4. Set targets per team before it starts. Pick a handful of measures that reflect what that team needs and agree a target for each. Google’s SRE book advises choosing a small number of indicators based on what users care about, rather than what is easy to measure, and advises against setting targets from current performance (Service Level Objectives, Google SRE book). A loose target you can tighten later is better than none.

5. Name the owners. One business owner and one technical owner for the whole system, and one named contact inside each team. When something drifts, everyone knows who decides.

How do you add teams?

One at a time. Give each new team a start date, a rollback plan and a review at the end of its first weeks, with the targets from step 4 on the table. If a team misses them, you know which team, which data and which change to look at.

Two limits help. Do not add a team while the previous one is still above its agreed error level. Do not change the system and the team at the same time without saying so.

When is the pilot not ready to scale?

When the workaround list cannot be moved into the system, when no one can be named as owner, or when the first team’s own measures were never written down. Knowing that early is a useful result. It tells you what to settle before you spend on a wider rollout.

Where OrNsoft fits

Our forward-deployed engineers work directly with your business and technical teams, learn how your operation actually runs, determine what has to be built, and move the system from pilot to production. You explain the objective once. A named senior lead responsible until it runs.

OrNsoft is an advanced AI engineering firm, founded in Florida in 2009, with applied AI in its software since 2015. If your pilot works for one team and has stopped there, bring us the problem. One call is enough to know whether we can help.

Questions

Why does an AI pilot work for one team and stall for the next?
The first team quietly compensates for gaps: it fixes inputs by hand, knows who to ask and tolerates odd outputs. A second team has none of that, so the gaps become failures.
How many teams should we add at once?
One. Each new team shows a gap the previous ones did not, and a single new team keeps the cause easy to find.
What should we measure per team?
A small number of indicators that reflect what that team needs from the system, each with a target agreed before the team starts. Google's SRE book recommends the same discipline for services.
Who owns the system once several teams use it?
One named business owner and one named technical owner for the whole system, plus a named contact in each team.

Bring us the problem.

Book a call