10 Nov 2026

From pilot to rollout: navigating AI implementation pitfalls

By balancing big-picture vision with in-the-weeds rigor, tech leaders can build AI initiatives that look beyond the pilot phase to create value in the long run.

Dean Clark

Chief Technology Officer

blogAbstractMinutes

blogAbstractTimeReading

Digital Transformation

2026

AI Modernization

contact

An AI pilot that works doesn’t necessarily tell you much about what the tech would look like in enterprise-level action. Pilots operate in relatively ideal conditions: they get clean data, are driven by keen volunteers, and can often avoid plugging into awkward legacy systems.

Rolling out AI initiatives after pilots conclude is a whole different story. That requires sorting out messy data and, potentially, winning over employees who never asked for the change. It also means hooking the tech into old systems – which is where complications frequently arise.

As we found in The Future of AI, Unfiltered, our global survey of 945 CIOs and CTOs at $500M+ organisations, 84% of respondents have scrapped at least one AI pilot or project because of limitations in their legacy systems. Over 20% have cancelled more than one.

Beyond the challenges posed by legacy systems, the sums of AI implementation also change once you go big. Pilots are cheap and easy to show off, whereas a full rollout means bills can grow unexpectedly with every new user and each new development. The numbers that got the pilot approved rarely persist at scale.

For all the methodology it requires, moving from pilot to rollout can be a messy process. Successful AI initiatives must face and hold up to scrutiny. They must have their pain points ironed out and their value maximised through thorough and relentless testing.

That’s how value gets created. And here’s how tech leaders can get there.

1. Plug the pilot into your ugliest system first, not your cleanest. If it survives the 20-year-old ERP and the spreadsheet nobody owns, it’ll survive anywhere. Integration is one of the key things a pilot should prove.

2. Price the rollout before you approve the pilot. Work out what it costs at 10 times the users, with the integration work and the ongoing per-seat bills included, and put that number in front of the board instead of the pilot number. If it only stacks up at pilot prices, that’s your answer, early and cheap.

3. Stop asking for volunteers. When staffing a pilot, pick the team that you know will voice their criticisms, give them a real say in how it works, and measure whether they’re still using it after three months. Adoption is the number that matters, not accuracy on a tidy demo dataset.

4. Give it an owner. Pilots belong to the innovation team – and rollouts need someone in the business whose job gets harder if it fails. That accountability and buy-in are crucial.

Anyone who has shipped software knows that a successful pilot is a starting point, not a guarantee. But by balancing big-picture vision with in-the-weeds rigour, tech leaders can build AI initiatives that look beyond the pilot phase to create value in the long run.

I’m looking forward to exploring all things digital transformation at this year’s AWS re:Invent, where the GFT team and I will showcase our latest AI-native solutions and host a range of events to connect with the global tech community.

Download The Future of AI, Unfiltered for more insights ahead of the conference, learn more about our programming for AWS re:Invent, and reach out if you’d like to catch up there.