Why Industrial AI Pilots Stall
Plenty of pilots, too little payoff. That was the refrain when seven manufacturing and industrial leaders sat down in London on 8 October to debate how Agentic AI can deliver tangible business value. The stakes are not abstract: thousands of vehicles stranded off the road for a month at a time, decades of engineering expertise heading for retirement, and bid teams spending weeks on a single tender response.
The conversation took place at The Art of Agentic AI, hosted by DataArt and Google Cloud. Seven senior technology and operations leaders gathered for an executive roundtable at Google’s London office, followed by a Chef’s Table dinner at a separate venue. The guest list was a cross-section of British industry: fleet operations, industrial engineering, commercial distribution, automotive retail, construction, energy and life sciences. Job titles ranged from AI engineer to IT director to global head of digital.
Guests ran their estates on established cloud environments. Most were past the beginner stage of AI, using tools day to day, but sat at a medium-to-low level of maturity. Two themes came up again and again: their data was a mess, and too many pilots had ended without a clear use case or a clear return.
Lillian Zhang, DataArt’s Google Cloud GTM Lead, emphasised a business-first approach to AI: start with a clear operational challenge and validate ROI upfront. She asked each guest about their business needs, existing agents, and what was working or falling short. Their answers shaped the discussion and demo around real-world priorities.
The framework was clear: define measurable problems, build agents that act, establish trust through data and governance, and measure ROI from day one.
Bringing together leaders from different industries also proved valuable. Without direct competitors in the room, participants could speak more openly about their challenges, creating space for more candid and practical discussions.
Start With a Painful, Measurable Problem
The strongest AI use cases in the room did not start with model capabilities. They started with an operational problem the business could already measure.
- Seeing breakdowns before they happen. A fleet director described a problem of scale: at any moment, roughly 5,500 vehicles are off the road, each for 32 to 33 days. Many are adapted for wheelchair users, so downtime is especially disruptive, and a workshop can take three to four weeks just to inspect the car. The bill runs to tens of millions of pounds a year. His team is working with manufacturers to feed sensor data and diagnostic trouble codes into predictive models, so repairs can be booked before a vehicle fails.
I don’t think we’ve got an analyst that can consume all the data and metrics that we’ve got. I want something far smarter.
- Keeping expertise when experts retire. An industrial engineering group relies on around 15,000 application engineers who act as trusted advisers to customers. Training one can take up to a decade, and many senior engineers are close to retirement. The group is grounding models on decades of internal engineering literature, and is even exploring delivering that expertise to customers as a service. The obstacles are familiar: legacy data in poor shape, and user adoption that stalls below 30%.
What if every sales engineer could gain the knowledge of the very best engineer within three to six months, rather than taking a decade to learn it?
Digital leader, industrial engineering grou
- Tenders in minutes, not weeks. A distributor serving airports, hospitals and public bodies used to spend at least 20 days drafting each tender response, working across more than 56,000 documents and 30,000 products, with each sector using its own terminology and its own margin rules. Its internal AI system now allows experts upload a request for proposal and returns a complete draft in about five minutes. The team uses zero-data-retention APIs to keep sensitive material private.
In each case, the starting point was an operational problem the business could already measure, which is what made it possible to ask whether AI was actually moving the number.
Aim for Agents That Act, Not Just Answer
The challenges discussed were action problems, not information problems: booking repairs, drafting tenders, or delivering expertise to engineers when they need it.
One distinction framed everything that followed: the difference between a model that answers and a system that acts. To anchor the conversation, the session offered a simple picture. A raw large language model is a librarian: knowledgeable, but it can only draft a static itinerary from memory. An agent is a travel agent: it checks live fares, handles booking errors, and completes the reservation.
For industrial companies, that means moving beyond summarising maintenance manuals to identifying failing components and scheduling repairs.
A live demo using Google’s Antigravity and Gemini platform brought this to life. A supply chain manager agent coordinated specialised agents monitoring logistics data, shipping lanes, and news to flag potential disruptions. Antigravity translated natural-language prompts into working code and infrastructure, while Gemini provided the runtime and governance.
The business goal: shorten the gap between detecting a problem and acting on it. Because the demo was tailored to the challenges guests had shared, it demonstrated relevant applications rather than a generic scenario.
Don't Wait for Perfect Data, but Earn Trust
Agents that act raise the stakes, and the room was candid about what stands between a promising pilot and measurable value. Four tensions surfaced, and not all were resolved neatly.
- Data first, or agents first? Several guests described data scattered across systems. One asked whether agents can work through a messy landscape directly, instead of waiting for a multi-year clean-up that would delay value by two to five years.
- Can we trust the answer? Hallucinations and data accuracy were raised early. The response was practical: ground the AI in the company’s own documents with tools such as Gemini Notebook and Gemini Enterprise connectors, and attach a source to every answer so users can always check it themselves.
- Whose data is it? Guests asked how a hyperscaler would handle their information. The answer combined evidence and options: Google Cloud already supports highly regulated organisations, including defence and public-sector environments, enterprise data is not used to train models, on-premises deployment is available, and zero-data-retention APIs delete data at inference instead of holding it for up to 30 days.
- Lock-in, but not the kind you expect. Vendor lock-in was not the concern. Leaders wanted to build on existing cloud investments, exploring specialised capabilities beyond standard productivity copilots while prioritising adoption and cost control. The key question was where a second cloud could add value, making cross-cloud expertise essential.
Measure ROI From Day One
If one phrase summed up the afternoon, it was return on investment:
- In automotive retail, where there is little room for inefficiency, one leader highlighted the commercial impact of processes that don’t work as expected. His group sees 70% of deals contain errors, and vehicles can wait up to four months for parts. He came to find out how others are turning AI into measurable gains.
- An energy company is piloting 50 Gemini Enterprise seats and wants proof of adoption before committing further, while it reports that work once needing six people for six months can now be done by one or two people in one month.
- A construction IT director, keen to manage adoption and cost, wants a hackathon to pick the most valuable use cases, then move from pilot to production.
Lillian’s advice to every executive was the same: start from the pain, not the technology. Then prove the value upfront: estimate what a project should return before building it, not after the pilot ends. In the closing minutes, Roman Sorocan, Head of the Google Cloud Unit at DataArt, with more than 10 years of experience within the Google Cloud ecosystem, showed the ROI calculator the company built from almost 30 years of client delivery. Rather than relying on vendor estimates, it helps executives estimate what an agentic project could deliver and compare it with what it costs. The guests stayed to see it, and several asked for more.
Put together, the argument is straightforward: start with a cost you can quantify, build agents that act rather than just answer, earn trust through strong data and governance, and use ROI modelling to demonstrate the potential return and build a business case before committing to the build, while validating the actual benefits through delivery and deployment.
Attendees left with valuable insights and knowledge, and several have already booked follow-up deep-dive sessions. One Senior AI Engineer described the event as “one of the best industry events I have ever attended.”
Develop Your Custom AI Roadmap
The guests left with a clearer view of how AI agents could create business value. You can gain the same clarity through an executive prioritisation workshop with Google and DataArt specialists.
Getting the right AI strategy starts with understanding your business: its priorities, processes, data, and the outcomes it needs to achieve. Together, we'll identify high-ROI opportunities, assess them against your data and governance requirements, and define the next steps with our AI engineering team. We'll also explore how Google Cloud can integrate into your existing environment as a specialised layer.
With deep technology and industry expertise, DataArt helps organizations connect business priorities with the right data, platforms, and engineering capabilities, turning promising AI opportunities into solutions designed to deliver measurable results.
Ready to turn AI opportunities into an actionable roadmap? Let's identify where agents can deliver the greatest business value.












