I spent Thursday at the Institution of Civil Engineers for a Rail in AI event. Four speakers, good turnout, honest questions.
But the room had a mood. Every Q&A came back to the same concern: is AI going to take our jobs?
It's a fair question. I have been in signalling for 14 years. I have seen what happens when someone who does not understand the system trusts an AI output. It is not pretty. I have also seen what a good engineer can do with the right tool and deliver twice as fast.
The data we have now is clearer than the mood in that room suggests.
Employment grows where AI is adopted seriously
The strongest evidence comes from the Ramp Economics Lab and Revelio Labs, who published a study in June 2026 linking actual AI vendor spending, real money flowing to OpenAI, Anthropic, and other providers, to headcount records across 21,559 US firms.
High-intensity AI adopters grew total headcount 10.2% over the two years after adoption. Entry-level roles grew 12%. The effect was broad: engineering, sales, administration, customer service, finance. Not just technical roles.
The study has honest caveats. Adopters are already larger and faster-growing before they adopt. The clearest gains show up in the Information sector. But the paper is the first to measure adoption through actual payments, not surveys or occupational exposure scores. It counters the claim that AI inevitably drives job losses.
The PWC Global AI Jobs Barometer 2026 tells a consistent story. The most AI-exposed companies see faster headcount growth than the least exposed (52% vs 36%) and higher wage growth (24% vs 17%). Productivity growth is 40% higher in the most exposed companies.
The gap between piloting and delivering
80% of organisations run AI pilots. Only 9% have mature deployments.
MIT's Project NANDA published "The GenAI Divide: State of AI in Business" in July 2025. The headline finding: despite an estimated $30-40 billion in generative AI investment, about 95% of organisations reported no measurable P&L impact from AI pilots.
The number went viral for a reason. But it is worth reading what the study actually said. The 95% includes pilots that saved time without that time translating to a financial outcome. It measures impact on the profit and loss statement, not technical success or user satisfaction. Critics note the methodology was preliminary and self-reported.
The more interesting finding in the same report: pilots built with vendors or partners succeeded 67% of the time. Pure internal builds succeeded only one-third as often. The divide is not about technology. It is about approach.
The gap UK rail is missing
Jay Shu at the event talked about operational research in UK rail. The key point: UK rail underinvests in OR compared to aviation and logistics.
Aviation has used AI crew scheduling for years. The major carriers run sophisticated platforms from specialists like Sabre and Jeppesen. Route optimisation at UPS saves $300-400M a year. 10 million gallons of fuel. 100 million fewer miles driven. DHL's robot-assisted sorting increased capacity 40%.
UK rail has network modelling tools. But the investment, data quality, and cultural commitment to operational research lag behind.
That is not an AI problem. It is a prioritisation problem.
The IPO incentive
One thing nobody mentioned at the event: AI companies have a financial incentive to look powerful and dangerous.
Anthropic confidentially filed for IPO with the SEC on June 1, 2026 (Fortune, WSJ). Its Series H funding round in May 2026 valued the company at $965 billion. The company told investors its annualised revenue run-rate had passed $47 billion, up from $9 billion at the end of 2025.
A valuation of that size depends on a narrative. AI as a force reshaping entire industries supports it. AI as a tool that helps engineers draft documents faster does not.
I am not saying the jobs concern is manufactured. It is real. But the incentive to amplify the danger narrative is also real.
What actually needs to happen
Senior engineers can spot when AI is wrong. Juniors can not.
That is the real risk. Not replacement. A generation of engineers who do not build the judgement to challenge an AI output because the AI did the work before they had to.
The solution is not stopping AI. It is good leadership, structured mentorship, and letting juniors make mistakes in environments where the consequences are contained. The same approach every good engineering manager already uses.
One speaker at the event used electrification as an analogy. Swapping a steam engine for an electric motor gave limited gains. Real productivity came when factories were redesigned around the motor. Same with AI. Bolt it on to existing processes and you get marginal improvements. Redesign the process and you get step changes.
The practical question from the event was direct: how did you do your timesheet last month? That is the level to start at. Find the repetitive, high-volume, low-risk work. Automate that. Prove the methodology. Scale.
Someone at the event told me they forwarded AI-generated information that turned out wrong. After that, they stopped trusting the tool entirely. One bad output. That is how trust works in safety-critical environments. It takes months to earn and one mistake to lose.
That is not an obstacle to adoption. It is the right response.
What I think
The data is consistent: serious AI adoption correlates with employment growth. The companies that invest in it deeply grow headcount faster. The projects that fail do so because of organisational problems, not technical ones.
But data describes averages. It does not describe your project or your railway. What matters is how you apply AI to the work you actually do. Start small enough to finish. Run it alongside the existing process. Compare quality and time saved. Prove it before scaling it.
The gap in rail is not an imagination problem. It is an adoption and trust problem. The sector knows what AI could do. It has not built the confidence, the skills, or the organisational capability to do it well.
That is the work. Not the technology.
Sources
- Ramp Economics Lab / Revelio Labs: "A New Look at AI's Impact on Jobs: Firm-Level AI Spending and Workforce Adjustment". Kharazian, Simon & Stevens (June 2026). ramp.com
- PwC 2026 Global AI Jobs Barometer: Global findings (2026). pwc.com
- MIT Project NANDA: "The GenAI Divide: State of AI in Business 2025" (July 2025). valtao.com
- Fortune: "Anthropic confidentially files for IPO after a $965 billion valuation" (June 1, 2026). fortune.com
- The Wall Street Journal: "Anthropic Files to Go Public in Blockbuster Year for IPOs" (June 1, 2026). wsj.com
- UPS ORION: Route optimisation savings. UPS corporate communications.
- Research and Markets: Railway AI market report (February 2026). researchandmarkets.com
- Dataintelo: AI crew scheduling market and savings (March 2026).
- techUK: UK rail data fragmentation and innovation barriers.
- UK Government: Digital skills gap in rail.
Image credits
- Header image: Institution of Civil Engineers, One Great George Street. CC BY-SA 2.0, by Derek Harper, via Geograph / Wikimedia Commons. Source file
