# Rebase by Elsewhen: Putting AI to Work

> Rebase is Elsewhen’s invite-only series for senior leaders driving AI adoption — a space for open, practical discussions on scaling AI and unlocking enterprise impact.

**Published:** 18 November 2026
**Last Updated:** 18 November 2026

**Source:** [https://www.elsewhen.com/blog/elsewhen-rebase-putting-ai-to-work/](https://www.elsewhen.com/blog/elsewhen-rebase-putting-ai-to-work/)

---

On November 12 we brought a room of technology and operations leaders together for Rebase, a morning of two conversations held under the Chatham House Rule. The first was about putting AI to work across a private equity portfolio, where every project is judged against earnings. The second was about automation in air traffic control, where the software runs 24 hours a day and a mistake is measured in lives.

The two sound far apart. By the end of the morning they were describing the same job from opposite ends: the model is the easy part, and most of the work lies in the system you build around it. Here is what we took away.

## The model debate is a distraction

The private equity session began with a comparison to the ERP and CRM rollouts of the past 20 years. Those programmes succeeded or failed on who owned the data, how well the change was managed and whether people changed the way they worked. The same holds for AI, and an argument over whether Claude or an OpenAI model is better tells a business very little about whether it will see a return.

One figure stayed with us: up to 90% of the effort in an AI rollout goes on change management, which leaves the technical plumbing as the small part. Staff have to learn to prompt in rounds and to review what comes back with a critical eye, and the shape of a firm changes as they do. We heard that the familiar pyramid may give way to something closer to a dumbbell or a barrel.

The same realism applied to tooling across a portfolio. Mandates from the centre tend to fail, whereas two or three master agreements, negotiated in advance with volume pricing, let each company pick its own tool and still get the benefit of scale.

## Boards want earnings

Boards have run out of patience with pilots. We heard that open-ended experiments are losing their funding and that the phrase "AI strategy" now invites suspicion, because it describes activity and promises no result. The projects that survive are tied to a business KPI such as a margin ratio or a measure of operating efficiency, with an executive sponsor and a named owner on the line who answers for the number.

Where to begin is fairly clear. Routine operational work, such as scheduling field engineers or planning delivery routes, can be handed to software with confidence. Decisions that touch tax, statutory accounts, ethics or a conversation with a customer still need a person to sign them off, with strict limits on what the system is permitted to do.

Build versus buy is open again too. Teams can now prototype a working copy of a software product quickly, which makes building in-house a serious option once more. The advice on contracts followed from that: think hard before signing a five-year SaaS deal, because the economics of software are moving too fast to lock in for that long.

## Safety is designed in

Air traffic control is safe mostly because of decisions made long before an aircraft takes off. The simplest example we heard: aircraft heading one way fly at odd flight levels and those heading the other way fly at even ones, which gives 1,000 feet of vertical separation without a controller having to do anything. The design of the airspace removes most conflicts, and the controller's tactical calls deal with what remains.

That design carries human limits inside it. The fixed "motorways in the sky" exist because a person can only keep track of so many aircraft on predictable lines. Software has no such limit, so it can work out flexible routes in three dimensions that no human planner would attempt.

The systems themselves were described as Trigger's broom. Their lineage goes back to the 1960s, and nearly every component, layer of infrastructure and set of requirements has been replaced at least once since. Safety used to mean multi-year waterfall projects and long test cycles. Cyber threats have ended that, since a vulnerability has to be patched fast, and the answer has been modular systems with continuous integration.

One project trained AI agents on a decade of radar data to imitate the decisions controllers make, so that complex airspace designs can be stress-tested in simulation.

## A human in the loop is a weak control

The uncomfortable finding of the morning was about oversight. Studies cited in the session found that people supervising an automated system accept its recommendation about 70% of the time without really checking it. A sign-off given that easily protects nobody, and yet "human in the loop" is the control most enterprise AI programmes lean on.

The alternative offered was a deterministic safety envelope. The AI can propose any trajectory by whatever reasoning it likes, and a separate layer of plain mathematics rejects anything that breaks a fixed rule, such as five miles of separation between aircraft. Nobody has to explain how the model reached its answer, because the envelope holds either way.

Trust still has to be earned from the people doing the job. An early tool that advised controllers to clear aircraft straight up to 37,000 feet met deep suspicion, and it took time before they followed it. We were also warned that AI can be a flattering and persuasive liar, which is a good reason to keep engineers who doubt it.

The stakes explain the caution. An airspace outage of two or three hours takes about five days to recover from, and the effects reach as far as gridlock on the roads around a major airport.

## What we took away

Both conversations arrived at the same advice. Decide where a machine may act alone and where a person must decide, and write those limits into the system itself so they do not depend on someone paying attention. Tie the work to a number the board already cares about. Spend most of the budget and most of the patience on the people whose jobs will change.

That matches what we see in our own work at Elsewhen. The applied AI programmes that reach production start with one costly workflow, a measured return and limits agreed at the outset. If that is where your business is heading, [get in touch](https://www.elsewhen.com/contact). There will be another Rebase, and we would be glad to see you there.

[Video: YouTube - 778guOyk5q0](https://www.youtube.com/watch?v=778guOyk5q0)