Most workplace AI starts with the wrong question. It asks what can be automated. Ross Barnes wants the first question to be what should not be automated.
Barnes comes to the subject with unusual baggage, and he treats that as an advantage. He is a former global CTO at one of the world’s largest agencies, the builder behind Galahad, and a person recently diagnosed with ADD and autism. His daughter has ADHD. That matters because his argument is not abstract. He says the world of work was not built for the way he thinks, so he first builds Galahad’s operating system around himself. Now he wants the same kind of cognitive support available at work and at home.
The clearest version of the idea sits inside Galahad’s Ikigai framework. The framework takes an existing workflow and separates the work that needs a person from the work an AI agent can take on. In marketing, that might mean wiring an agent into Google Ads to deliver the right content to the right consumer in the right voice. In finance, it might mean connecting to a billing system or rules engine. For a CEO, Barnes asks for a job description, a week of calendar data, and an interview, then builds a circle of agents that can flag the meeting that should not exist, the task that needs support, or the dissenting view no one on staff is paid to give.
That last point is small but telling.
A company may not want to employ someone whose job is to disagree all day. An agent can do that without turning every meeting into a fight.
The safety problem is just as concrete. Barnes is blunt about shadow AI: staff are already uploading material into personal ChatGPT accounts when company tools are too clumsy. Galahad has a beta system that inserts canaries into company content, then queries ChatGPT to see whether those canaries appear. With a large FMCG client, the answer is not to lock everything down. It is to make the internal system easier to use than the risky workaround. He gives Unilever only as an example of the sector, not as the client.
His guardrail is plain: "just because we could doesn’t mean we should." Galahad’s four H’s decide when AI should step back. If a process needs someone to be here, to show hunger and resilience, to bring heart and empathy, or to use a hunch, it stays human.
The most surprising claim is not about avatars, though Barnes predicts digital versions of people will attend and act in meetings within 18 months. It is about content. Large language models, he says, do not want to feed on their own food. As AI output spreads, truly human work becomes more valuable, not less.
