.png)
Daniel Hulme has a one question test before anyone in a 100,000 person company builds an agent: would you hire an intoxicated graduate to do this job? If the answer is no, do not build the agent. It is the fastest way he has found to kill the pilots that were never going to work.
Who is on this episode
Dr Daniel Hulme is Chief AI Officer at WPP, the world's largest advertising group, and the founder and CEO of Satalia, the AI company WPP acquired to get him. He holds a doctorate in computational complexity from UCL, where he also directed the business analytics programme, and he has been applying AI for two decades rather than two years. At Satalia he ran some of the most radical structural experiments anywhere: no managers, decentralised decision making, algorithms allocating work and peers setting salaries.
What they covered
Daniel opens with the three things an enterprise has to get right to see any return. Give employees access to the tools, because they cannot innovate without them. Then accept that tool access alone differentiates nothing, since your competitors have the same tools and similar data, so the real edge comes from pairing domain experts with deep AI experts. And third, stop building what you should be buying, because employees across the market are rebuilding software they could license for a fraction of the cost.
On agents and vibe coding he is blunt. Writing the code is the easy part. Testing, security, scale, resilience, documentation and support are the job, and nobody is putting testing frameworks around agents at all. Imran adds the operator's version of the same point: companies are making the same mistakes with agents that they make with new hires, dropping them in without context or access to the right systems.
"Creating the software is 5 % of the problem. The next challenge is how do you make sure that it's safe, secure, it operates at scale, is resilient, is robust."
Daniel Hulme
The Satalia salary experiment gets a full walkthrough. People recommended their own salary with a dashboard of internal and external comparisons, then peers voted, with each vote weighted by how closely that person had worked with them over the past year and how well they understood the domain. Interns outweighed Daniel on some decisions. There were no caps. The lesson they landed on was that people are bad at putting a number on themselves, so peer assessed skill level turned out to be the better proxy for pay.
Then the part most useful to a CFO. Daniel argues that measuring AI adoption is like measuring internet adoption, and that any such metric gets gamed the way lines of code and story points were. The alternative is unglamorous: list every friction in the organisation, across your supply chain, your back office and your workforce allocation, prioritise them on data availability, cost, return and reuse, and knock them off one at a time. Some will be solved by process, some by software you buy, a few by software you build.
Key insights
- The agent litmus test: if you would not hire an intoxicated graduate for the task, do not build an agent for it. And if you would, test it, because Daniel puts 80% of the effort in shipping software into testing.
- Most organisations have somewhere between 50 and 500 identifiable frictions. That list, prioritised by a simple business case, is a better AI roadmap than any adoption dashboard.
- Objective setting exists to police the 5% of bad actors, and everyone else pays the tax. Daniel's example of the better pattern is expenses made publicly visible so peers self police, which is less efficient and more effective.
- Weight the vote by proximity. Peer input on pay or performance is only useful from people who actually worked with the person and understand the domain.
- Do not replace the word manager with a single new title. Daniel split it into roughly eight groups of responsibility and distributed them, so gaps like team cohesion get owned explicitly rather than assumed.
- His number that changed his mind: 500. The months he worked out he had left to live, in his early twenties.
Watch the full conversation on YouTube, or listen to the audio version on Apple Podcasts.
