Daniel Hulme talked to Digiday about how WPP is approaching the verification of its AI agents.
Digiday interviewed my Conscium co-founder Daniel Hulme, in his role as WPP's chief AI officer, about how the company manages AI agents at a scale most businesses haven't come close to yet. The number alone is worth sitting with: WPP has more than 28,000 AI agents running across media planning, content generation, analytics and optimisation. The more interesting part of the conversation, though, was Daniel's argument that deployment at that scale isn't actually the hard problem. Verification is.
Agents that assist, not agents that act alone (for now)
Right now, WPP's agents work alongside human employees rather than operating with full autonomous access to systems and decisions. Daniel was clear that this is a deliberate, current-stage choice rather than a permanent ceiling, and that the company expects to expand what these agents are trusted to do on their own as confidence and verification tooling improve. That sequencing, capability before autonomy rather than the other way round, is the responsible order to do this in, and it's not the order most companies rushing to deploy agents are actually following.
Why Daniel wants adaptive systems, not static software
The more technical argument Daniel made is about the kind of system being built in the first place. He's pushing WPP, and Conscium's research more broadly, toward what he calls goal-directed adaptive systems, agents that learn and evolve through experience rather than running fixed, hand-coded logic. The risk he flagged is specific: an untested adaptive agent operating independently can develop and spread flawed logic across every system it touches, faster and more invisibly than a static piece of software ever could.
Teaching values through environments, not rules
This is where Conscium's own research connects directly to what Daniel is doing at WPP. Rather than hard-coding ethical rules into an agent, which tends to produce exactly the kind of brittle, gameable guardrails that sophisticated systems learn to route around, the approach is to build simulated environments where cooperative behaviour is what actually leads to success, so agents learn values the way organisms learn behaviour, through consequence rather than instruction. Moral Me, the app we built to crowdsource human judgment on moral dilemmas, feeds into this same effort, and WPP has connected it to Open, its privacy-first platform for campaign planning.
An analyst's outside view on why this matters commercially
Digiday also brought in William Blair analyst Ralph Schackart, who made a point that's easy to miss if you're focused purely on the safety angle: proprietary CRM data gives large agencies a real edge over smaller competitors and off-the-shelf tools, something like a 10x step up in the data available to train and run these systems well. His firm's own survey found two-thirds of advertising executives already rate generative AI as highly effective, which tells you the commercial pressure to deploy faster isn't going away, making Daniel's insistence on verification alongside deployment more urgent, not less.
My take
What Daniel's argument boils down to, and what I'd underline for anyone deploying agents at any scale, is that agents need constant verification and constant testing, not a one-off check before launch. An agent that passed evaluation last quarter isn't guaranteed to still be behaving correctly today, especially if it's the adaptive, learning kind rather than static software. That's the entire premise Conscium's agent verification work is built on.
Read the full interview on Digiday.