AI Agents Without a Portfolio Add Digital Clutter
By Professor Ashley Braganza
An organisation can add AI agents and end up with clutter instead of capability. An agent for underwriting, another for finance and a third for one team's knowledge work can each pass a proof of concept and leave the organisation no more capable than before. When those agents sit on top of fragmented processes and dysfunctional data, they add a further layer of digital clutter. Use cases, pilots and agents built by individuals or departments are activity. They are not a workplace capability, and they will not deliver a return on AI.
My guest on The AI Adoption Podcast is Ericson Chan, Group Chief Information and Digital Officer at Zurich Insurance Group, and digital clutter is his phrase. This is the second of a two-part conversation. The first sets out three regions building AI for different ends and Zurich's AI360 strategy for working across them. The second moves inside the organisation, to the reasons some will earn a return on AI, which Ericson calls ROAI, and others will not.
Unbounded agents turn unreliable answers into unreliable actions
Ericson draws a distinction that boards should hold onto. A chatbot produces an answer; an agent interprets a goal, plans, calls tools, retrieves data and takes action. The concept and the technology emerged barely two years ago, in Ericson's account, and the shift from content generation to execution already changes the risk profile. In his words, an unreliable answer can become an unreliable action inside a business process.
Zurich's response has two parts. The first is the workflow. Ericson separates deterministic orchestration, in which each agent follows a predefined path and hands work to the next in a designed sequence, from non-deterministic orchestration, in which an agent is given a goal and finds its own route, sometimes a creative one that is not ideal. For now, Zurich uses the deterministic approach so that it knows the results are reliable.
The second is what he calls an operating envelope. Each agent is ring fenced by identity, permissions, data access, the actions it may take, the threshold at which it escalates and the way it is monitored. Approving which model to use is a small part of this governance; defining what each agent can and cannot do is the larger part. Moreover, whether agents that call tools and act inside live business processes can be held reliably within such an envelope is an open question.
An organisation that approves models but not envelopes has governed the wrong thing.
Judging every AI project on its own return rules out the moonshot
Ericson's answer to the question of value is a portfolio. The board, he argues, must govern AI as part of enterprise strategy and capital allocation, testing where AI changes the economics, the risk profile and the competitive position of the business. Reviewing a pipeline of projects does not do that, and neither does demanding an instant ROAI from every experiment.
Some initiatives in any portfolio will be mundane and will deliver efficiencies. Those are useful in their own right. The danger lies in a leadership team that looks for nothing else. Ericson's reason for the portfolio is the moonshot, an uncertain bet on the future that survives only when it is judged alongside safer work. Managed project by project, in his words, the moonshot "will never happen".
Here Ericson is more optimistic than I am. He notes that organisations were running pilots a year or two ago and are rolling out sequences of projects today, and he expects boards to reach portfolio thinking because reviewing projects one at a time will not produce measurable results for the whole organisation. He is "not too worried". He does call holding a moonshot within the portfolio the tricky part, and that is where the weight belongs. This is a point I put to boards and senior managers I work with: in large, complex organisations, silos pull AI back into piecemeal review, and a portfolio assembled from departmental initiatives is not a strategy.
A team that never sleeps exposes the culture leaders already have
Ericson describes the employee's role shifting from using a passive tool to directing semi-autonomous collaboration. Employees need the confidence to delegate and enough scepticism to verify, intervene and take ultimate responsibility. He calls this a big ask. The two qualities pull in opposite directions, and an organisation that rewards speed alone is likely to get the confidence without the scepticism.
Managers face a harder version of it. Today a manager leads a team of ten or fifteen. Going forward, Ericson says, "half the team will never sleep". On my reading, that changes what a manager is accountable for: work produced overnight by members who cannot be coached in the usual way still carries the manager's name. Expertise changes too. It has long meant producing analysis and content personally; in an agentic world, contribution depends more on problem framing and judgement.
His account of roles is precise. AI takes tasks away, he says, and "They don't take jobs away most of the time." Many roles combine repeatable cognition, relationship work and judgement work, and AI affects each differently, so tasks are redesigned and jobs rebundled while the title may stay the same.
Culture follows from this. Ericson calls AI a force multiplier that makes an organisation's existing culture more apparent, and he names adaptability and curiosity as more important than before. An organisation that is slow to adapt is likely to find its agents amplify that slowness.
Automating junior work removes the place where judgement is learned
Ericson calls AI literacy a basic workplace capability: knowing what a system can do, where it fails, how data is handled and when human review is required. The emphasis moves from producing AI output to interrogating it.
The harder problem is the route into expertise. Junior tasks such as research, drafting and reconciliation are how people, Ericson included, learned the patterns, the standards, the context and eventually instinct. He stops short of calling the apprenticeship model at risk; he says it needs to be rethought. His proposals are simulation, shadowing and rotations, with AI used to build simulation and augmentation tools so that junior employees still gain pattern recognition. In his phrase, organisations use AI to manage this AI change.
Leaders are under pressure for productivity, efficiency and cost reduction, and removing entry level tasks answers that pressure. Ericson put the same question to me about universities, and the answer from a university matches the answer from an insurer. I see the same tension with students almost daily. They can find an explanation of almost anything in seconds, yet ten videos offering ten approaches still leave them to judge which combination fits their context. That judgement rests on subject expertise, and there is no shortcut to it.
Zurich's Hyper Challenge hackathons, which bring together hundreds of employees from 20 to 30 countries to build prototypes, point to the practical answer: people learn how agents work by building them.
Ericson describes a two-speed transformation: exponential technical possibility, constrained by institutional change capacity. He does not offer a single solution, and he is candid that "We're all learning together at this juncture." A portfolio of agents inside envelopes, judged on ROAI, is a governance design. Whether an organisation can learn as fast as its agents improve is a separate matter, and it is the one that decides which organisations earn a return on AI. Learning to learn about AI is the capability leaders need to build in themselves and in every colleague.
Listen to the full conversation
Comments