Last week I moderated a panel at the ALIGN AI Executive Summit called Scaling AI That Works: Proving Value, Managing Cost, and Operationalizing Agentic AI, with leaders from HSBC, Genentech, UnitedHealth Group/Optum, and Alteryx. On paper, those four organizations have almost nothing in common. Different regulators, different data, different risk tolerance, different definitions of what “wrong” costs you.
I went in expecting those differences to shape the conversation. Twenty minutes in, it was clear they weren’t going to. We were describing the same obstacles from four perspectives.
When banking, biotech, health, and analytics independently arrive at the same list of issues, those obstacles aren’t sector specific. They’re structural. Which is good news for anyone still assuming their industry is uniquely experiencing struggles in builds and adoption.
Nobody on that stage was debating whether AI works. Operational complexity compounds. Here are my take aways from the conversation.
Cost is an architecture decision
Spend accumulates across four surfaces: the data you prepare and maintain, the models you select, the data those models are permitted to use, and the inference you run every time the system thinks. Agentic architectures multiply the last two, because an agent that plans and retries doesn’t consume a predictable unit of anything. A workflow that looks inexpensive in a pilot can look very different in an agentic workflow.
The teams handling this well are making deliberate choices upfront: routing simple steps to smaller models, capping how far an agent may reason before it hands back to a human, and instrumenting cost per completed task. Cost discipline isn’t procurement’s job to negotiate after the fact. It’s an engineering constraint you design towards.
Good data is essential
AI agents are only as good as the foundation they sit on. You cannot unleash AI across bad data and expect deterministic business outcomes. Real enterprise impact happens when AI is grounded in trusted data, clean instructions, and strong human context.
It’s anchored in four properties every workflow should carry: visible, understandable, repeatable, auditable.
That’s the most portable idea from the session, because it converts an abstract governance conversation into a checklist. Can someone see what the system did? Can they follow why? Does it behave the same way tomorrow? Can it be reconstructed for a regulator, a client, or a general counsel? All questions that need to be answered before launch.
The human in the loop
The most consequential design choice in any agentic system is not the model. It’s human intersection.
The organizations getting this right are specific rather than philosophical. Humans hold the judgment calls, the genuine ambiguity, the moments of consequence, and the point of external release.
That is the argument for the human in the loop that has nothing to do with compliance. The client relationship history, the reason a strategy was abandoned in 2019, the regulator who reads things a particular way, the sensitivity that isn’t in any document, that context lives in people. An agent operating without it will produce output that is fluent, confident, internally consistent, and wrong in ways only a human will catch.
Designed well, oversight is an accelerant rather than a brake. It’s the mechanism that lets you increase autonomy over time instead of guessing at it. Track where humans intervene and why, and the intervention rate becomes your roadmap: the places people stop overriding are the places you can safely widen the aperture. The places they keep overriding are telling you something about your data, your logic, or your use case that no evaluation set will.
Keeping humans in the loop is not a hedge against the technology. It is the condition under which the technology gets to scale.
What I’m taking back to our work
I sit in communications, which is obviously not banking or biotech. That was the whole point of the afternoon: it didn’t matter. Four verticals, four risk profiles, one shared set of constraints.
Our work is now being read, ranked, and summarized by generative engines as often as by people. If we can’t ground an audience insight in real data, explain how a recommendation was produced, reproduce it next quarter, and audit it when a client asks, then it isn’t a capability. It’s a story about a capability.
The teams that pull ahead over the next eighteen months won’t be the ones with the most agents. They’ll be the ones who can prove which agents are working, defend what they cost, and show exactly where a human is standing when it matters.
So, start by anchoring AI in trusted data, keep people in the loop by design rather than by exception, and take it one step at a time.
Thanks to David Sabow (HSBC), Deepa Vulupala (Genentech), Haris Dindo (UnitedHealth Group / Optum), and Sharon Martin (Alteryx) for an operator-level conversation — and for being willing to talk about what hasn’t worked, not just what has.
Heather Hughes is SVP, AI & Innovation at Golin Ketchum, where she leads generative engine optimization, agent development, synthetic audience development, and AI product strategy.
