What It Takes to Get an AI Agent Into the Core Business

One task, one room, three weeks: how Sun Life turned a big AI plan for claims into a working agent.

Sun Life flew its US claims team from Massachusetts and its technology team from Ireland to Seattle, a city where none of them had an office, and kept them there for three straight weeks. Their job: build an agent that checks claim eligibility across the company’s US products, work that separate teams had been doing by hand, one product at a time.

Dana Parsons, who leads innovation for Sun Life’s global business, set it up this way because of what usually happens to big AI plans. A plan to bring AI into a whole function goes to executives around a table, then to another table in another city, then another, and nothing moves. Her team’s answer is to cut the plan down to one task, give it to a small crew with the authority and time to finish it, and test every decision on three threads: client impact, technical feasibility, and business value. She calls it the Triple Thread Strategy.

When Dana joined Sun Life from an innovation agency, about two months before ChatGPT launched, the company's innovation teams in Canada, the US, and other business groups were all working on improving the existing business. Her global team went where nobody else was: new business models, disruptive ideas, and strategic plays. It kept its projects away from the core, because bringing an idea in before it had all the answers mostly drew detractors.

Over the past six months, the team's job has changed drastically. Its focus has moved from building new businesses on its own to helping the core business change. As Dana puts it, "now everybody's got to be an innovator."

Dana uses the classic innovation matrix to show what AI brings to each of its four corners:

All four corners now run on the same loop of data, AI, action, and learning. In her view, that makes rebuilding the core and creating new businesses one activity: "the same capabilities that improve the core create new businesses."

Once a company's data is aligned, "the data should tell you a story about where to go next." The new competitive advantage, she argues, comes from "reinventing the core to create what did not previously exist."

AI runs on a company's existing systems, data, and processes, which puts the biggest AI opportunities inside the core. In an insurer, AI could personalize a client's experience in real time, from the first visit to the website to the moment the client files a claim. Every step of that journey runs on core systems, so the innovation team has to work inside them, alongside the people who run them.

Many of those people have been in their roles for years, some for 30. They won CEO awards and rose from senior manager to AVP to VP by improving what already exists. When they needed new technology, they bought software, rolled it out, and improved it slowly over the next ten years. "That is all out the window," Dana says. AI asks organizations to rethink their processes altogether, yet the default is to apply AI to the practices already in place.

Data makes that default risky. AI improves in a loop: data feeds the agent, the agent acts, and the results become new data. Many global companies collected client and partner data across countries and business units long before anyone set a global data strategy. Dana sees this as a really big barrier to agentic AI. A team can build a working agent, but without data governance and a plan for how to collect data going forward, she says, the agent "is going to just make a bad process worse. It's just going to make it faster."

To deal with stalled plans, old habits, and uneven data, Dana's team built a program that follows the same steps every time. It picks one narrow task inside a large AI plan, puts a small group of business, technology, and data people on it, builds an agent for that task, and hands the agent over to the business. Keeping the task small lets the group start right away. With business people in the group, the process gets rethought as the agent takes shape, and with data experts there from day one, data problems surface early. The program's stated goal is raising the company's innovation maturity and AI adoption.

Getting the program off the ground came down to two early choices: where in Sun Life to start, and who to bring onto each team.

Sun Life is, in Dana's words, a business of businesses: real estate, Medicare and Medicaid, dental, and nine family wealth offices in different countries across Asia. The core technology team is building the shared AI infrastructure for all of them, on the idea that a high tide lifts all boats.

Dana's team wanted visible progress faster than that. So instead of one large project, it started small pilots in several businesses at once: claims in the US, underwriting in the corporate group, and actuarial work in Asia. The idea was to light "little fires everywhere." If those small slivers made progress, leaders in each business would see the team was on to something.

Each pilot depends on the people working on it, and Dana's team looks for three things in them.

Each pilot runs on a small, cross-functional core team that owns the use case from start to finish and makes decisions close to the work, instead of sending them up through the usual governance.

The product owner has the biggest job. This person owns the problem, the outcomes, and the trade-offs, defines what success looks like at each milestone, and makes the day-to-day calls on scope, speed, value, and feasibility while consulting the rest of the team. Agentic AI involves a lot of ambiguity and experimentation, so Dana's team looks for product owners who are comfortable with both.

The role differs from a traditional product owner or delivery lead in three ways:

As more of the business adopts AI, Dana expects this role to sit at the center of how the company’s structures change. In her view, the product owner “will be central in the transformation of business structures in the wake of AI adoption.”

The rest of the core team covers the business, the build, and the data. A business owner sets the business goal and problem statement with the product owner, and a technical lead turns that into a delivery plan. Business experts check the design against real workflows and what clients say. A solution architect designs the technical solution against company standards, developers build it, and an external partner of AI specialists helps shape and build it. A data expert knows what data exists, how reliable it is, and how it has to change before an agent can use it. Teams add a UX designer or business analyst when the use case needs one.

Each pilot moves through three phases: scope, where the team picks one small task and checks that it's worth doing; build, where the team makes a working agent; and scale, where the business takes the agent into daily work. The US claims pilot shows how each phase works.

Claims has many steps, and building AI for all of them at once would be slow and expensive. So the team's first job was to find one task it could build and test quickly.

Scoping takes a significant portion of each member's capacity. They come off their day-to-day work and their roles are backfilled, with the plan that they never go back. They hold daily working sessions, sometimes several a day, to align on the problem, validate the opportunity, and set a direction for the build. The program attracts people who care more about impact than about climbing the corporate ladder, so the group can be frank with each other, something that doesn't come naturally in corporate culture.

Every candidate task has to pass three tests, the threads. At this stage they work as gates, in order:

Eligibility checking in the US business was manual and siloed, with separate teams reviewing claims for each product. One agent that could look across all products at once was the first task to pass all three threads. Its small scope also made it a quick, inexpensive way to test the company's multi-million-dollar claims ambitions, and it gave the team a clear goal to build toward.

Technology teams usually build a minimum viable product (MVP) on their own, then toss it back over the fence to see if the business will use it. Dana's team wanted the people who would use the agent along for the ride, shaping it as it was built.

So the build phase moves everyone to one place, full time, for three weeks, somewhere none of them has a home office. For the claims pilot, that was Seattle.

There, the team builds what Dana's team calls a minimum lovable product (MLP). An MLP adds adoption to the test: the product has to work, and people have to want to use it.

Working side by side also changes how trade-offs get made. The product owner makes the day-to-day calls, but the bigger solution choices have to work for all three threads, with no group given preferential consideration. If a cheaper or faster technical option would significantly hurt the client outcome, the team doesn't take it. And because the easiest build can also be the most expensive, the business acts as what Dana's team calls an impact marshal, making sure the build that goes ahead is the best one for all three groups. As Dana puts it, "everybody has to be in agreement for any trade-offs that we make."

In Seattle, the claims team built its MLP: an agent that looks at eligibility horizontally, across products. Dana calls it one step into agentifying claims.

When the team members come home from their build, the rest of their business unit is still working the old way. In Dana's words, they're "being put in front of the firing squad": they have to get the rest of the unit on board with what they want to do next. Scaling means aligning three things: the resourcing, the operating model, and the business ownership of the agent.

Because their old roles are already backfilled, the people who built the agent keep working on it instead of slipping back into business as usual.

The team still checks all three threads in this phase, but each one now covers different work:

All three groups are responsible for the outcome and incentivized on it, which keeps the cross-functional team together after the build wraps up. An agent only changes the business once people use it, so Dana's team also sent the group to the Harvard Data Science Institute to learn a change management methodology. The US claims pilot is in this phase now.

Adoption takes time, so Dana's team gives the scale phase a three-to-five-year horizon, in a company used to quarterly results and quarterly ROI. The program has also moved some of the company's best people out of their old jobs for good. Neither choice is easy to justify in a quarterly review, but together they keep the people who built the agent working on it until the business adopts it. "This is a huge change," Dana says, "especially for something that's taking some of the best talent from the company and deploying it in something brand new."

Innovation teams pulled into their company's AI transformation, especially when a big plan keeps circling leadership meetings, can start the same way. They can pick one task inside the big plan, give a few people the authority and the time to finish it, and test every decision against client impact, technical feasibility, and business value.

For any innovation team, the path from a big AI plan to a working agent begins with one check: whether clients, the technology team, and the business all have a reason to build it.