Human handoff

Intuit Expert Platform is where tax and bookkeeping experts help customers in live sessions. After ChatGPT launched, we started shipping AI features to support expert workflow — starting with Q&A and meeting summary.

Summary turned out to be more than a convenience feature. The live conversation — what the customer clarified, what the expert tried, what changed in the moment — was the richest context the platform had. We called that ambient context.

This case study is about one specific application: using ambient context to preserve continuity when a human handoff happens. But the same layer also made AI answers more specific, and eventually became a shared service other product teams could build on.

I designed the handoff system that made ambient context portable at the moment it was needed most.

The platform already had customer data — structured history, stored and queryable.

What it did not have was the session itself. What the customer clarified, what the expert had already tried, what changed in the moment. I called that ambient context.

Without it, AI could reason from history but not from the live situation. When a lead stepped in, they were starting from scratch — an expert on average spent 6 minutes re-explaining a case the system had already witnessed.

The design problem was clear: make ambient context portable.

Previously, I had launched meeting summary on the platform. From the start of every call, AI transcribed the session and assembled ambient context in the background — the conversation was no longer ephemeral. It was being preserved, continuously, in case it was needed.

To make that visible to the expert, the platform surfaced a single cue: Intuit Assist is taking notes. A quiet signal that the system was listening and preparing.

Capturing context was only half the problem. At the moment of handoff, ambient context had to be distilled.

/lead is a perfect summary trigger — the expert's gesture carries clear intent. We tuned the prompt to be a targeted distillation of why this case needed a human.

I designed a handoff package that carried the live conversation forward. The package combined:

• A concise conversation summary, with access to the full transcript
• Key customer data, with access to the full customer record

At launch, the model was newer, context windows were smaller, and our prompts less refined. Hallucination was a real concern. So the package included an edit option: a deliberate human-in-the-loop moment at the point of highest risk, giving the expert a chance to correct the summary before it reached the lead.

This turned handoff from a reset into a continuation.

Once ambient context worked at the handoff, the same shape kept reappearing. Transfers needed it. Multi-session conversations needed it. Downstream product teams kept asking for the same capability, scoped to their moment.

That pattern argued for treating summary not as a feature but as a service.

I joined Intuit's Global Engineering Day hackathon and built Summary as a Service with a team of engineers in a week — a shared capability other product teams could call into, parameterized by moment: handoff, transfer, session resumption. The handoff was the proof of concept. The service was the consequence.

Ambient context stopped being a concept. It became infrastructure.

The handoff started as a workflow problem. It exposed a system constraint.

AI can only act on the context the platform can actually unify. Three layers had to come together: customer data for history, system knowledge for grounded domain reasoning, and ambient context for what was actually unfolding in the live session. Without all three, the system could answer from the record but not from the situation.

That made the handoff more than a feature. It became a proof point for a broader argument: the quality of the AI experience depends on the platform's ability to capture, preserve, and connect context across time. Summary as a Service made that argument concrete — ambient context stopped being a layer in a diagram and became a primitive other teams consumed.

I used key-moment prototypes and an end-to-end vision film to help senior leadership see this not as infrastructure work, but as product quality. Every system boundary that context could not cross was a limit on what the model could do. Fixing the handoff without fixing the foundation would leave that ceiling in place.

That argument helped drive investment in the underlying data foundation.

The limiting factor in high-stakes AI is rarely the model.

It is whether the system can preserve the context well enough for either AI or a human to act with judgment.

That was the problem here — context continuity under transfer of responsibility.