Turning conversations into context, in seconds.

Advocates were documenting the conversation while trying to have it. That created two kinds of cognitive load:

How might we give advocates enough context to act, without asking them to read everything that came before?

They weren't reading transcripts. They were hunting for signals. Watching advocates work, we saw them scan every conversation for answers to the same three questions:

1. 1

   What happened?

   What is the member contacting us about?

2. 2

   What's been done?

   What troubleshooting, resolution or promises already happened?

3. 3

   What's next?

   Is anything unresolved or needs follow-up?

InsightAdvocates didn't need a shorter transcript. They needed the conversation reorganized around decisions.

1. Compress
2. Structure
3. Prioritize

1. V1 · Paragraph

   We started by compressing the conversation into one concise paragraph. It cut reading, but advocates still had to scan prose for what mattered.

   Shorter wasn't the same as scannable.

2. V2 · Structured

   We reorganized around the advocate's mental model: predictable sections, Caller's concerns and Resolution provided, so they could jump straight to what they needed.

   Structure mattered as much as summarization.

3. V3 · Signals

   Topic tags and visual hierarchy surfaced the key signals before advocates read a word of detail.

   The goal wasn't to read faster. It was to decide what deserved attention.

AI could write the summary. The advocate still owned the record. The question wasn't only “can AI summarize accurately?” but “what should happen when it doesn't?” So nothing is written into the member's permanent history without a person.

1. 01

   Generate

   AI drafts the summary from the conversation.

2. 02

   Review

   The advocate sees it first, with a clear note that it's AI-generated.

3. 03

   Edit

   Fix anything missing or wrong in place, no regenerating.

4. 04

   Confirm

   Only “Add to incident” saves it to the record.

AI proposes → Human verifies → System records

Failure wasn't an edge case. It was part of the interaction model.

Incomplete

The advocate adds what's missing

Incorrect

Corrected before it's ever saved

Unhelpful

Cancel, and the original conversation is still there

AI-generated

Always labeled as such

Permanent record

Requires an explicit human confirmation

Less documentation. Faster context. Human control intact.

The ~30% isn't a failure rate. It's the editable workflow doing its job: advocates kept control where the AI fell short, and those edits show where the summaries can get better. Closing tickets with AI also took 7 seconds off average handling time.

“Another win for our contact center advocates and members. I love the path we are on with new self-serve and gen AI capabilities.”

Contact center leadership, at go-live

From summarizing the past to helping with what happens next.

Smarter structure

Summaries that adapt to each type of interaction

Quality signals

Learn why advocates edit, and feed it back into the experience and model

Cross-channel context

Continuity across chat, calls and past incidents

Next best action

Go beyond what happened to what to do next

The future isn't a better summary. It's an advocate experience that understands context, surfaces what matters, and moves the conversation forward.