How Definity Cut 3.5 Minutes Off Every Insurance Call

Every minute of insurance call-centre time is a unit of cost, an indemnity-loss vector, and a regulatory liability stitched together. Multiply by call volume across a national broker network, multiply again by adjuster headcount, and the unit economics start to dictate strategy.

That's the framing Definity Insurance — the Canadian property-and-casualty carrier formerly known as Economical — brought to its multi-year Google Cloud build. Working with Google Cloud and Deloitte Consulting, Definity put automated call summarisation, sentiment analysis, agent assist, and contact-centre authentication into production. The first quantified result: 3.5 minutes removed from every customer call, roughly 40% of average handle time, within about a month of go-live.

Definity isn't a tech company. It's a 150-year-old insurer that listed publicly in late 2021 and has been rebuilding its digital stack ever since. The contact-centre work is one slice of a transformation that also includes a 10-month migration to BigQuery and Vertex AI, an enterprise-wide Gemini rollout, and a strategic alliance with Deloitte for the AI build itself. Read together they're a useful reference architecture for any insurance operator evaluating contact-centre AI in 2026 — both what to ship first, and what data foundation it actually demands.

The 3.5-minute call summarisation win

Definity didn't lead with a virtual agent. They led with the most boring possible AI use case: writing the post-call summary so the human agent didn't have to.

That choice is more strategic than it looks. Ron Mills, then VP of API, data platforms and infrastructure at Definity, framed the diagnosis at Google Cloud Next 2024:

"What were the long poles in those experiences — a lot of it having to do with call wrap-up, having to do with authentication."

Wrap-up is where adjuster productivity goes to die. Every customer call ends with the agent typing notes, classifying the interaction, updating Salesforce, sometimes flipping into Guidewire to log a claim event. Anand Nimkar, who leads Deloitte Consulting's gen AI practice and worked the Definity build directly, described the worst case: "The most I've ever seen in one flow was like over 30 applications that an agent had to do to execute an interaction."

The summarisation pipeline replaces most of that. According to Google Cloud's case study material, audio gets ingested through Dataflow, transcribed (an open-source model also runs on Dataflow with GPU acceleration), routed through Cloud DLP for PII redaction, then fed to Vertex AI large language models that produce a structured summary. The summary writes back to Salesforce as the system of record. The agent reviews and corrects rather than authoring from scratch.

The outcome — 3.5 minutes per call, 40% of the average handle time — is striking because it's both small (one workflow) and load-bearing (the agent's time-on-task at the expensive moment). It's also fast: the system was identified as a quick win, scoped, built, and producing measurable savings inside a month.

The lesson for any insurer evaluating contact-centre AI: don't spend the first six months arguing about virtual agents. Ship summarisation first. The economics show up immediately, the agents like it, and the audit trail of redacted transcripts plus structured summaries is exactly the kind of evidence base regulators and internal compliance teams ask for next.

The full-stack CCAI rollout: authentication, agent assist, sentiment

Summarisation was the first phase. The Deloitte-led contact-centre build kept going.

The next deployment was virtual-agent authentication — the front-of-call identity check that gates every customer support interaction. Ron Mills was explicit about the goal: "How can we drive more efficiency away from the call agent? When you actually want to reach someone, you want to have that person-to-person conversation." Authentication is exactly the kind of work no one wants done by a human, and exactly the kind of work where small accuracy gaps (false rejection, escalation churn) compound at scale.

In parallel, Definity put real-time agent assist into the servicing path. As described in Insurance Business Canada, the AI "listens" to broker and customer calls, pulls relevant data from the multiple back-end systems an agent would otherwise traverse, and surfaces compliant guidance — what to say, what regulatory disclosures to include, what the policy actually covers. The same pipeline runs sentiment analysis on customer communications and flags upset customers for prioritised handling. On the claims side, Tatjana Lalkovic, SVP and Chief Technology Officer at Definity, has connected this work to two harder-to-move metrics: Net Promoter Score and indemnity-loss ratios.

This is the architecture every insurance contact-centre roadmap eventually converges on:

1. Call summarisation at the wrap-up step (low risk, immediate ROI).
2. Authentication automation at the front of call (removes a hated human task).
3. Real-time agent assist during the call (knowledge retrieval, regulatory prompts, next-best-action suggestions).
4. Sentiment analysis as a quality and triage layer (escalation routing, supervisor coaching, NPS reconstruction).
5. Claim-level downstream automation that ties the contact-centre signal back into the policy administration system — Guidewire, in Definity's case.

Notice the order. The riskier, more model-dependent capabilities sit on top of the foundational ones. Each layer is testable in isolation before the next is added. That's how you build trust with operations leaders who have seen too many gen-AI demos.

The 10-month data foundation that made it possible

None of the contact-centre work happens without the data. Insurance generates a lot of it: policy admin, claims, broker comms, telematics where applicable, third-party fraud signals, regulatory filings. Definity's pre-migration estate was running at 80% capacity on legacy infrastructure — the kind of operating constraint that makes any new analytics workload impossible to land safely.

The migration to BigQuery and Vertex AI took 10 months, finished about 50% faster than industry benchmark for a project of that scope, and was delivered with Google Cloud partner Quantiphi rather than Deloitte (the partnership map for Definity's overall transformation is layered: Quantiphi did the data migration, Deloitte did the contact-centre AI build). Numbers from Google Cloud's case study material:

• 200 TB of compressed source data, roughly 1 PB uncompressed.
• Time to first insight on a new question dropped from days to about 4.5 hours.
• Business release frequency doubled year-over-year, with 2–5 production releases per day during migration sprints.
• Test automation coverage rose 30%; deployment cycle time improved 63%; new infrastructure provisioning got 10× faster.
• Net Promoter Score against the platform's internal users: 9.9 out of 10.

Ron Mills's framing of the BigQuery choice is the part most worth borrowing: "BigQuery's serverless architecture has been a game-changer. The 'nothing to manage' approach is a huge differentiator. For enterprises like us that are migrating from on-prem clusters constantly running at 80% capacity, it's like night and day."

The takeaway isn't that BigQuery is the only answer. It's that contact-centre AI in insurance is a data-foundation problem before it's a model problem. The summarisation pipeline only works because Cloud DLP can redact PII at speed. The agent-assist surface only works because real-time data access to the back-end systems isn't a 90-second wait. The sentiment routing only works because Salesforce gets written to and read from quickly enough that the supervisor desk has live context. Build the foundation first or none of the customer-facing AI will land.

What insurance buyers should take from this

Definity is a useful proof point because it's specific. Named insurer, named partners, named products, real metrics, real timeline. If you're a head of digital, COO, or CIO at an insurance carrier evaluating contact-centre AI right now, four things stand out:

Lead with summarisation, not virtual agents. It's the cheapest path to measurable handle-time reduction, the easiest to govern, and the one that builds organisational trust in the AI stack before you ask for permission to ship anything customer-facing.

The architecture is layered, not monolithic. Summarisation → authentication → agent assist → sentiment → downstream policy automation. Each layer is a separately ship-able milestone with its own ROI case. Don't let a vendor sell you all of it as one programme.

Pick partners by scope, not by brand familiarity. Definity's stack has Google Cloud as the platform, Deloitte as the contact-centre AI integrator, Quantiphi as the data-migration specialist, and Salesforce + Guidewire as the operational systems of record. Different vendors, different scopes, deliberately layered. "One throat to choke" is rarely the right procurement model for insurance AI.

Budget for the data foundation explicitly. The contact-centre wins are visible. The 10-month migration that made them possible is invisible to executives but essential. If your board only funds the customer-facing layer, the project will stall in pilot.

The 3.5-minute number gets the headline. The infrastructure decisions underneath are the part actually worth copying.

If you're scoping a similar build, our claims and contact-centre AI work for insurers sits in this layered-architecture pattern, and the customer-support agent capability and workflow automation pieces are how we deliver the agent-assist and downstream-system layers respectively. Want to see what your equivalent of the 3.5-minute number could be? Send us a brief.