TriZetto, a Cognizant company, ran a six-month engagement with KUNGFU.AI that paired two live pilots with an executive-endorsed playbook, proving that requirements clarity, not tooling, was the binding constraint on delivery speed.
We build production AI for healthcare and life sciences: architecture that controls what inference costs to run, governance that keeps automated decisions defensible under review, and a use case order that puts the business case before the build.
Scrum teams in scope
Scrum teams in scope
Scrum teams in scope
North-Star AI-Driven SDLC
Operating Model
AI Tooling Procurement and Cost Strategy
Role Evolution Guidance
6 years
Two pilots on live roadmap work showed the constraint was requirements clarity, not code. Leadership endorsed a phased path to scale the fix.
TriZetto builds the software health insurers use to review claims and prior-authorization requests. When a doctor’s office needs to know whether a surgery will be paid for before performing it, TriZetto’s products answer that question. The company grew through acquisition and delivers like it, with product lines operating as near-independent units on legacy Java stacks and little cross-team communication. Leadership’s ask wasn’t efficiency or headcount reduction. It was speed, quality, and capability sufficient to outpace competitive pressure, plus a shared model for how software gets delivered so that any gain could scale past the team that found it.
Requirements documents stood in for a design phase. In one instance, a product manager watched a dev team demo a completed feature and asked why they had built it. Gaps in requirements surfaced late, driving rework, defects, and delays, and changing a single line of code triggered a manual QA cycle measured in weeks, capping how fast any team could move regardless of tooling. A year or more typically passed between a feature being finished and reaching a client environment, often after the engineers who built it had left, and roughly 90% of engineering effort went to maintenance rather than innovation. Client leadership framed the stakes in economic terms: approximately $300 million in annual labor spend, with about 6% of it tied to innovation. AI adoption already existed, but only in pockets, with teams optimizing what they individually owned and defining value on their own terms. Effort was duplicated and results didn’t compound, because a tool any competitor can license could not show where TriZetto’s own delivery broke.
KUNGFU.AI ran a six-month engagement spanning discovery, pilot design, pilot execution, and methodology development. Discovery combined more than 40 interviews and 20 surveys, from the C-suite to individual practitioners, across nine product families and both onshore and offshore teams, producing a shared picture of current-state friction alongside an honest read on the organization’s emotional climate: practitioners expressed fear about leadership’s AI decisions, and leadership expressed urgency about falling behind.
Rather than test one hypothesis, KUNGFU.AI ran two pilot tracks in parallel, both on live roadmap work rather than exercises: 44 participants, two kickoffs, 10 standups, six mid-point check-ins, six retrospectives and demos, two pulse surveys, and 44 friction logs. Track one layered agentic techniques into two established teams’ existing sprint structure while they continued real feature work, with KUNGFU.AI coaching rather than co-building, to test whether targeted AI techniques amplify how a working team already operates. Track two stood up a purpose-built cross-functional pod executing a redesigned lifecycle from concept through working software, with KUNGFU.AI shaping roles, workflow, tooling, and cadence directly, to test whether a fundamentally different operating model produces materially better outcomes.
An AI coding assistant served as the agentic coding tool for both pilot tracks, with one track running a spec-driven framework inside it to move from prototype to codified requirements to implementation. Coaching engineers on collaborating effectively with the tool was a named part of KUNGFU.AI’s delivery role, distinct from tool enablement. The pilots validated a set of patterns carried into the final playbook: making requirements concrete before build as a UI mock or API contract rather than a document; running development and QA in parallel against a single contract instead of sequentially; authoring validations before code and treating them as the source of truth; keeping units of work small and decomposed, since agents degrade on large, loosely scoped features; documented design patterns and technical standards that keep AI-generated code consistent across teams; and treating heavyweight, spec-driven frameworks as optional rather than default, since they proved too costly for small work items and for teams new to agentic coding tools. The most consequential finding wasn’t about code generation at all. The binding constraint was requirements clarity, and prototyping addressed it directly, with pilot teams surfacing requirement gaps in hours rather than across sprint cycles. Pilot findings became a six-part playbook: a north-star AI-driven software development life cycle (SDLC), synthesized pilot findings, an AI tooling procurement and cost-management strategy, an operating model, role evolution guidance, and a framework tying AI adoption to value creation across product and engineering.
The engagement delivered a validated methodology and an endorsed path to scale it. The pilot ran on live roadmap work across three teams, with both tracks completing full iteration cycles and working demos delivered from every pilot team. Leadership aligned on a target operating model, endorsed the phased sequence to reach it (now, next, later), and had the support cost of the new team model quantified before any wide rollout, so the organization could size the investment rather than discover it. Communities of practice were established for product managers, engineers, and designers to move learnings across the silos that caused the original problem, with all six playbook sections handed over as an actionable blueprint.
Running two tracks in parallel also ruled something out. The purpose-built pod under a rigid framework forced sequential work, idled while waiting on plans, created friction between newly defined roles, and would have required high-touch support to sustain. That answer was only available because both tracks ran side by side, rather than after a failed enterprise rollout. The recommendation that came out of the pilot is a hybrid adaptive SDLC, and agentic-coding adoption has continued expanding enterprise-wide since, with teams using managed agents and skills to streamline requirements gathering, ticket generation, rapid prototyping, and automated test suite creation.
Early results, self-reported by the track-one pilot team: cycle time down 24%, an 80% test-coverage target met through automated test generation, requirement-gap identification down from sprint cycles to hours using a managed agent, and routine endpoint and bug work down from sprints to hours after the team built a new API generation framework.
Any competitor can buy the same tools. None of them can buy a delivery model built from TriZetto’s own roadmap and the finding its own teams produced.
