Deploying Generative AI at Scale Across 5,000 Developers

liberty_it · Insurance · 2026 · source Agentic: no · promotional bias: low · source-backed · extracted with prompt 0.5.6 (schema v2)

Generated from extraction.json — do not edit by hand. Every claim cites verbatim source evidence; stated means the quote directly asserts the claim, implied means it supports it indirectly.

Problem

Domain: Enterprise generative-AI adoption

Liberty IT needed to deploy generative AI across a large regulated enterprise without confusing rollout with adoption, losing control of cost, or trapping teams behind a single model or tool.

Constraints:

The parent company spans legacy COBOL/C++ systems and a large AWS footprint.
A workforce-scale deployment had to serve all employees and thousands of developers.
Model availability, regional rules, data sovereignty, egress, and internal protections complicated rollout.
Different models have different costs and capabilities.
Adoption requires fluency, training, and measurement beyond switching tools on.

Architecture

Liberty IT rolled out Liberty GPT as an internal ChatGPT-like tool, GitHub Copilot for developers, role-specific GenAI tools, a centralized platform with shared rails and standardized APIs, a supported catalog of 28 models, dashboards, cost controls, alerting, and observability. The program paired platform governance with autonomy, education, adoption tracking, and cost-aware model selection rather than assuming deployment equals usage.

Diagram confidence: reconstructed (reconstructed from prose).

flowchart LR
  employees-and-developers["Employees and developers"] --> liberty-gpt["Liberty GPT"]
  employees-and-developers --> github-copilot["GitHub Copilot"]
  employees-and-developers --> role-specific-genai-tools["Role-specific GenAI tools"]
  liberty-gpt --> central-ai-platform["Central AI platform"]
  github-copilot --> central-ai-platform["Central AI platform"]
  role-specific-genai-tools --> central-ai-platform["Central AI platform"]
  central-ai-platform --> model-catalog["Model catalog"]
  central-ai-platform --> dashboards-and-cost-controls["Dashboards and cost controls"]
  learning-and-adoption-program["Learning and adoption program"] --> employees-and-developers

Drag to pan · scroll or use +/− to zoom · reconstructed from the source's described component topology

Components

Component	Role	Interacts with
`Employees and developers`	Enterprise users of Liberty GPT, Copilot, and role-specific GenAI tools.	`Liberty GPT`, `GitHub Copilot`, `Role-specific GenAI tools`
`Liberty GPT`	Internal ChatGPT-like assistant rolled out broadly across Liberty Mutual.	`Central AI platform`
`GitHub Copilot`	Developer coding-assistance tool rolled out to Liberty IT.	`Central AI platform`
`Role-specific GenAI tools`	Additional tools for specialized employee or developer use cases.	`Central AI platform`
`Central AI platform`	Shared rails, standardized APIs, governance, and autonomy layer.	`Model catalog`, `Dashboards and cost controls`
`Model catalog`	Supported set of models optimized by task, performance, and cost.	`Central AI platform`
`Dashboards and cost controls`	Usage, cost tracking, alerting, notifications, and observability.	`Central AI platform`
`Learning and adoption program`	Training and support for effective use rather than nominal access.	`Employees and developers`

Data flow

Employees and developers use Liberty GPT, GitHub Copilot, and role-specific tools through shared enterprise rails. The central platform exposes standardized APIs and supported models, while dashboards, cost-tracking middleware, alerting, and observability monitor usage and spend. Adoption telemetry and learning programs help teams move from access to effective use.

“we had to come up with a centralized platform. Uh we had to come up with our own standardized APIs. Uh and that involved dashboards and cost controls”transcript, stated

“provide dashboards automate alerting and notifications. uh observability matters as much here as anywhere else. In fact, probably more.”transcript, stated

Why this architecture

Liberty IT chose a centralized platform with model choice, dashboards, cost controls, and adoption support because enterprise GenAI deployment is constrained by cost, model access, internal protections, regional concerns, and uneven fluency. The design keeps teams autonomous while making cost, usage, and model selection visible.

“supporting 28 different models uh in house for uh different use cases and we're we're optimizing against performance and cost.”transcript, stated

“shipping things isn't adoption. It's really hard to get the adoption past about 70%. And just because you've given it to somebody doesn't mean they're they're using it effectively.”transcript, stated

Models used: Liberty GPT · GitHub Copilot · 28 supported in-house models · GPT-4.1 · GPT-5 · Claude Opus 4.1

Scale

Liberty Mutual is described as a Fortune 100 company with roughly $50 billion in revenue.
“they're a Fortune 100 company with like 50 billion in revenue.”transcript, stated
The talk frames the rollout as leading 5,000 developers into GenAI-era software engineering.
“we're here to talk about leading 5,000 developers into the next era of software engineering.”transcript, stated
Liberty GPT reached 70% of Liberty Mutual employees.
“We got Liberty GPT ruled out to 70% of Liberty Mutual.”transcript, stated
GitHub Copilot adoption in Liberty IT moved from 60% to 90% in the first year.
“we got GitHub co-pilot to 60% adoption, 90% adoption after the first year”transcript, stated
The platform supported 28 different models in house.
“we're currently supporting 28 different models uh in house for uh different use cases”transcript, stated

Alternatives considered (roads not taken)

Deploy the latest frontier model for every use case

Rejected because: Model costs and capabilities differ, so the platform optimized model choice against performance and cost.

“we don't just say automatically use the latest and greatest or to use the technical term frontier model.”transcript, stated

Measure rollout as adoption

Rejected because: The team found that access does not mean effective or fluent use.

“just because you've shipped it or you've got it out there doesn't mean it's actually being used effectively.”transcript, stated

Let each team absorb GenAI cost independently

Rejected because: A bad model choice can create large cost exposure, so dashboards, middleware, alerting, and education are needed.

“one bad choice uh, can end up costing you an awful lot.”transcript, stated

Patterns

`model-routing` (model-ops)

Liberty supported many in-house models and selected among them by use case, performance, and cost rather than defaulting to the newest model.

“supporting 28 different models uh in house for uh different use cases and we're we're optimizing against performance and cost.”transcript, stated

`cost-aware-routing` (operations)

The platform used dashboards, cost tracking, alerting, and model-cost education to prevent excessive spend.

“build your own dashboards, build cost tracking middleware and so on.”transcript, stated

`observability-tracing` (operations)

Enterprise GenAI usage needed dashboards, alerting, notifications, and observability because cost and adoption failure modes were operational.

“provide dashboards automate alerting and notifications. uh observability matters as much here as anywhere else. In fact, probably more.”transcript, stated

`safety-policy-enforcement` (governance)

A centralized platform and shared rails imposed governance while still preserving engineer autonomy.

“Gen AI uh in a company like Liberty Mutual uh needs shared rails. you know, we had to come up with a centralized platform.”transcript, stated

Tradeoffs

Balance centralized platform governance with engineer autonomy

Chose Shared rails, standardized APIs, dashboards, and cost controls over A fully decentralized tool rollout.

Rationale: Enterprise GenAI needed governance and visibility without eliminating local experimentation.

“we had to come up with a centralized platform. Uh we had to come up with our own standardized APIs. Uh and that involved dashboards and cost controls”transcript, stated

Support many models

Chose A 28-model in-house catalog over Automatic use of latest frontier models.

Rationale: Different models suit different tasks and have very different cost profiles.

“models are not equal. Yeah. Uh different models are good at different things and my goodness they cost the costs are different.”transcript, stated

Invest in adoption work after rollout

Chose Measure real usage and fluency over Assuming enabled accounts equal impact.

Rationale: The team found that rollout percentages hid whether people were using GenAI effectively.

“just because you've given it to somebody doesn't mean they're they're using it effectively.”transcript, stated

Gotchas (what bit them)

Shipping tools is not the same as effective adoption.

Root cause: Users may have access but lack fluency, training, or real workflow integration.

Resolution: Measure real usage and support adoption beyond the rollout percentage.

“just because you got it out there uh and it's being used effectively doesn't mean it's being used fluently.”transcript, stated

The last part of adoption is hardest.

Root cause: Localization, data sovereignty, egress, and internal protections create barriers after the easy users are onboarded.

Resolution: Treat adoption as asymptotic and keep solving the regional/control-plane blockers.

“Yeah. So adoption can be asympto. >> Yes we got there [laughter] six rehearsals. Yeah. Uh the closer to the the you know the end result you get the harder it becomes”transcript, stated

Model access and roadmaps can lag enterprise demand.

Root cause: Providers had quota and availability limits during fast GenAI growth.

Resolution: Support a plurality of models and avoid single-provider/single-model dependence.

“getting access to the models was through different requests and forms and whatnot and you know a lot of that was due to the the providers themselves not having enough quota.”transcript, stated

Higher model capability can create uncontrolled cost exposure.

Root cause: More powerful models and high-volume processing can incur large token spend.

Resolution: Use cost dashboards, middleware, alerts, and model-selection education.

“increasing performance increases the ability of your system to incur cost.”transcript, stated

Evaluation

The source emphasizes adoption telemetry, usage measurement, dashboards, cost tracking, alerting, and education rather than model-quality benchmark scores.

Metrics: Liberty GPT rollout percentage · GitHub Copilot adoption percentage · Model cost by provider/model · Usage fluency beyond switched-on access · Dashboard/alert visibility into cost and usage

3 supporting quotes

“We got Liberty GPT ruled out to 70% of Liberty Mutual.”transcript, stated

“we got GitHub co-pilot to 60% adoption, 90% adoption after the first year”transcript, stated

“you want to measure the real usage and not just the fact that you've switched it on.”transcript, stated

Outcomes

Rolled Liberty GPT out to 70% of Liberty Mutual.
Reached 90% GitHub Copilot adoption in Liberty IT after the first year.
Standardized around shared APIs, dashboards, cost controls, and a catalog of 28 models.
Identified adoption fluency, cost management, and model selection as ongoing operational work.

The deployment operated inside a Fortune 100 insurer with roughly $50 billion revenue, 5,000 developers, 900 technical staff in three Irish locations, and a 28-model in-house catalog.