Enterprise-Scale AI Agent Deployment in Insurance

wakam · Insurance · 2025 · source Agentic: yes · promotional bias: high · 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: Insurance enterprise AI agents

Wakam needed to overcome knowledge silos across insurance operations, legal, data, customer service, and business teams while maintaining data governance and driving broad employee adoption beyond a technical prototype.

Constraints:

Knowledge lived across Notion, SharePoint, Slack, Excel, Snowflake, HubSpot, contracts, regulatory documents, and other systems.
A five-person data science team could build a prototype but not sustain platform velocity alone.
Sensitive insurance, legal, financial, and regulatory data required permission controls.
Business users needed to build and use agents without relying on the AI Engineering team for every use case.
Autonomous agents in insurance must stay within validated and auditable boundaries.

Architecture

Wakam pivoted from a custom in-house RAG chatbot to Dust's AI agent platform. The implementation used platform connectors to internal knowledge sources, dual-layer permissions for agent-to-data and human-to-agent access, executive sponsorship, weekly success-story reinforcement, employee training, Slack support, build days, a meta-agent for agent creation, adoption dashboards, and a roadmap from knowledge assistants to action agents and eventually bounded autonomous agents.

Diagram confidence: reconstructed (reconstructed from prose).

flowchart LR
  enterprise-knowledge-sources["Enterprise knowledge sources"] --> dust-spaces["Dust spaces"]
  dust-spaces --> dual-layer-permissions["Dual-layer permissions"]
  dual-layer-permissions --> ai-agents["AI agents"]
  ai-agents --> employee-agent-builders["Employee agent builders"]
  employee-agent-builders --> ai-engineering-support["AI Engineering support"]
  ai-agents --> adoption-dashboards["Adoption dashboards"]
  ai-engineering-support --> adoption-dashboards["Adoption dashboards"]

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

Components

Component	Role	Interacts with
`Enterprise knowledge sources`	Notion, SharePoint, Slack, Snowflake, HubSpot, contracts, procedures, and regulatory documents.	`Dust spaces`
`Dust spaces`	Permissioned data containers that control what agents may retrieve.	`AI agents`, `Dual-layer permissions`
`Dual-layer permissions`	Controls both agent-to-data access and human-to-agent access.	`AI agents`
`AI agents`	Knowledge assistants, legal agents, productivity agents, and emerging action agents.	`Employee agent builders`, `Adoption dashboards`
`Employee agent builders`	Business users trained to create agents for their own roles and workflows.	`AI Engineering support`
`AI Engineering support`	Provides training, Slack support, build days, documentation, and help for complex integrations.	`Adoption dashboards`
`Adoption dashboards`	Tracks activity rates, valuable agents, and productivity impact.	—

Data flow

Enterprise data sources are organized into Dust spaces with access controls; employees interact only with agents whose required spaces they can access; teams build knowledge assistants or action agents over Notion, Outlook, SharePoint, Slack, HubSpot, web search, and other tools; adoption and productivity are tracked through dashboards and support programs.

“Agents only retrieve information from their assigned spaces, and users can only interact with agents if they have access to all the spaces those agents require.”article, stated

“Harvey (Legal Agent) operates across the corporate legal team's entire digital workspace. With access to Notion, Outlook, web search, SharePoint, and calendar tools”article, stated

Why this architecture

Wakam chose a platform-and-change-management architecture because the in-house chatbot was technically possible but did not solve adoption, platform maintenance, permissions, and employee self-service at enterprise scale.

“maintaining and extending it required constant engineering effort.”article, stated

“successful AI implementation is fundamentally a change management challenge, not only a technology challenge.”article, stated

Models used: Dust · GPT-3.5 · RAG · AI agents · Notion · SharePoint · Slack · Snowflake · HubSpot

Scale

Wakam reached 70% employee adoption and 136 deployed agents within two months.
“70% employee adoption and 136 deployed AI agents within 2 months of launch”article, stated
Wakam employs 250 people across five European countries.
“employing 250 people across 5 European countries”article, stated
Wakam reaches customers through 100 distributor partners across 30+ countries.
“through 100 distributor partners across 30+ countries.”article, stated
The legal team cut contract analysis time by 50%.
“the legal team cut contract analysis time by 50%”article, stated
Of 136 agents, around 40 were AI Engineering-built and 96 were employee-created.
“approximately 40 were built by the AI Engineering team while 96 were created by employees”article, stated

Alternatives considered (roads not taken)

Build and maintain a custom chatbot in-house.

Rejected because: The five-person data science team could prototype it, but maintenance and adoption were limiting.

“their five-person data science team began building a custom AI chatbot with RAG capabilities”article, stated

Let technical teams build agents for every business problem.

Rejected because: Wakam found employees understood their own business challenges better and should build agents themselves.

“The AI Engineering team couldn't possibly understand every business challenge well enough to build optimal solutions.”article, stated

Make AI usage optional and peripheral.

Rejected because: Leadership framed agent use as strategic and the preferred method for information retrieval and task automation.

“positioned AI agents as the preferred method for information retrieval and task automation, not as an additional option”article, stated

Patterns

`rag-retrieval` (retrieval)

Wakam's initial and platform approach used company data sources to answer employee questions.

“answering questions using company data and browsing the internet”article, stated

`privacy-preserving-processing` (governance)

Dust spaces enforce both data access and user access for agents handling sensitive enterprise information.

“controlling both agent-to-data and human-to-agent access, enabled Wakam to democratize AI while maintaining strict data governance.”article, stated

`tool-use-orchestration` (orchestration)

Action agents use workplace tools such as Outlook, Slack, Notion, HubSpot, SharePoint, calendar, and web search.

“MoneyPenny retrieves emails, prepares meetings, synthesizes weekly activity, writes to Notion pages, and summarizes Slack mentions.”article, stated

`observability-tracing` (operations)

Wakam tracked adoption, valuable agents, and productivity impact with dashboards.

“Wakam built internal dashboards to track adoption metrics, they monitored user activity rates by team”article, stated

Tradeoffs

Buy a purpose-built agent platform.

Chose Dust for platform capabilities and business adoption. over Maintaining a custom RAG chatbot from scratch..

Rationale: The platform let Wakam focus on business impact instead of matching dedicated AI-platform velocity.

“Purpose-built solutions let companies bypass years of trial and error and focus on driving business impact”article, stated

Empower employees to build agents.

Chose Training and self-service agent creation. over Centralized AI Engineering delivery for every use case..

Rationale: Employees were closest to their business needs and created most agents.

“the most critical aspect of their approach was empowering every employee to build their own agents.”article, stated

Progress from assistants to action agents cautiously.

Chose Human-initiated agents first, then bounded autonomous agents. over Immediate autonomous job-function automation..

Rationale: Regulated insurance workflows require risk assessment and organizational acceptance.

“This evolution from human-initiated to agent-initiated workflows requires careful change management and risk assessment.”article, stated

Gotchas (what bit them)

A technically working chatbot can still fail organizationally.

Root cause: Without change management and training, the custom chatbot stayed mostly with technical users.

Resolution: Use executive sponsorship, training, support, and employee-owned agent creation.

“it remained a tool used primarily by technical team members.”article, stated

Permissioning must cover both data and the agents that use it.

Root cause: Giving agents sensitive knowledge access can expose data to the wrong humans.

Resolution: Use dual-layer permissions for agent-to-data and human-to-agent access.

“how do you give agents access to information they need while ensuring sensitive data doesn't end up in the wrong hands?”article, stated

Autonomous agents in insurance require auditable limits.

Root cause: Regulated workflows can create legal, risk, and security exposure if agents act beyond validated boundaries.

Resolution: Subject autonomous capabilities to legal, risk, and security assessments with periodic reviews.

“every autonomous capability is subjected to legal, risk, and security assessments before deployment, with periodic reviews.”article, stated

Evaluation

Wakam evaluated the rollout through monthly/weekly adoption targets, user activity by team, valuable-agent identification, productivity impact, legal-analysis time reduction, self-service intelligence impact, and readiness assessments for autonomous capabilities.

Metrics: 70% monthly adoption · 90% monthly usage ambition · 70% weekly usage ambition · 136 deployed agents · 50% legal contract analysis time reduction · Agent usage by team

“their ambition is to reach 90% monthly usage and 70% weekly usage.”article, stated

“Wakam built internal dashboards to track adoption metrics, they monitored user activity rates by team, identified the most valuable agents, and tracked productivity impact across different use cases.”article, stated

Outcomes

70% employee adoption and 136 deployed AI agents within two months.
Legal contract analysis time reduced by 50%.
96 of 136 deployed agents were created by employees across business units.
Wakam moved from knowledge assistants toward action agents such as Harvey and MoneyPenny.

Wakam is a 250-person B2B2C insurance company across five countries, serving customers via 100 distributor partners in 30+ countries.

“Dramatic decrease in data analysis time through self-service intelligence”article, stated

“Wakam is now building agents capable of taking actions autonomously vs just providing information.”article, stated