The Results:
3 Days → Minutes
~3 Months
Contract update time
From development to production launch
Client
The client is a global leader in pharmaceutical contract lifecycle management, whose highly customized, multi-tenant SaaS platform is used by more than 150 customers worldwide and underpins over $1 trillion in managed contract revenues. In this industry the contract is not a document it is the source of truth for downstream billing. Every clause, quantity, and price feeds revenue calculations, so an error introduced during contract entry propagates directly into inaccurate billing and lost revenue.
The Challenge
The platform’s depth was also its constraint. Years of customization made the core powerful but hard to change: the product could safely ship only about two core releases per year, so any new user-interface or workflow capability sat behind a long release train. Meanwhile the day-to-day work of maintaining contracts remained stubbornly manual.
- Highly complex platform, slow to change
UI and feature extensions were bottlenecked behind roughly two core releases per year. - Up to 3 days per contract update
Onboarding or updating a single contract was a multi-day manual effort. - Manual master-data mapping
Mapping fields between incoming contracts and the system’s master data was done by hand. - Hundreds of variations per section
Each contract section could take one of hundreds of structural and language variants. - High-impact errors
Manual mistakes flowed straight into billing, creating direct revenue leakage and rework.
The mandate was clear: extend the platform’s UI and AI capabilities on top of the legacy code adding modern, revenue-relevant functionality without destabilizing a core that customers depend on and that can only be released twice a year.
The Wizr AI Solution: Modern AI Workflows on Top of Legacy SaaS
Rather than modify the core, Wizr built a modern AI and UI layer that runs alongside the existing platform and integrates with it through APIs and MCP-based connectors. This decoupling is the key move: new capability ships continuously on the AI layer, independent of the twice-yearly core release cycle, while the legacy system of record stays untouched and stable.
At the center is an AI Contract Extractor built on multi-agent intelligence. It targets the highest-friction use cases quick AI-assisted contract onboarding, an adaptable sectional user interface for the hundreds of section variations, a distinctive human-in-the-loop review experience, confidence scoring, and CXO dashboards and forecasting delivered on three build layers:
- Application UI. A custom, conversational, multi-modal interface (text, documents, images) purpose-built for contract extraction and review.
- AI agent & workflow builder. Orchestration across agents, LLMs, embeddings, tools, and functions, plus data transformation, feature engineering, and entity resolution.
- Data & integration pipelines. Real-time and batch ingestion, wired to the current SaaS application through MCPs and APIs.
How It Works: Multi-Agent Contract Extraction
The extractor processes each contract section through a coordinated team of specialized agents, keeping a human as the final approver throughout.
- Upload & extract. A user uploads a contract through the new UI; a Document Intelligence step extracts the content into hierarchical, relationship-preserving text (markdown/JSON).
- Identify the template. The system identifies which contract format/template applies based on the content, drawing on a prompt library of template variants.
- Run the per-section agentic workflow. Controlled for general and pricing sections, a Supervising Agent coordinates specialists: an Entity Extractor Agent pulls key fields into structured JSON; a Mapping Agent checks each field against master data and produces a mapped JSON; a Review Agent & Confidence Scorer validates completeness, data types, and mandatory fields and reverse-checks confidence; and an Error Handler Agent re-extracts mismatches and routes exceptions.
- Human-in-the-loop review. Results surface in the UI with color-coded confidence green for high-confidence extractions, yellow for fields that need a second look so reviewers focus their attention where it matters and adjust or correct before finalizing.
- Export to the legacy platform. Approved output is generated as structured XML and exported back into the legacy SaaS, with every action written to audit trails.
Architecture & Technology
The solution is cloud-native and multi-tenant on AWS, built to scale securely and to integrate with the legacy platform without altering its core. Users reach a web application (Amplify/CloudFront) via Amazon API Gateway; microservices run on ECS on Fargate, with AWS Step Functions orchestrating the agents. Reasoning is served by Amazon Bedrock alongside external LLMs, document extraction by a Document Intelligence service, and semantic retrieval by a Pinecone vector database. Data persists across RDS Postgres, DocumentDB, and S3, with Cognito for authentication, EventBridge and SNS for eventing, Secret Manager for credentials, and CloudWatch for observability.
Human-in-the-Loop & Trust
Because contract terms drive financially material billing, the design keeps people in control and every action explainable. Each extraction is confidence-scored and reversible; reviewers remain the decision owners and can correct any field before it is committed; and end-to-end audit trails make every step reconstructible for internal audit and compliance what makes an AI layer safe to run against a system of record in a regulated industry.
Impact
| Dimension | Before | After |
| Single contract update | Up to 3 days of manual work | Minutes of review-and-approve on near-real-time extraction & mapping |
| UI / feature changes | Gated by ~2 core releases per year | Shipped continuously on a decoupled AI/UI layer |
| Master-data mapping | Manual and error prone | Automated, with confidence scoring and exception handling |
| Billing accuracy | High risk of manual errors → revenue leakage | Reduced error and rework exposure; revenue protected |
| Contract onboarding | Slow, specialist heavy | Fast, guided, human-in-the-loop |
| Time to value | — | Production launch in ~3 months |
Directional outcomes reflect the design targets and realized improvements of the engagement; exact figures vary by contract type and customer configuration.
Engagement Model
Wizr delivered through a four-stage model built to move from assessment to production quickly:
Strategise: Assess business and process challenges and objectives, and define a technology roadmap.
Design: Design the multi-tenant cloud architecture, AI agents, data model, UI/UX, and secure, scalable infrastructure.
Engineer: Implement in agile, sprint-based development with a production launch in ~3 months
Innovate: Partner on ongoing AI-driven business-transformation initiatives beyond the initial launch.
Why It Matters
This engagement shows how an established SaaS leader can add revenue-relevant AI capability quickly and safely extending the core, not replacing it. A decoupled, multi-agent AI layer with human-in-the-loop review turned a multi-day, error-prone process into an auditable, confidence-scored one freeing feature delivery from the twice-yearly release cycle and protecting revenue-critical billing, all in a single quarter.
