The problem: promises scattered across paperwork
Microsoft's account says obligations can sit in master service agreements (MSAs), statements of work (SOWs), task orders, milestone schedules, service-level commitments and reporting requirements. Legal, finance, taxation, sales and delivery teams all shape these documents. After signing, delivery managers had to interpret the legal language and operational commitments and then validate them with several stakeholders.
Ramasubramanian S, Birlasoft's Senior Vice President and Delivery Head for BFSI (banking, financial services and insurance), says adding a maker-checker (two-person approval) or a few more reviewers would not have made that process foolproof. Birlasoft wanted more than document automation. The stated goal was to make each commitment visible, tie it to the right stakeholders and keep checking whether it had been fulfilled.
How Contract.ai is built
Birlasoft chose several agents over one because, in Ramasubramanian's words, one agent could not handle every nuance of different contract types. According to the feature, the system has:
- Six specialist agents, coordinated by a planning agent.
- Three jobs for those agents: extracting obligations, validating outputs and flagging inconsistencies.
- 335 metadata elements used to interpret contract context, extract obligations and link each requirement back to its original source.
The Azure services named, and the role each plays, are:
| Azure service | Role described |
|---|---|
| Azure Data Factory | Ingestion |
| Azure Data Lake | Data storage |
| Azure Databricks | Analytics and reporting |
| Azure Key Vault | Access management |
| Azure AI Content Safety | Supporting responsible AI practices |
Ramasubramanian says Azure won on scale, security and predictability of execution. Those are his criteria, and the feature does not describe a head-to-head comparison with other platforms.
His example of the system's value is a link between language on page two of a document and an obligation referenced on page 25. He says a person reading sequentially would struggle to spot that quickly.
Humans stay in the loop
The most useful detail for IT and governance readers is the risk boundary. Ramasubramanian says these obligations carry financial and legal implications, and Birlasoft cannot let the agents "run rogue." The described functions are extraction, validation and flagging. The feature does not say the agents approve contracts, settle legal interpretations or act on commitments themselves.
This matches how the wider contract-AI market describes the pattern. One vendor guide calls AI contract management a layer on top of contract work under attorney supervision. Another says agents escalate material changes while final approval stays with the human reviewer. Those are general industry descriptions, not evidence about Birlasoft's design.
Vimal Kant, a Birlasoft programme manager, adds that the system gives real-time visibility into obligations that are due, overdue and completed. He says this lets teams concentrate on delivery instead of contract interpretation. He works at Birlasoft and speaks within Microsoft's own feature, so this is a second attributed view, not outside verification.
The reported numbers
Microsoft's feature says Contract.ai delivers:
- A realized annual benefit of approximately 4,000 person-hours saved.
- Processing of more than 1,000 SOWs a year.
- An average of more than 10 obligations identified per project.
The article adds that SOW volume, delivery-manager numbers and hours saved are rising as the company grows. It gives no measurement period, baseline or calculation method for any of these figures. I would not turn "approximately" or "more than" into exact totals. I would not derive per-document savings either, because the inputs don't support it.
Ramasubramanian himself ranks productivity second. His main point is confidence that no obligation is being missed, which he links to financial outcomes and Birlasoft's reputation as a delivery organization.
What the story does not say
A few gaps matter if you are weighing a similar build:
- Models and orchestration: the feature names no language models, prompts or agent framework, and no deployment date.
- Accuracy: there are no error, recall or false-positive rates, so "identifies more than 10 obligations per project" says nothing about how many were correct or how many were missed.
- What the metadata elements are: the 335 elements are not defined, and the article does not explain how cross-page links are made.
- Security detail: Key Vault is named for access management and Content Safety for responsible AI. The feature does not claim compliance with any law or standard, and it says nothing about encryption settings, private networking or identity design.
- Review workflow: it does not say how humans approve or correct agent output, who signs off, or how disagreements are logged.
- Cost and users: no cost, user count or independently validated risk reduction is given.
Searches for independent Birlasoft reporting on Contract.ai turned up nothing beyond Microsoft's feature. The wider results were generic contract-AI vendor material, so treat the claims as Birlasoft's account.
Practical takeaways for IT and delivery teams
- Treat it as a pipeline, not a chatbot. The described flow is ingestion, storage, analytics, specialist extraction, validation, then tracking. It uses ordinary Azure data services with agents on top.
- Keep source traceability. Linking each extracted obligation to its original text is what lets a reviewer check an agent's output quickly. That is a design choice worth copying.
- Specialize by document type. Birlasoft's rationale for six agents plus a planner was that contract types differ, a reasonable pattern to test against your own documents.
- Define the human checkpoint first. For anything with legal or financial weight, decide who reviews and approves before you automate extraction.
- Measure accuracy, not just hours. Hours saved are easy to report. Recall on missed obligations is what protects the business, and the story doesn't publish it.
Bottom line
Birlasoft's account is a concrete example of Azure-based agentic AI applied to contract governance. It puts people in charge of consequential decisions and uses agents to find and connect obligations. The reported results are encouraging but self-reported and unaudited. Organizations considering something similar should run their own accuracy tests and treat the 4,000-hour figure as one company's claim.