Most agreements are trapped in PDFs. While PDFs ensure that a document looks the same on every device, the format is fundamentally ill-suited for the era of AI-driven automation.
The industry has responded with extraction pipelines to pull data from these documents, which infer what a contract says from the way it looks. However, that’s not the same as understanding what a sequence of characters mean, especially when you want to search across millions of agreements, integrate agreement data into other systems, or automate contract revisions or renewals.
The Accord Project takes a different, but complementary approach. Its three-component template architecture keeps the contract computable from the start, and adds legal guardrails. A single template produces both the human-readable document (PDF, DOCX, HTML) and the structured data layer that AI agents can work with.
The Three Components
Markdown Text
Accord Project contract text is written in TemplateMark, a minimal extension of CommonMark Markdown. Markdown is the lingua franca of large language models; the vast majority of LLM training data is Markdown or plain text. An agent drafting, reviewing, or summarizing a TemplateMark document is working in a format it already knows well.
TemplateMark adds just two things to standard Markdown: {{variable}} placeholders for data, and {{% expression %}} for computed values. The same template that an agent reads and populates is the one that renders to a PDF for signature.
Concerto Schema
Concerto is the schema language that describes the structure of contract data; the layer that keeps agent output honest.
When an agent populates template variables, the Concerto schema enforces that {{rate}} receives a Double and not the string "January". Invalid values are rejected at validation, before execution. It’s the same principle as JSON Schema for an API: type constraints catch hallucinated values structurally, rather than relying on prompt instruction alone.
TypeScript Logic
Contract logic in Accord Project is written in TypeScript. Agents can generate TypeScript for contract calculations and the compiler will catch type errors before execution. Concerto auto-generates TypeScript interfaces from the schema, so the types the agent codes against come directly from the contract’s data model.
The compiler sits between agent output and deployment. Errors in agent-generated logic surface as compile failures — not silent bugs in payment calculations or clause conditions.
Two Paths to Intelligent Automation
Organizations can adopt this stack through two primary strategic paths:
1. The Legacy Bridge: Standardizing Extraction
For existing documents where metadata is unavailable, Accord Project templates serve as the standardized target for AI extraction.
The Accord Project whitepaper identifies structured, computable formats as the missing foundation for reliable LLM-based contract workflows – what it calls the “contract gap” between a legal document and an automated system. We expect LLMs to continue to improve the quality of extractions, nevertheless, knowing that the value must be a decimal is legally superior to guessing that it should be. Enterprise software vendors and research teams have demonstrated this in practice:
- A joint team from National Taiwan Normal University, UCL, and HSBC Business School showed automated conversion of natural language contracts to Accord Project templates using NLP,
- Northwestern University and Adobe Research published a pipeline using AI-extracted Obligation Logic Graphs mapped to Accord Project templates, presented at ACM ICAIL.
- Docusign’s Iris AI transforms unstructured agreements into data-rich, searchable assets by mapping extractions to Concerto models, providing the foundational scale needed to transition legacy archives into structured implementations.
2. The Native Future: Deterministic Agreements
While extraction is a necessity for the past, the Native Future belongs to deterministic agreements. By using templates as the source of truth, semantic information is recovered with 100% certainty rather than inferred through probabilistic AI guesses. A single template produces both the human-readable document (PDF or DOCX) and the structured data layer simultaneously. This ensures that semantic information is recovered deterministically rather than inferred through probability.
Interoperability and Standardization
Standardization is the key to scaling AI agents across the enterprise and between organizations. By adopting a unified protocol, you ensure that a contract generated in one department is immediately actionable by another’s automated systems.
To facilitate this, the Accord Project Agreement Protocol (APAP) exposes these templates over REST. It also includes a Model Context Protocol (MCP) endpoint, allowing MCP-compatible AI clients to integrate contract data directly into their workflows.
Close the Gap Today
The transition from static text to executable logic is no longer theoretical.
-
Explore: Use the Template Playground to modify data models and watch them render to documents in real-time with no installation required.
-
Integrate: Follow the AI & Agent Workflows guide to bridge the gap between your LLM outputs and verifiable legal structures.
-
Standardize: Build your next agreement on the stack that AI agents already understand.