AI contract intelligence case study
Building an AI contract intelligence platform: from obligation extraction to contract creation.
How Changeable designed and built a governed AI contract intelligence platform in house, growing it from a single obligation extraction use case into a multi-product system covering contract review, contract reform and contract creation for New Zealand businesses.
Project overview
A contract sitting in a shared drive is not a managed obligation.
Most organisations sign contracts, then lose track of what they actually agreed to. Renewal dates slip past, indemnities go unnoticed, and obligations sit buried in clauses nobody rereads after signing. Changeable built this AI contract intelligence platform to turn that unstructured risk into a governed, structured obligation register that a business can actually work from.
The build started as a single capability: structured extraction of obligations, dates, risks and financial terms from an uploaded contract. It grew into a small product family covering contract summary, deeper contract intelligence review, reform of existing agreements, and guided creation of new ones. The same document intelligence methods now inform Changeable’s client builds.
Governed extraction
Every obligation, date and risk flag traces back to its source clause, with human review built into delivery.
Tested against complexity
The extraction schema was stress-tested against a dense, multi-clause commercial agreement before client use.
Product family, not a single tool
Grew from one extraction use case into five tiers covering summary, review, reform and creation.
Continuous lessons capture
A running lessons log turns every extraction pass into a refinement of the underlying schema.
The problem
Contracts are read once, then forgotten.
Most small and mid-sized organisations do not have a contract management system. Agreements live in inboxes, shared drives and filing cabinets. Nobody owns the obligation register, and nobody is watching the calendar for notice periods, price reviews or renewal deadlines until something is missed.
The commercial risk is not usually a single catastrophic clause. It is the slow accumulation of missed dates, unclear rights, and terms nobody has looked at since the day of signing. That is a process problem as much as a technology one.
What a business typically cannot answer
- What are we actually obligated to do under this agreement?
- When does this contract renew, and what is the notice period?
- Which clauses expose us to financial or compliance risk?
- Are our commercial terms in line with market norms?
- What has quietly changed since the last variation was signed?
Research and testing
The extraction schema was proven before it was trusted with client work.
Before the platform was used on a real client agreement, the AI contract extraction approach was stress-tested against a deliberately complex, dense commercial contract designed to surface edge cases: long-stop dates, liquidated damages, uncapped indemnities, aggregate liability caps and non-renewal notice windows.
That single test pass surfaced a large number of refinements to the extraction schema and workflow, each logged, reviewed and applied before the next iteration. It is a disciplined lessons process rather than a one-off build: every extraction is a chance to sharpen the underlying method.
What the testing process covered
- Obligations, prohibitions, rights and warranties extracted into a structured register
- Key dates, recurring obligations and tracked timers identified separately from narrative text
- Financial terms, remedy chains and risk flags surfaced for human review
- Anomalies and gaps against normal market drafting practice flagged, not silently corrected
- Every finding logged against a growing rules file so future extractions inherit the lesson
How it was built
A phased build, from single extraction to a governed product family
Rather than build a full platform upfront, the contract intelligence platform was developed in deliberate stages, each one tested and proven before the next was layered on.
Start with the decision, not the tool
The starting point was not “build a contract AI tool.” It was a defined business use case: help a business understand what it is actually obligated to do under a signed agreement, without paying for a full legal review every time a question comes up. That framing shaped every decision that followed, including which extraction fields mattered and where a human needed to stay in the loop.
Prove the method against complexity before scale
A structured extraction schema was built to convert unstructured contract text into a register of obligations, rights, prohibitions, warranties, key dates, financial terms and risk flags. Before this went anywhere near a client contract, it was tested against a dense, purpose-built commercial agreement designed to expose weaknesses in the schema. The test surfaced a substantial set of lessons, each applied to the extraction rules and instructions before the next pass. This is the same discipline Changeable applies in AI data model engagements more broadly: prove the model against difficult, realistic data before it goes near production use.
Make the workflow reliable, not just the extraction
A structured extraction is only useful if it reaches the client as a clear, usable deliverable. Changeable built a controlled automated intake workflow so a client can purchase a tier, receive a private upload link tied to their order, and submit a contract securely. Files are never handled as open email attachments or stored in a public location; delivery is admin-controlled and every submission is logged. The extraction output then feeds a defined build step that produces the client-facing deliverables for that tier, whether a plain-English summary, a full obligation register, or a governance-ready contract intelligence report.
Apply the same discipline to drafting new agreements
Once the extraction and delivery pipeline was proven, the same underlying capability was extended into two further product lines: reforming an existing agreement that no longer fits how the business operates, and guiding a client through creating a new contract from a structured brief. For contract creation, a client answers a series of structured questions about the parties, purpose, key terms and risk areas. An AI assistant helps them turn rough notes into clearer answers for each field, with the client always able to edit or write the answer themselves. The assistant is scoped tightly: it improves clarity of the client’s own input and does not offer legal advice or invent facts. The structured brief is then used to produce a full draft agreement for professional review.
What the platform does
Five tiers across two product lines
The AI contract intelligence platform covers both sides of the contract lifecycle: understanding what you have signed, and drafting what you sign next.
Contract Summary
A clear, plain-English summary of what a contract says, for a fast first read.
Contract Intelligence
A structured obligation register with key dates, financial terms and flagged risks.
Contract Intelligence Plus
The full governance-ready output, including anomalies, remedy chains and negotiation notes.
Reform Contract
Reworks an existing agreement that no longer matches how the business actually operates.
Create Contract
A structured brief and AI-assisted intake that produces a full draft agreement for review.
Human review, always
Every tier ends with a professional-ready deliverable, not an unreviewed AI output.
Governance
Built with the same controls Changeable recommends to clients.
The platform was designed with the same AI governance principles Changeable applies in client engagements: clear data boundaries, controlled access and a defined human review point before anything reaches a client.
Questions
Questions about AI contract intelligence?
Common questions about how this contract intelligence platform was built, and how the same approach applies to your organisation.
What is AI contract intelligence?
AI contract intelligence uses large language models to read a contract and turn it into structured information: obligations, rights, key dates, financial terms and risk flags. Instead of rereading a whole agreement every time a question comes up, a business works from a register it can search, track and act on, with a person reviewing the output.
Why start with extraction rather than the full product?
Starting with a single, well-defined use case made it possible to prove the method before investing in a wider platform. Extraction was the highest-value, most testable capability, and every other product tier was built on top of a proven extraction foundation.
How was the extraction approach tested before real client use?
The schema was stress-tested against a deliberately complex commercial agreement designed to surface edge cases such as long-stop dates, uncapped indemnities and aggregate liability caps. Findings were logged and applied to the extraction rules before any client contract was processed.
Which contract terms can AI extract reliably?
Clearly drafted obligations, parties, key dates, renewal and notice periods, payment terms and liability caps extract well. Terms that depend on cross-references, defined terms used inconsistently, or obligations implied rather than stated need more care, which is why the schema flags anomalies for human review rather than resolving them silently.
Does AI contract intelligence replace a lawyer?
No. The platform produces a structured, human-reviewed starting point: a clear register of obligations, dates, risks and terms. Every deliverable is designed for professional review, not as a substitute for legal advice.
How does the contract creation product work?
A client answers a structured set of questions about the parties, purpose, key terms and risk areas. An AI assistant can help the client turn rough notes into clearer answers for individual fields, but manual entry always works and the client can edit any AI suggestion before submitting. The completed brief is then used to produce a full draft agreement.
What guardrails are in place for the AI assistant used during intake?
The assistant is scoped to improving the clarity of what the client has already written. It does not offer legal advice, does not invent facts, and does not generate enforceable claims. It fills gaps with clear placeholders rather than guessing.
How is client data protected during the process?
Uploaded contracts are stored in a controlled location rather than a public media library, are never sent as email attachments, and are only accessible through admin-controlled, order-linked delivery.
Why did Changeable build this platform in house?
Building our own products means the advice we give clients comes from delivery experience, not theory. This platform is a working example of applying use case development, AI data modelling and AI governance to a real product, not just advisory work.
Can Changeable build a contract or document intelligence tool for our organisation?
Yes. The same extraction, review and governance approach applies to contracts, leases, policies and other document-heavy work. We start by defining the decision the tool needs to support, then design, build and test it around your documents and systems. See our AI app development service, or book a discovery session to talk it through.
Need contract or document intelligence built for your organisation?
Talk to Changeable about applying the same use case-led, governed approach to your own contracts, documents or workflows.