The Backend System for In-House Legal Work.
“A living source of truth that represents your legal standards, connected to the systems and AI you already work with.”
A growing business negotiates faster than its institutional memory can keep up. A familiar clause returns in a new contract. Sales asks which fallback it can offer. A lawyer searches an earlier deal to remember why an exception was approved.
The pressure becomes visible when commercial agreements arrive every week, outside-counsel bills keep returning for familiar markups, or a fundraise, acquisition or restructuring exposes gaps in the record. The problem is the same: useful legal knowledge has not become a repeatable way of working.
The work is done, but much of its value stays in the inbox. The next request still needs someone to reconstruct the context. Outside counsel can become part of that repeated review when advice has not been turned into a reusable position.
We think the next step for in-house legal is to make its knowledge usable: standards, business facts and approval limits that people and tools can work from. Legal work should build on what the business has already learned.
Put your standards
in charge.
Language models help interpret a request and turn the result into useful wording. Your approved rules set the boundaries: which terms are acceptable, which fallbacks are available, and who can approve a deviation.
Keeping those jobs separate makes the decision easier to check. The same checked facts, under the same rule version, lead to the same rule result. Missing facts, conflicting wording and questions outside the playbook need review; a confident answer is not a substitute for a supported one.
Your team retains authority. A repeated exception may suggest a better standard, but it does not change the standard. You approve the next version before it is used.
Your team gets a repeatable process: test a decision before rollout, trace it to the approved rule, and update it when your standards change.
Accurate input equals accurate output.
See how each approach handles the work: getting the answer right, avoiding needless review and keeping approval boundaries intact.
| SETUP | ACCURACY | UNNECESSARY REVIEW (FALSE POSITIVES) | MISSED SAFEGUARDS (FALSE NEGATIVES) |
|---|---|---|---|
| Represent | 96%correct on every repeat96% correct answers · 100% repeatable | 4.2%3 / 72 answers | 0%0 / 72 answers |
| Code & DB | 92%correct on every repeat97% correct answers · 92% repeatable | 1.4%1 / 72 answers | 1.4%1 / 72 answers |
| RAG, DB, & LLM | 58%correct on every repeat72% correct answers · 79% repeatable | 19.4%14 / 72 answers | 0%0 / 72 answers |
| Skills, Plugins, & RAG | 54%correct on every repeat63% correct answers · 67% repeatable | 29.2%21 / 72 answers | 0%0 / 72 answers |
| Prompt Layers | 4%correct on every repeat4% correct answers · 92% repeatable | 79.2%57 / 72 answers | 4.2%3 / 72 answers |
| Prompt | 4%correct on every repeat4% correct answers · 83% repeatable | 91.7%66 / 72 answers | 4.2%3 / 72 answers |
HOW TO READ THE RESULTS
Setup Details
- Represent
- Playbook and business facts, with executable decision checks.
- Code & DB
- Connected playbook and business records. AI interprets the rules.
- RAG, DB, & LLM
- Markdown playbook plus emails, documents and CRM context. AI decides.
- Skills, Plugins, & RAG
- Review instructions and a prose playbook; no business records.
- Prompt Layers
- Expert instructions; no company playbook or business records.
- Prompt
- A clause and a question; no company playbook or business records.
Test Structure
24 cases × 3 runs = 72 answers per setup. The same fictional or composite cases are checked against an answer key; AI judges assess key issues. Answer and error rates use all answers; consistency measures use cases.
Legend Details
- Correct answers
- Right decision, approver, required suggestions and key issue.
- Correct on every repeat
- Correct on all three runs of a case. The main accuracy figure.
- Repeatability
- Identical decision, approver and suggestions across runs, right or wrong.
- Unnecessary review · false positives
- An unnecessary referral to counsel or a senior approver.
- Missed safeguards · false negatives
- An unsafe concession, missed walk-away or skipped or incorrect approval.
Keep more than the signed contract.
The record should explain how the business got there: the clause, the rule applied, the exception, the approval and the promise made to close the deal.
Linked records make it possible to look across negotiations instead of relying on recollection. They can reveal recurring pushback, repeated escalations and commitments that need attention.
They also give an incoming lawyer, an auditor or a due-diligence adviser a clearer starting point, with sources and evidence gaps visible.
This is the lasting value of the work: decisions that leave the business with more usable knowledge than it had before.
Built around the team you have.
Extend an established legal team, or build the foundation your business is missing.
Optimisation & maintenance.
Your team sets the standards. We turn them into connected workflows and keep them current as the business changes.
- Approved positions turned into rules
- Workflows connected to existing tools
- Testing, optimisation and ongoing care
Legal support & engineering.
Hands-on support fills the gaps, alongside a system that takes recurring work off your team as you grow.
- Missing contracts and policies put in place
- One-to-one commercial legal support
- Specialist lawyers when you need them