Skip to content
Koderead

Commercial law firm

AI Precedent Search

We keep paying twice for analysis we have already done.

A professional practice reviewing a paper file

A commercial practice whose accumulated expertise sits in several thousand matter files that are, in any practical sense, unsearchable. A fee earner asks a colleague whether the firm has handled something similar; the colleague half-remembers a matter from years ago; finding it means trawling the document system by client name. Frequently the work is simply redone.

The product

AI Precedent Search

The queue, the approval screen, the exception list. Every number is readable and every handover is visible.

Built by Koderead to show the operating system. Figures are a worked example.

Watch one record

Where the handover actually happens.

A screenshot shows a state. This follows a single record through the system, including the point where it stops and waits for a person.

One question, start to finishWorked example
  1. Fee earner asks a question11:04

    Previously this meant interrupting a senior colleague.

  2. Authorised corpus searched11:04

    Only what the firm decided to index, never a shared drive.

  3. One result withheld11:04

    A barrier applies. Its existence is not disclosed either.

  4. Source read, judgement applied11:05

    The passage and its date, not a summary to trust blindly.

A person is doing this

Question
Change-of-control on reorganisation
Practice area
Corporate
Usual route
Ask someone who might recall
Usual outcome
Often redone from scratch

The system retrieves and cites. It never advises, never summarises into a conclusion, and never drafts unsupervised

Deeper into the build

The screens where a person is still in the loop.

A queue is easy to show. What matters is what happens at the record, at the approval, and at the exception — because that is where the design either holds or does not.

Access check — restricted matterBarrier enforced
Query
Change-of-control on reorganisation
Results in corpus
4 matters
Returned to this user
3 matters
Withheld
1 matter
Reason
Information barrier
Enforced at
Data layer, not prompt

System recommendation

One matter is withheld because this fee earner sits the wrong side of an information barrier. The existence of the matter is not disclosed, the count is not inflated, and no passage from it appears in the answer.

  • Barrier enforced at data layer
  • No leakage into the summary
  • Access logged
ApproveEditReject

Conflicts checking remains a supervised human process. The system never advises.

Twenty-three barrier conflicts blocked this month. Enforcement lives in the data layer, never in an instruction the model could be argued out of.

Reuse and gapsMonthly
Questions asked
894
Answered from precedent
617

69%

Median time to source
14s
  • No precedent in corpus187Logged as a gap — the firm has not addressed this
  • Precedent exists but predates a rule change54Flagged with the date for a human to weigh
  • Withheld by information barrier23Not disclosed, not counted in results
  • Ambiguous question13Clarification requested

277 items awaiting a person, grouped by cause so one decision clears a class rather than a case.

The gap list is the second product. It tells the firm precisely which questions it has never answered in writing, ranked by how often someone needed one.

The change

Where the work moved.

Amber marks a step the system performs. Solid ink marks one a person still performs. The counts above each column say the same thing in numbers.

Today

6 human steps
  1. Fee earner has a question

    Performed by A person
  2. Asks a senior colleague

    Interrupting them

    Performed by A person
  3. Colleague half-remembers a matter

    Performed by A person
  4. Document system trawled

    By client name

    Performed by A person
  5. Often not found

    Performed by A person
  6. Analysis redone from scratch

    Performed by A person

After

3 human steps
  1. Fee earner asks the question

    Performed by A person
  2. Authorised corpus searched

    Barriers enforced at data layer

    Performed by Automated
  3. Source passages returned

    With matter and date

    Performed by Automated
  4. Reads the source, not a summary

    Performed by A person
  5. Applies professional judgement

    Performed by A person
  6. Gap logged where nothing found

    Performed by System
  • Automated
  • A person
  • System
  • Customer

Scope of authority

What this system is not allowed to do.

On a real engagement this is a document your operations lead signs before anyone writes code.

  • The system retrieves and cites. It never advises, never summarises into a conclusion, and never drafts unsupervised
  • Privilege and information barriers are enforced at the data layer, including between matters in the same firm
  • Conflicts checking remains a supervised human process
  • Where the corpus does not cover a question, the system says so rather than composing something plausible
  • Every output carries its source so it can be verified before it leaves the firm

Built from

The solutions behind this build.

Your version of this

This took a mapped workflow before it took any code.

If this looks like your commercial law firm, the place to start is the same place we started here.