Services / Legal AI Engineering
Legal AI engineering, from first workflow to daily use
I build AI workflows for law firms and legal teams inside the tools they already use, from contract review in Word to intake and firm knowledge, with human review, checks you can verify, and any AI model you choose.
No legal advice. A person approves every consequential step.
DEMO-1042 · Services agreement
Select a clause to review it
DEMO-1042 · Services agreement. 7. Limitation of liability: Supplier's total liability shall not exceed the fees paid in the prior three months.
- get_playbook(limitation_of_liability): position found, 210ms
- retrieve_precedents(DEMO-1042): 3 sources, 640ms
- check_permissions(user): allowed, 90ms
- write_audit(decision): saved, 120ms
Cap below firm position. Suggested redline: twelve months' fees, per the firm playbook. Sources (3).
Accepted. Lawyer accepted the redline. Model: any.
- 3+ years building legal technology
- Contract review inside Word, with any model
- Microsoft 365 and SharePoint integrations
- Security-tested before launch
01 / Who this is for
For the people who own legal work, not just the AI budget
Law-firm innovation, KM and operations leaders
You know which tasks AI should help with, but you don't have the engineering capacity to connect firm sources, add review steps and keep workflows working as models change.
Legal AI advisers and consultancies
You set the strategy. You need an engineer who can build the workflow, integrations and checks behind your recommendation.
In-house legal and legal ops teams
You need repeatable workflows for contracts, requests and summaries that respect permissions and keep a person accountable.
Building a legal-tech product? See product and systems engineering
02 / Works around your tools
Works around the tools you already have
Many firms already use Microsoft Copilot or a legal AI platform. The gap is usually the work around the tool: the right sources, permissions, review steps, checks, and keeping it working.
- 01 · Configure
When it fits
Your approved tool can already do the job.
What I do
Structure prompts, templates and sources; define review steps; document how attorneys use it.
- 02 · Extend
When it fits
The tool needs your firm's data or systems.
What I do
Connect Microsoft 365, SharePoint, document stores, case data and research sources such as LexisNexis; add checks before output reaches attorneys.
- 03 · Build
When it fits
No tool fits the workflow or your control requirements.
What I do
Custom workflow software or Word add-ins with retrieval, approvals, audit trail and a model-agnostic harness.
Where lawyers work · Legal platforms, models and data
Where lawyers work
WordAdd-ins and redlines
OutlookEmail intake and drafts
TeamsApprovals and alerts
SharePointFirm knowledge- OneDriveMatter files
- ExcelTrackers and schedules
- Microsoft GraphMicrosoft 365 data
- Entra IDSign-in and permissions
- Microsoft CopilotConfigure and extend
Microsoft AzurePrivate hosting- E-signatureSigning steps
Legal platforms, models and data
- LexisNexisResearch sources
- HarveyLegal AI platform
- Case managementMatter data by API
- Court noticesCM/ECF and e-filing
OpenAIGPT models
AnthropicClaude models
GeminiGoogle models
LlamaPrivately hosted models
Mistral AIOpen-weight models
LangChainAgent workflows
PostgreSQLAudit and matter data
RedisQueues and caching
If your current tool's builder is enough, I'll tell you. Not every task needs custom engineering, and some don't need AI.
Product names and logos are trademarks of their owners. Shown to identify integrations; no endorsement implied.
03 / Contract review in Word
Contract review where lawyers already work
A Word add-in that brings your firm's knowledge into the document: playbooks, preferred positions and clause libraries. It reviews a contract against them, and the lawyer accepts, edits or rejects each suggestion. A harness layer sits between Word and the model, so the firm can use OpenAI, Anthropic, Google or a privately hosted model and change models without rebuilding the workflow.
- Firm knowledge in the side panel, not a separate app
- Any model through one harness: permissions, sources and checks stay the same when the model changes
- Every suggestion is the lawyer's call, and decisions are recorded
Delivered for law firms · under NDA · walkthrough available on a call
How firm knowledge is organisedWord add-in
7. Limitation of liability
Liability is capped at the fees paid in the prior three months.
Review against playbook
Harness
Same controls for every model
Firm knowledge
Playbooks, clause library
Permissions
Matter access only
Checks
Agreed test set
Audit
Every decision logged
Checks run on the answer
Model
OpenAI
Answered
Pick a model
Changes each run
Word: lawyer decides
Suggested redline · Sources (3)
Capped at three months' twelve months' fees.
- Accept
- Edit
- Reject
04 / Workflows to start with
Six legal workflows to start with
Each one keeps a person at the checkpoint before anything leaves the firm.
- Delivered · NDA
Contract first-pass review in Word
Lawyers spend first-pass time checking clauses against positions the firm already agreed.
- Contract
- Review against firm playbook
- Lawyer accepts or editsPerson
- Recorded
- Delivered · case study
Client intake to attorney review
Calls and emails arrive unstructured, and staff re-key the same details.
- Call or email
- Structured matter data
- Staff reviewPerson
- Attorney queue
- Delivered · NDA
Answers from firm knowledge
The answer exists somewhere in the firm, but finding it takes longer than redoing it.
- Question
- Permission-aware retrieval
- Reviewer checks sourcesPerson
- Answer with sources
- Delivered · NDA
Matter follow-ups and summaries
Case activity piles up faster than anyone can summarise it or follow up.
- Case activity
- Summary or follow-up draft
- ApprovalPerson
- Sent and logged
- Example scope
Drafting from firm templates
First drafts start from the right template but miss facts nobody noticed were absent.
- Matter facts
- Template draft
- Attorney editsPerson
- Missing-info flags resolved
- Built to your process
Custom workflows around how your firm runs
Your process doesn't match a product's template. The AI layer should fit the way your firm already works, not the other way round.
- Your trigger: matter, email or document
- Firm rules, sources and AI steps
- Named reviewer approvesPerson
- Into your systems, logged
05 / How an engagement runs
How an engagement runs
- 01
Understand
Sit with the people who do the task; map sources, owners and risks.
→ Workflow map
- 02
Scope
Agree deliverables, access, review points and the checks that define good output.
→ Acceptance criteria and test set
- 03
Build
Prompts, retrieval, templates, integrations and review steps.
→ Working workflow in a test environment
- 04
Verify and hand over
Run the checks, fix what fails, document limits and walk the team through it.
→ Check results, documentation, walkthrough
- 05
Maintain
Re-run checks when models, sources or practice needs change.
→ Change log and updated prompts and sources
06 / Quality you can check
Checks agreed before the build, run again on every change
Illustrative checks for a contract-review workflow. Each one is agreed with your team before building and re-run whenever something changes.
- Source support
- Each point traces to a firm source or retrieved passage.
- Playbook position
- Suggestions match the firm's agreed position for that clause type.
- Missing information
- Gaps are flagged, not filled in.
- Defined terms
- Redlines keep the contract's defined terms and cross-references intact.
- Permission boundary
- Users only see sources they're allowed to see.
- Confidential data
- Client details stay out of logs and prompts that leave the firm's boundary.
- Escalation
- Defined cases route to a person.
- Planted instruction
- A hidden instruction inside a contract is ignored.
- Repeatability
- The same contract gives the same review on every run.
When the model or platform changes, the checks run again before attorneys see the change.
How I test AI before launch- Passed
Source support
24/24 points cited
- Passed
Playbook position
11/11 clauses matched
- Passed
Missing info
3 gaps flagged
- Passed
Defined terms
0 broken references
- Passed
Permission boundary
2 restricted sources hidden
- Passed
Confidential data
0 client details in logs
- Passed
Escalation
Indemnity routed to partner
- Blocked
Planted instruction
Hidden text ignored
- Passed
Repeatability
5 runs, same review
Control
- What the AI may do, and where a named person approves
- Which sources it uses and who can access them
- Every action and approval recorded
07 / Evidence
Public proof, and the rest on a call
Case study
From first contact to attorney review
An implemented intake system for a US personal-injury firm. AI prepares the next step; an attorney approves it before anything runs.
3 channels
Voice, email and documents
Every action
Approved before it runs
Full trail
Action-level audit log
Most delivered legal work is under NDA, so firm names aren't shown. I can walk through architecture, prototypes and decisions on a call.
Questions
Before you book a call
We already use Copilot or Harvey. Why would we need an engineer?
To connect them to your sources and systems, add review and checks, and keep them working. If the tool's own builder is enough, I'll tell you.
Can we use our own AI model?
Yes. The harness layer works with OpenAI, Anthropic, Google or a privately hosted model, and the permissions, sources and checks stay the same when you switch.
Does the Word add-in send our documents to an AI provider?
That depends on the model and hosting you choose, and it's agreed before build. A privately hosted model keeps processing in your environment. See AI and data governance.
How do you decide whether a task should use AI at all?
I look at volume, whether reliable sources exist, how the output will be checked, and the cost of an error. Some tasks are better served by rules or templates.
How do we know it works?
The checks are agreed before building and run before handover, then again whenever the model or sources change.
Do you give legal advice?
No. Your attorneys make every legal decision. I build and test the system that supports them.
Can we see past work?
Most is under NDA. The intake case study is public, and the rest can be walked through on a call.
Next step
Pick one legal task. We'll start there.
Tell me the task, who does it today and what a good result looks like. I'll tell you whether it fits and how I'd approach it.
Book a call(opens in new tab)30-minute intro call · or email hkhantanoli@gmail.com(opens in new tab)
No legal advice. A person approves every consequential step.
Last reviewed