A 12-attorney firm replaced its intake stack in 14 days.
Intake-to-engagement cycle dropped 62%. Partner time on routine NDAs went to zero.
- Mid-Atlantic litigation boutique
- 12 attorneys, 4 paralegals, 2 admins
- Inbound lead intake → conflicts check → engagement letter
- 14 days, fixed-fee
- Clio, Google Workspace, Make.com, Supabase, Anthropic Claude
“We didn't realize how much of our intake was a senior attorney pretending to be a paralegal until the system started doing it. The first week was a relief. The second week, two associates asked what else we could automate.”
Intake was a partner pretending to be a paralegal.
The firm was profitable. Reputation was strong, referrals were steady, and matter quality was high. The bottleneck was below the work itself.
Inbound leads landed in a shared inbox. A senior attorney triaged them, ran the conflicts check by hand against Clio, drafted a response from memory, and looped in a paralegal to put together the engagement letter. Median time from inbound email to signed engagement was 9.4 days. The firm was losing maybe one in five qualifying leads to a competitor that had responded inside 24 hours.
The partner doing this triage billed at $850/hour. He was spending six to nine hours a week on it.
Diagnose, then build the smallest thing that closes the gap.
Two days of working sessions: one with the triage partner, one with the paralegals. We watched five real intakes happen, mapped the decision tree, and identified the four points where AI could replace human transcription without replacing human judgment.
We rejected the obvious "build a chatbot for the website" pitch. Inbound leads arrive over email, not a form. The win was internal: a triage layer that did the prep work behind the scenes so a human could approve in 90 seconds instead of doing the prep work themselves.
Architecture written up in a single document the firm could read in 20 minutes. Fixed fee for the build. Two-week target.
Five components. In production at day fourteen.
-
A relational source of truth in Supabase.
Leads, conflicts checks, engagements, matter records, and time entries all flow through a single PostgreSQL schema. Clio remains the system of record for active matters; the new schema is the layer that talks to everything else.
-
Make.com workflows for intake routing.
Lead form submissions hit a triage scenario that classifies (matter type, geography, urgency), runs the conflicts check against Clio, and routes to the right partner with a pre-drafted engagement letter ready to review.
-
Claude as an invisible Chief of Staff.
Claude reads the inbound message + the conflicts result + house style guides, and drafts: (a) the partner-to-prospect response, (b) the engagement letter, (c) the matter brief for the assigned associate. Drafts only — sends require human approval until the firm trusts the path.
-
A standup-style review queue.
Every morning the assigned partner sees one screen: drafts awaiting approval, conflicts edge cases, lead scoring outliers. Approve, redline, or reject. The queue replaced a partner Slack channel that had 200 unread messages on day one.
-
Audit trail and evals.
Every Claude run is logged with input, tools called, citations, and output. The firm runs a weekly eval against a held-out set of 40 historical leads to detect regression. Two prompt revisions in week three came directly from the eval results.
Measured against the criteria from week one.
- −62%
- −74%
- +28%
- 14
From 9.4 days median to 3.6
Routine NDAs are now zero-touch
Faster response → fewer cold leads
Diagnosis through production rollout
We write this section on every engagement. Few consultancies do.
What we’d do differently next time.
-
We underestimated the partner-trust ramp.
We assumed two weeks of shadow mode would be enough before turning on auto-send for routine NDAs. It took four. The right call was the partners' — they wanted to see more samples than we expected. We've baked a longer shadow window into our default scope since.
-
House-style training data is messier than firms admit.
The "house style" turned out to be three different house styles, depending on which senior partner had drafted the most recent template. We resolved it with a working session, but we should have asked for redlines on five real engagement letters in the diagnosis week, not the build phase.
-
Conflicts check edge cases need their own queue.
About 6% of leads turned up conflicts edge cases that required human judgment. We initially routed those into the same partner queue as drafts; it created noise. Splitting them out into a dedicated conflicts queue with its own SLA was cleaner and is now part of our default architecture.
Got a workflow you’d like written up like this?
Send the messy version. We’ll spend 30 minutes on a call, sketch the architecture if there’s one to sketch, and tell you whether we’re the right shop to build it.