AI automation for regulated work
AI that does the busywork.
A human still makes the call.
I build automation for firms and practices where a wrong answer has consequences. The AI sorts, extracts, drafts and routes. The judgment stays with your people, and every step leaves an audit trail.
I build it inside the systems and the compliance you already have, on accounts you own. If we stop working together, you keep all of it.
The friction is rarely the software. It is the manual glue between it.
Most firms and practices I meet already run good systems. What costs them time is everything happening by hand in the gaps: the intake retyped into a second tool, the referral chased from memory, the email that got missed because the right person was not copied, the report that runs wrong because a name is spelled four ways.
Work that arrives faster than anyone can triage it
Leads, referrals and requests land in three places at once. Whoever is free grabs them. Nothing records why a decision was made.
The good stuff buried in the routine stuff
The urgent case looks identical to the tyre-kicker until someone reads it properly, and reading everything properly is the bottleneck.
AI that nobody trusts enough to turn on
The tools can draft and summarise. What stops adoption is not capability, it is the question of what happens when it is confidently wrong.
How I keep AI out of the judgment call
The model reads. It does not decide.
This is the part that makes the rest safe to adopt, and it is an architectural choice rather than a promise. In the intake systems I build, the language model is used only to pull facts out of unstructured text. The decision that follows is made by ordinary code with rules your professionals wrote and can read.
The model extracts. Nothing else.
It turns a messy message into structured facts: what happened, when, where, who else is involved. It is never asked whether something is good, valuable, or worth taking.
Rules you own make the decision.
Deadlines, thresholds, conflict checks and routing run as plain deterministic code. Same input, same output, every time. Your people can read the rule and change it.
It fails toward a human.
If the model is unavailable, uncertain, or the input is strange, the work routes to a person rather than being guessed at or dropped. An outage never silently loses anything.
Everything leaves a record.
An append-only log of what came in, what was extracted, which rule fired, who acted and when. If anyone ever asks how a decision was made, the answer exists.
In regulated work this distinction is not academic. It is the difference between a tool that assists your professionals and one that quietly practises on their behalf.
What I actually build
Intake and qualification
Every enquiry captured in one place, facts extracted, graded against your own rules, routed to the right person with a response clock attached.
The manual glue removed
Shared inboxes that route themselves, trackers that stop being spreadsheets, weekly chase lists, one-click summaries, data cleaned so reporting is trustworthy.
Acquisition wired end to end
Paid search, profile and call tracking connected through to what actually became a client, so spend is judged on signed work rather than on clicks.
Findability, including by AI
Structured data, technical fixes, and the files answer engines read. Increasingly your next client asks an assistant before they ask Google.
Reporting with a human gate
Internal digests daily, client-facing reports drafted for your approval before anything is sent. Nothing goes out over your name unreviewed.
Compliance treated as a deliverable
Advertising rules, messaging registration, disclaimers, records retention, and sign-off from your own counsel where it is warranted.
Work I can point at
Client engagements are described without naming the client. Where a number is not shown, it is because it is theirs to share rather than mine.
Legal
A New York personal injury firm
Graded intake, response clocks, an audit trail built for a regulator, and paid acquisition connected to signed cases. The grading architecture above came from this build.
Read the case studyMedical
A specialist practice, 18 years of records
What the audit found: referrals chased from memory, email as the coordination layer, and reporting quietly wrong. The whole first phase could be built inside an agreement they already held.
Read the case studyMy own product
Enkily, live on the App Store
An AI capture app I designed, built and shipped myself. Speech handled on device, model calls gated behind explicit consent. I build AI products, not just wire up other people's.
Earlier local search work, including a Houston fencing company, is written up in the Fenced Up case study.
How engagements work
Almost everyone starts with the audit. It is the only way either of us can scope a build honestly, and what you pay for it comes off the build if you go ahead.
Step One
The Workflow Audit
$2,500
Two weeks. I map where the manual work actually is, what is safe to automate, what has to stay human, and what a build would cost. Yours to keep either way.
Step Two
Build the system
$6,000 to $25,000
One or two workflows taken end to end into production. Scoped at the audit, billed against milestones, built on accounts in your name.
Step Three
Run it with you
from $1,500 per month
Monitoring, tuning, reporting and the next phase. Three bands depending on how much of the operation you want me holding.
Start by finding out what is worth automating.
Fifteen minutes is enough for me to tell you whether there is anything here worth paying for. If there is not, I will say so.