AI implementation
We build the workflows
we can measure.
Measure the work. Build the automation. Measure what changed. One job at a time, each one running before the next begins.
- Runs on its ownCheck the delivery dateTracker flags dates that slip836emails
- You approveChase a late orderChase drafted, Theo approves412emails
- Stays with youAgree a new date with the clientRosa decides and replies96emails
- Runs on its ownClose it in the projectTracker closes the order690emails
The order of work
Implementation follows measurement.
- 1SelectOne ranked opportunity out of observed work.
- 2DefineRoute, systems, rules, owners, approvals, exceptions.
- 3BaselineCapacity, turnaround boundary, period, evidence.
- 4Build and testNormal, exception, permission and failure paths.
- 5Release with controlNamed owner, escalation and a manual route.
- 6Measure againSame definitions, comparable period, context disclosed.
Categories, not a product catalogue
What gets implemented.
Runs on its ownConfirm the order
You approveChase a late order
Runs on its ownLog it in the project
Workflow automationTrigger to completed output, across approved systems.
What is our policy on change orders?
Change orders over $2,000 need Maya's sign-off before ordering.Source · Operations handbook, p. 4
Knowledge systemsApproved internal information, findable, with its source attached.
Draft · Weekly client updateHi Dana, this week the cabinets arrived and…
ApproveEdit
AI-assisted workPrepare, summarize, classify or draft. A person decides.
Read new supplier emailsdone
Match to open ordersdone
Checkpoint: date slipped?yes
Draft the chase for Theowaiting
Bounded AI agentsA defined multistep task with named tools, checkpoints and an owner.
Answered in a day79%Waited a week31
Operating visibilityState, exceptions, capacity and turnaround, where the measures exist.
What stays human
Named before anything runs.
- ✓The business owner and the system owner
- ✓The actions the workflow may take
- ✓The information and systems it may access
- ✓The decisions and approvals that stay with a person
- ✓Exception and escalation conditions
- ✓Stop conditions and the manual route
This week · Fernhill Studioevery send waits for a yes
Supplier order tracker22 chases drafted
Weekly client update2 wait for Rosa
Site visit desk1 waits for Dev
Invoice reminderspaused by Maya
Higher autonomy is not better.
Use the simplest level
- ✓Fixed rules where fixed rules hold
- ✓A person at every real judgment call
- ✓A bounded sequence only where the route is stable
What we do not build
- ✕Tools chosen before the work is understood
- ✕Open-ended autonomy where a bounded route will do
- ✕Professional judgment, regulated advice or client approval
- ✕Release without an owner, tested exceptions and acceptance
- ✕Activity reporting presented as business improvement
Common questions.
It is the usual start. The map and the walkthrough tell us what is worth building, and what is not.
No. It is scoped separately, with its own test cases, acceptance criteria, exception handling and manual route.
The tool comes after the work. We build on what you already use where we can, and we name every system in writing before we start.
Then it is not automated. A recommendation not to build is a real outcome.
A named business owner and a named system owner, agreed during the build.
Not sure what to build yetStart with your map and the walkthrough.Get your free inbox map
Already knowWork with us monthly, as your AI Sherpa.See the retainer