AI owns the
grind.
Humans own the judgment. The art is in the handoff. We draw the line on day one — before a single prompt is written — and document it in the same shape for every engagement. This is the artifact that keeps AI projects from turning into political grenades when they meet the operations team.
AI projects fail on the handoff, not the model.
Every AI pilot we've seen stall has stalled for the same reason: the team never agreed on what the AI was allowed to do unsupervised. The model was fine, the data was fine, the integrations worked. But somebody senior walked by the dashboard, saw the agent drafting a client email, and asked "who approved that?" Nobody had. The pilot got paused.
The framework prevents this by making the boundary explicit, named, and signed off before build. It becomes a one-page policy document that the CEO, the head of operations, and the compliance officer all see and agree to. It ships with every AI engagement.
If a senior operator cannot tell you, in one sentence, what the AI is allowed to do without a human in the loop, the project is not ready to build. Get that sentence first.
What AI does. What humans do.
This is the template we fill in for every engagement. The specifics change per firm; the two-column shape does not.
What gets automated
Every item here is a task where the "right answer" is deterministic enough to audit, and the volume is high enough to matter.
- MonitoringDaily polling of e-signing platforms, inbound folders, third-party webhooks. No human in the loop.
- ExtractionStructured reading of source documents (code sheets, decision letters, contracts). Pure reading, no judgment.
- ReconciliationComparing two data sources and flagging disagreements. The flag is automatic; the resolution is human.
- Form populationCorrect form per item status, filled from verified data. The form is prepared, not filed.
- DraftingClient emails, status updates, prep notes. Always drafts — never sends.
- CalculationInvoice amounts, rate changes, timeline projections. Pure math.
- RoutingClassify the inbound document, move the deal to the matching stage. Deterministic policy on top of AI classification.
What stays with people
Every item here requires professional judgment, has legal weight, or is where the firm's reputation lives. No AI.
- StrategyListening to the client. Deciding which case to file. Setting the condition list. This is the firm's craft.
- ReviewApproving AI-drafted forms and emails before anything leaves the system. Named senior operator on every case.
- SubmissionFiling upstream to the system of record (regulator, tax authority, court, client). The AI prepares; the human submits.
- AppealsWhen a case is lost, a named senior operator decides whether to appeal. Not a classifier.
- ApprovalInvoice approval, refund approval, exception approval. Humans in the money loop, always.
- EscalationExceptions that the AI flags but can't resolve. Route to a named human, not a queue.
How we classify every task.
Every manual task in the firm gets sorted into one of three buckets. The bucket determines the engineering shape and the governance shape at the same time.
AI-owned (no human)
High volume, deterministic, auditable. The AI runs unsupervised and reports exceptions. Example: daily e-sign monitoring. The "right answer" is binary — document is complete or it isn't. No judgment involved.
AI-prepared, human-approved
High volume, requires judgment before it leaves the system. The AI prepares the artifact (form, email, invoice); a named human approves before submission. Example: form population + client emails.
Human-owned (no AI)
Low volume, high stakes, professional craft. The AI does not touch the task, full stop. Example: strategy sessions, appeal decisions, invoice approvals. We sometimes add AI context (e.g. "here is the case history") but never AI recommendations.
The AI never submits.
Across every engagement, in every framework on this site, the same rule holds: the AI prepares and proposes; a named human approves and submits. The AI has no authority to file upstream — to a regulator, the tax authority, a court, a billing system, or a client email inbox. That authority lives with an operator.
Why this is a hard line: the moment the AI can submit, the review discipline collapses. Operators stop reading the output carefully because they know it went out regardless. Holding the submit button gives the review real weight — and makes every downstream mistake caught by a human, not by a client.
The AI never clicks submit. Not on contracts, not on filings, not on outbound emails, not on invoices. That authority lives with a named human, always.
Draw the line first. Build after.
We run this framework as a single working session, before any engineering starts. Output is a signed one-page policy. Cost of getting it wrong is an AI pilot that gets paused in month three.