How it works

Decide which work AI should do. Keep the rest human.

That single principle drives everything JDK builds. The hard part isn’t the AI. It’s deciding, task by task, which side of the line each piece of work belongs on.

Try the line

Where does the work belong?

Pick a task to see how we’d call it, and why.

VERDICT: AI

High-volume, rule-based, easy to review, and fully reversible. A triage agent handles it and routes the exceptions.

The allocation framework

Six signals on each side of the line

GOOD FOR AI

Repetitive: the same shape of task, again and again

Rule-based: the policy can be written down

High-volume: enough throughput to matter

Data-rich: the inputs live in systems, not heads

Easy to review: a person can check the output fast

Reversible: mistakes can be caught and undone

KEEP HUMAN-LED

Ambiguous: the right answer depends on context

High-stakes: errors are expensive or public

Relationship-driven: trust is the actual product

Exception-heavy: every case is its own case

Tacit knowledge: the rules were never written down

Irreversible: no undo button

The process

Five steps, each with a deliverable on the table

DISCOVERY SPRINT

Discover

Map how work actually gets done: interviews, process mapping, data readiness.

→ PROCESS MAP

DISCOVERY SPRINT

Score

Rate every task for automation fit. Rank the backlog. Name what stays human.

→ RANKED BACKLOG + KILL LIST

90-DAY PILOT

Prove

One agent, one real workflow, measured against the pre-agent baseline.

→ MEASURED PILOT VERDICT

ON APPROVAL

Deploy

Run it in your systems with identity, approvals, audit, and cost controls.

→ AGENT IN PRODUCTION

ONGOING

Manage

Monitor, evaluate, tune, and expand agent by agent on the same foundation.

→ MONTHLY SCORECARD

The first two steps are the Business AI Analyst. Steps three through five run on the Agent Framework.

Start by drawing the line for your own workflows.

Start an AI Work Discovery Sprint