M2 · FIRST FUNCTIONS · P06

The Agent Migration Playbook

Onboarding the digital employee. Treat agents like new hires, not features.

10 SLIDES · ~10 MIN · PLAYBOOK OVERVIEW
Key Takeaways
  • An agent is a new hire, not a feature.
  • 95% — of enterprise AI agent pilots stall before reaching production.
  • 4 — phases of structured onboarding — map, spec, shadow, bound — before autonomy.
  • 90 d — from kickoff to a production-grade agent on a workflow that already works.
  • Name one workflow that already works end-to-end with humans — write it down.
MODULE 2 · PLAYBOOK 06

The Agent Migration Playbook

Onboarding the digital employee. Treat agents like new hires, not features.

BuildClub Academy
01 · THE HOOK

95% of agent pilots stall. The 5% that ship are onboarded like employees.

95%
of enterprise AI agent pilots stall before reaching production.
MIT NANDA 2026
4
phases of structured onboarding — map, spec, shadow, bound — before autonomy.
90 d
from kickoff to a production-grade agent on a workflow that already works.
THE THESIS

An agent is a new hire,
not a feature.
Onboard it on a workflow that already works.

The Thesis
02 · THE MIGRATION PLAYBOOK

Five steps from workflow to agent — the same arc you use to onboard a human.

1
2
3
4
5
1
Map the workflow
Find one that already works end-to-end with humans. Document every handoff.
2
Spec the JTBD
Write the job description before you write the prompt. Title, inputs, outputs, KPIs.
3
Shadow (weeks 1–4)
Agent watches and proposes. Humans still act. Measure proposal quality.
4
Approve with confirmation
Agent acts but human approves every consequential output.
5
Bound autonomy
Remove approval gates only after 30 consecutive clean outputs. Set blast radius limit.
03 · WHAT AGENTS INHERIT

Agents do not go rogue. They inherit — your permissions, your paths, your stale facts.

1
Permissions
Over-permissioned by default. Industry research suggests most agents inherit far more access than they need — the same problem as service-account creep, accelerated.
2
Approval paths
The agent calls the same chain you do. If the chain is broken, so is the agent.
3
Stale knowledge
The agent’s facts are as old as your last refresh cycle. Stale RAG produces confident wrong answers.
04 · WORKFLOW ARCHAEOLOGY

Humans are the middleware. That seam is the workflow.

BEFORE — HUMAN AS WORKFLOW GLUE SYSTEM A CRM HUMAN copy / paste / reconcile the workflow seam SYSTEM B ERP AFTER — AGENT MIGRATES THE SEAM SYSTEM A CRM AGENT routes · validates · writes SYSTEM B ERP The human seam is where agents land first.

Find the seam: the human spending 40% of the day moving data between two systems. That seam is the agent's first job.

05 · THE AGENT JOB SPEC

Write the job spec before the prompt. Six fields. No exceptions.

1
2
3
4
5
6
1
Title
What the agent is called.
2
Reports to
Human owner.
3
Inputs
Exact data sources, format, freshness SLA.
4
Outputs
Exact deliverables, format, acceptance criteria.
5
KPIs
How you know it is working.
6
Blast radius limit
What it cannot touch without human approval.

If any field is blank, the agent is not ready to be built. The prompt comes last.

06 · THE TASKS-NOT-JOBS FRAME

You do not replace Susan. You fold the 10% spreadsheet task across 8 people into one agent.

8 people × 10% task
0.8 FTE distributed across the org. Invisible on any single org chart.
1 agent × 100% task
0.8 FTE consolidated. Visible, measurable, governable.
Look across departments
Not within roles. The 10–15% activity nearly everyone does.
Clear inputs and outputs
That is your first agent target. Tasks first. Jobs later.

Agents do tasks, not jobs. This reframe changes how you brief HR.

07 · THE 90-DAY ARC

Three phases. Ninety days. One agent in production.

Takeaways
  • 1 Days 1–30 — Map + Spec. Select the workflow, write the job spec, identify the owner, name the metric.
  • 2 Days 31–60 — Shadow. Agent runs alongside the human. Measure output quality, not speed.
  • 3 Days 61–90 — Bound + Ship. Approval gating off on clean outputs. Agent owns the workflow. Human reviews exceptions.
MONDAY-MORNING ACTIONS

Five moves to ship your first agent in 90 days.

Monday Morning
  1. 1
    Name one workflow that already works end-to-end with humans — write it down.
  2. 2
    Identify the human who is the middleware between two systems — that is your first spec.
  3. 3
    Write the six-field job spec before any prompt or vendor conversation.
  4. 4
    Name the 90-day owner — one person, weekly check-in, outcome metric defined.
  5. 5
    Set the blast radius limit on day one — what the agent cannot touch without approval.
P06 · Questions CEOs Ask

Frequently Asked Questions

What is the core idea of The Agent Migration Playbook?
An agent is a new hire, not a feature.
What does the data say a CEO should pay attention to?
95% of enterprise AI agent pilots stall before reaching production. 4 phases of structured onboarding — map, spec, shadow, bound — before autonomy.
What should a CEO do Monday morning after reading The Agent Migration Playbook?
Start here: Name one workflow that already works end-to-end with humans — write it down; Identify the human who is the middleware between two systems — that is your first spec; Write the six-field job spec before any prompt or vendor conversation.
What are the steps in The Agent Migration Playbook?
1) Map the workflow; 2) Spec the JTBD; 3) Shadow (weeks 1–4); 4) Approve with confirmation; 5) Bound autonomy.
Where do the claims in The Agent Migration Playbook come from?
The playbook cites MIT NANDA; McKinsey AI deployment research (2025–2026); OWASP Agentic Security Initiative.
P06 · Sources

All Sources in This Playbook

00 / 10