Worksheet: The Agentic AI Business Model Canvas

One page that forces the strategy to be specific. Fill it for one initiative: a workflow, a department, or a product line where agents will carry real work. Under each prompt sits a micro-example from Meridian Facilities Group, a 900-person building services company, so you can see the altitude: concrete, numbered, small enough to be true.

The nine boxes:

1. Value proposition	2. Workflows in scope	3. Data and systems required
4. Agent roles	5. Human roles	6. Guardrails and controls
7. Governance	8. Metrics and evidence	9. Risks and failure modes

Initiative name: ____________________________ Owner: ____________________________

1. Value proposition

What outcome improves, for whom, by how much. Not what the technology does: what the business gets.

Example: work orders confirmed in under 5 minutes instead of 9 hours; technicians arrive with the right parts on the first visit 90 percent of the time, up from 71.

Yours: __________________________________________________________________

2. Workflows in scope

Name the workflows and their edges: where each starts, where it ends, and what is explicitly out.

Example: work-order intake from email and portal, triage, and schedule proposal. Out of scope: emergency dispatch, pricing changes, anything on a client's own ticketing system.

Yours: __________________________________________________________________

3. Data and systems required

Each system, the access level, and why. Minimum necessary: access you cannot justify is access you remove.

Example: maintenance system (read and write, work orders only), parts inventory (read), client contract terms (read), technician calendars (read).

Yours: __________________________________________________________________

4. Agent roles (from ARDs)

The agents by name and goal, one line each. Every agent listed here has a completed Agent Role Definition. The canvas summarizes; it does not replace.

Example: Intake Agent (codes and files every work order inside one hour). Schedule Proposer (drafts visit plans inside contract windows; never confirms).

Yours: __________________________________________________________________

5. Human roles

What stays human, and why: judgment, accountability, relationships, or law. Write the why. Teams that skip the why delegate by accident later.

Example: the dispatcher approves any schedule touching a penalty-clause client, because an SLA is a promise with a signature. Account managers own all client-facing commitments, because relationships compound and apologies do not delegate.

Yours: __________________________________________________________________

6. Guardrails and controls

Caps, gates, and forbidden actions across the initiative, with numbers.

Example: parts auto-orders capped at $500 per order and $5,000 per day; any message leaving the building requires a human send; emergency keywords route straight to the dispatcher and the agent stands down.

Yours: __________________________________________________________________

7. Governance

Who owns the agents, who may change their instructions, decision rights, and the escalation path with names.

Example: the operations director owns the agent fleet; instruction changes require the director plus one reviewer; the escalation log is read weekly; disputes end at the COO.

Yours: __________________________________________________________________

8. Metrics and evidence

How you will know it works, and what evidence you would show a board. Pair every speed metric with a quality metric.

Example: median time to confirm (target under 5 minutes), first-visit fix rate (target 90 percent), escalation share (healthy band 5 to 12 percent), weekly audit sample accuracy (target 97 percent), incident count and cost.

Yours: __________________________________________________________________

9. Risks and failure modes

What breaks, how you would know, and what you would do. Name the failure before it names itself.

Example: stale inventory data schedules technicians against phantom parts, detected by a first-visit fix dip, mitigated by nightly reconciliation and a data-age alarm. Tone drift in client messages, detected by the audit sample, mitigated by a template boundary and human send.

Yours: __________________________________________________________________

The reading test

Hand the completed canvas to someone outside your team. If they can answer four questions (what improves, what the agents do, what humans keep, what happens when it breaks), the canvas is done. If not, the gaps they find are the work.