Choose the work before granting autonomy

Compare agent opportunities at an HR software company, then assign authority to individual actions with explicit promotion and stop conditions.

LESSON 4 OF 22 · 20 MIN ESTIMATED STUDY

On this page 6 sections

TL;DR: Choose a workflow where the outcome is valuable, observable and supported by available tools. Assign autonomy separately to each action, based on authority and consequence. Expand only when evidence shows the whole workflow and its recovery path can support the additional responsibility.

Where you are. Your strategy names where Corvane should invest. This section tests what an agent would actually be allowed to do there.

The sales promise is larger than the product

“An AI employee that handles onboarding,” says the fictional Corvane sales deck. Corvane serves 1,200 businesses with six product teams. Onboarding can mean answering a policy question, collecting forms, ordering equipment, granting system access or deciding whether somebody is eligible to work. Those are different jobs.

Start with evidence collection and coordination. A new-hire coordinator can identify missing documents and prepare an equipment request without deciding employment eligibility or granting sensitive access. This gives customers a useful result while keeping consequential authority explicit.

Compare opportunities using value, task clarity, tool availability, outcome verification and harm. Avoid collapsing them into one score that lets high savings compensate for an action you cannot legally or contractually perform.

Candidate job	First product decision
Answer policy questions	Retrieve permitted current policy; escalate ambiguity
Collect missing onboarding information	Ask for specific permitted fields and show completion status
Order equipment	Prepare a purchase within approved catalog and budget; approval before commitment
Grant payroll administrator access	Keep under the existing privileged-access process
Rank new hires by predicted performance	Exclude from the coordinator; requires a separate purpose and risk assessment

Read diagram description

A matrix of five onboarding actions against four levels of autonomy. Answering policy questions from current permitted policy, escalating ambiguity, and collecting specific missing information both act alone. Ordering equipment from the approved catalog within budget is prepared by the agent and approved by a person first. Granting payroll administrator access stays in the existing privileged-access process. Ranking new hires is out of scope and needs its own purpose and risk assessment.

Write the autonomy contract

For each action record the actor, permitted resources, conditions, maximum scope, approver and completion evidence. “Can order laptops” is incomplete. “Can prepare one standard laptop request for a named employee, from an approved catalog, within the assigned budget, with manager approval before purchase” is inspectable.

Reversibility is more than a cancel button. A purchase may be refundable but still expose a home address or create a vendor commitment. A system-access change may be technically reversible while exposing information during the interval. Evaluate consequences already incurred.

Use the lowest autonomy that delivers the intended benefit. A fixed workflow can collect a missing form and validate required fields. Model discretion may help decide which policy paragraph answers an unusual question. Adding agents to predictable steps increases the number of behaviors you must test.

Promotion is a decision, not a calendar milestone

Suppose 300 equipment requests were approved without edits. That is evidence about those requests and reviewers. It does not prove reviewers caught every error, nor does it cover unusual purchases absent from the sample. Audit accepted requests, include negative cases and verify the controls independent of the model before expanding authority.

A promotion rule should name the eligible slice, representative evaluation, prohibited failures, reviewer effectiveness and monitoring owner. A demotion rule might disable purchases after any unauthorized charge while preserving read-only status updates. The exact thresholds depend on harm and exposure; write them as your proposed policy, not an industry standard.

Do this before moving on

Write the contract for four onboarding actions: read a policy, request a missing form, purchase a laptop and grant access. For each, name one event that lowers autonomy immediately. Then answer a CEO who asks why the competitor's “fully autonomous” claim should not set your launch scope.

Expected reasoning: compete on completed, acceptable onboarding work and time saved. Demonstrate where approval is quick and where it protects an important boundary. Ask what evidence the competitor's claim actually supplies. A marketing adjective does not establish that the workflow is authorized or reliable.

A strong leadership answer also identifies the reusable investment: scoped identity, approval records and verified task outcomes can serve several teams. It does not make the platform so broad that the first useful workflow never ships.

Go deeper

Autonomy levels provides a vocabulary for the action-by-action contract.
Risk and reversibility explains why an undo button may not reverse the full consequence.
Workflows and agents helps remove unnecessary model discretion from predictable steps.
The tool execution envelope turns the autonomy contract into enforceable permissions and limits.

Key takeaways

Select an observable job before promising broad autonomy.
Write authority at the action and resource level.
Audit human approvals before using them as evidence for automation.
Promote and demote specific capabilities rather than the whole agent.

Check yourself

Answer out loud first, then reveal. Retrieving it yourself is what makes it stick; reading ours and agreeing does very little.

1Do 300 approvals without edits prove purchases can be automated?
2Why can a refundable purchase still be consequential?