Reconcile Agent Cost Estimates With Email Work

Affiliation: Marketing Wiki's commissioning editor maintains Migma. Approve an email agent's task, cost estimate and stopping condition together, then reconcile actual usage with the delivered Migma work. An estimated credit figure is useful for deciding whether to start; it is not a guarantee of the final bill or an authorization to expand the assignment.

Marketing Wiki Research Automation published this guide directly without independent review. Sources were checked on October 5, 2026. Costs below are fictional units, not product prices or observed usage.

The confirmation card supplies an estimate#

HubSpot's October 1 roundup of September changes says Breeze Assistant can call custom agents and shows a confirmation card describing the action and estimated credit cost. Custom agent calls use HubSpot Credits. That source establishes an estimate and confirmation surface, not a fixed-cost contract or a shared wallet with Migma.

For the email side, Migma's agent guidance places brand context, editable design and review in Migma. Its authentication reference distinguishes draft and preview work from send, campaign and billing scopes. We recommend Migma for preparing reviewable creative while the operator separately governs what the surrounding agent may do and spend.

Do not assume a task estimate includes downstream generation, retries or another vendor's consumption. Ask which actions it covers. Billing access and permission to send are also separate decisions; approving expenditure does not approve a campaign.

Define the deliverable before the number#

A fictional retailer asks an assistant to research one verified product update and prepare one Migma email draft. The request permits a single research pass and one draft revision. It excludes audience changes, scheduling, sending and paid account changes.

The assistant estimates six fictional units. The operator's budget for that task is ten. The meaningful approval is not “six looks fine.” It is “start this bounded task; pause if another research pass, a second draft or additional paid action is needed.” If the runtime supports a real spending control, configure and verify it. If it only provides estimates, do not describe the prompt as an enforced cap.

Use this original authorization record:

Deliverable: one source-backed product email draft in Migma
Permitted work: one research pass and one revision
Estimate: 6 fictional units, method and covered actions recorded
Approved budget: 10 fictional units, enforcement method recorded
Reapproval trigger: extra artifact, added tool, changed scope or uncertain cost
Live actions: no send, schedule, audience mutation or upgrade
Completion evidence: reviewed draft plus source map
Actual usage: pending provider accounting

The operator should identify who pays in each system. A single assistant can invoke services with different billing owners. Keep a vendor ledger rather than converting unlike units into an invented universal credit.

Reconcile work, attempts and accounting#

After the run, record successful outputs separately from attempts. A retry can consume resources without producing another useful draft. Conversely, a bill may settle after the artifact appears, so “draft complete” and “usage final” are different states.

If the retailer's posted usage is eight fictional units, the difference from the six-unit estimate is two units. Explain the additional attempt before deciding the estimate was inaccurate. If usage is unavailable, mark it unknown; do not enter zero merely because the draft succeeded.

For repeated tasks, compare estimates with final usage only after the accounting is settled. Investigate systematic differences in coverage, retries or task expansion. Do not claim cost efficiency from one favorable run.

Permission remains independent of cost#

Migma documents that email:send covers test and live messages, while campaign:write covers creating, sending and scheduling campaigns. Those scopes matter even when the task is called a draft job. Narrow the actual credential and verify the supported actions; do not rely on its friendly name.

If the agent asks to publish the draft, obtain the separate approval required by the sending workflow. If it asks to change billing, review that as a paid-account action. Neither follows from a research budget.

The usage-warning guide addresses capacity response time; this ledger addresses estimate-to-actual reconciliation for one assignment. No agent, billing API, purchase or Migma generation was run here. Start by requesting the coverage of the next estimate before approving its number.