# Who keeps your business agent working after setup?

Give a small-business agent a named owner, backup, change log and pause procedure. Includes a handover card and maintenance test sheet.

Canonical: https://agentreport.io/maintain-small-business-ai-agent/
Author: Agent Report Editorial
Published: 2026-10-10
Modified: 2026-10-10
Evidence: Editorial guide
Disclosure: Written with Codex assistance. Sources checked October 10, 2026. This is an original handover and maintenance proposal. No vendor-specific pause, revocation or recovery procedure was tested.

Before the person who built your agent steps away, ask a coworker to explain its job, find its unfinished work and pause its next action. If they cannot do those things, the setup is still dependent on its builder.

[Goldman Sachs' March 2026 research](https://www.goldmansachs.com/community-transformation/10000-small-businesses-voices/insights/ai-presents-a-major-opportunity-for-small-businesses) describes an implementation and training gap among participating small businesses. The survey covered 1,256 program participants, not a representative sample of all small firms, and concerns general AI use. Our maintenance proposal is an editorial response to that issue; the study did not test this checklist or measure agent failure rates.

## Give the workflow an owner and a backup

Name the person responsible for the business outcome, then name a backup who can cover an absence. The software supplier can help repair a connector, but someone in your business still has to notice that customer callbacks stopped arriving.

Write the job narrowly enough to inspect: “Prepare replies to opening-hours questions for the support lead to approve.” Avoid “run customer service,” which hides several decisions behind one phrase.

Use the [handover card](https://agentreport.io/worksheets/small-business/agent-handover-card.csv). Fill in the queue where work starts, what a completed result looks like, who approves it, the systems it can access and how a failed case becomes visible. Record where authorized staff obtain credentials through your existing access-management process. Do not put passwords or tokens in the card.

## Preserve the business facts it relies on

Assign an owner and revision date to the reference material: opening hours, service area, product details or whatever this particular workflow uses. Record where that material lives and how the agent receives an update.

When a business fact changes, update the authoritative record first. Then refresh any copies or indexes the integration uses. Run a known question that should produce the new answer. Seeing the document change is not enough; check the resulting draft or action.

If you cannot establish which version the agent used, pause work that depends on the changed fact until somebody can verify it.

## Schedule checks around the consequence of a miss

During the first pilot, review every completed item and every exception. Use what you learn to choose an ongoing cadence. An agent handling same-day customer calls needs a different response deadline from one compiling a weekly internal summary.

For each review, inspect accepted work, unclaimed exceptions, usage charges and the latest successful run. A dashboard showing no errors does not establish that expected work arrived. Compare the queue with the incoming source when possible.

Repeat the relevant tests whenever a connector, model, permission or business rule changes. Keep a short change log naming the change, its date, who reviewed it and which cases were rerun.

## Rehearse three failures with the backup

Use a test environment or invented cases. The [maintenance test sheet](https://agentreport.io/worksheets/small-business/agent-maintenance-tests.csv) records expected outcomes, actual results and unresolved work.

| Drill | What the backup should establish |
| --- | --- |
| The reference answer changes | New drafts use the approved revision; older queued drafts are reviewed |
| An integration loses access | Failed work remains visible and a named person receives it |
| A bad action requires a pause | Future work stops through the actual control; queued and in-progress work is accounted for |

Write the exact control names from your chosen system into the handover card. A button labelled “pause” might stop scheduled runs while leaving queued actions active. Determine its real effect in the test environment. If you cannot safely verify it, record the limitation and keep that action out of unattended operation.

## Restore ordinary work before debugging

For each workflow, define how a person can continue without the agent. An inbox might return to manual assignment. A call workflow might return to the previous routing. Invoice preparation might return to the existing reviewed process.

When something goes wrong, first establish what already happened. Preserve run IDs and the relevant records. Separate completed, failed and uncertain items. Avoid replaying a whole batch merely because the run reported an error.

The owner should reconcile uncertain items against the destination system before deciding what to retry. Restoring access does not itself prove that repeating a send, booking or charge is safe.

## Make the handover demonstrable

Give the backup the card and an invented failed case without a live explanation from the builder. Ask them to find the exception, identify the last confirmed action and show how they would restore the manual process. Record every missing instruction.

When the backup completes that rehearsal, you have evidence that another person can follow the handover. You still need to observe the workflow in use. Include the ongoing review and maintenance time in the [ROI ledger](https://agentreport.io/measure-ai-agent-roi-small-business/) so the agent's real operating cost remains visible.

Keep case-level review time in the case ledger. Put only maintenance beyond individual cases in the costs sheet under “Maintenance outside case work”; do not count the same minutes twice. The [worksheet instructions](https://agentreport.io/worksheets/small-business/README.txt) explain these categories.

## Ask the vendor to demonstrate the handover

For the products in our [receptionist](https://agentreport.io/ai-receptionist-small-business-trial/#three-receptionist-vendors-to-investigate), [inbox](https://agentreport.io/ai-shared-inbox-human-approval/#choose-the-inbox-product-before-the-agent) and [invoice](https://agentreport.io/ai-invoice-drafts-spreadsheet/#two-accounting-assistants-worth-investigating) comparisons, ask who can change instructions, inspect failed work, revoke access and stop charges. Have a second team member perform those steps in the trial. A reassuring sales answer does not replace a working handover, and a positive product review does not tell you who maintains your configuration.