Playbook · Interactive
Building a back-office automation agent
Scheduled and event-driven work that runs without a person watching: the six steps of a run, a cost model with a waterfall, and a rule spec and run receipt to copy.
What the automation agent does#
A back-office agent runs a written rule on a schedule or an event, with nobody watching each run. The rule, the locked list and the receipts do the work a reviewer would do in a copilot.
Every weekday at 6:00, send a reminder for invoices over 30 days overdue and open a CRM task for the account owner.
Step 1 of 6
A schedule or an event starts the run
A rule fires on a schedule or an event, like a new invoice or a changed record. The run checks that the trigger still holds before doing anything.
Levers
- TriggerWeekdays at 6:00, or a record change
- Freshness checkRead the record again before acting
- OverlapSkip a run while the last one is still going
Patterns
In the cost model: Runs per monthStep 2 of 6
Scope the run and lock the list
The run lists the exact items the rule covers and locks the list. New matches wait for the next run instead of slipping into this one.
Levers
- ScopeInvoices over 30 days overdue, one region
- CapAt most 200 items in one run
- New matchesWait for the next run
Patterns
In the cost model: Items per runStep 3 of 6
Act only within written rules
For each item, the agent takes the actions the rule allows: a reminder, a CRM task. Anything past the limits becomes an exception, not an action.
Levers
- Allowed aloneReminder email, CRM task
- LimitsUnder $10,000; one reminder per customer a week
- Acting accountA named service account with only these grants
Patterns
In the cost model: Cost per runStep 4 of 6
Send exceptions to a person
Items outside the rules go to a named owner with the reason and the record. The owner decides; the agent doesn’t retry around the rule.
Levers
- ReasonsOver the limit, disputed, missing data
- OwnerThe accounts-receivable lead
- Response timeSame business day
Patterns
In the cost model: Exception rateStep 5 of 6
Write a receipt for every run
The run records each item as changed, held, skipped, failed or unknown, with before and after values. A failed item is retried once, then reported.
Levers
- Item outcomesChanged, held, skipped, failed, unknown
- RetriesOnce, after checking the change didn’t already happen
- Where receipts liveThe rule’s run history, not an inbox
Patterns
In the cost model: Failures and retriesStep 6 of 6
Monitor and test the rule
Someone samples finished items against the record, and every bad item becomes a saved test case. A kill switch stops new runs and keeps the receipts.
Levers
- Sampling5 items a day, checked against the record
- AlertsException rate over 20%, any wrong recipient
- Kill switchStops new runs and further writes
Patterns
In the cost model: Checked results1 / 6 Trigger
The cost model#
Change any lever. The waterfall shows the hand work checked results remove, and what the platform, exceptions, failures and checks add back.
Example inputs, not benchmarks. Every default is made up, except the published platform prices under “Platform price”, which link to the vendor’s page. Replace each number with the team’s own.
Where the monthly cost goes
- Total
- Lowers cost
- Adds cost
- Hand work today$18,000
- Checked results−$15,543
- Platform price+$1,137
- Extra time on exceptions+$2,160
- Cleaning up failures+$745
- Monitoring and checks+$920
- Net monthly cost$7,420
The axis starts at $0.
Point at a bar to see how the step is calculated.
How each bar is calculated
| Step | Calculation | Monthly |
|---|---|---|
| Hand work today | 6,000 items × 4 min × $45/h | $18,000 |
| Checked results | 5,181 checked results × 4 min × $45/h | −$15,543 |
| Platform price | 11,371 actions × $0.10 (failed tries billed) | +$1,137 |
| Extra time on exceptions | 480 exceptions × (10 min − 4 min) × $45/h | +$2,160 |
| Cleaning up failures | 66 failures a retry didn’t fix × 15 min × $45/h | +$745 |
| Monitoring and checks | (15 h upkeep + 109 samples × 3 min) × $45/h | +$920 |
| Net monthly cost | Today’s hand work plus every step above | $7,420 |
Cost per result
The platform, exceptions, failures and checks come to $4,963 a month. Divided by attempted items, then by checked results, next to a person doing the item by hand.
- Per attempted itemCounts items that failed or weren’t confirmed.$0.90
- Per checked resultThe number to manage.$0.96
- By handMinutes by hand × hourly cost.$3.00
The platform price alone is $0.22 per checked result.
People’s time
260 h freed a month.
- Exceptions and unfixed failures stay with people: the hand work stays, plus the extra minutes.
- Freed hours lower the bill only when the team hires less or takes on other work.
What counts as a checked result
In this model, a checked result is an item the run completed and a check confirmed in the system of record. A green run status is not a check.
- Zapier counts a task only when Zapier completes a unit of work; failed actions are not counted. Autoreplay retries failures caused by temporary errors.Zapier pricing
- Salesforce now has a model judge whether a session was deflected, and fixed a bug that counted some sessions as deflected.LLM-based deflection and abandonment metrics in Agent Analytics
- Airtable tracks which AI-field cells are outdated or have errors, and a bulk run can target just those.Using Airtable AI in fields
Templates#
Copy or download. The spreadsheet carries the numbers from the cost model above, with live formulas.
Standing rule: rule name
Highlighted text is a blank to fill in. The words inside are an example.
1. The rule
- Trigger
- Weekdays at 6:00, or when an invoice turns 30 days overdue
- Scope
- Invoices over 30 days overdue, EU region
- Owner
- AR lead, with a named backup when away
- Version
- v7, changed date by name
2. What the agent may do alone
| Action | Limit | Acting account |
|---|---|---|
| Send a reminder email | Once a week per customer | A service account with only these grants |
| Open a CRM task for the account owner | One open task per invoice | The same account |
| Add a note to the invoice | Never edits amounts | The same account |
3. What needs a person
- Invoices over $10,000
- Disputed invoices
- A customer contacted in the last 7 days
- More than 200 items in one run: stop and ask before acting.
4. The list
- The run lists the exact items and locks the list when the run starts.
- New matches wait for the next run.
- An item that changed since the list was made is checked again before any action.
5. Failures and retries
- A failed item is retried once, after checking the change didn’t already happen.
- An item that fails twice goes to the owner with the error.
- A retry never sends an email a second time.
6. Kill switch and alerts
- Kill switch
- Pause on the rule’s page: stops new runs and further writes; receipts stay.
- Who can pull it
- Owner, backup, ops on-call
- Alerts
- Exception rate over 20%, failures over 5%, any wrong recipient.
- Owner leaves
- The rule pauses until someone takes it over.
7. Success metric
- Metric
- Cost per checked result under $X.
- Checks
- 5 finished items a day, compared with the record.
- Review
- The rule and its limits, every quarter.
Run receipt: rule name · run #212
Highlighted text is a blank to fill in. The words inside are an example.
1. The run
- Rule
- Overdue invoice reminders, version 7
- Trigger
- Weekday 06:00, fired at 06:00:01
- Acting account
- The rule’s service account
- List locked
- 41 items at 06:00:02
- Cost
- 82 actions
2. Items
| Item | Outcome | Before → after | Note |
|---|---|---|---|
| INV-1043 | Changed | No reminder → reminder sent, task opened | None |
| INV-1057 | Held for a person | No change | Over $10,000 |
| INV-1061 | Skipped | No change | Paid at 05:58, after the list was made |
| INV-1070 | Failed | Reminder sent, task not created | CRM timeout; retried once; sent to the owner |
| INV-1072 | Unknown | Email accepted by the mail service | Delivery not confirmed |
3. Totals
- 37 changed · 2 held · 1 skipped · 1 failed · 0 unknown
- Exceptions sent to the AR lead: 2 items
- Checked afterward: 5 sampled against the record, 0 problems
4. What can be undone
- CRM tasks: the owner can close them.
- Reminder emails: sent, and can’t be unsent.
- Invoice notes: can be removed.
0 of 24 checked
- Every input from the cost model above, with the numbers on screen now.
- Every output as a live formula: change an input in the sheet and the rest recalculates.
- Opens in Google Sheets and Excel. The file is made in the browser; nothing is uploaded.
What the real products do#
Three automation platforms from the teardowns, and one finding from each.
- Zapier Agents & AI by ZapierZapier
An approval step reports “Success” after a decline
Zapier’s Request Approval step can keep running after a decline or a timeout, and lists all three outcomes as Success. Only the step’s Decision field tells them apart.
- Airtable Omni & field agentsAirtable
Creating an AI field approves every future value
With Run automatically on, an AI field fills every record and runs again when its inputs change. One column can then hold older-prompt values, current ones and a person’s edits.
- Agentforce ServiceSalesforce
A time-ordered record of every step
Agentforce observability keeps every processing step in order, with per-step timing and error text. Deflection, by contrast, is scored by a model reading the session, not checked against the record.
How the numbers were sourced
The published prices in the calculator come from the vendors’ own pages, read on September 29, 2026: Salesforce Agentforce, $500 per 100,000 Flex Credits, and a standard action uses 20 credits and Zapier, $49 a month for 2,000 tasks, billed annually (Professional). Every other calculator number is an example input. The product findings come from the teardowns, whose sources were checked September 28, 2026.
Inside the rule the agent acts; outside, the agent asks the owner.