Data-health watchdog
Scans the operating databases every night for orphaned records, missing fields and mismatches. Writes findings to a tracked queue with owners.
- Read sourcesYes
- ProposeYes
- Write to queueYes
- Write to source of truthNever
Five layers, four scoped agents, a human approval gate and a read-only shadow mode. The AI is useful because it is fenced.
I led, governed and specified this system and its rules. It was built with a small data-and-automation team.
AI proposes.
A person approves.
The source of truth is never silently overwritten.
The architecture
Each layer has one job. The AI sits in the middle, where judgment is useful, and is fenced on both sides: by what it is allowed to read, and by what a person must approve before anything is written.
An event starts the work: a deal is signed, a record changes, a schedule fires, or a person asks a question in chat.
Read from the systems of record through service accounts, never personal logins. Normalize each record and tag it with where it came from.
Extract, classify, match, summarize, draft and score. Every output carries a confidence level and its sources. It stops here and waits.
Anything that touches money, customer status or a production workflow waits for a named person. Approve, edit or reject, in one tap.
A deterministic workflow engine does the approved step: create, update, notify, route. Logic that must be exact lives in code, not in a prompt.
A named owner for every workflow, an audit log, monitoring, rollback and a recovery path. Success rates sit on the executive report card.
The agents
Narrow jobs, narrow permissions. Each agent has a named owner, a mode and a written reason it exists.
Scans the operating databases every night for orphaned records, missing fields and mismatches. Writes findings to a tracked queue with owners.
Answers questions in chat by assembling status from the systems of record. It drafts the answer and the next step. A person decides.
Audits the automation estate itself: ownership, stale credentials, schedules and failure rates. Flags drift before it becomes an outage.
Runs revenue validation without writing anything. It logs what it would have done and is compared with what people did. Write access is earned, never assumed.
Earning trust
New automation starts as a read-only observer. It is promoted one rung at a time, only after a clean record against what people actually did.
Proposes only. Output is compared with what people actually did.
Drafts the work. A person sends or saves it.
Acts after a named person approves each step.
Low-risk, reversible steps only, with an audit log and a kill switch.
Promotion rule used here: four clean weeks with no false positives on any record that touches money or customer status.
Workflow health · illustrative
Every workflow is inventoried, owned and measured. The numbers below are recreated, fictional data in the shape of the real report card.
Failures are triaged by an owner within a day. A workflow with no owner is a defect.
| Domain | Flows | Success, last 7 days | % | Owner |
|---|---|---|---|---|
| Revenue operations | 14 | 94.3% | RevOps lead | |
| Delivery & installs | 12 | 92.2% | Delivery lead | |
| Support | 9 | 93.9% | Support lead | |
| Finance & billing | 8 | 95.4% | Finance lead | |
| Reporting | 10 | 91.3% | Operations | |
| Platform & utilities | 7 | 90.6% | Platform owner |
Owner: Platform owner · rotate to a service account within 2 days
Owner: Operations · retry logic added, review at next weekly check
Owner: Delivery lead · write runbook this sprint
Design rules
More in AI & Automation
See all tools