← BackSee all toolsAI & Automation Architecture: Five Layers, Four Scoped AgentsWorking demo · sample data · shows a method I used, not a real event
Architecture · workflow health

AI & Automation Architecture

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

Five layers, and a gate you cannot route around.

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.

  1. 01

    Trigger

    An event starts the work: a deal is signed, a record changes, a schedule fires, or a person asks a question in chat.

    WebhookScheduleChat commandManual run
  2. 02

    Data collection

    Read from the systems of record through service accounts, never personal logins. Normalize each record and tag it with where it came from.

    CRMBillingSupport deskProject boardData warehouse
  3. 03

    AI reasoning

    Extract, classify, match, summarize, draft and score. Every output carries a confidence level and its sources. It stops here and waits.

    ExtractMatchSummarizeDraftScore
  4. ◆

    Human approval

    Anything that touches money, customer status or a production workflow waits for a named person. Approve, edit or reject, in one tap.

    ApproveEditReject
  5. 04

    Action

    A deterministic workflow engine does the approved step: create, update, notify, route. Logic that must be exact lives in code, not in a prompt.

    CreateUpdateNotifyRoute
  6. 05

    Governance

    A named owner for every workflow, an audit log, monitoring, rollback and a recovery path. Success rates sit on the executive report card.

    OwnerAudit logMonitoringRecovery

The agents

Four scoped agents. None can write to the source of truth.

Narrow jobs, narrow permissions. Each agent has a named owner, a mode and a written reason it exists.

Read-only + report

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
Assist

Operations command center

Answers questions in chat by assembling status from the systems of record. It drafts the answer and the next step. A person decides.

  • Read sourcesYes
  • ProposeYes
  • Write to queueNo
  • Write to source of truthNever
Read-only + audit

Integrity auditor

Audits the automation estate itself: ownership, stale credentials, schedules and failure rates. Flags drift before it becomes an outage.

  • Read sourcesYes
  • ProposeYes
  • Write to queueYes
  • Write to source of truthNever
Shadow mode

Revenue shadow agent

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.

  • Read sourcesYes
  • ProposeYes
  • Write to queueNo
  • Write to source of truthNever

Earning trust

Shadow mode first. Autonomy last.

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.

  1. 1

    Shadow

    Proposes only. Output is compared with what people actually did.

  2. 2

    Assist

    Drafts the work. A person sends or saves it.

  3. 3

    Approve

    Acts after a named person approves each step.

  4. 4

    Autonomous

    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

What the weekly health panel shows.

Every workflow is inventoried, owned and measured. The numbers below are recreated, fictional data in the shape of the real report card.

60workflows audited
40active
20scheduled
150runs a week
93%run success
20+ hrsmanual work returned weekly

Run success, last eight weeks

target 95%W1W2W3W4W5W6W7W8

Failures are triaged by an owner within a day. A workflow with no owner is a defect.

Health by domain

DomainFlowsSuccess, last 7 days%Owner
Revenue operations14
94.3%RevOps lead
Delivery & installs12
92.2%Delivery lead
Support9
93.9%Support lead
Finance & billing8
95.4%Finance lead
Reporting10
91.3%Operations
Platform & utilities7
90.6%Platform owner

Needs attention

  • High
    Credential near expiry on a scheduled sync

    Owner: Platform owner · rotate to a service account within 2 days

  • Medium
    Reporting flow failed 3 of the last 7 runs

    Owner: Operations · retry logic added, review at next weekly check

  • Low
    Two scheduled flows have no documented recovery step

    Owner: Delivery lead · write runbook this sprint

Design rules

What I would not let AI do.

  1. Overwrite a record of truth. It writes to a queue; a person or a deterministic rule writes to the record.
  2. Decide on money or customer status. It can recommend and explain. The decision has an owner.
  3. Automate a broken process. Fix the process, then speed it up.
  4. Run without an owner. No owner, no schedule.
  5. Hide its reasoning. Every output carries sources and a confidence level.
  6. Act without a way back. Every action is logged and reversible, or it needs approval.

More in AI & Automation

See all tools