Part A · How DMA AI works
A.2The AI workers and their autonomy
Built in the POC- No.
- A.2
- Relates to DMA's book
- Decision Book p.7, p.8
- Part
- A · the foundation
What this part does · Decision Book p.7, p.8
DMA's book rates every capability on five autonomy levels: observe, recommend, prepare and wait for approval, act within rules, and autonomous. The AI is the default operator; people step in on defined exceptions.
Why DMA needs it: Clear levels make it obvious what the AI may do on its own, what waits for a person, and who is accountable for each step.
How we do it
- The work is split into ten AI workers, each with one job: the website salesperson, reactivation analyst, outreach writer, reply handler, CRM scribe, briefing agent, escalation router, revenue reporter, site reviewer and the team's workspace assistant.
- Each worker has a ceiling on its autonomy (for example, the outreach writer is L2: its emails always wait for approval) and a list of the only tools it may use. The ceiling is checked in code by the rule engine, not by asking the AI to behave.
- Each worker's instructions, model and tool list are a version. A new version is tested in the AI Lab and runs only after a person at DMA approves it; every decision records the version that made it, so a bad change can be rolled back in one click.
- Conversations and writing use Claude Sonnet 5; bulk work such as reviewing records, keeping notes and classifying replies uses the faster Claude Haiku 4.5. All of them run on Anthropic's Claude Agent SDK.
The flow
It plays on its own while it's on screen; hover or use the controls to pause or step through.
A version is tested and approved
A change to a worker's instructions runs its test suite in the AI Lab; a person at DMA approves it before it goes live.
- Stored data
- AI
- Rule in code
- Routing
- Screen
- Person
Where the data goes
The same six stages on every page. Nothing reaches Nutshell except through the write gate.
- 1Any request
Source
A chat, an email reply, a call transcript, a nightly review, a team question.
- 2Sync
Copies Nutshell changes into the bridge database
Not involved in choosing the worker.
- 3Bridge database
A copy of the CRM data, plus what the AI works out
Holds the workers, their versions, their tool grants and every run.
- 4AI
Claude models, only through DMA's own gateway
The worker's model, through the AI gateway.
- 5Write gate
The only way back into Nutshell: checked, approved when needed, sent once
Anything a worker proposes for Nutshell goes through the write gate.
- 6Nutshell
The system of record
Workers never call Nutshell directly.
The tables behind it
Drawn from the POC's database catalogue: structure only, no data.
PK primary keyFK reference the database enforces (solid line)ID reference kept by id (dashed line)
| Table | What it holds | Columns |
|---|---|---|
| agent.agents | The ten workers: job, where they work, and the autonomy ceiling. | 10 |
| agent.agent_versions | Each worker's versions: model, instructions and who approved it. | 19 |
| agent.tool_grants | Which tools each version may use, and at what autonomy. | 5 |
| agent.runs | Every run of a worker, with its tokens and cost. | 17 |
How the tables connect (6)
| Column | Points to | Kept by |
|---|---|---|
| agent.agents.active_version_id | agent.agent_versions | The application (by id) |
| agent.agent_versions.agent_id | agent.agents | The database (foreign key) |
| agent.tool_grants.agent_version_id | agent.agent_versions | The database (foreign key) |
| agent.tool_grants.tool_id | agent.mcp_tools | The database (foreign key) |
| agent.runs.agent_version_id | agent.agent_versions | The application (by id) |
| agent.runs.conversation_id | conv.conversations | The application (by id) |
Worked example
Synthetic demo data: every name, business and number is made up.
The website salesperson, as a worker with a version and limits
What you see
1.Each worker, its ceiling and its version
DMA's AI settings list every worker with the most it may do on its own and the version that is live.
What happens in the system
- Bridge database
agent.agentsThe website salesperson: autonomy ceiling L4 for the chat itself; the reactivation analyst: L3.
- Bridge database
agent.agent_versionsEach runs version 1, the approved starting version. Newer drafts wait until a person approves them.
Right-hand values come from a real run of the POC on September 25, 2026 (trace tr_muh5tbo52jgpqd0smgrf8, tr_muh5tbqv18g3rgfbfdhda…). Rows marked “sample” are illustrative.
The ten workers as configured in the prototype today.
The AI workers
| Worker | Its one job | Works in | Most it may do alone | Model |
|---|---|---|---|---|
| Website salesperson | Holds the website conversation, diagnoses, books the call | Website chat | L4 · autonomous (chat only) | Claude Sonnet 5 |
| Site reviewer | Reviews a visitor's website: three things to fix, with evidence | Behind the chat | L0 · observe | Claude Sonnet 5 |
| Reactivation analyst | Reviews and scores the Nutshell records; tiers and next best action | After each Nutshell sync | L3 · act within rules | Claude Haiku 4.5 |
| Outreach writer | Writes one-to-one reactivation emails | Approvals | L2 · prepare, wait | Claude Sonnet 5 |
| Reply handler | Reads replies: opt-outs, out-of-office, interest; drafts answers | Inbox | L3 · act within rules | Claude Haiku 4.5 |
| CRM scribe | Turns calls and chats into CRM notes, fields, tasks and follow-ups | Team workspace | L3 · act within rules | Claude Haiku 4.5 |
| Briefing agent | One pre-call brief per account | Team workspace | L3 · act within rules | Claude Sonnet 5 |
| Escalation router | Routes cases to the right person and writes the brief | Escalations | L3 · act within rules | Claude Haiku 4.5 |
| Revenue reporter | The daily money report | L1 · recommend | Claude Haiku 4.5 | |
| Workspace assistant | Answers the team's questions in the DMA AI workspace | Team workspace | L2 · prepare, wait | Claude Sonnet 5 |
In the running POC
1 / 4DMA's AI settings: each worker, its autonomy ceiling and its live version.
- 1The website salesperson.
- 2Its live version and approval.
- 3The reactivation analyst, at a lower ceiling.
Dashboard · Build → Agents. Captured from the running POC on September 25, 2026; synthetic data.
See it live
Opens the exact screen in the running POC, in a new tab.
| Open | Where | What to try |
|---|---|---|
| AI workers | DMA AI dashboard | Each worker's ceiling, versions and tools. |
| AI Lab | DMA AI dashboard | Test a version before approving it. |
Status
Built in the POC. This capability runs in the POC today, on synthetic data, end to end.