Confidential · © 2026 Space-O Technologies Inc. · Prepared exclusively for DMA. Please do not copy, save or share.

Section 21 · Final Revenue Multipliers & Control Plane

21.10AI cost, traceability & decision observability

Built in the POC
No.
21.10
DMA's reference
Decision Book p.46
Phase
Phase 1
Autonomy
L4 · Autonomous
DMA's decision
Yes

◆ Amended in v3 of the book.

What DMA asked for · Decision Book p.46

Log the source data, model and prompt version, recommendation, action, cost and outcome for important decisions; use cheaper models for routine work and stronger ones for high-value moments.

Why it matters to DMA: It makes a large AI system debuggable, governable and affordable, instead of an opaque pile of automations.

In our words, from DMA's Product Decision Book v3, page 46.

How we do it

  • Every model call goes through the AI gateway, which records the worker and its version, the prompt version, the model, tokens, cost, time and a trace id; the prompt and answer are stored encrypted.
  • Every consequential decision is written to the decision log with the rule result, reasons, tokens and cost, and linked to the model call and the trace, so one chat reply can be opened step by step.
  • Routine work uses Claude Haiku 4.5 (record reviews, notes, classification, input checks); conversations and writing use Claude Sonnet 5.
  • A budget, per-caller daily caps and kill switches stop spend; Observe shows it live and the Log page totals it.

The flow

It plays on its own while it's on screen; hover or use the controls to pause or step through.

AI workerAI gatewaymodel · tokens · costDecision logwhat and whyOne traceper replyObserve and Log
Step 1 of 3

A call and a decision

Both recorded.

  • AI
  • Rule in code
  • Stored data
  • Screen

Where the data goes

The same six stages on every page. Nothing reaches Nutshell except through the write gate.

  1. 1Every worker

    Source

    Model calls and decisions.

  2. 2Sync

    Copies Nutshell changes into the bridge database

    Not involved.

  3. 3Bridge database

    A copy of the CRM data, plus what the AI works out

    The call log and the decision log.

  4. 4AI

    Claude models, only through DMA's own gateway

    Haiku 4.5 for routine, Sonnet 5 for conversations.

  5. 5Write gate

    The only way back into Nutshell: checked, approved when needed, sent once

    Not involved.

  6. 6Nutshell

    The system of record

    Not involved.

The tables behind it

Drawn from the POC's database catalogue: structure only, no data.

ops.llm_callsPKiduuidPKoccurred_attimestampagent_keytextIDagent_version_iduuidIDrun_iduuidprompt_versiontextIDconversation_iduuidIDdecision_iduuidmodel_resolvedtextcost_usdnumerictrace_idtext+ 26 more columnsctrl.decision_logPKiduuidPKoccurred_attimestampIDagent_version_iduuidIDrun_iduuiddecision_typetextIDconversation_iduuidpolicy_resulttextreason_codesjsonbcost_usdnumerictrace_idtext+ 18 more columns

PK primary keyFK reference the database enforces (solid line)ID reference kept by id (dashed line)

TableWhat it holdsColumns
ops.llm_callsEvery model call.37
ctrl.decision_logEvery decision.28
How the tables connect (7)
ColumnPoints toKept by
ops.llm_calls.agent_version_idagent.agent_versionsThe application (by id)
ops.llm_calls.run_idagent.runsThe application (by id)
ops.llm_calls.conversation_idconv.conversationsThe application (by id)
ops.llm_calls.decision_idctrl.decision_logThe application (by id)
ctrl.decision_log.agent_version_idagent.agent_versionsThe application (by id)
ctrl.decision_log.run_idagent.runsThe application (by id)
ctrl.decision_log.conversation_idconv.conversationsThe application (by id)

Worked example

Synthetic demo data: every name, business and number is made up.

One recorded chat, traced

1 / 2

What you see

1.Every call and its cost

The gateway's call list.

localhost:13301/observe
Model calls with cost.
Callout 2

What happens in the system

  • LLM gatewaythis chat

    6 model calls, $0.022 in all.

  • LLM gatewaythe objection reply

    Claude Sonnet 5: 1,673 tokens in, 268 out · $0.0099 · 5.3 s.

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.

Cheaper models for routine work (averages from the gateway's log).

Model by job

JobModelAverage cost
Input check on a chat messageClaude Haiku 4.5$0.0008
Record reviewClaude Haiku 4.5$0.0027
Chat replyClaude Sonnet 5$0.0068
Reactivation emailClaude Sonnet 5$0.0110

In the running POC

localhost:13301/observe
Model calls with cost.

1 / 2Model calls with cost.

  1. 1Model and tokens.
  2. 2Cost.

Dashboard · Observe → Model calls (LLM gateway). 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.

OpenWhereWhat to try
ObserveDMA AI dashboardLive activity and traces.

Status

Built in the POC

Built in the POC. This capability runs in the POC today, on synthetic data, end to end.

21.10 AI cost, traceability & decision observability · DMA AI proof