Proof

How We Actually
Run This

Every consultancy claims AI expertise. This page is the version of that claim you can check: the agent org that operates Service Robotics, the controls that keep it in bounds, and counters pulled straight from the running system.

01 / Live counters

From the Running System

Aggregates from our own operational tables, covering August 2026 unless the card says otherwise. Counts and durations only. No client data, no spend figures, no agent output.

5
Agents in the org

Each one has a defined role and limited tool access.

16
Tasks completed

Closed by agents this month on real operational work.

9
Approvals decided

Reviewed by a person this month. Pending ones don't count.

22
Posts published

All time. Only counted where an approval record exists.

7
Days since a cap stop

Last time the daily cap halted autonomous runs.

Cached hourly · last computed 10 Aug 2026, 19:56

02 / The org

An Org Chart, Not an Assistant

One general assistant with access to everything turns out to be the wrong unit. Several narrow roles work better, each holding only the tools its job requires, so a failure stays inside one lane instead of reaching the whole business.

Service Robotics agent org chart The human owner sits above a CEO agent, which coordinates CFO, COO and CMO agents. All outbound actions route through an owner approval queue. HUMAN OWNER decides · approves · can stop everything CEO AGENT delegates · sets priorities CFO AGENT read-only financial access COO AGENT operations · scheduled jobs CMO AGENT drafts only · never publishes every outbound action routes back up through the approval queue
03 / The governance stack

Five Controls, All Load-Bearing

This is the layer that rarely survives the trip from demo to production, because none of it shows up in a demo. It's also what decides whether a business can depend on the system or ends up quietly abandoning it.

Owner Approval Queue

One queue where agents park anything needing a human decision: publications, spend, client-facing actions. Approvals stay in one place instead of scattering through chat, and nothing leaves the building unapproved.

Daily Cap, Predictively Gated

A batch gate stops work once the day's spend hits the cap. A second gate runs before each individual agent run and skips it when that run's predicted cost would breach the cap, so a batch that starts under it can't run its way through.

Stop Switch

One owner-only flag halts every scheduled job and work cycle before a token gets spent, whatever the budget says. No agent holds a tool that can read or flip it, so nothing can un-pause itself.

Model Routing Per Job

The model tier gets set per unit of work rather than per role. Routine tasks run on the cheapest capable model and critical ones on the strongest, so spend follows the value of the work instead of the seniority of the agent.

Read-Only Financial Access

Agents that need financial context get reads and nothing more. None of them holds a tool that can move money, issue a refund or change a payment setting. The capability doesn't exist for them to misuse.

Why This Is the Whole Ballgame

Pilots rarely die because the model was bad. They die at the first question a responsible person asks. What happens when this is wrong? Who checks it? What does it cost if it runs all weekend, and how do I stop it? A prototype has no answers to any of that, and building the answers afterwards usually means building the thing again.

So we build them first, and we run on them ourselves. That's a stronger claim than it sounds, because it means our own operations break when this layer is wrong. The counters at the top of this page come out of the same tables our internal dashboard reads. Nobody types them in each quarter.

Hire us for an implementation and this stack ships as part of the build rather than as a hardening phase you pay for later. That isn't generosity. We've simply never seen a system reach production without it.

Next step

Want This Shape of System?

The assessment maps what you should build and designs the oversight around it. $3,500, about two weeks, and the plan is yours whether or not we build it.

Book a Free Consultation