All case studies
LiveTeams agent in development

RBW Atlas

One system of record, and an agent that knows when to ask a person.

Problem
Paper, scattered data, no field alerts
Move
One system of record, AI on top
Agent
80% benchmark, ~70% live
Ops
1,200 locations watched 24/7
Role
Solo build with AI coding agents

Scattered to centralised

The question: could RBW run on AI workflows it controls? First: where does our data live?

Atlas
Customers on paperCRM + onboardingInventory untrackedInventory recordsDocs scatteredKnowledge base + agentNo field telemetryTelemetry + alertsFailures found by phone24/7 escalation

Click through it

Hit play, then click the pulsing dot.

atlas, sanitised demo, fictional dataPlaceholder

Atlas click-through

Sanitised wireframes, fictional data

add: the real click-through goes here

Six decisions

Tap any box for the reasoning.

Keep Jira, build the agent

Buy the commodity. Build where you need control.

ticketing, email, git links Jira + Bitbucket rejected Rovo built Our support agent Buy the commodity Build where you need control

Swipe the diagram →

Tap a box.

add: what keeping Jira separate costs

Measure before trusting

Four rounds of fix-and-re-run. Live is lower.

80% 37% 80% ~70% 4 rounds: fix gaps, re-run −10 pts Before fixes After 4 rounds Live trial benchmark benchmark production, early
80% 37% 80% ~70% 4 rounds: fix gaps, re-run −10 pts Before fixes After 4 rounds Live trial benchmark benchmark production, early

Tap a bar, or the −10 pts.

add: define “resolved” for each baradd: same tickets reused each round?add: trial size and dates

Steady beats cheap

Open-source was slower and unpredictable. Customers need to know what to expect.

slower faster days Open-source via gateway Low-cost OpenAI model ≈ +2 s Illustrative, not measured

Swipe the diagram →

Tap a label.

add: measured p50 and p95add: cost per conversation

Alert on the servers

Telemetry in every five minutes. Alerts out in under 200 ms after ingest.

1,200 locationsS1S2S3S4S55 serversevery 5 minAtlasingest + alert< 200 msTeamsEmailauto-escalates

Swipe the diagram →

Tap a box.

add: alerts in the last 90 days, and incidents caught before a customer noticed

A person approves drafts

Drafts wait for a person to review them.

  1. Gap found
  2. Draft written
  3. Email + task
  4. Person edits
  5. Published

Tap a step.

add: is gap detection automatic in production?add: review effort per article

Ask first, then build

Staff shape each feature before it’s built.

  • Ops weekly
  • Support weekly
  • Dev team varies, mostly the employee hub

Tap a row.

add: weekly users out of staffadd: one case where feedback changed the build

Next

Teams agent, in developmentadd: what I’d change, and what I’d test next

Next case study

Guardian Compliance: from paper records to 950 live sites

Read it