What an AI workforce actually does — when you build it for production.

Most "AI consultancies" demo slides. Capstone runs the system inside its own business. This page walks through the productivity story — what AI does for us daily, the high-level shape of the build, and where governance sits.

The proof, in numbers

A two-person team, running a six-function business.

No more late evenings on invoices. No more "the proposal is in the pipeline." AI handles the structured, repeatable work — humans focus on training, delivery, and relationships.

≈190hrs
Saved per month
Across inquiry, content, finance and ops — work that previously took human time.
24/7
Client coverage
First responses within seconds, in any time zone — no human availability dependency.
Throughput
Same two facilitators handling triple the previous volume of inquiries, proposals and content.

Before / after

A typical Monday morning, then and now.

Same business. Same client volume. Different shape of work.

Before AI ~ 4 hours of work
  • Read through weekend inquiries one by one
  • Draft replies, copy-paste from past threads
  • Open spreadsheet · update pipeline stage by stage
  • Manually log new leads
  • Draft this week's content from a blank page
  • Capture last week's expenses from emails
With our AI stack ~ 30 minutes — review only
  • Inquiries already answered overnight — review handovers
  • Proposals drafted from briefs — sign-off or edit
  • Pipeline updated automatically — read the digest
  • New leads logged with full context
  • Content drafts in voice, ready for review
  • Expenses captured, categorised, reconciled

The shape of the build

A simple, three-part architecture.

The implementation details are proprietary, but the high-level shape isn't unusual — it's the discipline that's hard. Three layers handle every workflow, every day.

01

Always-on layer

Production runtime

Where the AI workforce actually runs. Tasks execute 24/7, regardless of where the team is. Built for reliability, not demos.

Local + cloud Always-on Monitored
02

Build & delivery layer

Field workstation

The Mac that travels — where new agents get built, training demos happen, and live classroom prototyping takes place. Connects to the runtime remotely.

Dev studio Live demos Travel-ready
03

Command layer

Mobile approvals

The pocket interface for approvals, alerts, and "go / no-go" decisions. Critical actions never run without a human green light.

Approvals Alerts Human-in-loop
Nervous system

Google Workspace binds the three layers — Gmail, Calendar, Drive, Docs, Sheets and Chat are the connective tissue every agent reads and writes through. Nothing is locked in a single tool.

In the open

How Abigail (WhatsApp) actually replies.

The one agent we put fully in front of clients is Abigail — Capstone's WhatsApp inquiry agent. Here's the flow she runs every time someone sends a message.

The reply path

A message lands in WhatsApp Business. Within seconds, it's been routed, grounded in Capstone's knowledge base, and answered — or escalated to a human if confidence isn't high enough.

STEP 01
Client
WhatsApp message
STEP 02
Gateway
Routes & logs
STEP 03
Local LLM
On-prem reasoning
STEP 04
Knowledge base
RAG retrieval
STEP 05
Abigail
Composed reply

Three-tier governance

Abigail doesn't get to invent. She has explicit rules for when to answer, when to caveat, and when to step aside.

Tier 1 · Primary

Reply from documents

Search the verified knowledge base first. Reply only from grounded content.

Tier 2 · Secondary

Formulate, with caveat

Allowed only at 90%+ confidence. Reply is flagged as "her best read" — not authoritative.

Tier 3 · Escalate

Warm handover

Politely defer to a human, log the question, and alert the team. No invented answers.

For your business

What this could look like for your team.

Most teams we train don't need an eight-role workforce on day one. The transformative win — for an FA firm, an SME, an in-house L&D team — is usually one or two agents that take recurring, structured work off a human's plate.

Foundational · AI Launchpad 2 days
  • A working Claude Project, configured around your real work
  • A 30-day implementation plan tailored to your business
  • Prompt engineering fluency and workflow mapping
  • Responsible-AI governance framework you can adopt
Hands-on · Agentic AI Accelerator 1 day
  • A working agentic workflow, deployed in your own stack
  • Multi-agent orchestration patterns, hands-on
  • Walk through Arvin's actual production reference build
  • RAG & tool-use patterns you'll actually use

Want this in your business?

Start with a short brief. We'll match you to the right programme and the right pace — and tell you honestly if your use case fits.

Send us your brief Compare AI programmes