Company's AI — seven agents, zero token bill

Why

Every company wants AI these days, and almost every path to it starts with a bill: a seat per person, a token per question, a line item that grows exactly as fast as people find useful ways to use it — plus a quiet legal problem, because every prompt leaves the building.

I didn't want that for the company I run. I wanted AI in the workflows where it actually earns its keep — the follow-ups, the paperwork, the reports, the scheduling — at a cost that doesn't scale with curiosity.

So the question was: what does a small company's AI look like when you remove the meter entirely?

Company's AI is the answer: seven specialised agents, self-hosted, reachable on the channels my staff already use.

The approach

A team, not a chatbot

One general assistant does everything badly. So each agent has a job, and jobs hand off to each other:

Agent	Role
Ain	Chief AI / orchestrator — coordinates work across all agents, manages priorities
Shiela	Executive assistant — scheduling, follow-ups, documentation, nothing falls through
Lisa	Data keeper — collection, analysis, reporting, integrity of the numbers
Kifli	Safety — security, compliance, risk, validating outputs before they're relied on
Elizabeth	IT & innovation — infrastructure, integrations, evaluating new tools
Balqis	Field operations — logistics, on-ground coordination, client-facing work
Advik	After-sales — support, issue resolution, customer success

Ain can delegate to anyone; anyone can pull anyone else in. Balqis asks Lisa for field data analysis; Kifli audits anyone's output before it goes out. That's the part a single chatbot can't do: work that crosses departments, with a named owner at each step.

Where the staff actually meets them

Agents that live in a browser tab don't get used. Each one is reachable on all three channels at once:

• Email — formal requests, documents, reports, approvals
• Telegram — quick queries, alerts, coordination
• WhatsApp — field communication, mobile, client-facing

Same context follows the conversation across channels, so a question started on Telegram can continue on WhatsApp without anybody re-explaining.

Beyond chat

The agents do work, not just answers: paperwork and documents, professional email, marketing visuals and diagrams, training and promo video, scripts and workflow automation. And they run on local inference — a model on our own hardware, orchestrated with a multi-agent framework, not a cloud API.

What it costs to run

This is the part that makes it a business case rather than a toy:

• Zero operational cost — no subscription, no per-token meter, no seat licences
• Data stays local — nothing leaves the building for a prompt
• No usage caps — nobody has to think twice before asking again
• Tunable to our workflows — each agent is shaped around how the company actually works, not a vendor's defaults

The agents aren't a slide deck: they run as live processes with scoped access to the tools they need, and every machine action they take is authenticated and logged (the same pattern that runs alfirus.my, where a nightly job publishes project news on a key that can do exactly one thing).

Honest caveat: this is infrastructure, not magic. Someone still has to design the roles, write the guardrails, and check the output — that's the actual work, and it's most of the value.

The stack

Layer	Choice
Orchestration	Multi-agent framework (Hermes Agent), self-hosted
Inference	Local LLM runtime — no cloud API dependency
Channels	Email · Telegram · WhatsApp, one context across all three
Collaboration	Cross-agent delegation with named departmental roles
Governance	Scoped access per agent + audit trail of machine writes
Cost model	Zero marginal cost after setup

Want the same for your company

The pitch is simple: your own AI team, running on your own machine, in the channels your staff already use — with a bill of zero after setup. That's one of the three things I do. The other two are field-ops software and web builds.

If that sounds like your company, request a quote. Four fields, two minutes, and I reply within one working day.