Skip to main content
FirstCoast.ai
All case studies
Case study: Professional services agency

Building an Agentic Operating System

A services company rebuilt its back office as a fleet of AI agents under one operating system: a version-controlled knowledge base as the shared brain, an approval surface in Telegram on the owner's phone, and more than twenty agents doing the daily work. Nothing reaches a client without a human tap.

How the agentic operating system is put togetherOne version-controlled git repository holds the company's services, pricing, voice, procedures, and client context, and every agent reads from it, so an answer drafted anywhere matches what the company offers and charges. More than twenty agents and automations run in production off that canon, across four functions. Sales handles lead intake and scoring on every new lead from HubSpot, and produces follow-up drafts. Billing runs invoice drafting from QuickBooks Online and retainer billing on schedule, and produces invoice drafts. Operations runs two agents: inbox operations over the day's mail in Gmail, which produces reply drafts, and meeting notes, which turns every Zoom call into a record. Reporting runs the morning brief on every active project before the workday starts, and produces a project status. The three client-facing streams, follow-ups, invoices, and replies, all converge on one approval queue in Telegram, where every outward-facing email, invoice, and proposal waits as a draft until one tap on the owner's phone releases it. Only released work reaches the client systems: HubSpot, QuickBooks Online, and Gmail, the same systems the agents read from. Nothing has ever reached a client unapproved, because the gate is architecture, not policy. The other two streams never face a client at all and never touch the gate: the meeting record, with its action items tagged by owner and confidence level, and the morning brief. Meeting records are filed back to the git repository, which is the system of record, closing the loop on the canon. The whole fleet costs a few dollars a month in metered AI spend, because every workflow is costed per run before it is built, many use no AI at all, and the rest use the smallest model that does the job: fast models for classification, stronger ones for drafting.THE SHARED BRAINGIT REPOOne version-controlled knowledge base, read by every agentServicesPricingVoiceProceduresClient contextso an answer drafted anywhere matches what the company offers and chargesMORE THAN TWENTY AGENTS IN PRODUCTIONSALESLead intake and scoringevery new lead, from HubSpotFollow-up draftsBILLINGInvoice draftingfrom QuickBooks OnlineRetainer billingruns on scheduleInvoice draftsOPERATIONSInbox operationsthe day’s mail, in GmailMeeting notesevery call, from ZoomReply draftsMeeting recordREPORTINGMorning briefevery active projectbefore the workday startsProject statusEVERY CLIENT-FACING DRAFTTELEGRAMDRAFTS ONLY, NEVER SENTOne approval queueEmail draftsInvoice draftsProposal draftsone tap on the owner’s phone releases itNEVER FACES A CLIENTMeeting recordaction items, every callOwnerConfidenceMorning briefevery active projectbefore the workday startsTHE GATE IS ARCHITECTURE, NOT POLICYClient systems updatedZERO UNAPPROVED, EVERthe same systems the agents read fromHubSpotQuickBooks OnlineGmailTOTAL METERED AI SPENDA few dollars a monthevery workflow costed per run before it is builtNo AI at allFast: classificationStronger: draftingFILED TO THE RECORD

The challenge

  • The owner was the operating system: every lead, invoice, follow-up, and status check ran through one person's memory and working hours
  • Admin work like billing, triage, meeting notes, and project status consumed hours that should have gone to clients
  • Off-the-shelf tools each solved one task but shared no context, so nothing knew what the rest of the business knew

What we built

An agentic operating system. Company knowledge lives in one version-controlled repository every agent reads from: services, pricing, voice, procedures, and client context. On top of it, a fleet of agents runs the day-to-day: lead intake and scoring, invoice drafting, retainer billing, Zoom meeting notes, inbox operations, and a morning brief that reports the state of every active project before the workday starts.

  • One knowledge base grounds every agent, so an answer drafted anywhere matches what the company actually offers and charges
  • A Zoom transcript becomes a meeting record with tracked action items, each tagged with an owner and a confidence level, so promises made on a call stop living in memory
  • Every outward-facing email, invoice, and proposal is created as a draft and waits for one-tap approval in Telegram
  • Each agent uses the smallest AI model that does the job, and many workflows use no AI at all

How it's built

A git repository as the system of record, an automation layer for scheduled and event-driven workflows, Telegram approvals, a Postgres database, and integrations with HubSpot, QuickBooks Online, Zoom, and Gmail. Model choice is per task: fast models for classification, stronger ones for drafting. Every workflow gets a per-run cost estimate before it is built.

Results

  • More than twenty agents and automations in production across sales, billing, operations, and reporting
  • Total metered AI spend: a few dollars a month
  • Zero automated messages have ever reached a client without human approval, because the gate is architecture, not policy

What would your business look like with an operating system?

The same architecture scales down to a single workflow. The Discovery Audit finds the right first agent.