The AI
Operating System.
How AI agents work, what they can automate for a 15-person firm, and honest math on the impact.
A reference · Written by Dara O’Beirne · Sacramento 2026 · 30 min read
Seven parts. Read what you need.
Seven parts. Each stands alone. Read the ones you need when you need them.
- What “AI agents” actually are: the vocabulary and mental model
- The 2026 landscape: major models, capabilities, cost trends
- Every workflow AI can automate for a small business: organized by function
- The math: realistic impact on time, cost, revenue
- How to start: the recommended sequence
- What to avoid: common failure modes
- Where to go deeper in your vault
Total read time: 30-40 minutes. Skim the parts that already feel obvious.
If you’d rather have the pieces without the scrollytelling: /parts/1 opens Part 1 as plain reading. /glossary, /workflows, and /math are direct paths to the parts most people search for.
Part One · The vocabulary problem
Most confusion about AI comes from vocabulary.
- LLM.
- The underlying neural network. Claude, GPT, Gemini, Llama, Grok are all LLMs. An LLM by itself does one thing: given text (and sometimes images), predicts what text should come next. Doesn't take actions. Doesn't have persistent memory. Doesn't run itself.
- Chatbot.
- An LLM wrapped in a conversational interface. You type; it responds. Answers questions but doesn't do things beyond talk.
- Assistant.
- A chatbot plus tools. Can read your email, access your calendar, look things up. Still primarily conversational: you ask; it does one thing; you ask again.
- Agent.
- An LLM that plans multiple steps, uses tools, observes results, revises its plan, and completes a job. Given a goal, works toward it across dozens or hundreds of steps without waiting for you to prompt the next one.
The shift from chatbot to agent is the one that matters. Chatbots save minutes. Agents save hours.
What makes an agent useful.
Three properties turn a chatbot into a working agent:
Persistent memory. The agent remembers earlier steps in the same task, and (where wired) remembers earlier tasks from days or weeks ago. Memory means the agent doesn’t ask the same question twice.
Tool use. The agent can call functions: hit an API, query a database, send an email, run a script. Every action the agent takes in the world is a tool call.
Judgment about when it’s stuck. A good agent recognizes when the LLM is uncertain (low confidence, contradictory tool responses, repeated failures) and escalates rather than confabulating. This is the property that makes 2026 agents deployable in a business setting where a hallucination has real consequences.
Add these together and you get: a system you can hand a goal to, walk away from, and come back to a finished result, with an audit trail of every step it took.
The agent loop
Every agent follows the same loop.
That’s the architecture. What makes 2026 agents useful: the loop runs for hours across hundreds of steps, and the LLM is smart enough to recognize when it’s stuck and escalate rather than hallucinate.
Part Two · The 2026 landscape
What the market looks like right now.
In July 2026, the space is dominated by a handful of frontier models and a rapidly-improving mid-tier. Two implications for small businesses: (1) You almost never need the frontier model for production workflows. (2) The cost curve has already made agents economically viable for tasks that were laughable in 2023.
The pattern that keeps costs sane.
The small-business pattern for model choice:
- Haiku 4.5 for anything classification-shaped: categorize this email, extract these fields, label this transaction. Cheapest, fastest, reliable when the task is well-defined. Roughly 90% of production LLM calls in a good workflow.
- Sonnet 4.6 for drafting content a human will send: emails to clients, summaries of documents, drafts of proposals. Best judgment-to-cost ratio. Roughly 9%.
- Opus 4.6 for reasoning-heavy tasks where quality is the entire product: legal analysis, complex strategic memos. Roughly 1%.
Route requests to the model that matches the task. Don’t route everything to Opus and don’t route everything to Haiku. This routing is the single biggest lever on your monthly bill.
The minutes-vs-hours shift
Chatbots save minutes. Agents save hours.
Chatbot
What’s this month’s revenue?
$47,232 this month.
Agent
Compile this month’s financial digest
- query stripe
- query quickbooks
- reconcile transactions
- detect anomalies
- draft summary
- format email
- send to cfo
- log to sheet
Chatbot: 1 message = 1 minute saved. Agent: 1 goal = 4 hours saved.
What agents can automate for a 15-person firm.
Below are ten of the most common small-business workflows agents can run in production. This is a sample; the full catalog lives at /workflows.
Each workflow is: a schedule or event that triggers it, the tools the agent uses, the deliverable, and the realistic time saved.
Weekly financial digest. Trigger: Monday 8am. Tools: Stripe + QuickBooks + email. Deliverable: 1-page email to the founder with revenue, expenses, cash runway, top 3 anomalies. Time saved: 2 hours/week.
Lead capture + CRM enrichment. Trigger: form submission webhook. Tools: enrichment API + CRM + Slack. Deliverable: contact + deal record + lead-scored routing to the right salesperson. Time saved: 3 hours/week.
Customer support triage. Trigger: new email in support inbox. Tools: LLM classifier + KB search + CRM. Deliverable: severity-scored ticket + KB-linked draft response for the agent to send. Time saved: 8 hours/week for a 3-agent team.
Invoice categorization. Trigger: new email with invoice attachment. Tools: vision LLM + QuickBooks. Deliverable: draft bill in the accounting system awaiting reviewer sign-off. Time saved: 6 hours/week.
Appointment reminders. Trigger: cron every 15 minutes. Tools: calendar + SMS + email. Deliverable: personalized 24-hour and 1-hour reminders to attendees. Time saved: eliminates a $2,000/mo no-show cost for many service businesses.
Meeting notes to CRM. Trigger: recording processed webhook. Tools: transcript + LLM summarizer + CRM. Deliverable: meeting summary + action items assigned + deal timeline updated. Time saved: 4 hours/week.
Review request automation. Trigger: purchase or service completion. Tools: LLM personalizer + email + review platform. Deliverable: personalized review request 48h after service. Time saved: 2 hours/week + 3x review response rate.
Social media cross-posting. Trigger: post ready-to-publish tag. Tools: LLM adapter + platform APIs. Deliverable: platform-tuned versions posted to LinkedIn, X, Threads, Bluesky. Time saved: 3 hours/week.
Contract expiration reminders. Trigger: cron daily 9am. Tools: CRM + LLM drafter + email. Deliverable: draft renewal outreach at 90/60/30/14/7 days before expiration. Time saved: prevents 2-4 lost accounts per year.
PTO request router. Trigger: Slack command or form. Tools: calendar + HR record + Slack. Deliverable: manager notification + coverage-conflict check + calendar update on approval. Time saved: eliminates a Slack subthread per PTO request.
Full catalog of 30+ workflows at /workflows.
Weekly digest · 2h; Lead capture · 3h; Support triage · 8h; Invoice categorization · 6h; Appointment reminders · $2k/mo; Meeting notes · 4h; Review requests · 3x; Social posting · 3h; Contract renewals · 4 accounts; PTO routing · 15m
Sixty-plus workflows in production somewhere at time of writing. The pattern is the same. The compounding is real.
The honest math.
Cost side (Anthropic + workflow runtime + hosting for a 15-person services firm running ~10 active workflows):
- Anthropic API total: $30-80/mo
- Self-hosted workflow orchestration (n8n on a $6/mo Hetzner box): $6/mo
- LiteLLM gateway (free, self-hosted): $0
- LangFuse observability (free tier): $0
Total incremental cost: under .
Time saved side (same firm, same 10 workflows):
- Bookkeeping cleanup: 6 hrs/week saved
- Customer support triage: 8 hrs/week saved for a 3-agent team
- Lead capture + follow-up: 3 hrs/week saved
- Appointment reminders + no-show reduction: 40% no-show reduction (industry norm)
- Meeting notes + CRM update: 4 hrs/week saved
- Weekly financial digest: 2 hrs/week saved
- Social + content cross-posting: 3 hrs/week saved
Total time saved: roughly .
If those hours were staffed by a VA at $25/hr, that’s $2,500/mo. If they were staffed by a bookkeeper at $75/hr, that’s $7,500/mo. The savings are real, and they are largely unaffected by whether you’re running Haiku or Sonnet under the hood, because you’re routing by task type.
The correct framing is not “AI is cheap” or “AI is expensive.” The correct framing is: for the specific workflows above, agents cost about 3% of the equivalent human labor at 25-hour-per-week volumes.
The crossover happens at 100 tasks/month for most workflows. Every workflow above that volume is agent-cheap.
The recommended sequence.
The single most common failure mode is trying to build a general-purpose “AI assistant” first. Don’t. Pick one workflow. Pick a workflow that (a) is painful this month, (b) has a clear deliverable, (c) has a bounded number of tools. Build that. Then build the next.
Weeks 1-2: Pick a workflow. Write the requirements as a paragraph. Not a spec. A paragraph.
Weeks 3-4: Deploy an n8n instance ($6/mo Hetzner or n8n Cloud starter). Wire the workflow. Every LLM call goes through LiteLLM with a per-workflow cost cap.
Weeks 5-6: Run the workflow read-only for a week. Route errors to a shared alert channel. Fix what’s flaky.
Weeks 7-8: Enable writes. Log everything. Watch for drift.
Month 3 onward: Add the second workflow. Then the third. Never try to build ten at once.
The pattern is: pick a small painful thing, ship it, prove it, repeat. Nothing else works reliably.
Part Six · Failure modes
1. Building a general-purpose assistant first.
Pick one workflow, ship it, learn from it. General assistants take 6 months to work reliably; specific workflows take 3 weeks.
2. Skipping the cost cap.
Every workflow gets a per-workflow monthly USD cap in LiteLLM. Skip this and you'll wake up to a $2,000 API bill.
3. Trusting LLM output for classification.
Use deterministic rules for classification wherever possible. Reserve the LLM for drafting and reasoning. This is the pattern behind Hedgi.
4. No human in the loop for consequential actions.
Every workflow that writes to a client-facing system (email, invoice, contract) has a human sign-off before the write. The workflow drafts; the human approves.
5. Ignoring the audit trail.
Every LLM call is logged in LangFuse. Every write is logged in your workflow log. If you can't answer "what did the agent do at 3:07pm last Thursday" you cannot deploy this in a business.
Where to go deeper.
If this reference clarified where AI fits in a small business, here are the deeper resources, all sites in Dara’s portfolio, all sourced from the same body of work as this document:
- The Repeatable AI OS at aios.dev: the platform behind everything above. Three-layer architecture, two-day installs.
- 100 Hermes Services at hermescatalog.com: the productized-service catalog. Each entry has a named buyer, budget, and Hermes-engine spec.
- Hedgi at hedgi.com: the safe AI-assisted accounting layer for QuickBooks-based firms.
- The Master Plan at themasterplan.co: Dara’s public strategy document.
- The Spatial Intelligence Stack at spatialintelligencestack.com: the book covering GIS + AI agents + Hermes in practice.
If you’d like to talk about how any of this applies to your specific business, book an intake call with Inish Labs.
Part Five · Sequence
The four-week starter path.
- Week 1: pick a painful workflow. Write it as a paragraph.
- Week 2: deploy an n8n instance. Wire LiteLLM with a cost cap.
- Week 3: build the workflow. Route to a shared alert channel.
- Week 4: run read-only for a week. Fix what's flaky. Enable writes.
- Month 3: add the second workflow.
The pattern is compounding. Every workflow you ship makes the next one easier. Every LLM call teaches you what the model is good at. In eighteen months you have a real internal AI capability.
Start with one workflow.
Ship it.
Then ship the next one.
If you want to talk it through, book an intake call with Inish Labs. Otherwise: pick the painful workflow and start today.