AI Chief of Staff for SMBs

Why this exists

If you run a small business, you already know the problem. You're drowning in email, follow-ups, calendar Tetris, CRM updates, and a hundred tiny decisions a day. You can't afford a full-time Chief of Staff — that's a $150,000+ hire — but you also can't grow without help.

An AI Chief of Staff is the realistic answer in 2026. Not a chatbot. An always-on assistant that drafts your follow-ups, prepares your day, keeps your CRM honest, and surfaces the one decision you actually need to make today. It learns your business over weeks. It coordinates sub-agents for specialized work (writing, voice calls, scheduling). And it costs less than a single SaaS subscription used to.

This playbook is the same approach Peak Agent AI uses to deploy assistants for its clients. The reference implementation, Stefan, has been running our CEO's day continuously since January 2026: 90+ days, 11 cron jobs, 221-contact CRM, autonomous follow-ups, integrated across Gmail, Calendar, WhatsApp, Brave/Chrome, and n8n. You can run the same pattern.

What an AI Chief of Staff actually does

A real AI Chief of Staff handles seven recurring patterns. Master these and you've built one.

1. Anti-repetition. Tracks what's been surfaced to you. Doesn't tell you the same thing twice. Escalates on day 3 if you haven't acted, archives on day 7 if it's no longer relevant.
2. WAT framework. Workflows, Agent, Tools. Every recurring task gets a written workflow (the SOP). The agent executes the workflow using its tools. No SOP, no automation.
3. Daily rhythm. Morning brief (priorities, key emails, calendar conflicts). Midday pulse (what's slipping, urgent follow-ups). Evening wrap (what got done, what's queued for tomorrow).
4. Draft, don't remind. When follow-up is needed, the assistant drafts the actual message in your voice — not a vague "you should follow up with Sarah."
5. Outcomes only. Reports results, not process. "Drafted reply to Sarah, sent to your inbox for review" — not "I checked your inbox and decided to draft a reply."
6. Silence is golden. Empty sections get skipped. If nothing's urgent, the morning brief is short.
7. Autonomous first. Acts when it has high confidence. Asks when it doesn't. Reports either way.

The core stack

You don't need every tool. You need this minimum viable AI Chief of Staff stack.

Layer 1 — The brain

The LLM. Claude is what we run on for business judgment and writing. Anthropic API for production. ChatGPT works too — pick one and stop debating.

Layer 2 — The runtime

Something that lets the brain actually do things on a schedule. For Peak Agent clients we use Hermes (our internal multi-tenant runtime). The open-source equivalent for SMBs is n8n — visual workflows with AI nodes, self-hostable on a $20/mo VPS. Zapier or Make.com if you're allergic to anything that looks technical.

Layer 3 — The channels

Where you actually receive the assistant's output. WhatsApp is the highest-signal channel — use Kapso for the official Meta Cloud API. Email is fine. SMS via Twilio works. iMessage if you're Mac-only. Pick one. Don't try to support all of them.

Layer 4 — The integrations

What the assistant reads from and writes to. Bare minimum:

• Email: Gmail API or Microsoft Graph.
• Calendar: Google Calendar API.
• CRM: GoHighLevel for service businesses, HubSpot or Pipedrive for B2B.
• Notes/docs: Notion or markdown files.
• Meeting notes: Granola or Fireflies.

Layer 5 — Sub-agents (optional, add as you grow)

For specialized work, your assistant should be able to call other agents:

• Voice calls: VAPI for inbound/outbound phone.
• Content engine: Claude or Jasper for blog/social.
• Image gen: Nano Banana 2 or Recraft.
• Video: Runway or Luma Dream Machine.

Deep dive 1: Inbox triage and drafted replies

The single highest-ROI capability. Most SMB owners spend 1.5 to 3 hours a day in their inbox. An AI Chief of Staff should cut that by at least 50%.

The workflow:

1. Every 30 minutes, the assistant pulls new emails since last check.
2. It classifies each one: needs reply, FYI, junk, calendar-related, follow-up-pending.
3. For each "needs reply" message, it drafts a response in your voice based on a SOUL.md file (more on this below).
4. Drafts are saved to your Gmail Drafts folder OR sent to your WhatsApp for review.
5. You approve or edit, then send.

What makes this work versus generic auto-reply: the assistant has read every email you've sent in the last 90 days. It knows your style, your common closings, your relationships. It drafts in your voice. You don't proofread — you skim.

Prompts to start with: Draft my follow-up (don't remind) and Morning brief from my AI assistant.

Deep dive 2: Calendar and follow-up management

The second-highest-ROI capability. If you've ever lost a deal because you forgot to follow up, this is your fix.

The workflow:

1. Every morning, the assistant looks at every conversation in the last 30 days that has gone quiet.
2. It cross-references with your CRM (or a simple "follow-ups" sheet).
3. For each one, it decides: still warm? Drafts a follow-up. Cold? Archives.
4. Calendar prep: 24 hours before any meeting, it generates a one-page brief on the attendee — last interaction, what they care about, the next step you should propose.

Prompts to start with: Pre-pipeline review brief and Strategic account 1-pager.

Deep dive 3: The SOUL.md file

This is the secret. SOUL.md is a markdown document that defines your assistant's identity, your patterns, your voice, and your business context. Every prompt loads it.

A good SOUL.md has:

• Identity. Name of your assistant. How it should refer to itself. Tone.
• Patterns. How you write. What you don't say. Closing styles.
• People. Your top 20 contacts, their context, the running thread.
• Business context. Your offers, pricing, sales process, current quarter goals.
• Operating rules. When to act vs. ask. What never to send without review. Privacy boundaries.

A weak SOUL.md is generic. A great one reads like your senior EA's onboarding doc. Plan to spend 4 to 8 hours on the first version. Iterate weekly for the first month.

Deep dive 4: Daily rhythm and digest

The assistant earns its keep with three daily touchpoints.

Morning brief (delivered 7am local time):

• Top 3 priorities for today (pulled from CRM/notes).
• High-signal emails awaiting reply.
• Calendar review with conflict flags.
• Follow-ups overdue.
• One decision worth making today, with the trade-off.

Midday pulse (delivered 1pm):

• What's slipped from this morning's plan.
• Urgent emails received during meetings.
• 3 prospects who responded — draft replies attached.

Evening wrap (delivered 6pm):

• What got done.
• What didn't, and why.
• Tomorrow's first focus block.
• One reflective sentence on the day.

Use these prompts as starting points: Daily brief, Evening wrap, Weekly business review.

Governance and risk

Things that will bite you if you skip them.

• PII in public LLMs. Never paste customer SSNs, account numbers, or health info into ChatGPT/Claude without an enterprise plan with zero-data-retention.
• Auto-send is a foot gun. For the first 60 days, every assistant-drafted message goes to your draft folder for review. Only after you trust the patterns do you start auto-sending low-stakes confirmations.
• CRM hygiene. Your assistant is only as good as your CRM. If you have stale contacts, dead deals, and missing fields, the assistant will hallucinate. Spend a weekend cleaning before you deploy.
• Approval thresholds. Define what the assistant can do autonomously vs. what needs your sign-off. Example: it can confirm a meeting. It cannot change a price.
• Logging. Every action the assistant takes gets logged with a timestamp. If something goes wrong, you can audit.

30-60-90 day rollout

Days 1-30: Foundation.

• Pick your LLM and runtime (Claude + n8n is our recommendation).
• Write your first SOUL.md (4 to 8 hours).
• Wire up email and calendar integrations.
• Deploy the morning brief and evening wrap.
• Run for 2 weeks. Tune daily.

Days 31-60: Expand.

• Add inbox triage with drafted replies.
• Add WhatsApp or SMS as a channel.
• Add CRM read access (start with read-only).
• Build your first 10 voice patterns into SOUL.md.

Days 61-90: Standardize.

• Move from drafts-only to autonomous low-stakes actions.
• Add the first sub-agent (content or voice).
• Document the SOPs for everything the assistant handles.
• Measure: hours saved per week, follow-ups recovered, response time on key prospects.

Maturity model

• Level 1 (you're here): The assistant drafts. You approve.
• Level 2: Autonomous on routine low-stakes work. Human in the loop for anything money-related or relationship-critical.
• Level 3: Multi-agent coordination. Your AI Chief of Staff orchestrates content, sales, and ops sub-agents. You spend your time on strategy.
• Level 4: The assistant runs the operating cadence of the business. You spend your time on people, partnerships, and big bets.

Most SMBs land at Level 2 within 90 days. Level 3 within a year if you keep investing.

What to skip

• Don't try to build a "general AI agent." Build for your seven specific patterns.
• Don't roll your own LLM. Pay Anthropic or OpenAI. Your moat is your workflow, not your model.
• Don't onboard 5 people at once. Get yourself working first. Then expand.
• Don't trust the assistant with your money in month one. Trust is earned.

Where to go next

• Browse the Prompts library filtered to "AI Chief of Staff."
• Look at the Operations workflows for ready-to-deploy patterns.
• If you'd rather have the whole thing built and run for you, peakagentai.com is what we sell. Starts at $149/mo.