Most writing about an AI chief of staff describes a single, impressive capability: a morning digest, a drafted reply, a dashboard that updates itself. In practice, the role is wider than any one trick. It is a catalog of small, repeatable jobs that would otherwise fall between a founder's calendar, a sales lead's inbox and a finance lead's spreadsheet.
What follows is that catalog. Twelve functions, grouped by the team that benefits most, each with a public statistic that explains the cost of leaving the work undone, and each tagged with a safety class so you know what the agent does on its own and what waits for a human. The classes are simple: read (looks only), reversible write (creates a draft, a task, a dashboard tile), and sensitive with approval (sends, publishes, deploys, spends, deletes, waits on a person).
Where should you deploy an AI chief of staff first?
Sales: The Pipeline That Goes Quiet
Sales leaks happen in silence. A deal stalls on day nineteen because the champion changed jobs; a lead sits unopened because the rep is in back-to-back demos; a renewal gets no touch because nobody owns it this quarter.
1. Pipeline review automation. The agent reads the CRM, flags deals with no activity in N days, missing next steps, or an inactive champion, and drafts the review doc your sales lead would have written on Sunday night. Safety class: read, with a reversible write for the draft. The cost of not doing it is well-documented: 80% of sales require at least five follow-up contacts, yet 44% of reps quit after just one attempt.
2. Follow-up drafts. Replies to specific threads, grounded in the last meeting notes and the deal stage, queued in the rep's drafts folder. Sensitive with approval. Email never sends without a person; the approval is bound to the exact body and recipient.
3. Inbound responder. New leads from forms or email get a routed, drafted reply with calendar links from your booking pages. Sensitive with approval. The statistic that keeps this one on the list: leads contacted within five minutes are 21x more likely to convert than those reached after 30 minutes, but the average company takes 42 hours to respond.
Marketing: The Dashboard Nobody Opened
Marketing teams produce more data than they can read. The agent's job is to read it for them and surface the two or three things that changed since yesterday.
4. Content operations. Briefs, drafts, outlines and SEO checks against Search Console data, staged inside your headless CMS. Publishing itself is sensitive with approval. Everything upstream is a reversible write.
5. Search Console and Analytics watch. A daily read of which URLs lost impressions, which queries moved, which pages broke. The agent explains the delta, not just the number. Read.
6. Ad budget guardrails. Google Ads pacing, creative fatigue, cost-per-acquisition drift. Budget changes are the sharp end of marketing work and never happen autonomously. Sensitive with approval, bound to the exact payload, as described in the piece on why AI agent approvals must be bound to the exact action.
Finance: The Invoice That Was Never Chased
Finance is a protected domain in Brigham. Access is granted per person, never by role, and no money moves without a human on the payload. The agent still does a lot of useful reading.
7. AR chasing drafts. Reads Stripe and the accounting system, lists invoices overdue by bucket, drafts a short, polite nudge for each. Reversible write for the draft; sending is sensitive with approval.
8. Monthly close and board pack. Pulls revenue, cash, pipeline and headcount into a draft deck or doc with the same section order every month. The CFO edits once; the agent remembers the edits. Reversible write. The reason it earns its keep: knowledge workers spend only 39% of their workweek on role-specific work, with the rest consumed by communication overhead and tool switching. The board pack is the archetype of that overhead.
Engineering: The Deploy That Rolled Back at 03:12
Engineering teams generate the clearest telemetry in the company. An AI agent for operations in engineering is a reader first and an actor only with explicit permission.
9. PR triage and reviewer nudges. Reads GitHub, surfaces PRs waiting on review for more than N hours, drafts the nudge, labels the stale ones. Writing code and opening PRs are reversible writes; merging and deploying are sensitive with approval.
10. Deploy and incident watch. Watches CI, flags failed deploys, assembles the incident doc. The scale of this problem is why it matters: the 2024 DORA report found change failure rates ranging from roughly 5% at elite teams to 40% at low performers. The agent does not make the deploy call; it makes the context ready when someone else does.
Operations: Status Chasing, Automated
The last three functions are the ones that compound across every team. They are also the ones founders personally resent most.
11. Morning briefings per person. One briefing, tailored to the person's responsibilities, delivered before their day starts. The format and discipline is covered in detail in the morning briefing checklist. Read, delivered as a reversible write to the person's private inbox.
12. Meeting prep, scheduling and business memory. Pulls the last three touches with the attendee, the open threads, the relevant notes. Business memory is what makes the briefing specific instead of generic: facts with sources, last-verified dates, and provenance you can audit. Read; scheduling that changes someone else's calendar is sensitive with approval.
The cost of not doing these last three is the biggest number in the whole piece. Asana's Anatomy of Work Index found the average knowledge worker spends 60% of their time on 'work about work': chasing updates, status meetings, searching, switching apps. That is the surface area an AI chief of staff is designed for.
How to Choose Where to Start
Twelve functions is a menu, not a mandate. The sensible path is to pick the one with the highest cost of neglect in your business this quarter, deploy it against a single team, and expand once the audit trail is boring. Three heuristics help.
- Pick a function with a known loss. If deals go dark for weeks, start with pipeline review automation. If invoices age, start with AR drafts. The point of pairing each function with a public statistic is to make the loss legible before you spend engineering time on it.
- Favor read before write, and reversible write before sensitive action. A morning briefing earns trust in a week. An ad budget change does not.
- Insist on policy in code and payload-bound approvals. Role-based access controls on their own are not enough once an agent can act. The requirement that high-risk AI systems technically allow for the automatic recording of events over the lifetime of the system is a floor, not a ceiling. Logs are what let you grant more authority later without flying blind.
If you want to see which systems the agent reads and acts on day one, the product page and the integrations list are the right next click. The functions above are the shape of the role. The safety classes are the shape of the trust you extend to it. Start with the function whose statistic made you wince, and leave the sensitive actions behind an approval until the audit trail reads clean.
