CHIEF AI OFFICER · OPERATING MODEL
AI Executive Assistants in the Enterprise
AI agents, human EAs, and how to scale the combination
Every large lab shipped a personal AI agent in 2026. Almost none of them shipped the controls an enterprise needs. This is the strategy view: what to give the agent, what to keep with your EAs, and how to roll it out without creating a new shadow-IT problem.
WHAT CHANGED
The 2026 agents, and why most are not enterprise products yet
Between May and October 2026, Google (Gemini Spark, now Task), xAI (Grok Bot), Meta (Muse), OpenAI (dots) and Anthropic (Claude Duet) each released an always-on agent with its own cloud computer. They check accounts in the background, draft and follow up, and ask before sending, paying or signing. For an individual, that is most of an assistant's queue work. For an enterprise, the question is plan eligibility, identity and data residency, and on those the field is thin.
| Agent | Business plans | Identity and approvals | Regions |
|---|---|---|---|
| OpenAI dots | Business Premium; Enterprise in beta, off by default | Own identity, credentials and tools per dot; Agent 365 integration in progress | Not EEA, UK or Switzerland |
| Grok Bot | X Subscription Teams and Enterprise; Team Bots since Sept 28, 2026 | Per-action approvals and review rules; bots share one computer per account | Computers run in the US |
| Claude Duet | Personal Pro, Max, Ultra only; Team and Enterprise later | Approval before send, pay, sign, delete; cites sources | US waitlist |
| Gemini Spark (Task) | Personal Google AI Pro and Ultra only; no Workspace accounts | Confirms before sends and purchases; scheduled tasks run unprompted | Not EEA, UK, Switzerland or Nigeria |
| Meta Muse | Consumer only | Approval before purchases; no admin console | US and Canada |
Status as published by the vendors through November 5, 2026. Product-by-product detail is on our partner site We The Flywheel's AI assistants hub.
AI VS HUMAN EA
What the agent takes, and what stays with your EAs
The split is not by seniority or by task type in the abstract. It follows two questions: can the decision rule be written down, and does the action carry trust that only a person can hold? Where the first answer is yes and the second is no, the agent owns the work. Everything else stays human, with the agent preparing.
| Work | AI agent | Human EA | Why |
|---|---|---|---|
| Inbox triage and first drafts | Owns | Spot-checks, approves sends | High volume, low judgment per item |
| Scheduling proposals | Owns, within written rules | Handles exceptions and declines that carry politics | Rules are executable; politics are not |
| Daily brief and follow-ups | Owns, with sources and timestamps | Decides priority | Agents check systems tirelessly but misjudge urgency |
| Travel and purchases | Researches and prepares | Approves and handles disruption | Every vendor gates payment behind a human |
| Board, investor and key-client relationships | Prepares material | Owns | Trust and representation stay human |
| Confidential HR, legal and M&A matters | Kept out by default | Owns | Data exposure risk outweighs the time saved |
Two failure modes decide where the line sits. Agents fail quietly on stale context: a cached figure or an out-of-date thread drives a confident wrong answer. Humans fail on hours and handoffs: work waits overnight, or the brief drifts from what the executive meant. Both are fixed the same way, by writing down the rule and making every claim carry its source and age.
OPERATING MODELS
Three models, matched to executive tiers
1. Agent-first for managers and individual contributors
People who never had an EA get an agent with a standard charter: triage, drafts, scheduling within rules, follow-ups. Approvals stay with the user. This is where most of the volume is, and where the risk of employees wiring work accounts into personal consumer agents is highest if you offer nothing.
2. EA plus agent for senior executives
The executive keeps a human EA, and the EA supervises the executive's agent. The agent drafts and prepares; the EA approves routine sends within delegated authority, handles exceptions and escalates the rest. The EA's job moves from doing the queue to owning the rules and the relationships.
3. A pooled EA team supervising agents for a leadership layer
For VPs and directors, one EA supervises agents for several leaders at once. This is where the cost case sits: the agent absorbs queue work, so a single person can hold judgment for more principals. Our partner We The Flywheel hires for a version of this role, agentic operators who coordinate AI agents and run operational tasks.
GOVERNANCE
Six controls before you scale
- Every agent is a non-human identity. Provision it like a service account, with its own credentials and an owner. See agentic AI security.
- Least-privilege connectors. Start with read access to mail and calendar; add write scopes per use case, not per vendor default.
- Approval tiers. Define which actions an agent may take alone, which need the EA, and which need the executive. Sending externally, paying, signing and deleting should never be in the first tier.
- Data boundaries. Keep HR, legal, M&A and board material out of agent scope by default. Check data residency; several agents exclude the EEA and the UK.
- Sources and timestamps. Require every briefing figure to name its source and age, so stale context is visible.
- A consumer-agent policy. Decide whether employees may connect work accounts to personal agents. Gemini Spark and Claude Duet run on personal plans, so this will happen unless the policy and a sanctioned option exist.
ROLLOUT
A phased rollout you can measure
- Baseline (weeks 1 to 2). Before any agent goes live, record the numbers you want to move: time EAs spend on scheduling and inbox, response time on executive email, briefings that had to be redone, and approvals pending at end of day.
- Pilot (90 days). One operating model, a small group of executives and their EAs, on one enterprise-eligible agent or your own platform. Charters written before launch.
- Review. Compare against the baseline. Count stale-context incidents and actions that needed correction, not only time saved.
- Expand by tier. Roll out model 1 broadly only after the identity, approval and consumer-agent controls are in place.
- Re-evaluate vendors quarterly. Enterprise plans for Duet and Gemini are announced but not shipped, and prices changed twice in October for Grok Bot alone.
Budget for usage. Background checks draw on the same quota as chat for Claude Duet and Grok Bot, so an always-on agent per executive costs more than a chat licence per executive.
For a first-person account of running an AI executive assistant alongside human support, see Thomas Prommer's AI assistant vs human executive assistant on his partner site prommer.net. For the org-design context, see the AI-native org chart.
Frequently Asked Questions
Will AI replace executive assistants?
What is an AI executive assistant?
Which AI executive assistants are ready for enterprise use?
What is an AI chief of staff?
How do you scale AI assistants across an enterprise?
