This question gets framed as a contest, and that framing causes bad decisions. A founder either assumes a monthly subscription replaces a senior hire, or dismisses software because it can't do the human half of the role.
A better approach is to break the role into its parts and assign each part. We'll do that using a published list of duties, then apply the result to three companies of different sizes. If you want the definition of the role first, read what a startup chief of staff does.
AI chief of staff vs human, duty by duty
a16z's hiring guide names the usual duties: building the operating cadence, bridging teams, protecting the executive's time, leading special projects, setting goals and metrics, and running the team during an absence. Here's how each one splits. The verdicts are our judgment, based on what the tools publish and the research below.
| Duty | What software can do | What needs a person | Lean |
|---|---|---|---|
| Operating cadence | Produce the weekly agenda, collect updates, chase missing ones, write the notes | Decide which meetings should exist, and stop a meeting that has gone wrong | Mostly software, with a human owner |
| Bridging teams | Spot that two teams are planning against different dates | Sit the two leads down and get them to agree | Person |
| Protecting time | Audit the calendar, flag conflicts with priorities, draft polite declines | Say no to a board member or a senior colleague in a way that holds | Split |
| Special projects | Research, benchmarks, first drafts of a plan or deck | Own the outcome across a month of changing facts | Person, helped by software |
| Goals and metrics | Pull numbers, track progress, flag slippage early | Decide which metric matters, and have the hard talk when one is missed | Split |
| Covering an absence | Keep the status visible and the routine reports running | Make calls on the founder's behalf with authority others accept | Person |
Read down the last column and a pattern shows up. Software does well wherever the work is information moving between systems. A person is needed wherever the work is getting other people to change what they do.
Evidence on what software can finish alone
Published evidence points to real but partial ability, and you should discount any vendor who implies more. These three pieces are the most useful we read.
- A simulated company. A research team built TheAgentCompany, a small fake software firm with internal sites and coworkers. The most competitive agent finished 30% of the tasks on its own. Simpler tasks could often be automated, and longer ones stayed out of reach. The paper dates from late 2024, and models have changed since, so read it as direction and don't treat 30% as today's figure.
- A real shop. Anthropic let Claude run a small store in Project Vend. It found suppliers well, but sold items below cost, gave customers a payment account it had made up, and was easily talked into discounts. The authors wrote that they "would not hire" it. In a later phase the business improved a lot, yet the agents "still needed a great deal of human support".
- The traits of the job. Bridgespan's guide lists what strong candidates share: getting things done, a strategic mindset, emotional intelligence, and loyalty, including being confident enough to be candid with the CEO. Software can do a share of the first two. The last two depend on being a colleague with something at stake.
Three companies, three answers
Size changes the answer more than any feature does. a16z ties the job to headcount: from about 20 people the focus is the executive, from about 75 it's the leadership team, and from about 150 it's the whole organisation. Here's how that plays out. All three companies are invented to show the reasoning.
Six people, pre-seed
The founder runs sales, investor conversations and hiring, with four engineers. There's no leadership team to align, so most of the human half of the role has nobody to work on. A senior hire would also absorb a large share of the runway, as the cost breakdown shows. Use software, or nothing. The realistic gain is a founder who stops missing things across email, pipeline and cash.
Thirty people, seed or early Series A
Now there are four or five team leads and a weekly leadership meeting that ends without clear owners. Here a person starts to pay off, because the problem has moved from noticing to getting people to act. If the budget is there and you can name the three problems you would hand over, hire. If not, try software with a named internal owner, such as an operations lead, for a quarter, and see whether the gaps are in information or in follow-through.
Ninety people, Series B
The leadership team needs a real operating rhythm, and the CEO can't personally resolve every cross-team dispute. A person is the answer here. Software then becomes that person's first-pass layer: reports assembled before they start, numbers pulled without asking three managers, and drafts ready to edit.
The risks on each side
Each option fails differently, and the failures matter more than the feature lists when you're choosing.
| Risk | Software | A person |
|---|---|---|
| Getting it wrong | States a wrong fact with full confidence. One research paper blames incentives that favour a guess over saying "I don't know" | Makes a bad call, but can usually explain the reasoning and take the feedback |
| Scope creep | An agent takes initiative you didn't intend, as Anthropic's incident log shows with deleted branches and a production database migration attempt | Becomes a shadow executive that people feel they have to go through |
| Vague mandate | Produces a tidy briefing nobody acts on | Fills the time with coordination. Underdog.io warns that if a CEO delegates tasks but not trust, the role collapses into overhead |
| Security | Can be manipulated through content it reads, and OWASP doesn't know of a sure fix | Holds sensitive information and can leave with it, which is why trust comes first in hiring |
| Retention | Depends on the vendor staying in business and keeping prices stable | a16z expects a 2 to 3 year stint, with burnout and career growth among the reasons |
The software risks shrink with one habit: review everything that leaves the company. That has a cost in attention, and human in the loop AI agents covers how to set it up without drowning in approval requests. The human risks shrink with a clear written mandate and a single reporting line, both of which a16z recommends.
Should you run both?
Yes, once you have a person, because the pairing removes the dullest part of their week. Picture a chief of staff who spends every Monday morning collecting updates from six managers. If software assembles that before they start, those hours go to the conversation one of those managers has been avoiding. That's the part of the job a good hire joined to do.
Two rules make the pairing work. Someone human owns every output, so a wrong number in the board pack belongs to a named person. And the person sets the policy for what the agent may do on its own. Don't let the tool's defaults decide that.
One warning for founders hoping to skip the hire with a stack of tools: a second agent checking the first isn't the same as a person checking. When Anthropic added an AI manager to Project Vend, it had much the same weaknesses as the shopkeeper it supervised, since both ran on the same model.
A simple way to decide this month
- List every recurring coordination task you did in the past two weeks, with rough hours next to each.
- Mark each as information work (collect, summarise, track, draft) or people work (persuade, confront, decide with others).
- Add up both columns. If information work is most of it, start with software. If people work is more than about a third, and you have at least 20 staff, begin scoping a hire. That one-third line is our rule of thumb and hasn't been tested in any study.
- Check the budget. Compare the monthly cost of a hire against your runway before you write a job description.
- Review after 90 days. Ask which gaps remain. If drafts are good and follow-through is still poor, software has reached its limit for your company.
If you try the software route, Nautis is one option, and there are others grouped by job in the tools roundup. Its plans cost $49 to $129 a month (as of October 2026), there's a 14-day trial, and its Chief of Staff defaults to asking before it acts. The company launched publicly in July 2026, so independent reviews are still few, and it doesn't replace a human chief of staff. It's also useful alongside one: the page for chiefs of staff covers that setup.
FAQ
Can AI replace a human chief of staff?
Not fully. It can take on monitoring, summaries and first drafts, but it can't hold authority with your team or manage a sensitive conversation. Many small companies use software because a senior hire isn't affordable, which is a different decision from replacing one.
At what company size should I choose a human over AI?
a16z's guide starts its chief of staff levels at about 20 people, and that's a sensible point to begin considering a hire. Headcount alone doesn't settle it. The stronger signal is that your problem has moved from missing information to getting people to follow through.
Is an AI chief of staff cheaper than a human one?
Yes, by a wide margin, but they aren't doing the same work. A person on a $185,000 base comes to about $224,000 a year once employer costs are added, on our estimate. Software usually costs hundreds or a few thousand dollars a year, and someone on your team still has to review its output.
Should a human chief of staff use AI tools?
Yes, for assembling reports, pulling numbers and first drafts. That frees their time for the work only a person can do. They should still own every output and set the rules for what the tool may do on its own.
Sources
- a16z: how to hire a chief of staff (checked Oct 31, 2026)
- TheAgentCompany: benchmarking LLM agents on real-world tasks (arXiv) (checked Oct 31, 2026)
- Anthropic: Project Vend, phase one (checked Oct 31, 2026)
- Anthropic: Project Vend, phase two (checked Oct 31, 2026)
- Bridgespan: what a chief of staff is and when you know you need one (checked Oct 31, 2026)
- Kalai et al.: why language models hallucinate (arXiv) (checked Oct 31, 2026)
- Anthropic: Claude Code auto mode (checked Oct 31, 2026)
- OWASP: LLM01 prompt injection (checked Oct 31, 2026)
- Underdog.io: how to hire a chief of staff for your startup (checked Oct 31, 2026)