A capable founder can get a decent morning brief out of a general AI assistant in an afternoon. That demo is what makes building feel cheap. The harder work begins in week three, when it misses something important, nobody knows why, and a colleague asks for the same thing with different permissions.
This guide covers three paths, then the parts every path has to supply, and finally an honest sum in hours. Prices are as of October 2026 from each vendor's own page. Hours are our estimates, reasoned from the steps involved, since we found no study measuring them. We make a product in the third group, so check our reasoning against your own situation.
The three paths
Path 1: Claude or ChatGPT plus connectors
You use a general assistant, connect your tools, write instructions, and schedule it to run. Claude's custom connectors link it to remote MCP servers, and its pricing page lists scheduled tasks on Pro, Max and Ultra. ChatGPT offers plugins, which bundle skills and MCP servers, and its pricing documentation lists connectors, MCP and scheduled tasks in the plan matrix. We couldn't confirm exact availability per plan from that page, so check the live table.
The strengths are real. You get a very capable model, you can ask it anything, and you're not tied to anyone's idea of what a chief of staff should do. The weakness is that it starts each job with no structure: you're supplying the sources, the ranking logic and the memory.
Path 2: An agent builder such as Lindy
These tools give you the frame for scheduled and event-driven work, along with many integrations. Lindy's site lists 1,500+ integrations plus MCP, routines such as a morning brief or a weekly report, and memory stored in plain files you can edit. It says setup takes minutes with no API keys, and its pricing page says anything with outside impact waits for approval. See Nautis vs Lindy for a closer comparison.
You skip most of the plumbing, but you still have to decide what it watches, how it ranks things and who sees what. Lindy's Slack-first setup is a good fit if your company already runs from Slack and your records are spread across tools you want to keep.
Vendor risk applies across this group. OpenAI's own Agent Builder documentation says the product is being deprecated, with shutdown scheduled for 30 November 2026. Anyone who built on it now has to move. That's a good reason to keep your instructions and logic in a form you can carry to another tool.
Path 3: Buy a product built for the job
Here, the sources, ranking, approvals and history come ready-made. Some products are built around email, some around company records. The trade-off is less control, and usually a requirement to keep your records in the product, because that's what it reads. If the way it ranks items doesn't suit you, you may only be able to tune it, and you can't rewrite it. The tools roundup groups the options by the job they do.
A note on writing code
There's also a fourth route, for a team with engineers: writing your own agent with an SDK. OpenAI's Agents SDK, for example, supports tools that need approval and runs that pause and resume. It gives you the most control. For a chief of staff use, it rarely makes sense unless operations tooling is close to your core product, because it uses the engineering time you'd otherwise spend on what customers pay for.
What every version has to supply
Whichever path you take, these seven parts have to exist somewhere. Use this table to see who's responsible for each.
| Part | Claude or ChatGPT plus connectors | Agent builder | Product |
|---|---|---|---|
| Access to company data | You connect each source and check its scopes | Connect from a catalog | Native records, sometimes plus connected tools |
| Schedule and triggers | Scheduled tasks, within plan limits | Built in | Built in |
| Ranking logic | You write and maintain the instructions | You design it | Built in, and you tune it |
| Memory of decisions and context | Assistant memory or files you maintain | Built in, editable | Lives in the product's records |
| Approval before writes | Confirmation prompts on write actions, and you must read them each time | Offered, and you configure the details | Usually built in. Check the default |
| Audit trail a colleague can read | Often just chat history, unless you build more. Check your plan | Varies by product and plan. Check before relying on it | Should be built in. Ask to see it |
| Team permissions | Depends on plan and connector | Varies by product and plan | Usually role-based |
The last three rows are where do-it-yourself setups tend to fall short. One founder working alone can live without them. A company with an investor asking who approved a message can't.
The honest sum in hours
Here's our estimate of the time for a founder at a 10-person company who wants a daily brief across email, calendar, tasks, pipeline and cash, and follow-up drafts. Real figures depend heavily on how tidy your systems are and your technical confidence. Change them to match your own case.
| Activity | Claude or ChatGPT plus connectors | Agent builder | Product |
|---|---|---|---|
| Finding and connecting sources | 6 | 3 | 2 |
| Writing instructions and ranking logic | 12 | 8 | 1 |
| Testing on real data for two weeks | 8 | 6 | 4 |
| Moving records into a new home | 0 | 0 | 15 |
| Setup total | 26 | 17 | 22 |
| Tuning and repairs, per week | 1.5 | 1 | 0.5 |
| Tuning over 48 weeks | 72 | 48 | 24 |
| Year one hours | 98 | 65 | 46 |
A product isn't quicker to start. Moving records takes real effort, and it's the cost vendors tend not to mention. If your records are already well organised in tools you'd keep, then the agent builder's total could beat a product's. The product pays back later, because you maintain less.
The sum in money
To compare the paths, price your hours. Assume $150 an hour for a founder, which is an assumption you should change. Most early founders would say their time is worth more than that during a raise.
| Path | Subscription for a year | Time at $150 an hour | Total |
|---|---|---|---|
| Claude Pro or ChatGPT Plus | $20 × 12 = $240 | 98 × $150 = $14,700 | $14,940 |
| Lindy Team, 3,000 credits | $29.99 × 12 = $359.88 | 65 × $150 = $9,750 | $10,110 |
| Lindy Team, 15,000 credits | $99.99 × 12 = $1,199.88 | 65 × $150 = $9,750 | $10,950 |
| Nautis Launch | $49 × 12 = $588 | 46 × $150 = $6,900 | $7,488 |
| Nautis Voyage | $99 × 12 = $1,188 | 46 × $150 = $6,900 | $8,088 |
Read this with care. Time drives every row, so a change in our hours estimates would change the order. For someone who likes tuning prompts and would do it anyway, the time isn't really a cost, and the first row is a bargain. Also, a Claude Pro plan costs $17 a month if paid annually, Claude Team starts at $25 a seat monthly or $20 billed annually, and ChatGPT Business is $25 a user monthly or $20 billed annually. A team on either pays more and needs its permissions thought through.
Usage limits matter too. Claude and ChatGPT plans have their own limits, ChatGPT's documentation warns that every MCP server adds context and uses more of your limit, Lindy sells extra credits at $10 per 1,000, and Nautis caps AI credits by plan. Run a full real week during any trial to see what you'd actually use.
The risks you take on when you build
Cost isn't the only thing that shifts to you. Three risks deserve a straight look, and the vendors' own documents warn about them.
- Unverified servers. Anthropic says custom connectors reach services it hasn't verified, that malicious servers may include hidden instructions, and that Claude can read, change or delete data in connected apps. OpenAI recommends not connecting a custom MCP server unless you know and trust the underlying application.
- Prompt injection. OpenAI's same page says that trusting the developer isn't enough, because you'd also need to trust all content the server can access. A brief that reads inbound email reads text written by strangers. This risk applies to bought products as well, but when you build, you're the person responsible for spotting it.
- Key-person dependency. If the person who wrote the instructions leaves, or just gets busy, the system decays quietly. Write down what it does, and give someone else access. Treat it as a small internal product with an owner.
If you build, adopt the approval table from our human-in-the-loop guide before you connect anything with write access. Anthropic's advice on building agents is worth following too: find the simplest solution possible, and only add complexity when you need it. A single scheduled prompt may do most of what you want.
Which build vs buy AI chief of staff path fits you?
- Start with Claude or ChatGPT if you're one technical person, you want to learn how this works, and your needs are mostly reading and summarising. It's cheap to test, and if you find you only use it for email, you've saved yourself a bigger decision. The Nautis vs Claude and Nautis vs ChatGPT pages cover how they relate.
- Choose an agent builder if your company already works well in tools like Slack, HubSpot and Notion, you don't want to move records, and someone will own the setup.
- Buy a product if the company needs shared records as well as an assistant, if several people will depend on it, or if you'd prefer your hours went to customers. Nautis fits this group: an AI business operating system with task boards, leads, investors, runway and more, plus a Chief of Staff with a default of asking first. It's $49 to $129 a month as of October 2026, with a 14-day trial. Its weaknesses are moving your records in, less control over the ranking logic, and a young product that launched in July 2026, with a dedicated tool often going deeper in any one function.
- Don't do any of it yet if you can't name what you'd want it to do tomorrow morning. A short list of what's missed today is a better starting point than a tool.
FAQ
Can I build my own AI chief of staff with Claude or ChatGPT?
Yes, for a useful single-person version. You connect your tools, write instructions and schedule a daily run. Expect to keep tuning it, and to add your own answers for shared approvals, audit history and team permissions if more people depend on it.
How long does it take to build an AI chief of staff?
We estimate about 26 hours to set up with a general assistant and connectors, and about an hour and a half a week after that. An agent builder may take about 17 hours. These are reasoned estimates and depend on how tidy your systems are.
Is it cheaper to build or buy an AI chief of staff?
Building has the lowest subscription, at about $20 a month. Once you count your time, the answer depends on how you value it and how much tuning you'd do. A founder who enjoys the tuning may find building cheaper, while a busy team often won't.
What is the difference between Lindy and a general assistant such as Claude?
Lindy is built around agents that do recurring work, with routines, integrations and team setup in Slack. A general assistant is broader, and you supply more of the structure yourself. Both can connect to tools through MCP, so the difference is how much you have to assemble.
Sources
- Claude: pricing (checked Oct 31, 2026)
- Claude Help Center: custom connectors using remote MCP (checked Oct 31, 2026)
- OpenAI: ChatGPT Work and Codex pricing documentation (checked Oct 31, 2026)
- OpenAI: plugins documentation (checked Oct 31, 2026)
- OpenAI: building MCP servers, including risks (checked Oct 31, 2026)
- OpenAI: Agent Builder documentation and deprecation notice (checked Oct 31, 2026)
- OpenAI Agents SDK: human in the loop (checked Oct 31, 2026)
- Lindy: product page (checked Oct 31, 2026)
- Lindy: pricing (checked Oct 31, 2026)
- Anthropic: building effective agents (checked Oct 31, 2026)