An operating system for AI startups, where compute is a business problem

Model your margin, because usage has a cost

a16z's The New Business of AI argued that AI companies often show gross margins of about 50 to 60%, against a 60 to 80% or better benchmark for SaaS, with 25% or more of revenue going to cloud. That was written in February 2020 and the authors called the figures anecdotal. Model prices have changed a lot since, so don't copy their numbers. Keep the question: what does one more customer cost you?

A standard SaaS template hides this. In the Financial Planner, each revenue item takes a cost of goods sold per unit. Put your inference cost per seat or per thousand calls there, and it flows to gross margin on the Dashboard.

1. Create a base model. Start from the SaaS template, set revenue items by plan, and enter a realistic cost per unit from your actual provider bills.
2. Clone it for your heaviest users. Imagine the top 10% of customers use ten times the median. Raise the cost per unit in the clone.
3. Clone it again for a price change. Your provider cuts rates by half, or raises them. Which plans survive either way?
4. Use Compare. It lays two projections side by side, so you can see which assumption moves margin most.
5. Write your reasoning in the Notes tab. Investors will ask, and a blended figure with no explanation won't help you.

Startup Health won't split this out for you. Burn is a single figure you enter monthly. It tells you how long the money lasts, not whether a spike came from a customer or a retraining run. For that level of detail you need your cloud provider's billing data.

Which parts of an AI company does Nautis not touch?

All of the technical ones. Nothing in its module list covers model evaluation, prompt versioning, experiment tracking, inference monitoring or cost per request across your own product. Keep your specialist tools. If those tools sit in GitHub, Linear or Slack, you can connect them under Integrations so agents can read them on demand.

This table is a useful way to draw the line with your team.

Put regulation on a calendar before someone asks

If you sell into Europe, the EU AI Act matters on specific dates. This public tracker of the implementation timeline, updated 31 August 2026, shows general-purpose model rules applying from 2 August 2025 and most of the Act from 2 August 2026, with some high-risk requirements starting on 2 December 2027 or 2 August 2028. Legacy synthetic-content providers have until 2 December 2026 for one transparency rule. Dates and interpretation can change, so confirm them with counsel.

In the US, NIST's AI Risk Management Framework is intended for voluntary use and organises the work around Govern, Map, Measure and Manage. Enterprise buyers may ask how you approach these, whichever framework you pick.

Nautis can't tell you whether you're in scope. It can stop these dates from living in one person's head. Make a Task Board called "AI compliance", add each date as a task with an owner, and let the Briefing read it alongside everything else. Keel, the startup legal advisor, can help you prepare questions. If your plan includes custom advisors, you can build one with your policies as knowledge files, then add a line under Never do such as "never give final legal advice". Don't treat either as legal review.

Investors have heard this pitch already

The a16z piece also warned about weaker moats, as models and data commoditize, and about edge cases and humans in the loop dragging on margin. Expect a sharp investor to ask about all three, so prepare before the meeting.

Use the Investor lens in Investor readiness to generate likely objections, each with the reason an investor would raise it, and write answers with evidence. The AI Voice Coach can drill you, and Rapid Q&A runs about 10 minutes. Then build a data room for technical diligence. Your deck goes in one folder and your architecture summary, model and data licences, plus your open-source list, in others. Law firm guidance on seed data rooms already asks for an open-source software list and IP assignments, and an AI company has more to say on both.

You'll want to test the agent harder than most

You know how these systems fail, so use that during the 14-day trial. The Chief of Staff defaults to Ask every time, with External and Internal labels per capability, and every action is audit-logged. Connected-tool writes ask for approval unless an admin sets that capability to Run without asking. Check the behaviour against what's described.

Try to confuse it. Ask about a record that doesn't exist. Give it an instruction that conflicts with your data. See whether drafts show exact recipients and text. It's an AI product, so it will rank something wrongly and write something in the wrong tone. Decide whether you can work with that.

Choose your model. Under AI Model Settings you pick an AI service, such as OpenAI, Anthropic or Gemini, and a model, and see credit rates per 1K, 5K and 10K tokens. That choice covers text features. Voice uses a fixed speech provider. You may also want to think about data handling across those providers, and ask Nautis about it before you put sensitive customer data in.

You can also connect Nautis to Claude or ChatGPT through an MCP connector, which may suit a team already working that way. It can read your workspace data, and it creates or updates records only if you allow that when connecting.

Plan and credits for a company that uses a lot of AI

Your team will likely use more credits than most. Everyone's chat, each advisor consultation and each board room reply draws on one shared balance, and technical people ask long questions with long attachments.

Launch at $49 a month as of October 2026 gives you 1,000 credits and 5 team members. Voyage at $99 gives you 2,000 and 10, plus the Cap table, which most funded AI companies need soon. Fleet at $129 gives you 2,600, 25 seats and 15 routines. Note that the credit gap between Voyage and Fleet is small, so pay for Fleet because of seats, not credits. See the pricing page for current details.

When an AI startup shouldn't use this

• You're two researchers before any customers. You have no business operations to run. Come back later.
• You could build it, and it's near your product. If agents for operations are your market, using a competitor makes little sense. Read build vs buy for the general case.
• You need strict data boundaries. If customer data or regulated data can't leave certain systems, get answers from Nautis first and don't assume any certification.
• You'd prefer general AI assistants and your own prompts. That can work well if you're disciplined. Compare Nautis and Claude or Nautis and ChatGPT.
• You want autonomy. Nautis asks first by default, and that's deliberate. If you want agents acting freely on outbound work, this isn't the right design for you.

FAQ

Sources

• a16z: The New Business of AI (February 2020) (checked Oct 31, 2026)
• EU Artificial Intelligence Act: implementation timeline tracker (checked Oct 31, 2026)
• NIST: AI Risk Management Framework (checked Oct 31, 2026)
• Klehr Harrison: building a data room for investor diligence (checked Oct 31, 2026)