Grok Bot, Dots or Duet? You're Comparing the Wrong Thing

Three always-on AI agents landed in a matter of months. xAI shipped Grok Bot. OpenAI shipped Dots. Anthropic opened the preview of Claude Duet. Every leader I talk to is asking the same question: which one do we buy?

It is the wrong question. I ran the pricing comparison, and the numbers are close enough that price will not decide this for you. What decides it is something none of the three sells: a safe way for the agent to reach your company's data.

• The problem. Leaders are comparing agent subscriptions seat by seat. Out of the box, all three agents can see an inbox and a calendar. None of them can see the business.
• The insight. The agent is the cheap, swappable part. The valuable part is one secured door onto your central database, built on the Model Context Protocol (MCP), the open standard agents use to connect to outside systems.
• The move. Build the door before you buy the seats. Five rules make it safe, and once it exists, any agent can use it.

The Pricing, in 60 Seconds

Here is where each product stands as of November 11, 2026. Prices move fast, so confirm with each provider before you budget.

• Grok Bot (xAI). Bundled into the Premium plan at $30 a month, Super at $100 and Ultra at $200. Team seats cost the same per seat and add single sign-on and central billing.
• OpenAI Dots. Your first Dot is included with ChatGPT Pro at $100, $200 or $500 a month, or with a Business Premium seat at $125 per user a month ($100 billed annually, two-seat minimum). Free, Go and Plus do not get one.
• Claude Duet (Anthropic). Included with Pro, Max and Ultra during a research preview. United States accounts only, by waitlist. Team and Enterprise plans are not part of the preview.

For a team of ten, that is $300 to $2,000 a month on Grok Bot, or $1,250 a month on Dots through Business Premium. Duet cannot be bought for a team yet.

That spread looks wide. Against a ten-person payroll it is a rounding error.

What All Three Have in Common

Look past the price and the three products are built the same way.

Each one comes bundled with the chat subscription you already know. Each one draws on a usage allowance, and background work eats into it. Anthropic says so directly: Duet's background checks count toward the same limits as the rest of Claude, which is why it checks your accounts every few hours rather than nonstop.

And each one becomes useful by connecting to other systems. Grok accepts a custom MCP server. ChatGPT workspaces accept custom MCP apps that an administrator publishes. Duet carries over the apps connected to your Claude account and draws on a marketplace of more than 6,000 connectors and plugins.

Same shape, three logos.

What None of Them Can See

Connect the standard apps and your agent can read email, check a calendar and draft a document. Useful. Also the least valuable information in your company.

The valuable information sits in your central database: customers, pipeline, projects, delivery, invoices. Getting it into one place is a leadership call of its own. Ask an agent "which clients are at risk this quarter?" and without that data it has two options. It guesses, or it asks you. Either way you have hired an intern who can only read the mail.

Four Ways Teams Get This Wrong

These are the patterns to watch for.

1. The master key. Someone pastes an administrator credential into a connector so the agent "just works". Now the agent can read payroll, and so can anything that tricks the agent.
2. The shared login. One connection for the whole company. Every person's agent sees the same data, and the log cannot tell you who asked for what.
3. The do-it-yourself connector. Each employee wires up their own. Nobody reviews them. OpenAI is plain that it does not vet custom apps; that job belongs to the customer.
4. The open door. Guides for Grok note that a custom server has to be reachable over the public internet. Publish that server without real sign-in and you have published your database.

None of these feels like a decision when it happens. That is the problem.

The Secure Door: Five Rules

One door, built once, that every agent goes through. Five rules make it safe.

1. Give the agent its own identity in the database, and start it at zero. A dedicated role that can read nothing until you grant it, table by table. A new table stays invisible until someone chooses to open it.
2. Redact by column, not by promise. Pay, bank details and identity documents are removed at the database, so no clever question can retrieve them.
3. Read first. Write later, and only with a human. Start with answers only. When you add actions, make each one a named tool that pauses for a person to approve. Where that line sits is a design decision per workflow.
4. Every person signs in as themselves. The agent inherits that person's access and no more. A tool they are not allowed to use is never handed to their agent in the first place.
5. Log everything, and keep one off switch. Every question, every answer, every approval. One place to revoke access when someone leaves or a vendor changes its terms.

How We Built Ours at Edge8

This is not theory for us. Edge8 runs on one central database, and our team has an assistant that answers questions from it.

The line we wrote into the design is "the security boundary is the database". The assistant reads through a dedicated role that is deny-by-default, with access granted table by table and in places column by column. Pay, identity records, people's private records, hiring, message logs, call transcripts, client money and client contacts are closed to that role at the database layer, whatever the model is asked.

There are five layers of protection, and each is meant to hold on its own. The team assistant cannot write anything at all, because no writing role exists for it. The sensitive data that people do need comes through narrow, named tools instead of open questions. Your own pay and reviews come back only to you. A client's invoices and contacts come back only if you are assigned to that client.

On the administrator side, an action pauses the conversation until a person approves or declines it, and that decision is recorded before the action runs.

One honest caveat. This assistant lives inside our own portal, not behind an MCP server. The rules do not change when the protocol does. The role, the redaction, the approval step and the log are the door. MCP is the standard fitting that lets Grok Bot, a Dot or Duet walk up to it.

One Door, Any Agent

Here is what the door does to the buying decision.

Prices will change. Limits will change. One of these three will leap ahead next quarter, and a fourth will launch. If your data access lives inside one vendor's connector settings, every change is a migration.

If it lives in a door you own, the agent is a seat you can swap. The pricing comparison stops being a strategy decision and becomes a procurement one.

Five Questions to Ask Before You Buy Seats

1. Where is our source of truth, and is it one place?
2. What should an agent never see, and is that enforced in the database or only in a prompt?
3. Does each person's agent act as that person, or as the company?
4. Which actions need a human to approve them?
5. If we cancel this vendor tomorrow, what do we have to rebuild?

If you can answer all five, choose whichever agent your team likes. If you cannot, the subscription is not your next step.

Deciding what your systems expose, and to whom, is leadership work, not a task to hand to whoever set up the account. It is the kind of work leaders prove by doing in the Leadership in the AI Era certification.