Dots, Team Bots, Muse, and Instinct: which agents fit your business?

A business owner wants someone to chase an unanswered email. An account team wants a colleague who remembers its process. An operations manager wants confidential documents processed inside the company network. All three may ask for an AI agent, but they are buying different things.

Our recommendation: choose a hosted agent for useful work you can delegate within its documented controls. Choose an agent you operate when the workflow, access rules, or data location justify owning the system. Start with a specific job and an acceptance test.

OpenAI announced Dots on September 29, 2026. Grok's Team Bots arrived on September 28. Meta expanded Muse with a small-business offering on September 29. Instinct takes a more personal-assistant approach. Their business value depends on how those different designs fit your work. OpenAI launch, Team Bots announcement, Muse for Small Business, Instinct.

This comparison uses first-party documentation reviewed September 29, 2026. The recommendations are Looski's assessment; we have not run a head-to-head performance benchmark. An undocumented feature is a question for the vendor, rather than evidence that it does not exist. The on-premises option below describes a system you would implement and operate.

Compare what you are actually buying

The hosted options reduce the amount of infrastructure you need to assemble. A custom deployment gives you more design choices, but each promised capability must be implemented. A server with a model installed is only one component of a working business agent.

Dots: persistent assistance, with an enterprise path still taking shape

Dots combine an OpenAI-hosted computer, connected applications, ongoing memory, and background work. The launch distinguishes a person's primary dot from specialist organization agents, which begin as focused enterprise pilots. Teams of personal dots working together are described as a future direction. Dots launch. OpenAI documents Custom Rules and action review; its proactive research mode uses read-only tools. Dots safety design.

Rollout currently covers eligible Pro and Business Premium users, with Enterprise beta access enabled by an administrator. Pro excludes the EEA, Switzerland, and UK; Business Premium covers supported ChatGPT regions. Local-computer access is optional and initially off. Connecting your laptop does not mean the model runs on it. Dots setup and availability.

For a business, a sensible trial is one employee's recurring preparation work: gather evidence, update a draft, and present the decision that needs attention. Before assigning a departmental responsibility, establish who owns the agent when that employee changes roles.

Team Bots: a shared role with individual working context

Team Bots package files, instructions, skills, plugins, API credentials, and memory around a team workflow. People can connect applications individually or use team connections. Private conversations retain separate user context and memories, while skills are shared; a bot can also participate in a Slack channel through its own handle. The release is a public beta on Teams and Enterprise plans. Team Bots.

This makes Team Bots a candidate for work such as preparing account briefs from an agreed set of sources. In a trial, verify that a junior employee cannot retrieve restricted information through a shared connection. Also test how a correction becomes shared knowledge: private context and team knowledge should have an explicit boundary.

For the broader Grok Bot offering, current documentation places network controls, SCIM, audit logs, and Action Recording in the Enterprise tier. Action Recording starts off, and administrative audit events are separate from recordings of bot actions. Confirm how those controls cover a shared Team Bot in your contracted plan. Teams and Enterprise controls.

Muse: a practical shortlist for an owner running connected business tools

Muse for Small Business adds connections to tools including QuickBooks, Shopify, Stripe, Canva, Slack, and Facebook and Instagram business accounts. Meta describes approval before publishing, sending, or spending. Its announcement builds on Muse's availability in the US and Canada. Small-business announcement.

An owner could evaluate it by asking for a weekly review of sales and campaigns, followed by proposed changes. Measure whether the recommendations reconcile with the source systems and whether approval clearly identifies the account, audience, and amount involved.

Meta describes a dedicated secure VM and protections around credentials, external content, and actions. It also acknowledges remaining prompt-injection risk: material an agent reads can contain instructions intended to redirect it. Those protections are meaningful, but the VM remains part of Meta's service. They do not establish that a business can deploy Muse on its own server. Muse security design.

The buying question is how well its owner-oriented workflow extends to your staff, access rules, and handoffs. Ask for the actual administrative controls you need.

Meta also describes an opt-out from training and separation from its advertising systems. Its more restrictive Confidential VM was announced for later this year; it should not be treated as an already available guarantee that Meta cannot access the data. Muse architecture and rollout.

Instinct: personal delegation needs a separate business evaluation

Instinct describes an assistant that connects to applications and devices and works through text and calls. Its examples emphasize following up, coordinating arrangements, and handling everyday tasks. That makes personal administrative work a reasonable starting point for evaluation. The public product page does not establish a company-wide operating model. Instinct product description.

Its privacy policy deserves a careful, specific reading. It permits some information to be used for model training, provides a prospective opt-out with a safety-review exception, and separately excludes data received directly through Google Workspace APIs from model training. Disconnecting an integration does not automatically erase previously collected data. Instinct privacy policy.

For company use, confirm the training settings, deletion process, account ownership, and any business-specific agreement before connecting a work inbox. Request evidence of shared administration and audit access if those are requirements. A convenient personal assistant may be valuable without becoming the system through which an entire department operates.

These data questions apply across the shortlist. OpenAI says Business, Enterprise, and Edu workspace content is not used for model improvement by default; its Dots help page also says disconnecting an app does not delete previously obtained information. Treat permission revocation, memory deletion, and model-training policy as three separate checks. OpenAI data policy for Dots, connection and memory lifecycle.

Self-hosted, on-premises, and fully local mean different things

Separate where the agent executes actions, where the model processes information, and where documents, memory, and logs are stored. You can control one without controlling the others.

A workflow that calls a cloud model is a hybrid, even if it runs on a Mac in your office. A locally running model can also send information out through email, web search, connectors, or diagnostics. Inspect the whole path.

For example, Ollama documents separate local and cloud model operation and a setting to disable its cloud features. That setting governs Ollama; it cannot stop another application in your workflow from making an external request. Likewise, the OpenAI Agents SDK documents tracing that can export run information, including sensitive inputs and outputs depending on configuration. Local execution alone does not settle data residency. Ollama FAQ, Agents SDK tracing.

On-premises control is useful when documents must stay inside a defined environment, an agent needs access to internal systems, or you need reproducible rules for a stable process. It brings responsibility for identity, patching, backups, monitoring, and recovery. A service provider can operate the equipment for you, but its remote access and contractual responsibilities still belong in the design.

What an on-premises agent must earn in a pilot

Consider an illustrative invoice workflow: read a document, extract its supplier and amount, compare it with a purchase order, and create an exception for a reviewer. An on-premises design could keep source documents and inference local, then send only an approved record to accounting.

That is a proposed architecture. You still need document extraction, a model suitable for the workload, restricted accounting access, a durable job queue, and evidence showing who approved each write. Deterministic checks should enforce totals, supplier identity, and duplicate prevention where those rules can be encoded. The model can help interpret ambiguous documents and explain exceptions.

Test difficult inputs, simultaneous jobs, and interrupted runs. A model that answers a short prompt well may behave differently with a long document or several concurrent requests. Ollama's documentation, for example, describes how parallel requests increase memory requirements and how requests may queue when capacity is insufficient. Ollama concurrency guidance.

A fully local system can continue local work without internet access only if every required dependency is also available locally. Email delivery, online accounting, and cloud identity may still need connectivity. Define the degraded mode before calling the workflow offline-capable.

Compare the cost of an accepted result

The relevant cost includes subscription or infrastructure, model usage, integration work, administration, human review, and correction of failures. Dividing that total by completed results that meet your acceptance criteria is more useful than comparing subscription prices with a machine's purchase price.

Launch offers complicate the hosted comparison. OpenAI's help page describes a one-month usage promotion with subsequent plan terms to follow; the launch page separately says delegated Work and Codex tasks consume their normal allowances. Obtain a workload-specific allowance and overage estimate. Dots usage details, launch billing distinctions.

Grok Bot's plans documentation describes weekly included usage and optional on-demand charges; an active run can exceed the monthly on-demand cap. Confirm how your Team Bots consume that allowance. Meta describes free use plus subscription options without sizing a business workload in the announcement. Instinct's product page does not provide a public business rate card. These gaps prevent an honest fixed-price ranking from the launch materials alone. Grok Bot usage, Muse, Instinct.

For an illustrative on-premises budget, spread a $6,000 machine over 36 months: about $167 per month. Add an assumed $100 for power, backup, and infrastructure, plus eight operating hours at $100: roughly $1,067 per month, before setup, software licenses, external APIs, and user review. These are planning assumptions, not hardware guidance, vendor prices, or a Looski quote. Change the operating hours and utilization before drawing a conclusion.

A hosted system can be better value when the workload is modest or maintenance would distract the team. Local inference can become attractive for steady, suitable workloads, but only after quality, throughput, and operating costs are measured. Neither deployment earns a savings claim merely by existing.

Choose a first job, then choose its agent

Use these starting hypotheses for a trial:

• Individual preparation and follow-through: evaluate Dots against the person's recurring deliverables.
• Shared departmental expertise: evaluate Team Bots against a documented process and staff access rules.
• Owner-led storefront and marketing operations: evaluate Muse against the actual applications and approvals involved.
• Personal administrative coordination: evaluate Instinct with a deliberately bounded set of accounts.
• Confidential, repeatable internal processing: evaluate an on-premises implementation against a written data-flow requirement and a realistic maintenance budget.

A mixed deployment can also make sense. A local agent might process confidential documents while a hosted agent researches public information. Define the exact fields allowed to cross between them; a summary can still disclose sensitive information.

Run the same representative cases through each shortlisted option. Include a normal task, a missing fact, a revoked credential, a duplicate event, and an interruption after an external action may already have succeeded. Measure accepted results, review minutes, elapsed time, and total cost. Require escalation when the evidence is insufficient.

Finally, have another employee take over the workflow. They should be able to see what happened, stop pending work, recover a failed run, and retrieve the records the business needs. The strongest choice is the one that completes useful work and remains understandable to the people responsible for it.

Looski builds private AI systems on Apple silicon. Our interest is in finding the workflows where that ownership pays for itself. For a starting point, compare the economics of a private deployment or bring us one workflow to scope.