There are now three ways to put an AI agent on a business process: subscribe to one, run it on a vendor's platform, or build it on an open harness. Most mid-sized firms will end up running all three, for different jobs. The choice for any one process depends on who the agent works for, where the process lives and who should own what the agent learns. The timing is not an accident. Anthropic opened a research preview of Claude Duet on 21 October, which means four of the largest AI companies now offer an always-on agent that can do business work, alongside OpenAI's dots, Meta's Muse and Microsoft's Copilot Autopilot. Behind them, OpenAI is piloting agents with their own company identity, and open-source harnesses such as OpenClaw and Hermes Agent have made a custom build far cheaper than it was a year ago. All three routes are capable. They suit different processes.
Three routes compared
Underneath, the three routes share the same anatomy: an agent with its own computer, a memory of how you work, connections to your software and rules for when it must ask first. As I wrote in always-on AI agents are here, that part has become a commodity. What differs is who the agent works for, who controls the rules around it, and who carries the work of running it.
| Subscribe | Platform | Build | |
|---|---|---|---|
| Examples | Claude Duet, OpenAI dots, Meta Muse, Microsoft Copilot Autopilot and Scout | OpenAI specialist dots, governed through Microsoft Agent 365 | Agents built on OpenClaw or Hermes Agent |
| Works for | One person | The company, under its own identity | The company, under your rules |
| Available | Now, in previews and paid plans | Enterprise pilots | Now |
| Set up | An afternoon | A vendor programme | Weeks, with a builder |
| Rules and memory held by | The vendor | The vendor | You |
| Who runs and secures it | The vendor | The vendor | You, or your builder |
| Best fit | One person's inbox, ledger and diary | A process that lives inside one vendor's software | A process that crosses systems, or where the rules are your edge |
Route one. Subscribe to an agent that works for you
This is the fastest route and the one most people will meet first. Duet is in preview for Pro and Max subscribers in the US. It runs on its own virtual machine, checks your connected accounts every few hours, sends a morning brief and messages you when something needs you. Before it sends a message, makes a payment, signs a document or deletes anything, it shows you exactly what it is about to do and waits (Anthropic, October 2026).
Personal does not mean trivial. Anthropic's own examples are business work: unpaid invoices in QuickBooks or Stripe, payroll in Gusto, follow-ups from HubSpot. OpenAI's dots reach Slack and Microsoft Teams and work in shared team spaces. About a third of Muse users had connected a business account, according to the Wall Street Journal (via VentureBeat). Microsoft's Scout even gets its own Entra identity in the corporate directory.
The limit is in the design. A subscription agent works for one person, so the process it runs is that person's share of the work. When they are on holiday, the agent's approvals wait with them.
Route two. Run it on a vendor's platform
The second route gives the agent its own seat in the company. OpenAI's specialist dots receive their own organisational identity, credentials and access to company systems, and a defined business responsibility. OpenAI has tried them internally on procurement, invoice processing, email marketing, customer support and contracting, and is integrating them with Microsoft Agent 365 so IT can govern them with tools it already uses (VentureBeat, September 2026). Meta is heading the same way. In late September it launched Meta Enterprise Platform to bring the Muse agent, Meta Business Agent, Muse API and Muse Code to businesses (Meta, September 2026), though it has shared few details so far.
This route is still young. Specialist dots are in focused enterprise pilots, so expect it to open up over the next few quarters. When it does, it will suit a process that lives inside one vendor's world, such as a firm that runs on Microsoft 365 end to end. The trade is control: the identity, the policy engine and the agent's memory belong to the vendor. I cover the governance side in agent sprawl is the new shadow IT.
Route three. Build it on an open harness
The third route is the layer the platforms themselves are built on. Microsoft runs Scout on OpenClaw, the open-source harness an independent developer released in late 2025, and is contributing its policy controls back to the project (The New Stack, June 2026). OpenClaw is built around reach: one agent answering across many channels, extended by a large marketplace of community skills. Hermes Agent, released by Nous Research under an MIT licence, is built around memory. It learns skills from hard tasks, refines them with use, runs on infrastructure you own and works with whichever model you choose (Nous Research).
Building means you, or the people you hire, write the rules, connect the systems and run the agent. It also means owning the security. Audits of OpenClaw's skills marketplace flagged 341 malicious entries (The New Stack, 2026), which is why a serious build vets every skill and runs the agent in a sandbox. In return, the rules, the memory and the audit trail are yours.
One process, three routes
Take credit control, which I covered in detail in credit control is a job for an AI agent.
Subscribe. The finance manager at a 60-person firm connects Duet to QuickBooks and Gmail. Each morning it lists the overdue invoices, drafts polite chasers and waits for approval. Where one person owns the ledger, that is most of the job, for the price of a plan they may already have.
Platform. Once specialist agents reach general release, a finance team of eight on Microsoft 365 could give one its own identity, read access to the ledger and send rights on the credit control mailbox, governed in Agent 365. It would chase on behalf of the company, so nothing waits for one person's holiday.
Build. A 300-person distributor reconciles across an ERP, a bank feed, a customer portal and three mailboxes, and its dispute rules differ by customer group. An agent built on a harness, running in the firm's own cloud, reads all of them, follows rules written for this business and keeps an audit trail the finance director owns.
The process, not the vendor, picked the route.
How to choose
Five questions settle most cases.
- Who does the agent work for? One person's work points to a subscription. A process several people touch points to a platform or a build.
- Where does the process live? Inside one vendor's software, wait for or use that vendor's platform. Across several systems, a build usually fits better.
- Are the rules your edge? If the way you triage claims or chase debt is part of why customers choose you, you want those rules in your hands.
- Who should own what the agent learns? Memory is becoming the real lock-in: a year of learned exceptions is costly to leave behind. With a build, that memory stays with you and survives a change of model.
- What will your auditor or regulator ask? If the answer is "show me every action and who approved it", make sure the log is one you can export and keep, whichever route produced it.
For most firms the answer is a mix: subscription agents for individual managers, a platform where one vendor already holds the process, and a build where the process is core.
What the three routes share
When OpenAI pilots specialist dots with enterprises, it sends its own engineers to work with the customer on each agent's responsibilities, permitted tools and approval steps. The scarce work is not the agent. It is the specification around it. A business with 50 to 500 staff will not get an OpenAI engineer, and it needs that work just as much.
I learned this building production agents at one of Europe's largest insurance brokerages, on a system that resolved 67% of customer service cases without a human. We sorted the data out first, designed the escalation logic with the people who knew the work, started at 10% of volume and grew only as the numbers justified it. In my view that discipline would have mattered just as much on any of the three routes.
That work is what I do, so let me declare my interest. I build on all three routes: on open-source harnesses, and on top of dots and Duet. The agent comes from whichever route fits the process. What I add is the layer that makes it perform: memory and retrieval so it knows the business, workflows organised around the process, evaluations that measure what it produces, and guardrails that put people in the loop at exactly the right points. I gain nothing from steering you towards one route. What I care about is that the one you pick is built to be trusted.
Where I would start
- Give your managers a subscription agent for their own work, where one is available to you. It builds the habit and costs little.
- Pick one process that matters, and measure its baseline: volume, time per item, error rate.
- Choose the route with the five questions above.
- Write the rules, the approvals and the escalation points in documents you own, not only in a vendor's console. That keeps every route open, including a move later.
- Run in shadow mode, then on 10% of live volume, then scale on the evidence.
If you are weighing these routes for a process of your own, I would be glad to hear how it works today. Let's talk.
Related: Always-on AI agents are here. The agent is now the easy part · Most late payments can be recovered. Credit control is a job for an AI agent · Agent sprawl is the new shadow IT. Your business needs a control plane · Your AI does not need a bigger model. It needs to know your business
