In the space of one week at the end of September, OpenAI, Microsoft and Meta all launched always-on AI agents. A ten-person firm can now have a persistent agent that chases invoices, watches its reviews, prepares the morning brief and drafts proposals, for between nothing and about $100 a month. The agent is now the easy part. What decides whether it helps is whether your records are clean, whether the rules for what it may do are written down, and whether someone checks the work. That is why these agents will widen the gap between small firms before they close it. The good news is that every one of those three things is fixable, and none of them needs a large budget.
Four agents, one design
The launches came over seven weeks, with the main cluster between 25 and 29 September. They look different on the surface and are strikingly similar underneath.
| Launch | What it is | Who can use it now |
|---|---|---|
| OpenAI dots (29 Sep) | Named, persistent agents on GPT-6 Astra, each with its own cloud computer and browser, connecting to more than 4,000 apps through plugins | ChatGPT Pro and Business Premium in eligible markets; Enterprise as an admin-enabled beta |
| Microsoft Copilot Autopilot (25 Sep) | A persistent agent with its own identity, memory and computer inside your Microsoft 365 tenant, which you @mention in Teams | Private preview; Home and Code roll out through the Frontier programme |
| Meta Muse (8 Sep) and Muse for Small Business (29 Sep) | Meta's consumer agent, extended to business, with connectors for Shopify, QuickBooks, Stripe, Slack, Canva and others, plus Meta ad accounts | Free with usage limits, paid tiers for more use |
| GrokBot (SpaceXAI with Cursor, Aug) | Several cloud agents, each with its own computer, coordinated by a "chief of staff" bot | Beta, through Cursor and SuperGrok plans |
Every one of them gives you an agent with its own computer, memory of how you work, connections to the software you already use, chat as the interface, and rules for when it must ask before acting. When it is idle, a dot does background research using read-only tools that cannot send, change or buy anything. The agent loop, the connectors and the approval flows have become a commodity. You will get them from whichever platform you already pay for.
What is still scarce
If the capability is a commodity, the advantage has to sit somewhere else. It sits in four places, and none of them is on the vendors' feature lists.
The first is clean, connected records. An always-on agent is only as good as the CRM, ledger, project files and shared drives it can reach. When Salesforce surveyed customers deploying its agents in 2026, only 31% had unified their data before launch. The rest were asking an agent to act on records that disagreed with each other. I explain why that breaks AI in your AI does not need a bigger model, it needs to know your business.
The second is clear rules: what the agent may send, spend, sign or change, and what always needs a person. The third is checking capacity, someone who reviews consequential output before it goes out. The fourth is what only your firm knows: its history, its judgement and the people who trust it.
A demo shows you the first two minutes of an agent's life. Your records, rules and reviewers decide the next two years.
Why the gap will widen before it closes
It is tempting to read this as the great leveller. The owner of a small firm can now delegate ongoing responsibilities, not just one-off tasks, at a price that used to buy a single software seat. Early evidence is encouraging, too. Gusto's payroll data (September 2026) found that small businesses adopting AI grew headcount faster than those that did not. For small firms, the first effect has been growth, not replacement.
But agents amplify what is already there. McKinsey's 2026 survey found the share of large companies (over $1 billion in revenue) scaling AI agents rose from 27% to 40% in a year, while smaller organisations stayed flat at 22%. A firm with a tidy CRM, a clean chart of accounts and organised project files will get compound returns from an always-on agent. A firm whose knowledge lives in three inboxes and one person's head will get confident mistakes, faster.
So expect agent-ready small firms to pull away from their peers, rather than small firms as a whole catching up with enterprises. Which side of that line you land on is a choice you can make this quarter.
An agent that works while you sleep also fails while you sleep
The most commonly reported failure in small-business automation is not dramatic. It is silent. A workflow reports success in the logs while doing the wrong thing: the chaser goes to last year's contact, the report pulls from a stale export, the refund message promises something nobody is processing. With an agent that works overnight and follows up forever, a silent failure can run for weeks before anyone notices.
The platforms have taken this seriously. Every one of them keeps credentials away from the model and asks before purchases, sends and irreversible changes. Keep those defaults on until you trust the agent's judgement on your own accounts, not the vendor's demo accounts.
The other thing that fails quietly is the bill. Microsoft's per-seat licence covers chat and Copilot inside Office, but Cowork, Code and Autopilot are all billed by usage, and the same pattern runs across the market. I explain why that makes cost a knowledge problem in why grounded AI agents are cheaper to run. Put AI spend on a cost centre with a cap and a named owner, like any other variable cost, and read the agent control plane piece before your third agent arrives.
What to hand over, function by function
The work worth delegating first has the same shape everywhere: the input is documents, the output is a structured result, the rules are known, and a person decides the exceptions.
In finance, that means invoice preparation with approval (OpenAI's own launch example is a dot that noticed a forgotten invoice, drafted it and sent it once its owner approved), collections chasing, reconciliation prep and close-pack assembly. Use the controls already built into your accounting system before adding a separate agent on top. In sales, it means inbox triage, follow-ups, meeting prep from CRM history and keeping records current, with approvals on anything that commits price or terms. In marketing, it means performance monitoring, repurposing and competitor watching, with a human on anything that makes a claim or goes out under a person's name.
One shift cuts across all three. When everyone's agent replies instantly and follows up forever, responsiveness stops being a differentiator. Relationships, credibility and specific proof take its place. And your customers are getting agents too, which I cover in headless is the new mobile-first.
AEC: bids are the first target
Architecture, engineering and construction firms will feel this first in business development. Microsoft's own launch list for Cowork starts with "an RFP response". Unanet, which sells software to AEC firms, reports proposal volume rising while win rates stay at about 50%, so agents that draft faster will mostly produce more proposals, not more wins. In AEC, winning depends on the quality of the past-project evidence, not the speed of the drafting.
That puts the pressure on project records: roles, dates, values, outcomes, whether the firm was prime or subconsultant, team experience and image rights. Firms that have organised them will get far more from any agent than firms that have not.
A disclosure before the next paragraph: my day job is Director of AI Engineering at OpenAsset, and I helped build Shred 3.0, so weigh my view accordingly.
This is exactly the problem OpenAsset's DAM and Shred 3.0 were built for. The DAM holds the firm's projects, people and imagery as one organised record. Shred reads the whole pursuit package, including addenda, recommends the best-matching past projects from the DAM, drafts with the source text in view, checks the draft against every requirement in the RFP before submission, and sends improved narratives and bios back to the DAM for the next bid. We built it to one standard: proposals that are accurate and relevant, built from the best of the firm's own content, that give every bid its best chance of winning, and that are fast and easy to produce. Whatever tools you use, the principle is the same. Keep a person accountable for every submission, because addenda, deadlines and mandatory forms are where silent errors disqualify bids.
A short playbook for the next quarter
If you run a small or mid-sized business, this is the order I would do it in:
- Pick one platform that matches your stack. Microsoft 365 firms start with Copilot, ChatGPT Business firms with dots, businesses that sell mainly through Instagram and Facebook with Muse. Resist running three at once.
- Start with read-only work. Morning briefs, monitoring, reporting and drafts for review. Then actions with approval. Only then autonomous actions within limits.
- Write the rules down. What the agent may send, spend, sign, change or delete, and what always needs a person.
- Cap the spend and name an owner before switching on any metered feature.
- Fix the records the agent will read. CRM hygiene, client and project files, the chart of accounts, document and image rights. A knowledge audit tells you where to start.
- Measure time to a checked, finished result, not time to a first draft.
- Redeploy, don't cut. The early evidence from small firms is growth among adopters, and the labour effect so far is fewer junior hires, not layoffs.
The timing matters. Most of the enterprise versions are still in preview, pilot or waitlist, and none of these agents has an independent reliability benchmark yet. That gives you perhaps two or three quarters before always-on agents are a default option in the software you already use. It is enough time to get your records and rules ready, and not much more.
If you would like help working out where an always-on agent would pay first in your business, and what has to be fixed before it can, let's talk. The Knowledge Audit is the short, fixed-fee version of that conversation.
Related: Your AI does not need a bigger model. It needs to know your business · Why grounded AI agents are cheaper to run · Agent sprawl is the new shadow IT. Your business needs a control plane · Headless is the new mobile-first. Agents are about to become your biggest user
