The consumer AI agent pricing war has officially arrived. OpenAI put Dots — its always-on agent tier — behind the roughly $100-a-month Pro price point, and Meta shipped Muse, a free assistant baked into WhatsApp, Messenger, and Instagram. If you run a small business, your feed is full of reviewers telling you which one is smarter.
Smart isn't the question you need answered. Your rent doesn't get paid by an agent that writes beautiful prose. It gets paid by an agent that finishes tedious jobs — while you sleep, without you babysitting it, and without handing your customer list to a data model you don't understand.
Here's the test that actually decides it.
The boring-work head-to-head
Don't benchmark these agents on coding challenges or trivia. Give both the same three tedious jobs, the kind every business has rotting on a to-do list:
- A calendar cleanup. Two weeks of double-bookings, zombie meetings nobody attends, and invites that should have been declined months ago.
- A subscription audit. Every recurring charge on the company card — flag what's unused, what's duplicate, and draft the cancellation emails.
- A refund chase. One overdue invoice or refund, end to end: compose the follow-ups, send them on schedule, escalate politely, report back.
Then score both agents on the only three columns that matter:
- Actions completed — did the job actually finish, or did it produce a nice summary of the job it could have finished? Reviewers love screenshots of plans. Owners get paid on completed actions.
- Guardrails — what did it ask permission for, what did it touch without asking, and could you see what it did afterward? An agent that emails your customers without showing you the draft first is not an efficiency; it's a liability.
- Data and access — where did the work happen, what did it read, and what happens to that context if you cancel next month?
This is the evaluation framework reviewers skip and owners can't. Dots reviewers got excited when it flagged a recurring weekend "zombie meeting" nobody attended and offered to fix it — that's the right kind of feature to get excited about. Not ambient intelligence for its own sake; ambient maintenance, the boring category of work that's genuinely worth paying for.
What the price tags actually tell you
The $100-versus-free framing is mostly noise for a small business, because the two products aren't selling the same thing.
Muse is free because it lives inside Meta's ecosystem — it's an engagement surface for WhatsApp, Messenger, and Instagram, where Meta's business model runs on attention and activity. The price you pay isn't on the invoice; it's where your work context lives and how it's used. For customer-facing busywork — drafting replies, sorting messages, following up in the same apps your customers already use — that ecosystem position is a real advantage, not a gimmick.
Dots is paid because OpenAI is selling capability as the product itself. The bet is that a more capable agent tier finishes jobs a free assistant summarizes instead. If the boring-work test shows Dots completing jobs Muse can only plan, the subscription is cheap. If both finish the same jobs, the $100 is a fan club fee.
The uncomfortable third possibility
There's a scenario almost nobody in the coverage wants to say out loud: both agents fail the test on most of your work, because consumer agents don't know your business. A generic assistant — free or paid — doesn't know which leads are hot, how you quote, which customers get the friends-and-family exception, or why you never schedule site visits on Fridays. It can clean a calendar. It can't run your follow-up process.
That's not a knock on either product. It's the difference between an assistant and an employee. The generic tiers are excellent at assistant-shaped work: inbox triage, scheduling hygiene, subscription hygiene, first-draft anything. The employee-shaped work — the processes that make your business yours — needs context those products don't have.
Which points at the honest conclusion for most owners: consumer agents are a $0-to-$100 monthly add-on, not a strategy. Run the boring-work test on the free one first. Upgrade only when the test shows the paid tier finishing jobs the free one doesn't. And keep the employee-shaped work in an agent whose context you actually own.
How to run the test without risk
- Use sandbox inputs. Not your real customer list — a test calendar, a personal card's subscriptions, one refund you don't care about.
- Keep a human checkpoint on anything outbound. First rule of agent work: nothing leaves the building without eyes on it, at least until a job has earned trust.
- Time-box it. One week, the three jobs, both agents. If neither finishes them, you've learned the market isn't ready for your use case yet — which is a real answer, and cheaper than a year of either subscription.
The pricing war is great news for small businesses, because the competition means capability keeps getting cheaper. But the price of the agent was never the real cost. The real cost is the work you hand it — and whether what it does with that work is something you'd sign your name to.
If you want the full framework for making any agent earn its seat — the permission ladder, the boring first assignment, the first-month inspection habit, and the one-page process format that keeps your context portable — it's all written down in The AI Agent Owner's Playbook. $49 CAD, instant download.
