Updated September 30, 2026
OpenAI Dots for small business could help with the work that keeps slipping between conversations: unanswered inquiries, unfinished handoffs, and jobs waiting to be invoiced. These agents can hold ongoing responsibilities, work across connected apps, and return with results while you are doing something else. OpenAI's announcement explains the product.
That raises a bigger question for an owner: what should happen to the time the business gets back?
Cedar & Loam's proposal is deliberately provocative: full market pay, a full working week, and no more than about 50% routinely committed to client delivery. Protect the other half for hands-on practice, critical thinking, improving systems, and building a company with other human beings.
Fifty percent is a starting point to test, not a measured optimum or an employment policy we have already adopted. The argument is that reliable automation could increase the firm's finished work while leaving its people more time to develop. Here is how Dots fits that argument, and what would have to be true for it to work.
What OpenAI Dots can do for small businesses
A dot can retain context about ongoing work, follow up as things change, and delegate substantial tasks to ChatGPT Work or Codex. It has its own cloud computer and uses the apps you connect. Access depends on the connections and permissions available in your account. OpenAI's computers and apps guide covers those boundaries.
“Summarize these emails” is one task. “Keep me informed when customers are waiting on us” is an ongoing responsibility. The second needs a definition of waiting, reliable sources, a review process, and a clear notification rule.
Three proposed workflows illustrate where a small business could start:
- Find leads that need a reply. Compare an approved sales inbox with a quote tracker. Return the last exchange, suggested next step, and source links. Someone checks whether a phone call or booking already resolved the issue.
- Prepare the invoice review. Compare completed jobs with invoice-status records. Flag possible billing gaps and missing purchase-order details. The bookkeeper confirms the figures and issues the invoices.
- Catch loose ends in handoffs. Compare approved customer correspondence with a project plan. Surface commitments that may be missing, with the original message and promised date. The project lead resolves conflicts before anything changes.
These are suggested uses, not Cedar & Loam deployment results. Keep existing automations that already work. A dot is worth testing where gathering context and interpreting exceptions takes meaningful human effort.
Why 50% capacity starts with the apprenticeship problem
Imagine a junior account manager whose day becomes reviewing drafts, approving summaries, and checking boxes. The business may be getting work out the door. But how does that person learn to build a campaign, diagnose a failed handoff, or recognize a convincing explanation that is wrong?
Good review requires understanding the work underneath it. If every unfamiliar task goes straight to an agent, we risk removing the experiences that develop that understanding.
There is relevant early evidence. In a randomized study involving 52 software engineers learning an unfamiliar Python library, researchers found lower immediate mastery scores among participants using AI assistance. The average speed improvement was not statistically significant. Some ways of using AI were associated with stronger comprehension than others. This was a short coding study, not a test of Dots or a long-term study of agency careers; it supports taking the learning question seriously, not a universal conclusion about AI. Anthropic's research summary and the paper describe the findings and limitations.
Our proposed response is to make development part of the paid operating model. For an illustrative 40-hour week:
- About 20 hours for client delivery: execution, client meetings, reviewing agent output, corrections, and approvals.
- About 20 hours protected from routine delivery commitments: hands-on practice, investigating failures, improving workflows, mentoring, and collaboration across the company. Essential internal administration also needs an honest place in this budget.
Reviewing client work counts as delivery. Calling it “learning” does not move it into the other half. An assignment can teach someone and still belong in the delivery budget; count the time once.
What meaningful practice looks like
A useful rep asks the employee to make decisions and explain them. Build a campaign plan before seeing the agent's version. Diagnose a broken reporting workflow before asking for the answer. Rework a process for an unfamiliar project, then defend the choices to a colleague.
Some practice should involve supervised responsibility for real assignments, with feedback from someone experienced. Exercises help, but people also need to see the consequences of their decisions. Over time, they should move from doing the work to reviewing it well to designing better ways of doing it.
The protected time also belongs to people learning together: pairing across disciplines, mentoring newer colleagues, and discussing mistakes without rushing back to a delivery queue. Culture needs shared time, not just a paragraph in the handbook.
The career test is whether an employee can apply what they have learned in a new environment. Their professional growth should be visible in their judgment, even when the familiar assistant and its memory are unavailable.
Could lower individual utilization produce more finished work?
It could, if automation reduces the total human effort needed for a completed outcome enough to make the allocation viable. An employee's percentage of time booked to delivery and the firm's total output measure different things.
Consider a deliberately simplified example. Today, one employee spends 30 hours of a 40-hour week on comparable deliverables. At two human hours per accepted deliverable, that produces 15. If a redesigned workflow reduces the human effort to one hour per deliverable, 20 delivery hours could produce 20 accepted deliverables, leaving 20 hours for the other work described above.
Those are invented numbers illustrating the hurdle. The one-hour figure must include preparation, review, corrections, and exceptions. Dots has not been shown here to achieve that reduction.
The same logic applies to bottlenecks. If work waits for the owner to collect information, a dot might prepare it earlier. If every prepared item then waits for that owner's approval, the queue has simply moved. Count completed customer outcomes at the required quality, and examine the whole path to completion.
The economics have to hold too. Full market pay remains a cost, alongside software, maintenance, and training. The business needs demand for additional output and pricing that supports the model. A firm billing strictly by the hour could lose revenue if it reduces billable time without changing the commercial arrangement.
The 50% proposal therefore comes with a management obligation: prove that the system can support it. Faster draft generation alone is insufficient evidence. Persistent overruns, growing review queues, or learning time routinely swallowed by client work mean the design needs to change.
How to start using Dots without losing the learning time
Start with one responsibility whose result a person can verify. Use our small business AI readiness assessment if the team first needs to agree on sources, ownership, and what success means.
For a lead-follow-up pilot, try:
Review the approved sales inbox and quote tracker for inquiries from the last seven days. Identify customers who appear to be waiting for our reply. Separate confirmed gaps from unclear records. Link the evidence and draft suggested responses here for my review. Do not send messages or change customer records.
Run it manually, check the sources, and correct mistakes before making it recurring. Then specify the schedule and notification rules:
For the next seven days, run this check every weekday at 9 a.m. America/Chicago. Put the review list here. Notify only me when a customer has waited more than one business day, the check cannot run, or you need my decision. Keep the same source limits and approval requirements. Confirm the schedule, time zone, end date, and notification destination you saved.
Replace the time zone and threshold with your own. Connecting an app does not establish a monitoring job; event-based monitoring depends on service support. OpenAI's task documentation explains the options.
At the same time, put paid practice and collaboration on the calendar. Give the employee a related skill to develop and someone to learn with. A first week can test whether the workflow is useful; a longer pilot is needed to assess learning and the economics of the capacity model.
Track delivery hours including review, accepted work completed, errors and rework, protected time actually used, and costs. Ask the employee to explain an unfamiliar case or diagnose a failure without outsourcing the reasoning. Count those practice hours too. If the new arrangement works only because people finish client work after hours, it has failed the test.
OpenAI Dots FAQ for small business owners
Who can get Dots, and what does it cost?
As of September 30, 2026, Dots is rolling out to eligible Pro users and Business Premium seats, with an admin-enabled Enterprise beta. Pro launch access excludes the EEA, UK, and Switzerland; Business Premium covers supported ChatGPT regions. Setup starts on desktop, and access may take time to reach an eligible account. Check the getting-started guide.
US Pro tiers are $100, $200, and $500 per month. Business Premium is $100 per person per month with annual billing or $125 monthly. Business requires at least two paid seats: one Premium plus one Standard totals $120 per month equivalent annually or $150 monthly, calculated from published prices. Only the Premium seat gets Dots through that arrangement. See the Pro guide and Business FAQ.
The first dot is included with Pro or Business Premium. OpenAI describes a deeper-work allowance with extended first-month limits; delegated Work and Codex tasks count toward their respective limits. Check the current terms when budgeting.
Does it work when my laptop is closed?
Its cloud computer can. Tasks needing your connected personal computer require that computer to be online with the ChatGPT app open. Local access is optional. Computer access documentation.
Can it act without asking, and how do I stop it?
Permissions and advance authorization can allow actions without a fresh confirmation each time. Review the plugin permissions and Custom Rules, as well as the task instructions. For the pilot, keep sending, publishing, and shared-record changes subject to human review. Pausing the main task does not necessarily stop delegated tasks or future schedules; manage those separately. OpenAI's controls guide.
What about privacy and keeping clients separate?
Connecting an app can make information available for proactive research beyond the task you just assigned. Disconnecting stops new access but does not erase what the dot already learned. Individual dot memories currently cannot be viewed, directly edited, or selectively deleted. Business data is not used for model training by default; Pro follows the relevant data-control setting. Dots privacy FAQ.
For agencies, a client name in a prompt is not an access boundary. Start with approved material for one client and verify which accounts and destinations the connection can reach. Do not assume separate conversations provide enforceable client isolation.
Is Cedar & Loam recommending everyone immediately cut delivery to 50%?
We are proposing a model to test as reliable automation earns back capacity. The number is intentionally bold. The commitment underneath it is concrete: retain full market pay and deliberately fund the practice, judgment, and human relationships the business will need next year.
If Dots makes your firm capable of more, decide in advance how much of that gain belongs to developing the people who make the firm worth hiring.
If you'd like help choosing a first workflow and measuring what it gives back, that is the work Cedar & Loam does with agencies and small firms. Proudly based in Kansas City. Say hello.
Further reading: Flavio Copes's A deep dive into OpenAI dots. Product facts were checked against official OpenAI sources on September 30, 2026. Availability, pricing, and controls may change. The capacity model and workflow examples are Cedar & Loam proposals, not measured results.