On May 29, 2026 — two days ago — OpenAI quietly shipped one of the most consequential updates in ChatGPT Business's history. The headline read like a developer-tools footnote: "Codex computer use is now available on Windows." Read that twice. ChatGPT can now see, click, and type in any Windows application running on your PC, and you can supervise the entire thing from the ChatGPT app on your iPhone or Android while you're in a meeting, on a flight, or watching your kid's soccer game.
For developers, this means agentic coding on Windows finally caught up to macOS. For small business owners, it means something far bigger: the gap between "AI that drafts a document" and "AI that actually finishes the work" just collapsed. Your QuickBooks, your Property Management software, your industry-specific ERP, the legacy Windows desktop tool your bookkeeper has been hand-clicking for fifteen years — all of it is now in scope for ChatGPT Business automation, with no API, no integration, no developer needed.
The release is also a direct shot across Microsoft's bow. Copilot has spent two years promising to "act in your apps." OpenAI just shipped it inside Windows itself, on a $30/month ChatGPT Business seat, with mobile remote-control out of the box. Here is the honest small-business read on what landed, the seven workflows that pay for the seat in week one, what the regional restrictions mean, and the 5-step pilot we're running for clients this week.
What Actually Shipped on May 29
The May 29 release bundled three things that most coverage covered separately. They matter as a system.
One. The Codex app, OpenAI's full agentic-coding desktop client, is now generally available on Windows 11. It runs natively in PowerShell using OS-level sandboxing built on Windows restricted tokens and filesystem access control lists — not a Linux abstraction layer or a virtual machine. That detail matters because it means Codex can interact with real Windows apps, real local files, and real Windows credentials with the same trust model as any other app the user installs.
Two. Codex on Windows now supports Computer Use — the visual GUI control mode previously available on macOS. With Computer Use enabled, Codex can see the active screen, move the cursor, click controls, type into fields, switch applications, and react to what appears on screen. This is the difference between an AI that asks you to "open Excel and click cell B2" and an AI that opens Excel and clicks cell B2 itself.
Three. The ChatGPT mobile app on iPhone and Android can now monitor and control an active Codex Computer Use session running on your Windows PC at home or at the office. You can kick off an agent on the PC, walk away, and watch it finish from the phone — intervening only when it asks for a decision. Combined with the Codex app's existing parallel-thread support, a single Windows machine can host multiple agents working different tasks simultaneously.
None of the three individually is unprecedented — Anthropic ships Computer Use, Microsoft has remote Windows desktops, mobile remote-supervision exists in IT-ops tools. The package, billed under your existing ChatGPT Business plan and built on the GPT-5.5 model that cut hallucinations 52.5% a month ago, is what makes this a small-business moment, not a developer-tools moment.
Why This Matters for Every Small Business, Not Just Developers
Most coverage framed the launch as a coding feature. That undersells it dramatically. The single biggest blocker for AI automation at small businesses isn't capability or cost — it's that the software a typical SMB lives in is not ChatGPT-friendly. It's a 22-year-old practice management system. It's a Windows-only QuickBooks Desktop install. It's an industry-specific ERP that has never heard of an API. It's a state Department of Insurance e-filing portal that needs three logins, an Excel upload, and a confirmation screenshot.
Until May 29, the only way to automate that work was: (a) pay a developer to build a Selenium robot, (b) buy a $200/seat RPA tool like UiPath or Power Automate Desktop, or (c) keep paying a human to click. Computer Use on Windows removes options (a) and (b) from the conversation for most SMB use cases. A bookkeeper, an office manager, or the owner can describe the task in plain English in ChatGPT, watch Codex perform it once, correct it, then schedule it to run nightly — the same way you'd train a new hire.
The economic shift this triggers for SMBs is the same one that ERPs triggered for the Fortune 500 in the 1990s, but compressed into months instead of decades and priced at a $30/seat subscription instead of a $400k implementation.
Computer Use vs Workspace Agents vs Goal Mode: Which Tool for Which Job?
OpenAI has now shipped three overlapping-but-distinct automation surfaces inside ChatGPT Business in the past 90 days. If you don't get the difference straight, you'll deploy the wrong one and blame the model. Here is the cheat sheet our team uses with clients.
| Capability | Codex Computer Use (May 29) | Workspace Agents (April) | Goal Mode (May 21) |
|---|---|---|---|
| Where it runs | Your local Windows or macOS PC | OpenAI cloud | OpenAI cloud |
| What it controls | Any GUI app on your desktop | 60+ SaaS apps (Slack, Salesforce, etc.) | Long, open-ended objectives |
| Best for | Legacy desktop software, local files, RPA-style tasks | Modern SaaS workflows, team-shared agents | Multi-day research or build tasks |
| Mobile supervision | ✓ iOS & Android | ✓ ChatGPT app + Slack | ✓ ChatGPT app |
| Included in Business? | ✓ No extra seat cost | ✓ + credit metering | ✓ + credit metering |
| EU/UK/Switzerland? | ✗ Blocked at launch | ✓ Available | ✓ Available |
| Right tool when | The system has no API | The system has a clean integration | The job has no fixed steps |
Said even simpler: Workspace Agents are for SaaS, Goal Mode is for open-ended outcomes, and Computer Use is for any pixel-level work a human currently does in a Windows or Mac app. The three are designed to be used together — a Workspace Agent can hand a sub-task to a Computer Use session and a Goal Mode plan can orchestrate both.
7 SMB Workflows Computer Use Pays for in Week One
These are not theoretical. Each is a workflow we've watched a small-business team prototype during the macOS Computer Use beta or the first 48 hours of the Windows rollout. All seven assume nothing more than ChatGPT Business and a single Windows 11 desktop the agent can run on overnight.
1. QuickBooks Desktop bookkeeping cleanup
The single most common SMB pain we hear. QuickBooks Desktop has no proper API for most workflows. A Computer Use agent can open QuickBooks, navigate to uncategorized transactions, apply category rules described in plain English, attach receipt PDFs from a shared folder, and produce an exception list for the owner each morning. Replaces the most expensive hour your part-time bookkeeper bills.
2. State portal e-filings (insurance, payroll, tax)
State portals are the graveyard of small-business productivity — multi-step logins, captchas, PDF uploads, confirmation screenshots. Computer Use can drive the portal, capture confirmation numbers, file the result in your shared drive, and email the owner a clean monthly summary. Especially valuable for insurance agencies, payroll services, and CPA firms.
3. Property management software data entry
Most property-management platforms (AppFolio Desktop legacy, Yardi, Buildium installs) have either no API or a heavily gated one. Computer Use can take a maintenance request from a Gmail label, open the property record, log the work order, assign a vendor from your rules, and update the resident with status. End-of-day report posts to Slack.
4. EHR / EMR clinical documentation prep
For practices that need every paid hour at the front desk, Computer Use can open the EHR, pull tomorrow's schedule, pre-load the relevant patient histories, and stage the documentation templates. Practitioners walk into appointments with everything queued. Combine with our HIPAA guidance for ChatGPT Business before deploying.
5. Inventory and price updates across legacy POS
For retail and food service running older Windows POS, Computer Use can pull vendor price sheets from email, open the POS admin, update SKUs, log changes, and post a Slack summary — nightly. The same job took a $35k/yr clerk three afternoons a week.
6. Litigation discovery file processing
Law firms drowning in PDF discovery can have Computer Use open documents in their existing review tool, apply OCR, run tag rules, and queue exception files for attorney review. Pair with our ChatGPT for law firms guide for security configuration.
7. Construction submittal and RFI logging
Most general contractors live in Procore, Sage, or a Windows-only Bluebeam workflow. Computer Use can extract data from submittal PDFs, log them in the construction management system, draft an RFI when fields are missing, and notify the project manager. See our construction playbook for vertical context.
The pattern across all seven: the agent automates the clicking and the typing, not the judgment. A human still owns approvals, exceptions, and customer-facing communication. That's the right line, and it's the line our 5-step pilot enforces.
The Cost Story Is Better Than You'd Expect
This is the part most coverage gets wrong. Computer Use on Windows is included in your ChatGPT Business seat — there is no separate "RPA license" the way Microsoft Power Automate Desktop bills, and no usage-based credit model the way Workspace Agents bill. The seat is $30/user/month on monthly billing or $25/user/month annual. If you already have ChatGPT Business, today's run-rate cost for Computer Use is zero incremental dollars.
The honest catch: Codex usage on the Business plan still consumes Codex credits when it runs background jobs, and Computer Use sessions can be long-running. OpenAI is also running a promotion through May 31 offering double Codex usage on the $100/month tier. For most SMBs, the right play is to stay on Business, set a single dedicated automation seat for the Computer Use machine, and monitor the credit consumption for the first two weeks before scaling.
Now compare that to the realistic alternatives. UiPath SMB seats start around $420/user/month for an unattended robot. Microsoft Power Automate Desktop's premium per-user plan is $40/user/month and Per-Flow plans run $150/flow/month. A dedicated bookkeeper or office admin to do the same clicks is $35k–$65k/year fully loaded. The math is not close.
The Microsoft Copilot Problem This Quietly Creates
Microsoft's whole pitch for Microsoft 365 Copilot was that it lives inside your apps and acts on your data — the implicit corollary being that ChatGPT does not. That story held up while ChatGPT lived in a browser tab. It does not survive Codex Computer Use on Windows. ChatGPT can now drive every Microsoft 365 app on the desktop, plus every non-Microsoft app on the same machine, including the ones Copilot has never heard of.
Microsoft also released their May Copilot update last week emphasizing better Edge integration and faster context loads. Useful, but defensive. The competitive position shifted: Copilot is now the integrated option, ChatGPT is the universal operator. For small businesses on a mixed stack — which is almost every small business — the universal operator wins. We made a similar argument in the Copilot meltdown post; this release strengthens it.
The wildcard is regional. Computer Use is unavailable in the EU, UK, and Switzerland at launch, almost certainly for AI Act and UK AI safety review reasons. SMBs in those markets get the announcement but not the product. Microsoft, with deeper EU regulatory relationships, is likely to use this gap as a Copilot talking point through Q3. North American SMBs have a clean window to deploy that European competitors do not.
Honest Limitations You Should Know Before You Pilot
To keep this useful and not breathless, four real frictions we have hit so far.
Computer Use is slower than RPA. A Codex Computer Use session that opens an app, clicks five buttons, and types a paragraph takes about 30–90 seconds. A traditional Selenium or RPA script does the same in 3–5 seconds. For high-volume, fixed-shape workloads (think: 10,000 records/night), RPA still wins on throughput. For variable, judgment-laden, low-to-medium volume work — the SMB sweet spot — the speed difference does not matter.
It needs a dedicated machine. Computer Use takes over the cursor and keyboard on whatever Windows session it is running. Practical implication: most SMBs will want a single $600 Windows mini-PC running 24/7 as an "automation seat," not an agent fighting the owner for control of the daily-driver laptop.
Captchas and 2FA still break it. Computer Use cannot solve image captchas reliably, and it cannot receive SMS 2FA codes. For state portals and bank logins that require step-up auth, plan for a hand-off-to-human checkpoint in the workflow.
Geographic gates are real. If your business has EU staff, EU customers, or runs in the UK, the Computer Use rollout has not reached you yet. Workspace Agents and Goal Mode are still available. Do not promise Computer Use to EU stakeholders this quarter.
Your 5-Step Pilot Plan (Run This Week)
This is the pilot we are putting in front of ChatGPT Business customers this week. Total elapsed time: about 6 hours over 5 business days.
Step 1 — Pick one workflow. Choose a task currently done by a human in a Windows desktop app, that runs at least weekly, has clear inputs and outputs, and would save at least 30 minutes per run. The bookkeeping cleanup or the state portal e-filing are usually the best first picks. Do not pick a customer-facing task in week one.
Step 2 — Provision an automation seat and machine. Add a dedicated ChatGPT Business user (call it "automation@yourcompany.com"). Install the Codex app on a Windows 11 machine that does not need to be used by humans during business hours. Sign Codex in to that automation user. Set OS-level user permissions to the minimum the workflow needs.
Step 3 — Demonstrate the task once, manually, with narration. Open ChatGPT, start a Computer Use session, and walk through the workflow yourself while describing what you are doing and why. The agent's first execution will essentially be re-running your narrated demo.
Step 4 — Run in shadow mode for 5 business days. Have the agent perform the task end-to-end but stop before any irreversible action (sending email, saving to production database, submitting filing). Compare its output to the human version every day. Refine the spec.
Step 5 — Cut over with a kill switch. Schedule the workflow to run on its production cadence. Set up a Slack or email alert when the agent completes. Document the manual fallback. Tell at least two people on the team where the kill switch is. Then measure for two weeks before adding a second workflow.
This is the same playbook our team uses with new Sayfe.ai customers, and it works inside a single business week.
The Honest Bottom Line for SMBs
Computer Use on Windows is the update that finally makes ChatGPT Business worth the seat for the small businesses that have been holding out — the ones whose work lives in legacy desktop software, the ones whose owner does the bookkeeping at 11 p.m. on Sundays, the ones who looked at AI agents in 2024 and 2025 and said "great, but my whole stack is on a Windows PC." That objection is now answered, on a $30/seat monthly bill, with mobile supervision out of the box.
The two-quarter window where this is meaningfully ahead of Microsoft Copilot is real. The European market is locked out for now and is going to feel that gap. The right move for a North American SMB this week is to pick one workflow, provision one automation seat, run the 5-step pilot, and measure. The right move is not to wait for the next release — OpenAI's release cadence in 2026 has been roughly one platform-shifting feature per fortnight, and the businesses that build automation muscle now are going to compound that advantage for the rest of the year.
Frequently Asked Questions
Yes. Codex Computer Use on Windows is included in the ChatGPT Business seat at $30/user/month (or $25/user/month annual). There is no separate RPA license required. Long-running background Codex jobs can still consume Codex credits over time, so we recommend dedicating a single automation seat for your Computer Use machine and monitoring credit consumption for the first two weeks before scaling.
Yes. As of the May 29, 2026 update, the ChatGPT mobile app on iOS and Android can monitor and control an active Codex Computer Use session running on your Windows 11 PC. You can kick off the agent on the PC, walk away, and watch it finish from your phone — intervening only when the agent asks for a decision. This is one of the headline changes in the Windows rollout.
Not at launch. OpenAI explicitly excluded the European Economic Area, the United Kingdom, and Switzerland from Computer Use on Windows in the May 29 release. This is almost certainly tied to EU AI Act and UK AI safety review timelines. Workspace Agents and Goal Mode remain available in those regions. If your business has EU staff or customers, do not commit to Computer Use deliverables this quarter.
Two key differences. First, training cost: traditional RPA tools require either a developer or a fairly technical operator to script every click. Computer Use can be trained by demonstrating the task once and describing it in plain English. Second, pricing model: UiPath SMB unattended robots start around $420/user/month, Microsoft Power Automate Desktop Premium runs $40/user/month, and dedicated per-flow plans add up fast. Computer Use is included in your $30 ChatGPT Business seat. RPA still wins on raw throughput for very high-volume fixed-shape work, but for the variable, judgment-laden tasks small businesses actually run, Computer Use is the better economic and operational fit.
Yes, during an active session. Computer Use drives the actual cursor and keyboard the way a remote IT support session would. For this reason, most small businesses will want to dedicate a low-cost Windows 11 mini-PC (around $600) as an "automation seat" rather than using the owner's daily-driver laptop. The mobile supervision feature reduces the friction of this setup — you can monitor and steer from your phone without going back to the automation machine.
Pick one weekly Windows-app workflow with clear inputs and outputs that saves at least 30 minutes per run — QuickBooks Desktop categorization, a state e-filing portal, or a property-management data-entry task are usually the strongest first picks. Provision a dedicated automation seat in ChatGPT Business and install the Codex app on a Windows 11 machine reserved for the agent. Demonstrate the task once, narrating as you go. Run in shadow/approval mode for 5 business days. Compare agent output to the human version daily. Only then cut over to scheduled production runs, with a documented kill switch and at least two people on the team trained on the manual fallback.
In principle, yes — Computer Use operates any Windows application by reading the screen and driving the cursor and keyboard. We've watched early users prototype workflows on QuickBooks Desktop, Yardi Voyager, AppFolio Desktop, Bluebeam, and a long tail of industry-specific Windows tools. The two failure modes to plan around are captchas (Computer Use cannot reliably solve image captchas) and SMS 2FA logins (it cannot receive SMS codes). For any workflow that hits those gates, build a hand-off-to-human checkpoint into the agent's plan.
Pilot Codex Computer Use on Your Windows Stack — in a Single Business Week
Sayfe.ai is an authorized OpenAI SMB Channel Partner. We provision your automation seat, pick the highest-ROI Windows workflow in your stack, run the 5-step shadow pilot, and stand up the kill switch — so you can measure the savings instead of building the plumbing.
Get Started TodayRelated reading:
- OpenAI Workspace Agents Killed Custom GPTs
- ChatGPT Goal Mode: Your First True Autonomous AI Employee
- GPT-5.5 Instant Just Cut AI Hallucinations 52.5%
- Microsoft Just Gutted Its Copilot Team: Why ChatGPT Business Is Winning
- ChatGPT Business vs ChatGPT Plus: Which Plan Does Your Team Need?
About Sayfe.ai: Sayfe.ai is an authorized OpenAI SMB Channel Partner. We help small and medium-sized businesses implement and optimize ChatGPT Business, ChatGPT Enterprise, and the OpenAI API. We make enterprise AI accessible to teams of any size.