For most smaller businesses the obstacle to automation was never intelligence. It was that the software running the business had no way in. No API, no webhook, no export worth the name — just a screen, and a person typing into it.
In the last fortnight that obstacle got a published price.
01 — What happened
At DevDay in San Francisco on 29 September 2026, OpenAI added computer use to its Agents API: an agent that clicks, types and navigates software the way a person would, with OpenAI running the infrastructure so developers do not host it themselves.
It arrived with GPT-6.1 Sol, a model aimed at coding and computer use that OpenAI claims approaches its larger GPT-6 Astra on several evaluations at a fifth of the token price.
OpenAI’s own documentation is plainer than the keynote. “Computer use lets a model operate browser and desktop interfaces.” The model either writes a script your application runs, or returns mouse and keyboard actions you execute and answer with a fresh screenshot, until it stops asking.
None of this is a new capability. What is new is that all three large platforms now sell it as a product with terms.
Microsoft’s Copilot Studio documentation states the proposition without decoration: agents “can complete tasks even when there’s no API to connect directly to the system. If a person can use an app or website, computer use can too.” It is generally available there on OpenAI’s computer-using agent model and on Claude Sonnet 4.5.
Google’s Gemini Computer Use covers browser, mobile and desktop actions, and is candid about maturity: a preview capability that “may contain errors and security vulnerabilities”, not for use where decisions are critical, data is sensitive, or mistakes cannot be corrected. And since May 2026, AWS has offered Amazon WorkSpaces as managed virtual desktops for AI agents, driving legacy desktop applications through computer vision and input simulation — without, as InfoQ put it, “application modernization or API integration”.
The detail that matters is the price. Microsoft bills each computer-use step at five Copilot Credits, or fifteen on a premium model. A footnote in the billing table records that computer-using agent usage is not included in the Microsoft Copilot user licence — unlike almost every other agent feature listed beside it.
02 — Why it matters
Many businesses run on vertical software: practice management, workshop scheduling, freight, dispatch, laboratory, lettings. It works, staff know it, and it has no usable interface for anything else. The choices were to replace the system, fund an automation project against it, or keep paying someone to retype.
Computer use removes the software vendor’s consent from that decision. You no longer need them to publish an API or answer your email. If a person can do it on screen, something else can do it on screen.
That is a real shift in leverage, and it arrives with a bill in two places. The first is access. An agent that works by using your software works as somebody. Microsoft’s documentation is explicit: with maker-provided credentials, anyone the agent is shared with “can act with the original author’s access on the configured machine”. Its access control is also narrower than it sounds — it “only prevents the model from taking actions on websites or applications that aren’t in the allow list. It doesn’t stop the model from opening them.” This is the access problem we covered in September, now reaching the desktop rather than the API.
The second is arithmetic.
03 — The economics
An integration and a computer-use agent have opposite cost shapes, and that single fact should drive the decision.
An integration is a fixed cost: you pay to build it once, and each run after that costs effectively nothing — a thousand runs a month cost the same as ten. The risk sits at the start.
Computer use reverses it. The build cost is close to zero: a page of instructions in plain language, no code. The cost sits in every run, and scales with steps rather than minutes. Microsoft’s worked example makes the unit visible — a four-step run that opens a portal, creates a record, fills three fields and submits consumes 20 Copilot Credits on a standard model, 60 on a premium one.
Put volume through it. Take an illustrative business re-keying 600 supplier invoices a month into a system with no import route, eight steps a run: 40 credits a run, 24,000 a month, every month, rising with growth. The same work through an export and an import is one build and then silence. On high-volume, structured work, metered clicking is the expensive answer — and it has an operational cliff as well as a financial one: reaching 125% of prepaid capacity disables Copilot Studio’s custom agents.
Then there is reliability, where the only controlled evidence does not flatter the marketing. Průcha, Matoušková and Strnad compared UiPath against Anthropic’s Computer Use Agent on data entry, monitoring and document extraction. Conventional automation “outperforms AACU in execution speed and reliability”, particularly in stable environments, while the agent “significantly reduces development time” and copes better with interfaces that change. Their verdict on current implementations: “not yet production-ready”, with promise for prototyping and light automation.
Read the two cost shapes together and the usual rule inverts. Conventional advice says start with your highest-volume repetitive task — and we have argued that hours saved is the wrong way to value an automation, with a rough floor of fifty runs a month before a build earns its supervision. Computer use lowers that floor, because there is almost nothing to build, and lowers the ceiling too, because cost and fragility accumulate per step. What it opens is the middle: processes too small to justify an integration, too tedious to keep doing by hand.
04 — What businesses should do
If you have been quoted a replacement system purely to get data in and out, test this first. Two weeks and one process is enough to learn whether a bridge buys you a year. A migration you can defer is worth more than most automations you can build.
Choose the first candidate by friction, not volume. The supplier portal nobody wants to log into, the monthly reconciliation, the fortnightly report assembled by hand. Eight runs a month at twenty minutes each is an excellent candidate; eight hundred is a reason to price an integration instead.
Count steps before you count hours. Cost scales with clicks, so a thirty-step run costs roughly four times an eight-step one regardless of how long a human took. Write the steps out first: past about twenty, the process needs simplifying more than automating.
Give it a dedicated machine and a scoped account. Microsoft’s own security guidance is the right baseline: an isolated machine, a least-privilege account, allow-listed sites and applications, credentials in managed storage rather than typed into instructions. Never the owner’s login.
Verify outcomes, not answers. OpenAI’s documentation advises checking the actual result rather than the model’s report of it. In practice: the agent writes to a staging record, or produces a list a person approves, until you have watched a few hundred clean runs.
Set a review date and a replacement plan on day one. Ninety days, one owner, two figures: runs completed and runs that needed a human. A bridge with no scheduled end becomes the architecture by default.
05 — Where technology fits
For most businesses this is not an AI purchase. It is a question about whether data can leave the system by a boring route: a scheduled export, a report emailed as a file, a read-only database copy, a supported import at the other end. If any of those exist, an integration beats an agent on every axis that matters: speed, cost per run, auditability, the ability to tell a customer what happened. Most of what looks like an automation problem is still the disconnection between systems rather than a shortage of intelligence.
Where there genuinely is no route out — and for much vertical software there is not — computer use is a legitimate bridge, worth building as one: narrow, supervised, documented, with the eventual integration or replacement still on the roadmap.
And the honest no: if the process runs a handful of times a month and the person doing it is not your bottleneck, you probably do not need an AI system for this yet. Fix the form, drop the step, or leave it alone.
06 — RevenueStack’s take
Our view: the announcement is not the capability, it is the price list.
Computer use has been demonstrable for two years. What changed in the last fortnight is that it became a managed service with a published rate per step, excluded from the licence you already pay for, and therefore a line item — and line items get bought in a way demos never do.
The consequence almost nobody is pricing: this is the first automation technology whose economics favour the long tail over the core. Everything before it rewarded volume; this punishes it. Most firms will point it at their core anyway, where the pain is loudest, and conclude it does not work. The businesses that get a return will use it on the twelve-runs-a-month jobs they wrote off years ago, then replace it with a proper integration the moment one of those jobs becomes important.
One last thing. Microsoft’s documentation says the technology adapts to interface changes and “continues working without breaking your flow”. The only controlled comparison says conventional automation was faster and more reliable, and that agents of this kind are not yet production-ready. Both cannot be your operating assumption. Build against the research and let the marketing surprise you.
Sources
- OpenAI Gave AI Agents Their Own Computers at DevDay 2026. Here's Everything It Announced — Decrypt, 2026-09-29
- OpenAI DevDay 2026 Recap for Developers — InfoQ, 2026-10-02
- Computer use — OpenAI API guides — OpenAI
- Automate web and desktop apps with computer use — Microsoft, 2026-09-18
- Billing rates and management — Microsoft Copilot Studio — Microsoft, 2026-10-01
- Computer use — Gemini API — Google
- AWS WorkSpaces Now Lets AI Agents Operate Legacy Desktop Applications without APIs — InfoQ, 2026-05-13
- Are LLM Agents the New RPA? A Comparative Study with RPA Across Enterprise Workflows — Průcha, Matoušková and Strnad (arXiv), 2025-09-04
