AI Agents 8 min read 23 November 2026

Computer Use AI Agents at Work: What to Hand Over

Computer use AI agents from Claude and Gemini now click through desktop apps. See which jobs fit, the rules to set and what it costs. Book a WizeB call.

Computer Use AI Agents at Work: What to Hand Over
Part of the Implementing AI Agents guideHow to Implement AI Agents in Your Business

Computer use AI agents are AI assistants that operate a computer the way a person does: they look at the screen, move the cursor, click buttons, type into fields and switch between apps. That makes them useful for one specific problem every growing business has: work trapped in software with no API, such as an old accounting package, a supplier portal or a desktop tool only one person knows how to drive. They are slower and less predictable than a proper integration, so they suit low-volume, rules-based jobs that a person can check, not high-volume or money-moving work. Use them as a bridge to legacy systems or a helper on someone's desk, and keep anything that runs thousands of times a month on APIs and workflow automation.

This post is part of our guide to rolling out agents, which starts at wizeb.com/blog/how-to-implement-ai-agents-in-business. If you are still working out what kind of tool you need, read wizeb.com/blog/ai-agent-vs-chatbot-vs-automation first.

The News: Computer Use Just Reached Business Plans

Two launches in the same week moved computer use from developer demo to something your staff can switch on. On 19 November 2026, Google started rolling out computer use in the Gemini app for Mac: Gemini can click, type and fill in forms across local files, desktop apps and websites, working in a separate background window with a live preview while the user keeps their own screen. It is limited at launch to US English, personal Google accounts and paid Google AI Plus, Pro and Ultra plans on Apple Silicon Macs (source: blog.google/innovation-and-ai/products/gemini-app/computer-use-mac/). On 20 November, Anthropic made computer use available in beta on Claude Team and Enterprise plans, rolling out to organisations gradually; it lets Claude see a user's screen and work in the desktop apps they allow (source: support.claude.com/en/articles/12138966-release-notes).

OpenAI and Meta already offer similar features. So the question for an owner or COO is no longer whether this exists. It is which jobs you let it do, on whose account, and with what limits, before staff decide for you.

The Problem: The Work That Integrations Never Reached

Every business has a few systems that sit outside the automation plan. The job costing tool from 2011. The bank or insurer portal that sends a one-time code. The shipping carrier site where someone re-keys addresses from an email. The government filing screen. Integrating them properly is expensive or impossible, so a person spends an hour a day copying data between windows. That is the gap computer use fills. It is also why it carries more risk than a normal integration: the agent works through the same screens and accounts a person does, so it can do anything that person can do.

Computer Use vs APIs vs Browser Agents

  • APIs and workflow automation: fastest, cheapest per run and most reliable. A tool like n8n or Make calls the system directly. Use this whenever the system has an API, which most modern SaaS does. See wizeb.com/services/automation.
  • Browser agents: an agent drives a website in a controlled browser, often recording a path once and replaying it. Good for web portals with no API. We cover this in wizeb.com/blog/ai-agent-browser-automation-legacy-portals-2026.
  • Computer use agents: the agent sees and controls the whole desktop, including installed apps that are not websites. The widest reach and the slowest, since each step means reading the screen and deciding what to click. Expect seconds per action and the occasional wrong click.

Rule of thumb

If a job runs more than a few hundred times a month, or moves money, build an integration. If it runs a few times a day in software with no API, and a person can check the result in a minute, computer use is worth a trial.

6 Jobs Where Computer Use Pays Off

  1. 1Re-keying between systems with no API. Copying customer, order or job details from email or a spreadsheet into a desktop accounting, ERP or job costing tool. The agent reads the source, fills the form and stops before the final save for a person to check.
  2. 2Pulling reports from legacy software. Logging into an old tool each Monday, running the same three reports, exporting them and dropping them in a shared folder. Low risk, because the agent only reads.
  3. 3Portal admin you can't automate any other way. Downloading statements, remittances or certificates from supplier, carrier or insurer portals that block bots or change their layout often. Keep the person on hand for one-time codes.
  4. 4Testing your own software and website. Clicking through sign-up, checkout or booking flows the way a customer would, and reporting where it gets stuck. Run it on a test environment, not live accounts.
  5. 5Desktop clean-up and file handling. Renaming, sorting and filing scanned documents, filling templates in desktop Word or Excel, and building a first draft of a monthly pack from several files.
  6. 6A bridge while the real integration is built. When a proper API connection is three months away, a supervised computer use agent can carry the work in the meantime, then hand over.

What doesn't fit: approving payments, changing bank details, sending customer emails without review, deleting records, anything in a regulated system where you need a clean audit trail, and high-volume work where a few percent of wrong clicks would add up to real cost.

The Rules to Set Before Staff Switch It On

  • Company accounts only. Gemini's computer use launched on personal Google accounts. If staff run it on a work Mac with a personal account, your data sits outside your admin controls and contracts. Say in writing which plans are approved; our post on wizeb.com/blog/shadow-ai-agents-dots-muse-2026 covers why this matters.
  • Allow apps, don't block them. Both products let the user choose which apps the agent may use. Make that list short and job-specific: the legacy tool and a spreadsheet, not email, banking and the HR system.
  • No credentials in prompts. The agent should work inside a session the person already signed into. Never paste passwords or codes into the chat, and never let it store them.
  • A person approves anything that leaves the building. Sending, submitting, paying and deleting stay as human clicks. Google's version already asks for review on sensitive actions; make that your policy for every tool.
  • Watch for instructions on screen. A computer use agent reads everything it sees, including text in an email or web page that tells it to do something else. That is prompt injection, and it is the main security risk here. Keep it away from inboxes and open web browsing unless a person is watching. The detail is in wizeb.com/blog/ai-agent-security-guide.
  • Log what it did. Keep screen recordings or step logs for any job that touches customer or financial data, so you can see what happened if something goes wrong.

What It Costs

  • Off the shelf: the feature comes with the plan. Claude Team and Enterprise and Google AI Plus, Pro and Ultra are per-seat subscriptions; check current pricing, since usage limits apply and each screen step uses model capacity.
  • A supervised desktop agent built for one job, with an allowed-app list, written steps, test runs and logging: about $6,000 to $15,000 and three to six weeks.
  • Replacing it later with a proper integration, once the job proves its value: about $8,000 to $20,000 for one job across two or three systems.

Those are WizeB's working ranges for growing businesses in the US, UK and Europe, not a price list. The full breakdown is at wizeb.com/blog/ai-agent-cost-guide.

Example: A 25-Person Distributor

This is an illustrative example, but the pattern is common. A UK wholesale distributor with 25 staff runs its stock and invoicing on a desktop accounting package with no usable API. Each morning an administrator spends about 90 minutes keying web orders from the online shop into it, and on Mondays another hour exporting stock reports for the buying team.

The business starts with the Monday reports, because the agent only reads. A computer use agent on a company Claude Team seat opens the package, runs the reports and saves them to a shared folder. After two weeks with no errors, it moves on to order entry: it reads each order, fills the sales order screen and stops before saving so the administrator can check it. Checking takes about 15 minutes instead of 90. Meanwhile WizeB builds a proper connection between the shop and a new cloud accounting system, and the desktop agent retires when that goes live. The bridge bought the business three months of saved hours without waiting for the migration.

How WizeB Helps

WizeB builds custom AI agents, workflow automations and voice agents for growing businesses in the US, UK and Europe. For computer use, we start by asking whether the job needs it at all: often an API or a browser agent does it faster and cheaper. Where computer use is the right tool, we write the job steps and the allowed-app list, set up company accounts and logging, test it against real past work, and plan the integration that replaces it. You own the instructions, test sets and accounts. Details are at wizeb.com/services/ai-agents, and integration builds are at wizeb.com/services/automation. For the step-by-step build process, see wizeb.com/blog/how-to-build-an-ai-agent-for-business.

Find the right tool for your stuck job

Book a free scoping call at wizeb.com/services/ai-agents. Bring the one task your team re-keys every day and the name of the software it lives in, and we'll tell you whether it needs computer use, a browser agent or a proper integration, what it costs and how fast it can go live. Not sure you're ready? Take the free AI-Native Index at wizeb.com/ai-native-index.

Ready to act on this?

We build exactly what this article is about.

Tell us about your situation — we'll come back with a realistic assessment.