A logistics broker I know runs his entire back office through a browser. Tracking shipments through carrier portals, filing customs paperwork through a 1998-vintage government site, reconciling invoices in a third-party platform that doesn't have an API and probably never will. Ten hours a week, every week, forever — because the work is just boring enough that he can't justify hiring a person and just specific enough that no SaaS will ever cover it.

He's the perfect customer for a browser agent. Eighteen months ago, the technology couldn't have served him. Today it can.

Screen-understanding accuracy on the OSWorld benchmark moved from 70% in late 2024 to 92%+ in 2026, driven by Claude Opus 4.7 and GPT-5's computer-use capabilities. Per-task agent cost fell from $0.50–$1.50 in 2024 to $0.05–$0.15 in 2026. An SMB with a $300/month budget can run 2,000–6,000 agent tasks a month. The economic threshold moved. Five workflows that were impossible to sell at a margin in 2024 are now obvious purchases at €500–€1,500/month.

What 92% Accuracy Actually Means

The OSWorld benchmark measures whether a browser agent can complete a real desktop task — open an app, navigate a UI, complete a multi-step workflow. Going from 70% to 92% sounds incremental. In practice, it's a phase change.

At 70% accuracy, the agent fails one in three tasks. Unusable in production. The human checking the agent's work is doing more cognitive labor than the human doing the work themselves.

At 92%, the agent fails one in twelve. The human moves from "doing the work" to "spot-checking the agent." That's a 70–80% reduction in human time per workflow. Now the agent has a margin to live in.

At 95–98%, which is where the trajectory is heading by year-end, the human checks samples and fixes exceptions. That's the same operational model as RPA — except RPA tools cost €5,000/month per process and require six weeks of consulting setup. Browser agents at €500/month are an order of magnitude cheaper.

The 92% line is where the math flips. Below it, agents are demos. Above it, agents are products.

The Five Categories That Just Came Into Scope

Some workflows have been waiting for this exact accuracy + cost threshold. They're now sellable.

Legacy SaaS administration. Filling out forms in SAP, Workday, NetSuite, or any ERP that hasn't been redesigned since the Bush administration. Buyers exist in every mid-market company. They will pay €800–€1,500/month for a workflow that runs nightly without complaint. Selling angle: "you'll never have to log into [vendor] again."

Marketplace and platform operations. Listing management for Amazon, eBay, Etsy, marketplace dashboards for delivery platforms, channel managers for hotel booking systems. The platforms hate APIs and refuse to make them complete. Browser agents bypass the limitation. E-commerce operators pay €499–€999/month per platform; multi-platform sellers pay multiples.

Government and compliance portals. Customs filings, tax submissions, registration renewals, license applications. Existed since the 1990s. Will exist in the 2030s. Nobody is going to build an API. Browser agents are the only viable automation. Tradespeople, importers, and small accounting firms will pay €299–€599/month to never touch one of these portals again.

Internal-tool data entry. Companies running 8–15 internal SaaS tools spend hours of human time moving data between them. Operations agents that watch one tool and update another are immediately useful, immediately measurable, immediately renewable. Sells at €699–€1,499/month per workflow.

Research and aggregation. Visiting 30 supplier websites to collect pricing, monitoring 50 competitor stores for inventory shifts, gathering due-diligence information across a list of companies. The agents that win here are domain-specific, not generic. Vertical buyers pay €499–€999/month for a workflow that would have cost €2,000+ in human time.

What Still Doesn't Work

Honest list. Not every workflow is in scope, and pretending otherwise is the fastest way to burn customer trust.

Visual creative work. Designing, image-editing, video review. Browser agents can navigate the tools but the judgment is still human.

High-stakes decisions. Medical triage, legal document review, financial trading. The 92% number means 8% failure. For high-stakes decisions, 8% is unacceptable. Stay away from these unless you're prepared to ship a high-risk-classification compliance package.

Workflows that depend on inconsistent UIs. Sites that A/B test their layouts, sites that geolocate-route to different templates, sites with CAPTCHAs every five clicks. Agents fail unpredictably here. Pick targets with stable, predictable UIs.

Workflows the customer can't articulate. If the customer can't show you the human version in under 20 minutes, the agent version will take three months. Walk away from these.

How to Sell a Browser Agent

The sales motion for browser agents is unusually clean because the value is observable.

Lead with a 90-second screen recording. Record yourself doing the customer's workflow. Then record the agent doing it. Show both side by side. The buyer's brain does the math without you saying a word.

Price against the human time, not the SaaS market. Your competitor isn't another browser agent — it's the part-time admin who currently does the work for €15/hour, ten hours per week. €600/month for the agent saves the customer €600 in admin time and the agent works at 3am. Frame the price against the human number.

Demo with the customer's actual data. Browser agents are uniquely demo-friendly because they run in a browser. Ask the customer for read-only access for the demo. The conversion rate on data-aware demos is 3–4x higher than the conversion rate on generic ones.

Sell month-to-month, but anchor the savings annually. "€699/month — that's €8,400 a year, against the €15,000 you're currently spending on the admin time." The annual frame makes the monthly feel small.

Include a kill switch. Every browser agent contract should give the customer a one-click pause. This isn't a feature for them — it's a sales weapon. "If anything goes wrong, you stop it instantly" handles the single biggest objection (loss of control) before it gets verbalized.

The Six-Month Window

Browser agents have an unusually short window. The accuracy is there. The cost is there. The platforms are there. What's missing is your specific vertical's leader — the operator who shows up to every dental practice, every roofing contractor, every customs broker, every Mittelstand HR team and ships the agent first.

That leader will become unassailable. They'll have the testimonials, the templates, the case studies, the vertical-specific prompts that took six months of customer iteration to refine. By Q4 2026, every category will have a clear winner. After that, the second-place player gets squeezed.

If you've been waiting for "agents to be ready," they are. The work is to pick one vertical, ship the boring version of one workflow, and start the customer conversations this week. The technology threshold is behind us. The market threshold is the only one left.