A browser agent is software that opens pages, reads what is on screen, clicks, types and submits forms under a policy. Buying one is closer to buying an operating process than buying a chatbot. The demo will look impressive; the decision hinges on what happens on the tenth thousand run, when a supplier portal changes its layout and nobody is watching.
This guide is for operations leaders who have a repetitive browser task and want to know how to buy the capability without handing over uncontrolled access to customer records, money or regulated data.
Write the task specification before you shortlist
Vendors will shape the conversation around their strengths unless you arrive with a written task. A usable specification fits on one page and covers:
- Task name and outcome: what must be true when the run finishes, expressed as a business result rather than a sequence of clicks.
- Start state: where the work comes from, who triggers it and what data arrives with it.
- Systems and access: which sites, portals or internal applications are involved, and which credentials or roles are needed.
- Decisions and exceptions: the branches a person currently handles, including the ones that require judgement or approval.
- Volume and timing: runs per day, peak periods and whether work can queue.
- Error tolerance: which mistakes are recoverable and which are unacceptable.
- Data classes: personal data, financial data, credentials and anything regulated.
- Completion evidence: the record, screenshot or confirmation that proves the task finished correctly.
If the task cannot be described without saying “the agent will figure it out”, it is not ready to buy. Narrow the scope first.
Know which of five things you are buying
The market uses one label for several different products. Ask each vendor which category they are in.
- Deterministic UI automation. Scripted steps with selectors, rules and scheduled runners. Predictable, but brittle when the interface changes.
- Browser automation platforms. Recorded workflows, managed runners and a control room for many tasks. Strong on operations and observability, weaker on open-ended judgement.
- Model-driven agents. The system plans steps at runtime using a language model. Flexible with messy pages, but harder to bound and test.
- Vertical agent products. A packaged agent for one domain, such as supplier onboarding or invoice portal entry. Faster to start, less adaptable outside the intended task.
- Build-and-run partners. A service that designs, builds and operates the workflow for you. Useful when the process spans several systems and your team lacks automation capacity.
Requirements that eliminate weak options
Put these in the request for proposal as mandatory requirements, not nice-to-haves.
- Scoped credentials. Separate accounts for automation, least privilege, no shared administrator passwords and a documented rotation process.
- Hard limits. Allowed domains, allowed actions, value caps, approval gates and an emergency stop that a supervisor can use.
- Step-level evidence. A log of each action, the page state, the decision taken, the model output where relevant, and a replayable recording.
- Exception handling. The run must pause, preserve context and hand a clear case to a person rather than retrying indefinitely.
- Change resilience. A test suite for the workflow, selector strategy, and an alert when a page structure changes.
- Data discipline. Minimum data collection, masking, retention limits, access controls and a written deletion process.
- Exit. Export of workflow definitions, logs and configuration, plus credential rotation and deletion on termination.
A vendor scorecard with evidence behind it
Score each vendor from zero to three on the same criteria and record the evidence behind the score. Weights should reflect your risk, not the vendor's slide deck.
- Task fit: does the product handle your actual pages, languages and exceptions?
- Control: can you restrict what the agent may do, and can you stop it?
- Auditability: can an auditor reconstruct who did what, when and on whose authority?
- Exception handling: how does work return to a human, and how quickly?
- Maintainability: who fixes the workflow when a portal changes, and at what cost?
- Integration: how do results reach your CRM, ERP, inbox or database?
- Commercial model: is pricing tied to runs, seats, workflows, implementation or model usage?
- Support and exit: what response commitment applies, and what do you get back if you leave?
Pricing questions that expose the real cost
Agent pricing is often split across several meters. Ask for a worked example using your own volume, including failures and retries.
- Is the charge per workflow, per run, per successful outcome, per seat or per hour?
- Are model tokens, browser minutes, storage and screenshots billed separately?
- What happens financially when a run fails or needs human review?
- Is implementation priced as a fixed scope, a day rate or a subscription?
- Who pays for re-recording a workflow after a supplier changes its site?
- What is the minimum commitment, overage rate and renewal increase?
- What is included in support, and what becomes a change request?
Run a pilot in shadow mode
Do not begin with live payment or admin access. Run the agent alongside a person, on real cases, in shadow mode: the agent proposes and records, the person completes the action. Compare completion without intervention, exception rate, time per case, cost per completed task and audit completeness.
Write acceptance criteria before the pilot starts. Include a stop condition for data exposure, repeated wrong submissions or an unexplained change in behaviour. A pilot that cannot fail is a demonstration, not a test.
Red flags in agent demos
- The demo only works on a clean test site with prepared data.
- No step-level log, or logs that cannot be exported.
- The vendor asks for a shared administrator account.
- “Fully autonomous” is used without explaining approvals, limits or escalation.
- Pricing is quoted per seat while run and model costs remain undisclosed.
- No process for site changes, and no test suite for the workflow.
- No exit plan, no data deletion commitment and no named support owner.
Decide by task type, not by brand
Buy a vertical agent when the task is common, the target system is stable and the vendor can show the exception path. Buy a platform when you have several browser tasks and an internal operations team to own them. Use a build-and-run partner when the process crosses many systems and judgement is frequent. Build only when the workflow is a genuine differentiator and you can maintain it as a product.
Start with a task that is repetitive, low-value and reversible. Earn trust before you let an agent touch money or regulated records.
If you want help mapping the process, running a controlled pilot and setting the approval model, start with Business Automation. The pricing page sets out how discovery, build and support are scoped. For the vendor side of the same decision, read how to evaluate AI automation vendors as your needs grow.
Sources
- NIST, AI Risk Management Framework (AI RMF 1.0), accessed 17 November 2026.
- OWASP, Top 10 for LLM Applications, accessed 17 November 2026.
- Gartner, Applying uniform governance across AI agents, accessed 17 November 2026.
- W3C, WebDriver specification, accessed 17 November 2026.
Responses (0)
No comments yet. Add the first one below.
Leave a response
Your email address will not be published. Required fields are marked *.