Research: Multimodal Agents Approach Human-Level Performance on Web Tasks
A benchmark of realistic web errands shows frontier multimodal agents completing tasks at roughly 70% of human speed and accuracy, up from under 20% a year ago.
Analysis
The gains come from better grounding, longer action histories and improved error recovery. Researchers caution that reliability on safety-sensitive tasks still lags.
Why it matters
Tracking the capability frontier lets enterprises time their pilots to when an agent can actually finish a job end to end, not just attempt it.
Source: AI Frontier
Related use cases
Autonomous Invoice Processing
An agent ingests invoices from email and portals, extracts structured fields, three-way matches against POs and goods receipts, and routes exceptions with a prepared resolution note to the right approver.
Customer Support Triage and Resolution
An agent classifies incoming tickets, retrieves relevant knowledge and prior resolutions, drafts a response, and resolves simple cases directly while escalating complex ones with a summary.
L1 IT Helpdesk Resolution
An agent handles common L1 requests through guided conversations, executes approved remediation scripts and opens tickets with full context when escalation is needed.
Automated Sales Lead Qualification
An agent enriches inbound leads, scores them against an ideal profile, drafts personalized outreach and books meetings, handing off warm conversations to reps.
Procurement Vendor Onboarding
An agent collects vendor paperwork, runs KYB and compliance screens, validates data against source documents and creates the vendor record once checks pass.
HR Employee Onboarding Orchestration
An agent orchestrates the onboarding checklist across systems, provisions access per role, schedules orientation and escalates blockers to the right owner.
Regulatory Reporting Automation
An agent assembles reporting data from source systems, validates it against rule logic, drafts the report with a full evidence trail and routes it for sign-off.
Field Service Dispatch
An agent ingests service requests, optimizes technician assignments against constraints, and notifies technicians with a prepared job packet.