AI Pilot Experience · Browser Control

Browser-Based Record Preparation

Read one authorized portal or inbox and prepare a source-linked candidate record, download list, or exception queue without external communication or consequential browser action.

Start Free Workflow Assessment Review Bounded AI Agent Pilot

Operating brief

Starting point: One authorized portal or inbox, restricted identity, approved records, read scope, purpose, test cases, and named reviewer.
Pilot output: Candidate records, source links or captures, download register, exceptions, run evidence, corrections, and reviewed acceptance state.
Authority: The browser worker can read and prepare inside the approved scope; people approve every write, communication, submission, payment, deletion, or permission change.

Pilot experience

See what the user does, what the system prepares, and what the pilot must prove.

Portals and inboxes without suitable APIs often leave staff manually reading pages, downloading files, and transferring facts into internal records.

Computer-use models can interpret changing interfaces while Playwright can handle stable deterministic steps. Both operate through a restricted browser identity, approved source list, allowlisted actions, and recorded stop conditions.

The pilot prepares candidates and evidence for human inspection. It does not bypass access controls or perform consequential external actions.

Visual walkthrough

Follow the workflow, review interface, evidence, and decision states.

Each view shows how source records, system outputs, exceptions, corrections, and human decisions remain connected during the pilot.

Authorized browser workflow. — Authorized browser workflow
Approved inputs remain connected to preparation, deterministic checks, human review, acceptance, and release.

Candidate record register. — Candidate record register
Stable source, run, output, version, evidence, correction, and review records remain visible.

Portal action review. — Portal action review
Reviewers inspect claims, actions, files, dependencies, geometry, behavior, and prohibited implications before acceptance.

Browser exception queue. — Browser exception queue
Uncertainty, failure, permission, security, rights, validation, and stop conditions receive named owners and states.

Reviewed preparation package. — Reviewed preparation package
Only accepted outputs move to the approved audience or next controlled environment.

User journey

Move from approved collection to source-linked review through explicit stages.

The interface can feel simple while the pilot preserves document identity, revision, permission, retrieval, uncertainty, and review evidence behind every result.

Authorize source

Isolate session

Use restricted credentials, managed browser, allowlisted domains, narrow downloads, redaction, logging, timeout, and teardown.

Map stable steps

Use Playwright for deterministic navigation, selectors, downloads, captures, validations, and retries where suitable.

Use computer use selectively

Route ambiguous visual interpretation through an approved model with screenshots, action proposals, and observation checks.

Prepare candidates

Create record fields, download entries, source URLs, timestamps, captures, file hashes, uncertainty, and exception states.

Stop on uncertainty

Escalate CAPTCHA, login change, unexpected domain, modal, missing source, ambiguous field, external write, or permission issue.

Human review

A named reviewer opens source evidence, corrects fields, rejects candidates, resolves exceptions, and accepts records.

Close session

Revoke temporary access where applicable and retain run, source, action, download, correction, acceptance, and incident records.

Technology options

Choose each component for a defined task and accepted output.

The final selection records exact versions, licences, data handling, cost, limitations, replacement options, and the evidence required before use.

Component	Examples	Pilot role	Boundary
OpenAI Computer Use	Current approved OpenAI computer-use tool and model	Interpret approved browser screenshots and propose bounded actions.	The application must validate proposed actions and keep consequential actions prohibited or human-approved.
Anthropic Computer Use	Current approved Anthropic computer-use tool	Operate an isolated desktop or browser inside explicit tool and action limits.	Provider tool use does not establish portal permission, source truth, or authority to act.
Gemini Computer Use	Current verified official Gemini computer-use capability	Remain a replaceable option after exact model, API, region, policy, and evaluation review.	Capability availability must be verified; screenshots and actions can still be wrong.
Grok Bot pattern	Monitored vendor pattern only after official identity, documentation, access, policy, and test review	Inform private monitoring or a future bounded comparison.	A product announcement, bot, wrapper, or demo does not qualify it for authorized portal use.
Playwright	Versioned browser, selectors, assertions, traces, downloads, screenshots, and tests	Perform deterministic browser steps where interface behavior is stable.	A passing script does not establish business correctness or continuing permission.

Client-owned pilot outputs

The pilot closes with evidence and a decision, not just a demonstration.

Portal or inbox, account, role, source, purpose, data, retention, domain, destination, and action authority definition
Restricted identity, credential, session, browser, network, download, logging, timeout, and teardown controls
Page, folder, message, attachment, record, source URL, timestamp, capture, and file-hash register
Deterministic Playwright and selective computer-use routing with action proposal and post-action observation
Candidate field, source evidence, uncertainty, download, exception, correction, and acceptance interface
CAPTCHA, login change, unexpected domain, modal, missing source, ambiguous field, external action, and permission queue
Stable, changed-layout, missing-field, duplicate, stale-session, restricted, timeout, and adversarial-content cases
Read, propose, stop, correct, reject, accept, incident, revoke, expiry, and withdrawal states
Quality, critical failure, reviewer effort, latency, and complete cost evidence
Runbook, ownership, monitoring, recovery, limitations, and client-owned handover

Evaluation evidence

Measure retrieval, citations, review effort, latency, and complete cost together.

A fluent answer is not the acceptance unit. The pilot tests whether a reviewer can reach the controlling evidence efficiently and detect important failure.

Source support

Trace outputs to approved sources, actions, files, references, versions, and evidence.

Task completion

Measure completion against frozen cases and acceptance criteria, including difficult and stop cases.

Critical failure

Record unauthorized action, false claim, security event, dependency issue, missing evidence, or prohibited use.

Reviewer effort

Measure inspection, correction, rerun, exception resolution, testing, and approval time.

Accepted output

Count outputs accepted by accountable reviewers, not processing completion or model agreement.

Complete cost

Include models, tools, compute, storage, preparation, testing, review, correction, monitoring, support, and replacement.

Cost drivers

Cost follows document condition, usage, controls, and accepted outcomes.

Source and output volume, formats, revisions, media duration, portals, repositories, assets, scenes, and update frequency
Model, API, browser, compute, GPU, rendering, storage, transfer, build, test, retry, and provider-plan cost
Permissions, rights, identity, environments, schemas, prompts, scripts, fixtures, validation, manifests, and acceptance preparation
Security, privacy, confidentiality, secrets, network, region, retention, training use, dependency, licence, and incident review
Process, technical, security, subject, communications, accessibility, legal, and release review
Corrections, reruns, exceptions, monitoring, support, provider changes, expiry, withdrawal, portability, and handover

Scaling factors

Scale one accepted pattern without flattening project and industry differences.

Volume

Use durable source, run, work-item, output, review, correction, acceptance, release, expiry, and withdrawal IDs with queues and checkpoints.

Permissions

Constrain source, identity, credential, tool, repository, file, network, provider, destination, audience, and action scope.

Environment

Version models, APIs, browsers, tools, packages, dependencies, engines, operating systems, hardware, settings, tests, and policies.

Provider replacement

Keep approved sources, prompts, scripts, records, assets, tests, evidence, corrections, and accepted outputs client-owned.

Concurrency

Measure queues, sessions, rate limits, rendering, build time, timeouts, retries, reviewer capacity, and cost.

Production promotion

Require separate security, integration, monitoring, recovery, service, rights, ownership, training, and deployment approval.

Cross-industry reuse

Reuse the retrieval pattern while changing documents, authority, and review.

The reusable capability is permission-aware multilingual retrieval across versioned documents with citations. Each industry keeps its own source authority, terminology, consequence, retention, and qualified review.