Workflow Observer · managed pilot
Bring the workflow.
Keep its history.
Enterprise work happens across people, systems and weeks. Bring permitted activity into one case, review what belongs, then let an agent attempt the workflow inside a world you can inspect and branch.
Start with Capture a human workflow in Console. Mac and Windows companions support the managed pilot; signed public installers are not yet available.
Opportunity 042
37 days · one case
Choose what can be observed
Approve the account, channel, pages and native window for this opportunity. Only approved sources and permitted evidence are retained.
Synthetic histories and simulated outcomes. This preview connects no sources and runs no agent.
From work to a testable world
A case can last 45 days. The history should hold together.
- 01
Define the boundary
Choose the business purpose, participants, sources, content modes and retention. An approved source is a precise scope, not permission to capture the entire desktop.
- 02
Follow the work
Keep sessions separate from the business case. Record permitted observations with source identities, timestamps and coverage gaps.
- 03
Review the history
Use AI suggestions to sort relevant activity, restore mistaken exclusions and resolve uncertainty. Automatic sorting needs measured calibration. The workflow owner verifies the exact history.
- 04
Run it. Inspect it. Branch it.
Review the proposed steps, add business rules and approve the evaluation rights. Create an executable Universe, watch an approved model attempt the workflow, and branch from a recorded checkpoint.
Choose what enters the case
Capture with a clear boundary.
Start with the sources that matter: an approved browser page, a selected desktop window, or specific email folders, channels, calls and CRM records. Keep content review separate from the decision that an event belongs to the workflow.
- Browser & desktop
- Exact-page browser metadata and explicitly started Mac or Windows sessions for one approved native window. Optional desktop continuation can record while up to three encrypted batches, totaling at most 50 MB, queue for upload. Permission must stay current within fifteen seconds; a full queue or unavailable permission pauses capture. Restart stays stopped. Manual encrypted export remains available.
- Reviewed screen text
- Request a screenshot, mask sensitive regions and review it. In Console, English OCR can draft a transcription for human correction and approval. An approved world uses that reviewed text with its source visibility.
- Service connections
- Read-only connectors for selected Slack and Teams channels, Microsoft mail folders, Gong calls and Salesforce opportunities. Durable checkpoints resume collection; Slack revisits retained threads it has already discovered.
- Managed review
- Encrypted project storage, approved receiver bindings, explicit content review and owner verification. Revocation and retention remove captured evidence and derived worlds. A late event or correction makes the earlier review stale.
- Pilot qualification
- Native distribution and each customer source need qualification. Screenshot masks and text redaction need human inspection. Reviewed rules define the simulation; capture does not prove complete coverage or reconstruct unseen business logic.
Understand the timing before you automate.
A captured history can help an operations team ask where a case slows down, which handoffs need investigation and what evidence is missing. Reviewed calendar intervals and timestamp gaps answer different questions; neither establishes active labor time or employee productivity.
Workflow efficiency analysis: inputs, evidence and pilot scope →Human work and agent work meet in evidence.
Observer brings reviewed human workflow history. Open-source Guard records covered agent activity. Governance controls which models, data and budgets an attempt may use. Universes execute approved workflow rules so you can inspect a path, replay it and explore a branch.
Evaluation answers a separate release question through an admitted task, environment and evaluator pipeline. A useful workflow world is the starting point for testing its declared rules; a benchmark grade requires that separate measurement contract.
For a precise read-and-decide question, Console connects the captured evidence to a model assessment: source permission, grader checks, Guard destination approval and independent human review. Freeze the conditions, run the approved model and compare matched attempts. Each result stays tied to its evidence and retention; one attempt does not establish general model performance.