Disclosure
Fundamental limits we don’t pretend to solve
These are domain-level limits, not bugs. Better models and more code won’t make them go away. Knowing what they are makes Onvoi safer to use.
Communication ambiguity
Human conversation routinely carries multiple valid interpretations. Onvoi picks one. It will sometimes pick the wrong one even when its confidence is green.
How it shows up
- "Let’s loop in design" — informational mention, or a delegation?
- "We should probably handle this" — commitment, or a discussion thread?
- "Send it to Alex" — which Alex, in which channel?
What you can do
- Read extracted tasks before dispatching.
- Use the assignee-confirmation modal when owner attribution is ambiguous.
- Edit task descriptions — your edits become learning signals.
Organizational knowledge bounds
Onvoi only knows what your meetings, documents, and chats have surfaced. Internal context that lives in someone’s head, a private channel, or a system you haven’t connected is invisible to it.
How it shows up
- Decisions made over coffee that never made it into a meeting.
- Project history from before Onvoi was connected.
- Tribal knowledge ("we never email that customer on Fridays") nobody wrote down.
What you can do
- Add facts via the Knowledge browser when you spot a gap.
- Connect Slack / docs / calendars so more context lands in the system.
- Treat the first month’s extractions as more error-prone — the corpus is still small.
Context dependence
The same sentence in two meetings can mean two different things. A "report" might be a Notion doc in one context and a board deck in another. Onvoi resolves context with the signals available; it can’t know what wasn’t said.
How it shows up
- "Update the doc" — which doc?
- "Same as last time" — last time being which time?
- "Get it to legal" — your legal team or a partner’s?
What you can do
- Resolve ambiguous terms via the clarification UI when prompted.
- Confirm new-entity resolutions on first use.
- Use deep-research ("Research this more") for low-confidence tasks.
Stale data is inevitable
People change jobs, tools change names, deadlines slip. Even a perfectly extracted task can become wrong by execution time. Onvoi tries to detect this (e.g. resolved deadlines, context-staleness warnings) but cannot eliminate it.
How it shows up
- "Next Tuesday" extracted in a meeting; dispatched two weeks later.
- A teammate listed as the owner has since left the project.
- A Slack channel named in the task was archived.
What you can do
- Confirm stale-deadline warnings before dispatching old tasks.
- Re-resolve context on tasks that have been sitting in the queue.
- Watch for amber "context stale" banners on task cards.
Value-alignment uncertainty
Onvoi optimizes for stated goals — it does not know what your team values when those goals conflict. "Ship faster" and "don’t cut corners" both look reasonable until the trade-off is forced.
How it shows up
- "Get this done by Friday" extracted as a literal commitment when the speaker meant "if it’s easy."
- Sending a customer-facing message under your name without a tone match.
- Optimizing for completion when the right answer was "stop and ask."
What you can do
- Use task-level dispatch constraints (read-only, draft-only, no external comms).
- Set the org communication tone in Settings.
- Keep high-stakes tasks in supervised or monitored agent mode.