AI acts directly,
when the action is cheap to undo
Native AI is the act zone of Onplana's three-zone decision model. Five operations that AI performs without waiting for an "accept" click, because each one is bounded in scope, cheap to undo, and auditable. Plan draft, natural-language parsing, status report first draft, portfolio Q&A, recommendations widget.
What the act zone is, and why these five
Onplana's three-zone model splits every AI operation into one of three buckets: act (AI decides for you), suggest (AI proposes, you decide), and stay-out (AI never touches). The full model is the subject of a long-form post; this page is the deep dive on the act zone.
The five operations in the act zone share three properties: the input space is bounded, the output is cheap to undo, and the work is high-frequency enough that a "are you sure" gate would be more annoying than useful. That combination is what lets AI act directly without forcing a confirmation step the user would learn to click through.
The boundary is set in code, not by the prompt. A misbehaving prompt cannot ask AI to extend its authority into the suggest or stay-out zones; the tool gates do not exist for the model to reach.
The five operations
For each, what you put in, what AI does, and what reverting looks like if the model gets it wrong. Plan tier listed top right.
Plan draft on kickoff
Input
A sentence, a paragraph, a meeting transcript, a sponsor email
Output
A real plan, not a preview: epics + tasks + subtasks + milestones + risks + estimated timeline
Reverting: Reverting the whole tree is one click. Editing any node is normal task editing. Regenerating the whole draft from a different brief is unbounded.
Natural-language parsing
Input
"Add a task for Sara to review the API spec by Friday"
Output
A task created directly: title, assignee resolved, due date calculated, inherited project context
Reverting: Wrong parses are noticed at a glance and edited like any other task. Per-team trust ladder lets the workspace flip the parse-preview gate off once acceptance rate is above 95%.
Status report first draft
Input
A project ID, optional period range
Output
A sponsor-ready draft saved (not published): exec summary, accomplishments, blockers, next-week plan, open risks
Reverting: The PM edits before publishing. AI never publishes on its own. The same engine powers the free Status Report Writer tool.
Portfolio Q&A
Input
"Which projects slipped this week?"
Output
Direct answer with cited rows. Reading, not writing.
Reverting: Nothing committed. Re-asking the question with a different framing is free.
Recommendations widget
Input
Implicit: your current dashboard context
Output
'What should I look at next' suggestions that refresh without confirmation. Hints, not state changes.
Reverting: Hint cards dismiss in place. Acting on a hint goes through the normal UI for that action.
Reversibility, counterfactual cost, auditability
The three tests an operation passes to land in the act zone.
Reversibility
Can the action be undone in a click, cheaply, and without anyone outside the team noticing? If yes, act-zone candidate. If no, at most suggest, more likely stay-out.
Counterfactual cost
If AI is wrong, what does the wrong outcome cost? A misparsed task wastes 30 seconds of edit time. A wrong baseline approval recalibrates a six-month commitment. The first is act; the second is stay-out.
Auditability
Does the operation produce a record explaining why the AI did what it did, what data it saw, and how a reviewer could check? Every act-zone operation produces an auditable trail by design.
The audit trail behind every act-zone operation
Per-project AI activity log, written on every act-zone decision.
Each log entry captures
- The prompt that triggered the action
- The retrieved context the model saw (including anything pulled in through Onplana's connectors to MS Graph, SharePoint, an MCP server, or an inbound webhook)
- The action taken
- The user who initiated the operation
- The timestamp
The log is filterable, exportable, and visible to project members by default. The "why" link on any AI-generated artifact opens the entry inline so the PM can see exactly what the model was reasoning over.
The trust ladder, widening the act zone over time
The act and suggest boundaries are designed to be tuned. The stay-out zone is designed not to be.
Auto-apply NL-parsed tasks
Default: preview the parsed task. Teams with acceptance rate above 95% over two weeks usually flip the preview off.
Auto-publish status reports on a fixed cadence
Default: save as draft. Teams with under 10% edit volume over two weeks usually flip to auto-publish with a 24-hour edit window before the report ships to stakeholders.
Auto-accept low-confidence risk dismissals
After the same false-positive dismissal repeats three times for the same signal class, the AI stops flagging that signal class for that project.
A quarterly suggestion in the admin console surfaces these candidates with the data backing each one. Admins accept or skip. Narrowings work the same way. The trust ladder is a control, not a one-way ratchet.
The other two zones
Native AI is the act zone. The suggest zone (AI proposes, you decide) and the retrieval substrate that feeds both live on their own pages.
You're here, Native AI
The act zone: AI commits directly because the action is cheap to undo.
Connectors
AI's retrieval substrate: MS Graph, SharePoint, MCP server, inbound email. The audit log captures what was retrieved.
Read moreAgents
The suggest zone propose-ratify model: risk flags, resource shifts, what-if, scope impact, baseline drift.
Read moreFrequently asked
Doesn't 'AI acts without asking' mean less control?
No, because the trade-off is calibrated. The five act-zone operations share three properties: bounded input space, cheap to undo, and high-frequency enough that an 'are you sure' gate would be more friction than safety. Anything that affects existing committed state (risk flags, resource shifts, scope-change impact, baseline drift) lives on the suggest zone with a preview-then-accept loop. Anything that creates legal, financial, or interpersonal cost (financial commitments, baseline sign-off, performance reviews) lives in the stay-out zone with no AI authority. The boundaries are the control mechanism.
How does the AI see my project data without a vector index?
Through typed tool calls scoped to your org's X-Organization-Id header. When AI needs context, it calls list_tasks, get_project, list_risks, etc. The tools return only that org's rows. No vector embeddings sit in a shared retrieval pool, no global index, no cross-tenant leakage path. The trade-off vs vector RAG: lower context recall on long-tail queries, perfect freshness, and tenant isolation that holds at the tool layer.
What happens when the AI gets it wrong on a plan draft?
Regenerate from a different brief, or keep the plan and edit the wrong assumptions like any other task. Plan drafts land as real plans because regenerating is unbounded and editing is normal. The act-zone framing makes the 'wrong' case cheap to recover from.
What's in the audit trail for an act-zone operation?
Every act-zone decision writes a per-project AI activity log entry: the prompt that triggered it, the retrieved context the model saw (including any context pulled in through Onplana's connectors), the action taken, the user who initiated it, the timestamp. The log is filterable, exportable, and visible to project members by default. The 'why' link on any AI-generated artifact opens the entry inline so the PM can see exactly what the model was reasoning over.
Which AI provider does native AI use?
Whichever your admin configured at the org level. Onplana ships with dual-provider support, Claude (Anthropic) and Azure OpenAI. The native AI surfaces use the tier abstraction (fast / balanced / powerful); per-org config maps that to a concrete model on the chosen provider. See the /ai pillar for the dual-provider section.
Can the act zone be widened or narrowed per team?
Yes, the trust ladder is configurable. A quarterly suggestion in the admin console surfaces candidates with data ('your team has accepted 47 of 49 AI-parsed tasks without edit over 14 days, auto-apply parsing on this workspace?'). Admins accept or skip. Narrowings work the same way: teams that find a specific operation noisy can pull it back without losing the rest. The stay-out zone is the exception, hard-coded off limits regardless of admin settings.
See the act zone on your own project
Sign up free, draft a project from a sentence, parse a task from one line of text, generate a status report in 30 seconds. The full three-zone model is visible from day one.