AI proposes,
the human ratifies
Agents are the suggest zone of Onplana's three-zone decision model. Five operations where AI proposes a change to existing state, surfaces the evidence inline, and waits for the human to accept, edit, or reject. Risk flags, resource shifts, schedule what-if, scope impact, baseline drift.
The propose-ratify pattern
Five suggest-zone operations all share the same shape. AI brings the evidence; the human commits the state change. The AI runs autonomously on retrieval, analysis, and drafting; it stays deferential on the actual mutation of existing committed state.
That shape is the propose-ratify model. It's the discipline that makes agentic work safe at the scale of real PMO governance. Without it, agents either over-promise (auto-apply mutations to existing plans, then break on the first edge case) or under-deliver (only return text, never act). The suggest zone is where most of the time savings show up over a quarter, because the AI handles the analysis and the drafting and the human handles the judgment.
The boundary is set in code, not by the prompt. A misbehaving prompt cannot ask AI to auto-apply a suggest-zone proposal without going through the ratification gate. The gate is server-side and not a prompt-engineering trust signal.
The five suggest-zone operations
For each: what AI proposes, what evidence shows inline, and what ratification looks like.
Risk flags
Proposal
AI flags a task or milestone as at risk and names the specific signal driving the call
Evidence shown inline
The signal is shown inline: an overdue dependency, no progress in 14 days, owner on PTO during the planned window, the actual row count behind the rule
Ratification
The PM accepts the risk into the register, dismisses it, or routes it to an owner
Resource shift proposals
Proposal
When the heatmap shows a 130% allocation, AI proposes a specific shift like 'move task X from Sara to Raj for the week of June 8'
Evidence shown inline
Sara's loaded calendar plus Raj's available capacity rendered alongside the proposal, with the projected post-shift load for both
Ratification
The PM accepts, edits the target assignee or date range, or rejects with a reason
Schedule what-if
Proposal
AI runs a scenario ('what if we push QA by a week') and computes the resulting finish dates and critical path
Evidence shown inline
The recalculated CPM path, the milestones that move, the dependent projects whose finish dates change
Ratification
Nothing changes until the PM commits the scenario. Multiple scenarios can be compared side by side before one is selected
Scope change impact
Proposal
Adding a feature mid-project triggers an AI estimate of downstream effects
Evidence shown inline
Which milestones move, which resources get overloaded, which dependent projects need a heads-up, with the calculation shown
Ratification
The PM uses the estimate to inform the change-control conversation. The plan does not auto-rewrite
Baseline drift alerts
Proposal
When the live plan diverges from the baseline by a configurable threshold, AI proposes a rebaseline
Evidence shown inline
The specific tasks whose dates or duration changed, the cumulative drift in days, the variance vs the configured threshold
Ratification
The PM rebaselines (the change record is the PM's, not the AI's), defers, or adjusts the threshold for next time
And the stay-out zone, hard-coded
Some operations don't move into AI authority no matter how the admin configures the org. The boundary is enforced in code, not policy.
Financial commitments
AI can summarise a budget burn rate and surface a risk that spend will exceed the approved amount. It cannot approve a PO, commit a contract, or change an approved budget number.
Baseline sign-off
AI can recommend a rebaseline based on drift. The sign-off itself, the act that says 'this is now the plan of record,' is a human authority.
Performance reviews
Onplana stores task completion data, comment history, assignment patterns. AI never assembles those into a review of an individual. The audit trail exists; the synthesis is yours.
Vendor selection
AI can summarise an RFP response. The decision to award is not an AI output.
Termination decisions
Closing a project, archiving a portfolio, removing a user from a role are human actions. AI can surface that they may be warranted; it cannot do them.
The principle: where the wrong decision creates a legal, financial, or interpersonal cost that a "reverse" button cannot fix, AI does not act. The stay-out zone is small and bounded specifically because the cost of misplacing its boundaries is asymmetric.
The trust ladder, in both directions
Suggest-zone proposals graduate into act-zone behaviour when team acceptance rate justifies it. Narrowing works the same way.
Widening
"Your team has accepted 92 of 100 baseline-drift rebaseline proposals over the last 14 days. Auto-apply at the configured threshold?" Admins accept or skip. Auto-applied proposals graduate into act-zone behaviour for that workspace; the audit trail and the one-click revert stay in place.
Narrowing
After the same false-positive dismissal repeats three times for the same signal class, AI stops flagging that signal class for that project. Teams that find a specific proposal noisy can pull it back without losing the rest. Not a one-way ratchet.
How the propose-ratify pattern stays honest
Three invariants hold across every suggest-zone surface.
Evidence by default
Every proposal surfaces its underlying rows inline. Asking a PM to trust an unsourced AI claim about their own project does not survive contact with the first wrong answer.
Ratification gate at the tool layer
Server-side. A misbehaving prompt cannot bypass the gate, because the ratification check is not a prompt-engineering trust signal, it's a database-level write guard.
Audit on every step
Proposals, accepts, edits, rejections all write to the per-project AI activity log. The change record on a ratified suggest-zone action is the PM's, not the AI's; AI's authorship is the proposal, not the commit.
Frequently asked
How is the suggest zone different from the act zone?
Act-zone operations change state directly because reverting is cheap and high-frequency (plan draft, NL parsing, status report first draft, portfolio Q&A, recommendations widget). Suggest-zone operations change existing state in ways that are not trivially reversible (risk register, resource allocation, schedule, baseline), so they propose with evidence and wait for human acceptance before committing. The same AI engine drives both; the difference is the boundary at the tool layer.
What does "evidence shown inline" actually look like?
On a risk flag, you see the specific rows that triggered the signal: which dependency is overdue, when the owner last touched the task, the actual gap in days. On a resource shift, you see the source person's loaded calendar and the target's available hours, side by side. On a what-if, the recalculated CPM with the changed nodes highlighted. The evidence pattern exists because asking a PM to trust an unsourced AI claim about their own project does not survive contact with the first wrong answer.
What if I reject a proposal?
The rejection writes to the AI activity log with your reason. For repeated rejections of the same pattern (e.g. you dismiss the same risk signal three times in a row), AI stops flagging that signal class for that project. The trust ladder works in both directions: teams can narrow the suggest zone the same way they widen it.
Can the suggest zone be widened so proposals auto-apply?
Yes, for individual operations, when the team's acceptance rate justifies it. The quarterly trust-ladder suggestion surfaces candidates: 'your team has accepted 92 of 100 baseline-drift rebaseline proposals over the last 14 days; auto-apply at the configured threshold?' Admins accept or skip. Auto-applied proposals graduate into act-zone behaviour for that workspace; the audit trail and the one-click revert stay in place.
What sits in the stay-out zone, and can that be widened too?
Stay-out is hard-coded off limits regardless of admin settings: financial commitments, baseline sign-off, performance reviews, vendor selection, termination decisions. The boundary is in code, not policy. The principle is consistent: where the wrong decision creates legal, financial, or interpersonal cost that a revert button cannot fix, AI does not act. The stay-out zone is bounded specifically because the cost of misplacing it is asymmetric.
Is this what people mean by 'AI agents'?
It's the propose-ratify shape of the agent pattern, applied to PM-tool work. An agent receives evidence, plans a state-change proposal, surfaces the proposal with its evidence, and waits for the human to ratify. The five operations above are the first-party agents shipping in Onplana today; new ones land as more PMO workflows mature into the propose-ratify shape. External MCP-compatible clients (Claude Desktop, Cursor, your own agent stack) drive the same surface, see /ai/connectors.
Does the agent learn from my corrections?
Within a conversation, yes, the agent sees your rejections and amendments and adapts subsequent proposals accordingly. Across conversations, no, the agent does not train on your data and the next session starts from the same base behaviour. The trust-ladder mechanism is how cross-session learning actually lands: pattern of acceptance over weeks becomes a tunable widening or narrowing, never an opaque drift in model behaviour.
The other two zones
Agents are the suggest zone. The act zone (AI commits directly) and the retrieval substrate that feeds both live on their own pages.
Native AI
The act zone: AI commits directly. Plan draft, NL parsing, status draft, portfolio Q&A, recommendations.
Read moreConnectors
AI's retrieval substrate: MS Graph, SharePoint, MCP server, inbound email. What the AI saw is in the audit log.
Read moreYou're here, Agents
The suggest zone: AI proposes with evidence, the human ratifies.
See suggest-zone proposals on your own project
Sign up free, import a .mpp or kickstart a project from a sentence, watch the risk flags and resource-shift proposals start arriving with their evidence inline.