“Who approved that?”
Phase two started before anyone signed the plan off. Now finance is unwinding it.
ConvOps runs recurring operations as governed AI agent workflows. Plans wait for sign-off. Phases run strictly in order. Reports have one shape. SOPs run as steps instead of sitting in a wiki.
claude · cowork · chatgpt · phases in order · sign-off before anything starts
new chat
Expense policy rollout
The steps are in someone's head or a wiki nobody opens. AI makes it faster to skip them.
“Who approved that?”
Phase two started before anyone signed the plan off. Now finance is unwinding it.
“Where did that decision go?”
Why the approval limit is 500 lives in a meeting nobody wrote down.
“What happens when someone leaves?”
The person who knew the month-end close is gone. So is the process.
A real change, run as a plan from Claude. Then the weekly report, on a schedule.
You describe the change in plain words. ConvOps creates a plan and drafts its phases. The work has a shape before anyone starts.
Each phase is its own ordered chunk of work. Phase two stays locked until phase one lands. The engine checks that order, not etiquette.
locked · waiting for sign-off
The plan stops at the approval gate. You read the phases, change them or sign off. Until then, no phase starts.
Sign off the expense policy rollout?
3 phases · pilot first · old form retired last
The pilot runs and lands. Only then does phase two open. The order is checked on every advance, whoever is driving.
phase 1 of 3 · running
The expense SOP stops being a wiki page. It becomes a workflow the AI follows step by step, with an approval where a manager decides.
The weekly report runs on a schedule. Each fire creates the task with the whole process attached, so every week has the same shape.
Phases run strictly in order, and nothing starts before you sign off.
The plan is described, scoped and broken down.
Child phases, each an ordered chunk of work.
No phase starts until a person signs the plan off.
needs a personEach phase must land before the next one opens.
The result is checked against the plan.
The plan closes only when every phase is done.
Kept by the business, not by whoever was in the meeting. Recalled when the work needs it.
Claims over 500 need a manager. Set in the Q3 review, with the reason.
why we did it
The policy doc, the expense form and the finance sheet. Found once, kept.
where it lives
Field-level history on every task and workflow run. Reviews stop being archaeology.
what changed
Every memory points back to the plan or task that produced it.
traceable
Phases run under a sequential_children gate. The next phase activates only when the one before it is done.
requires_approval holds the plan. The engine refuses to advance until it is met, and says what is missing.
Weekly reports run on an RFC 5545 schedule. Each fire creates the task, or resumes the same one where it stopped.
Field-level before and after on tasks and workflow instances, plus an activity feed. The agent label and the signed-in person stay separate.
Connect the AI tool your team already uses, install the workflows, and run one real piece of work through them.
Install the Business Ops pack: meeting notes, weekly report, SOP, decision memo
Run this week's report through the workflow: same shape as next week's
Move one real process into a plan: phases execute in order, after your approval
Write the SOP as steps the AI executes, not prose it might read
ConvOps is the operations layer for AI agents: an MCP server that holds your team's process as workflows, with approval gates, a shared memory and an audit trail. It runs no AI models.
That is the move: write the SOP as workflow steps with instructions, and the AI executes it instead of maybe reading it. The SOP Documentation template is the on-ramp; the result is process that runs rather than process that sits in a wiki.
Approval gates hold the run until the accountable human says go, and sequential ordering means phase two cannot start before phase one lands. Both are engine-evaluated, not etiquette.
A high-level activity feed plus field-level before-and-after diffs on tasks and workflow instances, each row separating the agent's label from the server-verified human. Reviews stop being archaeology.
No. Put the weekly report on an RRULE schedule and each fire creates the task carrying the whole process. You govern it at the gates.