The problem
The workflow is clear and the steps are known, but each item still waits days for a person to push it through, because every step is a copy and paste between systems. The one step that matters, the approval, sits in the same inbox as everything else, so it either waits with the pile or gets rubber stamped to clear it.
What goes wrong
- Full automation. The agent is allowed to write anywhere, so the first wrong change order posts to accounting before anyone sees it, and trust is gone on day one.
- Approval on everything. Every item needs a click, the approver gets forty cards a day, and the automation saved nobody any time.
- No record. Something changed in the schedule, nobody can say who approved it or what the agent saw, and the whole thing gets switched off.
The approach
- Map the workflow with the people who run it, and write down which steps change money, schedule, or a customer. Those get an approval. The rest do not.
- Let the agent read, classify, and draft first. Your team corrects it in a review queue until the proposals match what they would have done.
- Put the approval where the approver already works: one card in Slack, Teams, or Telegram with approve, edit, and reject. A named role owns it.
- Log every step with who approved what, when, and what changed, so the record exists before anyone asks for it.
When it works
When the workflow has a trigger you can point at, rules you can write down, and a small number of steps where a mistake costs real money or trust. Change orders, vendor invoices, permit submittals, and purchase requests all fit.
The step you keep
Most workflow automation fails at the same place. The whole thing runs on its own until the first item that should not have gone through goes through. A change order with a cost line posts to accounting. The team turns it off and goes back to the inbox.
The fix is not less automation. It is putting the approval where it belongs and nowhere else. Anything the agent does that moves money, changes a schedule, or reaches a customer or a vendor goes through a named person. Everything else runs.
What the agent does, and what a person does
Take a change order on a residential build. A framing sub emails that rain cost four days on lot 14. Today a coordinator reads the email, checks the subcontract and the schedule, drafts the change order, finds the project manager, and updates three systems once the PM says yes. With the agent in place, the split looks like this.
The agent reads and classifies
It picks the email up from the shared inbox, opens a task, and tags it: schedule-only change, no cost. It attaches the subcontract terms and the current schedule so the approver does not have to find them.
The agent drafts
It writes the change order the way your last fifty were written, fills the fields from the email and your files, and flags anything it could not find.
A person approves
The project manager gets one card on Teams: CO 0231, four days on lot 14 framing, no cost change. Approve, edit, or reject. One tap.
The agent updates
On approve, the schedule moves, the owner notice goes out, and the task closes with a record of who approved it and when. On reject, nothing writes, and the draft goes back to the coordinator with the PM's note. A change order with a cost line never gets a one-tap card. It goes to the PM and the estimator together, with the draft and the sub's email attached.
Where the approvals go
In the mapping session we walk the workflow step by step and ask one question at each: if the agent gets this wrong, what does it cost? If the answer is a few minutes of cleanup, the step runs. If it is money, a missed date, or an awkward call with a customer, the step gets an approval and a named role that owns it.
Keep that to the few steps where a mistake is expensive. Enough to protect you, few enough that the approver still saves time. If an approver sees more than a handful of cards a day, the rules are wrong, and we fix the rules. A full queue gets tapped through without reading, and a false record of review is worse than the inbox.
How we build it
First come the map and the baseline. We sit with the people who run the workflow, write down the trigger, the rules, the approval points, the systems, and the exceptions, and agree the one metric we are moving.
Next the agent reads real items and proposes what it would do. Nothing writes. Your team corrects it in a review queue, and the approval cards go live on your channel.
Then approved items start updating your systems. Exceptions route to a person with a reason. The agent runs beside your current process until the numbers match.
Then your team works from it. Monitoring switches on. You get the runbook and the first scorecard.
When it does not work
When the process lives in one person's head and changes every week. Write it down first, even badly, and bring that to the mapping session.
Where to start
If you know the workflow, start with the Workflow Agent. If you are not sure where the approvals go, ask Roo, Webaroo's AI: Which steps in our workflow should still need a person's approval?
Book 15 minutes with Connor