What "approval workflow" means when AI does the drafting
The enterprise version of this term means routing documents between sign-off layers. The small business version is leaner and more useful: AI watches the tools you already run on, prepares the next action those tools imply, and queues it for you. Each prepared action is a card: the draft itself, the source record it was built from, the reasoning, and a risk level. You approve, edit, or decline. Nothing executes without that click, which is why the model is called human-in-the-loop automation.
The point is not oversight ceremony. It is that composing is slow and deciding is fast, so a system that moves all the composing to AI and leaves only the deciding to you compresses admin work to minutes without ever letting an unreviewed action reach a customer or your books.
What flows through the queue
- Invoices and payment reminders. Drafted from QuickBooks and the email thread where the work was agreed; amounts, terms, and history cited. The class of action where the approval step earns its keep on the first wrong-amount catch.
- Customer and vendor email. Replies to pricing questions, order status requests, and follow-ups that would otherwise go cold, each drafted in your context, waiting for your read.
- Orders and fulfillment updates. Confirmations, restock notes, and shipping ETAs prepared from Shopify and ShipStation data.
- Payment links and refunds. Stripe actions with the exact amount pre-filled and flagged high risk by default, because money moving is exactly what a human should look at.
- CRM updates. Deal stages, logged threads, and follow-up tasks in HubSpot or Salesforce, prepared so the pipeline stays true without anyone doing data entry.
The loop, end to end
Link the tools over OAuth. flo.space starts read only: it can understand your context but cannot send or write anything.
Prepared actions appear as your real context implies them: an invoice to send, a thread to answer, an order to confirm. Each shows its draft, source, reasoning, and risk.
Read the draft against the cited source. Edit in plain language if something is off ("use net 30", "warmer tone"). Approve, decline, or reassign.
On approve, the action runs in the real tool: the email sends, the invoice posts, the link is created. Every decision lands in the audit trail with its evidence.
"Approving things is just more work"
It is the reasonable objection, and the arithmetic answers it. An invoice done by hand means finding the thread, checking the price, opening QuickBooks, filling the fields, writing the email: ten to fifteen minutes. Reviewing the same invoice prepared with its evidence attached: perhaps twenty seconds, most of it reading the number next to the source. Across a day's queue the composing time is gone and the deciding time was never the problem. What review actually costs you is the illusion of full autopilot, and for actions that leave the building with your name on them, the documented record of unattended AI prices that illusion pretty clearly.
For the tasks where autopilot genuinely is the right call, internal, reversible, low-stakes, use it; the sorting framework is in copilot vs autopilot. An approval workflow is for everything else.
A morning with the queue, concretely
Tuesday, 8:40 a.m., a wholesale operation with the inbox, QuickBooks, Shopify, and Stripe connected. The queue holds nine actions. Three are low-risk order status replies drafted from live Shopify data; each takes a glance and an approve. Two are invoice drafts from yesterday's confirmed quotes, amounts cited against the threads where they were agreed; one is right, one gets the edit "use the pallet price from Friday's email" and approves corrected. There is a payment reminder for an invoice that crossed 30 days overnight, a restock note flagged because stock on the top SKU crossed its floor, and a Stripe payment link for a new customer, flagged high risk by the amount, which gets an extra thirty seconds against the quote before approval. The ninth is a follow-up to a lead that went quiet twelve days ago, which nobody would have remembered. Nine decisions, about six minutes, and everything that left the building was read by the person whose name is on it. That is the entire product category in one sitting.
Common questions
A queue of AI-drafted actions, each reviewable and editable before it executes. The AI reads context from your connected tools, prepares the invoice, reply, or order update, and the action waits with its reasoning and source shown until a person approves it. Approval is the send.
Reviewing is faster than composing. The slow parts of admin, finding the context, cross-checking the numbers, writing the words, are exactly what the AI does before the action reaches you. What is left is a seconds-long decision made against the evidence on screen.
In flo.space, outbound actions never skip a human; risk levels change how much scrutiny each one gets, not whether a person is involved. Low-risk actions clear in a glance, high-risk ones (large amounts, contracts) are flagged for a deliberate look.
It does not execute, and the decision is recorded. You can decline outright, edit the draft in plain language and approve the corrected version, or reassign it to the teammate who has the missing context.
Gmail, Outlook, QuickBooks, Stripe, Shopify, Slack, HubSpot, Salesforce, Google Drive, Google Sheets, ShipStation, Twilio, DocuSign, and LinkedIn. Connections start read only, so the queue proves itself before anything can execute.
Put your admin on approve
Connect your tools read only and watch the queue prepare invoices, replies, and orders from your real context. Approving takes a click; nothing happens without it.