Approvals that stall the work — speed them up, do not remove them

The work is finished. The quote is written, the reply is drafted, the document is ready. It is sitting in a queue waiting for one person who is on site, in a meeting, or simply has ninety unread emails. The customer, meanwhile, reasonably concludes you are not that interested.

This is the workflow where AI produces the most obvious win and the most avoidable disaster, because the apparent solution — let the machine send it — is the one thing you should not do.

Why removing the approval is tempting

The arithmetic looks unanswerable. If the bottleneck is a person and you remove the person, the bottleneck disappears. Response time collapses from hours to seconds. In a demo it looks like pure gain.

What the demo does not show is the distribution of outcomes. Most automated replies are fine. A small fraction are wrong in ways that are expensive: a price quoted that you cannot honour, a commitment to a deadline you will miss, a confident answer about a policy that changed last month, a tone that reads as dismissive to someone who is already annoyed.

You will not find out from a log. You will find out from the customer, usually late, and by then the thing you automated has been repeating the error for a fortnight.

The distinction we use

Not everything needs approval, and treating every step as consequential produces a system so slow that people work around it. The line we draw:

No approval needed — anything reversible and internally visible. Classifying an enquiry, drafting text, summarising a thread, routing to an owner, preparing a report, flagging something for attention.

Approval required — anything that leaves the building or commits the business. Customer replies, quotes and prices, contractual dates, refunds and discounts, anything a regulator or a lawyer might read later.

The useful reframe: approval is about irreversibility, not importance. An internal report can be very important and still not need a gate, because if it is wrong you fix it and nobody outside knows.

What actually fixes the stall

If the approval stays, the bottleneck has to be attacked a different way. Four things that work, roughly in order of impact:

1. Make the decision small. Most approval delay is not deliberation, it is context-loading — the approver has to reconstruct what this is about before they can judge it. Put the draft, the original request, and the relevant history on one screen and the decision takes seconds instead of minutes.

2. Batch, do not interrupt. Twelve separate notifications produce twelve context switches and get ignored. One queue reviewed twice a day gets cleared.

3. Widen who can approve. Very often one person is the bottleneck because nobody defined a second approver. Role-based approval with a value threshold — anything under X can be approved by the team lead — removes most of the delay without removing control.

4. Auto-approve the boring tail, explicitly. Some categories genuinely do not need a human: appointment confirmations, "we received your message", scheduling acknowledgements. Carve those out deliberately, list them, and review the list. That is different from removing approval by default.

The pattern we build

Draft prepared automatically → shown with full context → a person approves, edits, or rejects → send happens → every step logged with who did what and when.

Two details matter more than they look.

Edit must be as easy as approve. If editing is awkward, approvers either rubber- stamp bad drafts or abandon the system. The edit path is the one to make fast.

A rejection should be recorded with a reason. Not for compliance theatre — because the pattern of rejections is the most useful signal you will get about where the drafting is weak.

Why the log matters commercially

An audit trail sounds like a compliance feature. In practice it is the thing that makes the system trustable enough to keep using.

When a customer disputes what they were told, "here is the message, who approved it, and when" ends the conversation in a minute. Without it, you are reconstructing events from memory and inboxes. That is also why we log automated steps in workflow automation generally — automation you can actually inspect covers the same principle applied to multi-step flows.

What this looks like built

Human approval runs through every build rather than sitting in one of them. The clearest expression is in the assistant work, where a drafted answer is shown with its source before anyone acts on it — document answers that resolve to a page. ClientPilot AI shows the customer-messaging side with the approval step in the path.

Honest limits

• Approval queues need a service level or they rot. If nothing enforces "cleared twice a day", the queue becomes the new place work goes to die.
• A human gate caps your response time. If you genuinely need sub-minute replies at 2am, this pattern will not deliver it, and you are choosing between speed and control. Say so out loud rather than discovering it later.
• Approver fatigue is real. Past roughly thirty items a day, people stop reading carefully. If volume is that high, the answer is better auto-approve carve-outs, not a more determined approver.

How to tell if this is your problem

Look at one recent quote or customer reply and ask how long it sat finished before it went out. If the answer is hours rather than minutes, and the reason was availability rather than deliberation, this is the workflow to fix.

← Back to AI for owner-led service businesses