The approval step decides whether an AI system can be trusted with real work. It separates what a model can do from what a business is willing to stand behind. In the systems we build, the rule is plain: the AI drafts, a person decides, and nothing sends itself.
That can sound like a limitation. It is closer to the opposite. The approval step is what lets a company put AI in front of customers, money and records at all.
Why not let it run on its own?
The models are good, and they are getting better. A wrong answer in a chat window costs a moment. A wrong answer sent to a client under your name costs something else, and the model does not carry that cost. You do.
Most businesses do not reject AI because it is never right. They hesitate because they cannot tell when it is wrong. An approval step answers that directly. Every output that matters passes in front of someone who knows the business, and that person can be held to it, which no model can.
It also changes how the team feels about the system. A tool that acts on its own is something to worry about. A tool that hands you finished drafts is something you want to open in the morning.
What should wait for a person?
Draw the line by consequence, not by how hard the task is. Reading a hundred invoices is hard and low risk. Sending one email is easy and can be high risk.
- Reading and classifying what comes in
- Extracting figures from documents
- Matching records and flagging gaps
- Summarising a thread or a day
- Drafting anything at all
- Anything sent to a customer or a client
- Any quote, price or commitment
- Any payment or order to a supplier
- Any public post or review reply
- Anything that cannot be undone
The left column is where the hours are. That is why this works: nearly all of the labour sits on the side that needs no approval, and the right column takes a minute each.
How do you stop approvals becoming a bottleneck?
A badly designed review queue is its own kind of inbox. A few things keep it short.
Put it where people already are. A short queue on a phone in the morning, not another dashboard to log into.
Show the source beside the draft. The figure next to the invoice it came from, the reply next to the review. A person should check in seconds, not redo the work.
Say what the system was unsure about. Highlight the one thing it could not find, so attention goes there instead of spreading across the page.
Make editing the easy path. Changing a line and pressing yes should be faster than rejecting and rewriting. Those edits are also the best teaching material the system gets.
Route it to one owner. Every approval has a named person and a backup. A queue owned by everyone is owned by nobody.
If approving takes as long as doing the work, treat that as a signal. Usually the drafts are not good enough yet, and the fix is in the rules and the context, not in the reviewer.
When can a step lose its approval?
Approval is not forever. It is how a step earns trust.
Keep a log of every draft, every edit and every yes. Over time, some steps will show a clear pattern: approved without changes, again and again, for a low-risk action that is easy to reverse. Those can move to automatic, one at a time, by a deliberate decision from the person who owns the work. Others should never move. A price to a client and a public reply stay with a person no matter how good the record looks.
Step by step, the system does more, and the business decides how fast. That decision belongs to you and not to the vendor.
What does this mean when you are choosing a system?
Ask three questions of anything you are about to put in front of customers. What can it do without a person? How would I know if it got something wrong? Can I see who approved what last month? If the answers are vague, the approval step was added at the end, or not at all.
It works better the other way round. Decide what must never leave without a person, then design the system around that line.
If you want to see where those lines fall in your business, our roadmap ranks the opportunities and what each one would take.