What Should AI Never Be Allowed to Approve?
Last Updated: October 2026
An automated approval is a decision your software makes and acts on with no one saying yes first. The invoice posts, the order ships, the record changes, and the first person to see it reads the report later. Most businesses meet this question the wrong way round. They ask how sure the tool has to be, long before they write down which calls it may touch at all.
AI Smart Ventures has guided growing businesses through AI implementation in finance, client service and back-office work. One habit shows up in nearly every project. Teams can say in detail what the tool does well, and no one in the room can name the calls it must never make.
Skip that list and you learn it case by case, in public. A credit note goes out that no one agreed to. A record changes and the client spots it before you do. Each slip looks small on its own, and together they cost you the thing that takes longest to rebuild, which is being believed.
Key Takeaways
- Sort by cost, not by count: the number of approvals tells you where time goes, while the cost of one wrong call tells you where harm goes, so build the list from the second.
- Four kinds stay with a person: calls that touch pay, a job or a lasting record; calls a regulator or a client contract reserves; calls you cannot take back; and calls that break a promise you made.
- Most teams cannot stop an agent mid-job: in the Kiteworks 2026 Data Security and Compliance Risk Forecast, a survey of 225 security, IT and risk leaders, 60% could not shut down a misbehaving agent, and 63% could not limit what it was allowed to do.
- Control bodies are writing this down: COSO put out Achieving Effective Internal Control Over Generative AI in February 2026, sorting AI uses into eight capability types so the controls match what a use can do.
Read those together, and the job gets smaller. You are not writing a policy about AI. You are writing one list of calls that stay human, and that list barely moves when the tools change, because it grows out of your business rather than out of a feature set. Human-first AI starts here, not in a tool review.
Why sort approvals by cost, not by volume?
Because volume and harm point in opposite ways. The approvals you make most often are the ones it costs least to get wrong, and the ones you make twice a year are the costly ones. Sort by count and the tool lands hardest where the stakes are highest, since those slow, awkward calls are what a busy team wants off its plate. Sort by cost and you guard the few calls that would really hurt.
Cost here is a question, not a number: what would you have to do if this call went wrong? If the answer is resend an email, the risk is small. If it is to phone a client, fix a filing, or explain yourself to a watchdog, it is not. Controls work now sorts the same way. COSO’s February 2026 AI guidance matched the depth of control to what a use can do, not to how much of it runs.
Which approvals should AI never make alone?
Four kinds, no matter what the score says. Work that touches a person’s pay, their job or a record that follows them around. Work a regulator or a signed client contract says a named person decides. Work you cannot take back once it has gone, such as a payment sent, a filing made, or a message reaching a client’s own customer. And work that breaks a promise you made, because the tool never saw the promise.
Those four are not ranked, and you do not get to pick three. Each names its own kind of damage: a person harmed, a rule broken, a step that cannot be undone, and trust lost. Where the rules of your trade settle the question, your compliance lead or counsel makes that call, not the vendor. Write their answer beside the category, so no one has to work out which of the four applies.
- Pay, jobs and records: hours, bonuses, hiring, discipline, and any note that lands in a staff file.
- Reserved by rule or contract: calls your regulator, your insurer or a client agreement says a named person makes.
- One-way doors: payments sent, filings made, data wiped, and work a customer has already seen.
- Promises you made: a due date, a service level, a pledge buried in a proposal the tool never read.
Which approvals can AI safely clear on its own?
The ones that repeat, follow a rule you can write in one sentence, and can be pulled back within the hour. Matching a receipt to a purchase order you agreed weeks ago. Routing a ticket to the queue its subject clearly belongs in. Booking a meeting room. Marking a form complete when every field passes its own check. None of those reach a client, and none of them are hard to undo.
A full example makes the shape clearer. A supplier invoice arrives, the tool reads it, checks the line items against the purchase order and the delivery note, then posts it when all three agree. Work that does not match stops and waits for a person. That is one automated workflow, and it is safe because the rule is checkable and a wrong posting can be undone the same day. Operational efficiency comes from the matching, never from dropping the stop.
How do you sort approvals into three lists?
Write down every approval your business made last week, then put each one in a single column: never, with a check, or on its own. Do it with the people who make those calls, in one sitting, and force a choice on every row. The arguments are the point. A row no one can place belongs in the middle column until someone owns it well enough to move it.
Keep the columns plain and the list short enough to read aloud. Ask two questions per row: who gets hurt if this is wrong, and can we take it back today? Those two settle most rows inside a minute. How sure the tool must be before it acts alone is the next question and its own job, so park it until this list is agreed. Deciding both at once is how teams end up with neither.
| Column | What goes in it | Who touches it |
|---|---|---|
| Never | pay, jobs, records, reserved calls, one-way doors, promises | a named person, every time |
| With a check | client-facing work, and any job with a date you promised | AI drafts, a person releases |
| On its own | repeat, rule-based, undone inside the hour | AI, with sampling later |
What keeps an approval list from going stale?
An owner, a date, and one rule: nothing new goes live until it has a column. Lists rot when tools change faster than the paper does, and AI features now arrive monthly inside software you already run. Give the list one named owner, review it on a fixed date each quarter, and make the column choice part of switching any new workflow on. That one rule beats any policy you could write.
Then test whether the list is real. Pick three things from the never column and try to make the system do them. If it will, your list is a wish rather than a control. Plenty of teams find out late: in that same 2026 forecast, 51% had agents in production while most could not stop one. AI gurus with zero operational depth sell autonomy levels and skip this list, which is why change management later gets blamed for a missing fence.
AI Smart Ventures builds AI Implementation around that list, so the approvals that must stay human are named and tested before a workflow goes live.
Frequently Asked Questions
What is an automated approval?
An automated approval is a call software makes and acts on with no person confirming it first. It covers everything from posting a matched invoice to sending a client a reply. The word approval is the part that counts: the system is not suggesting, it is committing your business. That is why the first question worth asking is which calls it may commit at all.
What are the different types of approvals in a business?
Most fall into four groups: money moving out, promises made to someone outside the business, changes to a record you must keep, and in-house housekeeping. The first three carry results you would have to explain if they went wrong. The last group rarely does. Sorting by group is far quicker than sorting row by row, and it hands you a default for new work.
How can I automate an approval process without losing control?
Start by naming what stays human, then automate the rest inside one narrow workflow. Write the rule the software will follow as a single sentence; if you cannot write it, the step is not ready. Add a stop for work that does not match, and route that case to a named person. Run it beside the manual path for a week before trusting it alone.
Does the law require a person to make some decisions?
Sometimes, and it turns on where you operate and what you sell. States moved fast in 2026: MultiState counted 155 bills on AI in the workplace that session, with two states passing broad AI employment laws. Your compliance lead or counsel decides which of them reach you. Treat their answer as a row in the never column, not a rule book of its own.
Can you undo an approval AI got wrong?
Sometimes, and you want that answer before you automate the step rather than after. Payments sent, filings made and messages a client has opened are one-way doors. In-house postings, draft records and queue routing can usually be pulled back within the day. Ask the question row by row while you build the list, because undoing is the cleanest single test you have.
Do customers mind when AI decides something about them?
Many accept it for the routine part and still want a person on the call. In a September 2026 survey of 1,000 US consumers by the mortgage firm Lower, 38% said AI should recommend while a human makes the final call, and 54.4% wanted a human the moment something went wrong. Comfort with speed is not comfort with judgment.
Who decides what AI is allowed to approve?
The people who own the outcome, not the people who own the tool. Bring in the one who answers to the client, the one who signs the filings, and the one who runs the workflow day to day. Your software vendor and your IT lead advise on what is possible. They should not set what is permitted, because they do not carry the fallout when a call goes wrong.
Should every team have the same approval list?
No. One company-wide list of never items keeps the floor steady, and each team adds its own rows on top. A client service team and a finance team meet their own one-way doors, so a single shared list either blocks safe work or allows risky work somewhere. Keep the shared part short, the local part exact, and both on one page.
How often should the approval list be revisited?
Quarterly for the whole list, and straight away when a new tool, workflow or client contract lands. AI features arrive inside software you already pay for, often with no notice at all, so a yearly review is far too slow to help. Put the review on one person’s calendar and keep a dated line at the top showing when the list last changed.
Where do we start with automated approvals?
Start with one week of real approvals and one sitting to sort them, with the people who make those calls in the room. Scope depends on how many workflows you run and how much of the sorting your own team can do, and most growing businesses get a first list agreed in a couple of sessions. Schedule a consultation to build yours around the calls your business really makes.
Executive Summary
Automated approvals should be limited by what a wrong call would cost, not by how many there are. Four kinds stay with a person whatever the tool reports: calls that touch pay, a job or a lasting record; calls a regulator or client contract reserves; calls you cannot take back; and calls that break a promise. Routine, rule-based work that can be undone may clear on its own. Sort last week’s approvals into never, with a check, and on its own, give the list an owner and a review date, then test the never column by trying to break it.
What Should You Do Next?
This week, list every approval your business made in the last seven days and sort each one into never, with a check, or on its own. Do it with the people who make those calls, and ask two questions per row: who gets hurt if this is wrong, and can we take it back today? Then try to make your systems do three things sitting in the never column.
AI Smart Ventures offers AI Implementation for growing businesses that want the limits agreed before a workflow goes live, with the workflow optimization and AI literacy to make them stick. Schedule a consultation to sort your approvals with clarity and confidence.
People Also Read
- AI Vendor Security Questionnaire: 25 Questions Owner-Operators Should Send
- AI on an Owner-Operated Shop Floor: Onboarding Staff Who Have Never Used a Chatbot
About the Author
Nicole A. Donnelly is the Founder of AI Smart Ventures and an AI Adoption Specialist with 20 years of experience as a founder and CEO and over a decade leading AI adoption initiatives. She helps businesses integrate artificial intelligence with clarity and confidence, driving innovation and sustainable growth. Nicole has trained over 20,217 professionals in Applied AI, delivered 624 workshops, and worked with close to 1,000 organizations across diverse industries.
Expertise: AI Transformation, AI Strategy, AI Implementation, AI Adoption, Applied AI, Marketing, Business Operations
Disclaimer: This content is for informational purposes only and does not constitute professional business or technology advice. Results vary based on industry, existing systems and implementation commitment. Contact AI Smart Ventures for a consultation regarding your specific situation.
