How much should AI do on its own?
Guided by Otto, written by Something Clever Co., updated
It depends on what a mistake would cost. For new or sensitive work, let AI suggest and have a person act. For most everyday jobs, let AI do the work and have a person approve it before it goes out. Only routine, low stakes work with a good track record should run on its own, with a person checking in on a schedule.
Why this is the right question
“Should we let AI do this?” is usually the wrong question. A better one is “how much should AI do on its own here?” The same tool can be perfectly safe drafting a reply and risky sending it. Treating AI as all or nothing leads businesses to either avoid it completely or switch it on and hope.
It helps to think in three settings.
AI suggests, you act
AI never touches your systems directly. It drafts an email, summarizes a call, proposes a price for a quote, or suggests which lead to call first. A person reads it and decides what to do.
This is where most people already are when they use ChatGPT or Claude. It is low risk because nothing happens unless a person does it. The downside is that a person still carries every task across the finish line, so the time savings have a ceiling.
Good fits: writing help, research summaries, first drafts of proposals, anything brand new.
AI does it, you approve
AI does the actual work in your tools, then waits. It might prepare the CRM update, draft the customer reply, or build the task in your project tool. A person looks it over and says yes before anything goes out or changes.
This is the sweet spot for many small businesses. You get most of the time savings, and a person still catches mistakes before customers see them.
Good fits: messages your customers see, CRM updates, website edits, anything where a mistake is visible or costly.
AI handles it, you check in
AI acts without waiting, but only on routine work with clear rules. Someone looks over what it did on a schedule instead of one item at a time. If you go this far, keep it to the tools that job needs, set a spending limit if the tool is billed by usage, and make sure someone hears about it when something breaks.
Good fits: work that comes in high volume, is clearly defined, is easy to reverse and already has a good track record with a person approving.
A quick comparison
| AI suggests | AI does, you approve | AI handles it | |
|---|---|---|---|
| Who takes the action | A person | The AI | The AI |
| When a person checks | Before every action | Before it goes out or changes | On a schedule |
| Time saved | Some | A lot | The most |
| Risk if AI is wrong | Very low | Low, caught before it goes out | Depends on how easy it is to undo |
| Best for | New or sensitive work | Most business jobs | Proven, routine work with little at stake |
When to give AI more room
Only with evidence. The job has been running for a while, people rarely change what the AI produces, and a mistake would be cheap to reverse. If any of those are missing, keep a person approving.
Going back is just as normal. If a tool changes, mistakes creep up, or the job starts touching something more sensitive, put a person back in front of it until it settles.
Common mistakes
- Letting AI run alone from day one. You have no track record to justify it.
- Approving everything forever. If people approve the AI’s work unchanged for months, the check may not be doing much.
- One setting for the whole company. Decide job by job.
- Running alone with no limits. Without a spending limit or someone watching, you are just hoping.
- Handing AI judgment calls. Pricing exceptions, complaints and anything legal should keep a person deciding.
