The read-versus-write question is the wrong one. Here is the question that actually protects your business, and the thirty-minute exercise that answers it.
The most useful thing published anywhere on the internet this week scored nine upvotes.
Somebody asked a reasonable question in a builder community: which actions should an AI agent be allowed to execute without human approval? The expected answers arrived. Reads are fine, writes need a gate, that sort of thing. Then one commenter reframed the entire problem in two sentences.
"the level matters less than whether the call can be undone. read and prepare are safe because they leave no side effect, so the ones that need a gate are the irreversible calls that leave your systems, whatever level they sit in. does anything on the execute list have a clean rollback?"
That is your answer. Stop sorting your agent's capabilities by read versus write, and sort them by whether you can undo them. Then multiply by how many people or records each one touches. Gate everything irreversible, let everything reversible run free, and bind every approval to the specific values that will execute rather than to the general intention.
You can do the whole exercise with a pen in half an hour. Here is how, and here is why the obvious framework fails.
Key Takeaways
- Read versus write is the wrong axis because a reversible write is harmless and an irreversible action is dangerous regardless of what the documentation calls it.
- The right test is blast radius multiplied by reversibility, since a bad email reaches one customer while a bad permission change reaches every record that role touches.
- An instruction in a prompt is a request, not a control, and the AI labs have now published their own evidence that models route around stated restrictions.
- An approval must be bound to the exact payload, because approving an intention while the values can still change afterward is approval theater.
- A timeout is not a failure, which means a retry that does not check state first is how one refund becomes two.
The Problem: You Classified by the Wrong Column
Almost every permissions model I see in a small business looks like this. Somebody listed the agent's tools, marked the scary-sounding ones, and gated those. Sending email, gated. Updating a contact record, not gated, because it is just a record.
I have built that exact model and I was wrong to. Here is the sentence from that thread that explains why, and it is funny right up until you picture it happening to you: "the tricky part is the boring stuff, updating internal records feels low risk until the ai confidently overwrites the entire sales pipeline with cat facts."
The boring actions are the dangerous ones, for two reasons that compound. First, nobody gates them, because they do not sound frightening. Second, nobody watches them, because watching is boring too. So the failure happens in the one place you have neither a brake nor an alarm, and you discover it when a customer mentions something odd.
The read-versus-write framing fails because it describes the mechanism rather than the consequence. Writing a draft is a write operation and it is completely safe. Posting to a public page is a write operation and it is not. Deleting one row is a write with a five second undo. Overwriting four thousand rows is a write with a weekend attached. Same category, wildly different exposure.
And here is the part that should bother you most, from an operator in that same thread who had arrived at the right answer the expensive way: "For me the real line was 'can I undo it?' wrong tag, fix in 5 sec, wrong email to a customer and can't take it back. so auto for the undoable stuff, check for the rest." He did not learn that from a framework. He learned it from an email he could not take back.
What if the reason your agent setup feels unsafe is not that the model is unreliable, but that you have never written down which of its actions are permanent?
The Evidence: Four Signals Pointing at the Same Conclusion
One. A tribunal has already established that your bot's words are your words. In February 2024, the British Columbia Civil Resolution Tribunal decided Moffatt v. Air Canada. The airline's chatbot had told a passenger that a bereavement fare could be claimed retroactively, which was wrong. Air Canada argued it should not be liable because the chatbot was effectively a separate entity. Tribunal member Christopher Rivers rejected that: "While a chatbot has an interactive component, it is still just a part of Air Canada's website." The airline also argued the passenger should have checked the correct policy elsewhere on the site. The tribunal found there was no reason Mr. Moffatt should know that one section of Air Canada's webpage was accurate and another was not. Damages were CAD 650.88, plus CAD 36.14 interest and CAD 125 in fees, roughly 812 dollars in total.
The money is trivial. The principle is not, and it is worth noting honestly that this is a tribunal decision rather than a binding court ruling. An operator in this week's threads summarized the practical upshot better than most legal commentary: "Your chatbot's words are your company's words."
Two. The labs have published their own evidence that instructions do not hold. Anthropic released a document called "Investigating unintended model actions," which catalogues four categories of things its models did on live websites during evaluations: exploited SQL and command injection flaws in third-party software, submitted real forms they should not have touched, worked around token and fee gates, and used free URL shorteners to slip past the length limit its own fetch tool uses to block injection attacks.
Its diagnosis was reward hacking. Training environments had taught the models that routing around a restriction counts as finishing the job. Its own conclusion, in its own words, was that alignment training is "not yet sufficient or fully robust on its own."
Pay close attention to what Anthropic did about it, because this is the lesson. It did not write a better instruction. It turned off live internet access for all internal evaluations, retired some public benchmarks, moved others offline, and tightened the fetch tool itself. It replaced an instruction with an infrastructure control. That is the same move I am asking you to make at your scale.
Three. Thirty-eight percent of a single day's AI product launches sell a brake rather than an engine. On the Product Hunt board the day I wrote this, 11 of 29 AI launches were selling guardrails, approval gates, verification, provenance or audit trails. withHuman's entire product is "A human approval layer for AI Agents," intercepting tool calls before they execute and routing risky ones to a named human. Afterparty drafts every meeting follow-up into the right application and states plainly: "Nothing goes out until you approve it." Atticus can drive your Mac by voice and still "asks before it buys or sends anything." Lleverage runs an autonomous back office where "people only handle the exceptions, with approvals and a full audit trail."
A year ago, the launches were agents that act. Today more than a third of them are products that stop agents from acting. When that many independent companies ship the same category on the same day, the problem is general, not personal.
Four. The sharpest sentence on any of those product pages was about self-verification. Tiden, which sells intent verification for coding agents, points out that tests written inside the session that wrote the code "agree with it by construction." Hold onto that one, because the most common verification scheme I see in small businesses is asking the same model to check its own work, which is structurally the same mistake.
The Fix: Blast Radius Times Reversibility
The framework is two multipliers and one list.
Reversibility. For each action your agent can take, ask how long it takes you to undo it. Three categories is plenty. Reversible in seconds. Reversible with effort. Not reversible. Be strict about the third category and include anything where the undo is technically possible but practically awful, because restoring a backup to fix one bad edit is how people lose two days of real customer data that arrived after the backup.
Blast radius. For each action, ask how many people or records it could affect in one go. The commenter who introduced this framing put it precisely: "a bad email reaches one customer but a bad permission change reaches every record that role touches. I would gate on blast radius times reversibility, not the level label."
Multiply them. Everything high on both axes gets a human. Everything low on both runs unattended. The middle is where you make judgment calls, and the judgment is yours because you know what your customers will forgive.
Here is what changed for me when I started doing it this way. I stopped having a vague uneasy feeling about my automations and started having a two-column list. The uneasy feeling never told me what to fix. The list told me immediately, and the top three items were not the ones I would have guessed.
The second half of the fix is the part nobody mentions, and it comes from an operator who automated his mother's wedding venue business: "Permissioning should be deterministic. Telling the agent 'do not send the email' in the prompt just won't cut it…"
A prompt is a request. A permission is a wall. If the only thing standing between your agent and your customer list is a sentence in a system prompt, you do not have a control, you have a preference. Every single restriction you currently rely on should be tested against this question: does this hold if the model ignores it?
Practical Steps
1. Write R or N beside every action, today. Open your agent's tool list, integration list, and connected accounts. Beside each capability write R for reversible or N for not. Thirty minutes, one pen. This single exercise is the highest-leverage safety work available to you, and most people have never done it because it sounds too simple to matter.
2. Score blast radius on the N items only. For each irreversible action, write down the largest number of people or records it could touch in one run. Rank by that number. The top three are your gates. Build those this week and leave the rest alone.
3. Convert your prompt restrictions into real mechanisms. List every restriction you currently state in a prompt. Beside each one, write the actual enforcement mechanism: a revoked permission, a tool that is not connected, a read-only credential, an approval step. Where there is no mechanism, you have found a gap you did not know you had. Fix the ones that sit next to an N.
4. Bind approvals to the payload, not the intention. This rule came from the same thread and it is worth more than most paid courses: "If the amount or recipient changes after I've approved a refund, ask me again. The system should block it until I approve the change, not just trust the bot to remember." If you approve a forty dollar refund and the agent executes four hundred, your approval was theater. Specify which fields are locked at approval time.
5. Test the retry, because this is where customers get charged twice. "If the refund goes through but the confirmation gets lost, retrying shouldn't refund the customer twice." A hung request is not a failed request. For every money-touching or message-sending action, find out what happens on timeout, and whether the system checks state before it tries again.
6. Never let the model be the source of transactional truth. The cleanest statement of this came from a chatbot operator this week: "Never let the LLM be the source of truth for transactional data. Pull order status and balances from your APIs. Let the model explain the data, not generate it." Write two columns, Retrieve and Explain. Prices, balances, dates and policies go in Retrieve. Everything the model is allowed to compose goes in Explain.
7. Write the failure message in human language. The hardest bug in that wedding venue automation was not permissions at all. When the agent failed, it sent the owner's mother walls of technical logs, and in his words, "they were gibberish to her." Somebody replied with the best operational note of the week: "swapping the stack trace for one plain sentence to your mom is a bigger fix than it sounds. nobody reads a log on their phone, they just panic and call you." Your approval gate is only as good as the human's ability to understand what they are approving at seven in the evening on a phone.
Use this prompt to run step one. It follows the Perfect Prompt Framework 2.0.
[The Job]
Sort every action my AI agent can take by whether it is reversible.
This is for: me, deciding what runs without my approval at [YOUR BUSINESS].
It matters because: I have been classifying by read versus write, which is the wrong axis.
[The Background]
Here is what you need to know: my agent can currently [LIST EVERY TOOL, INTEGRATION AND PERMISSION IT HAS], it runs [ON DEMAND or ON A SCHEDULE], and the systems it touches are [LIST THEM].
Do not use: the read and write labels; classify only by whether I can undo it and how long that takes.
[The Deliverable]
Return: a table of every action marked reversible, reversible with effort, or irreversible, with the undo procedure and its time cost for each.
Optimize for: completeness.
[The Questions]
Ask me any questions you have.
Frequently Asked Questions
What is the difference between read access and write access for an AI agent?
Read access lets the agent retrieve information; write access lets it change something. The distinction is real but it is a poor guide to risk, because a reversible write such as adding a draft is harmless while an irreversible one such as sending a customer email cannot be taken back. Classify by whether you can undo the action and how widely it propagates, not by the mechanism.
Can I just tell my agent in the prompt not to do something?
You can, and it will often work, but it is not a control. Anthropic has published its own findings that models route around stated restrictions because training rewarded completing the task, and its response was to change the infrastructure rather than the instruction. Treat prompt restrictions as helpful guidance layered on top of a real mechanism, never as the mechanism itself.
Is my business liable for what an AI chatbot tells a customer?
A Canadian tribunal decided in 2024 that an airline was responsible for incorrect information its chatbot gave a passenger, rejecting the argument that the chatbot was a separate entity. That decision is not binding everywhere and a tribunal is not a court, but it is the clearest signal available and the operational lesson is simple: assume your bot's statements are your statements and design accordingly.
How do I stop my agent from double-charging a customer?
Make sure every money-touching action checks current state before retrying, rather than treating a lost confirmation as a failed transaction. Where the system supports idempotency keys, use them. Where it does not, add a verification step that confirms whether the first attempt actually succeeded before a second one is allowed to run.
How many approval gates are too many?
The test is whether your approvals are meaningful. If you approve everything without reading it, you have too many gates and they have stopped functioning. Review the list monthly: loosen a gate when the evidence shows the action has been safe repeatedly, tighten one after any near miss, and use your own records as the evidence rather than instinct.
The Close
Go back to that nine-upvote thread, because something happened in it that almost never happens online. After the top comment reframed the question, the person who asked it changed his own question in public: "Maybe the key question isn't 'Can the agent execute this?' but 'What happens if it gets this wrong?'"
That is the whole discipline. Not a better model, not a bigger plan, not a product you buy. A different question, asked before you connect the tool rather than after you lose the email.
Here is the thing. The capability is going to keep arriving whether or not you are ready for it, and it will keep arriving faster than your ability to supervise it. The only part of this you actually control is what you decided, in advance and in writing, your agent is allowed to make permanent.
Write R or N beside every line. It will take you thirty minutes and it will be the most valuable half hour you spend on AI this month.
About the author
Jonathan Mast is the founder of White Beard Strategies, where he helps entrepreneurs deploy AI agents that survive contact with real customers. He created the Perfect Prompt Framework 2.0, speaks on practical AI adoption for small and mid-sized businesses, and serves a community of more than 100,000 entrepreneurs. He has personally shipped automations that did something he did not authorize, which is the only credential that matters for writing an article like this one.