{"course":{"slug":"ai-for-small-business","title":"AI for a Small Business","url":"https://addaly.com/learn/ai-for-small-business","level":"beginner","track":"work"},"lesson":{"slug":"the-permission-ladder","title":"Five rungs of letting go, and what licenses each one","url":"https://addaly.com/learn/ai-for-small-business/the-permission-ladder","position":58,"of":88,"minutes":9,"answer":"The rung you are on decides how long an error lives before anybody sees it.","keyIdea":"The rung you are on decides how long an error lives before anybody sees it, so climb only where the action is cheaply reversible and you have evidence about errors of the kind sampling can actually detect."},"body":{"format":"markdown","text":"## Autonomy is not one decision\n\nPeople talk about automation as though it were a switch: either a person does the job or the computer does. In practice there are five distinct positions, and most of the bad outcomes in small-business AI come from skipping two of them.\n\n**Rung one: suggest.** You are writing; the model offers phrasing, an outline, three subject lines. You type the thing that goes out. Errors cost seconds.\n\n**Rung two: draft.** The model produces the whole item. You read all of it, every time, and you press send. This is where the great majority of useful small-business AI lives, and a lot of businesses would be better off never leaving it.\n\n**Rung three: queue.** The model produces the item and prepares the action: the reply sits in a drafts folder, the invoice is created but unsent, the stock order is filled in but not placed. You approve. The difference from rung two is that the work of *doing* is already finished, so your job shrinks to judging.\n\n**Rung four: act, then tell you.** It happens. You get a notification and a window in which you can undo it. Nothing is lost if you are quick.\n\n**Rung five: act silently.** It happens. You find out from the monthly log, or from a customer.\n\n## What actually changes as you climb\n\nNot the error rate. The model is the same model on rung two and rung five; it makes mistakes at the same rate either way.\n\nWhat changes is **time to detection**. On rung two an error is caught in the ten seconds before you send. On rung three it is caught at the next approval sweep, so within a day. On rung four it is caught within your undo window. On rung five an error can run for weeks, which means it is no longer one error: it is one error multiplied by the number of times the job ran.\n\nThat multiplication is the whole risk. A 2% error rate is unremarkable at rung two and is roughly six wrong messages a week at rung five in a business handling three hundred enquiries.\n\n> Figure: Five rungs, read as time to detection. Layers, top to bottom: One, suggest (The model offers phrasing and you type what goes out. An error costs seconds.); Two, draft (It writes the whole item, you read every one, you press send. Caught in the ten seconds before sending, and where the great majority of useful small-business AI lives.); Three, queue (The action is prepared but not taken: the reply sits in drafts, the invoice is created and unsent. Caught at the next approval sweep, so within a day.); Four, act then tell you (It happens and a notification arrives with an undo window. Caught only if a human being genuinely reads that notification.); Five, act silently (You find out from the monthly log, or from a customer. One error becomes one error multiplied by the number of times the job ran.). The error rate is identical on every rung; only the time to detection changes. Two per cent is unremarkable at rung two and is about six wrong messages a week at rung five in a business handling three hundred enquiries.\n\n## The reversibility test, done honestly\n\nBefore any rung above three, ask what undoing this actually involves.\n\nCheaply reversible: creating a record, tagging something, filing a document, scheduling a task, drafting, adding a row, raising an unsent invoice.\n\nNot reversible, whatever the software claims: sending an email or a message, taking a payment, placing an order with a supplier, publishing anything, deleting anything, making a promise to a customer, submitting a form to a government body.\n\n\"Unsend\" features are a fourteen-second courtesy, not a control. Once a message has arrived in somebody's inbox on their phone, the fact that you later retracted it changes nothing about what they read.\n\n## What licenses a climb\n\nNot confidence. Evidence, and the evidence is boringly specific: a run of consecutive items where you corrected nothing.\n\nFifty consecutive clean items is a reasonable bar for moving from rung two to rung three on ordinary work. It is worth being clear-eyed about what that buys you: module five's arithmetic applies unchanged here. Fifty clean items is consistent with a true error rate of about 6%. It is not proof of safety; it is the point at which watching every single one has stopped being the best use of your attention.\n\nMoving to rung four needs two further things: an undo path you have actually tested, and a notification that a human being reads. A notification going to an address nobody opens turns rung four into rung five with extra steps.\n\n## The failure the ladder does not protect you from\n\nSampling estimates the rate of *independent* errors: the ones that happen one at a time for uncorrelated reasons.\n\nIt tells you nothing about correlated failure. If your facts sheet still says the old price, every single output is wrong, in the same way, at once. Fifty clean items last month is no evidence at all about the morning after the price changed, because the thing that changed was not the model.\n\nThis is why the climb is not permanent. Anything that changes the inputs (a price rise, a new product line, a model upgrade, a change in your terms) drops you back to rung two for a fortnight. Not as a punishment. As the only way to find out whether the thing still works.\n\n## Where to start\n\nPick the job you already trust most at rung two, and move it to rung three. Not to rung four. Rung three costs you almost nothing, because approving a prepared action takes about four seconds, and it gives back nearly all the time that reviewing was costing.\n\nMost small businesses that get value from AI over a year are sitting at rung three on two or three tasks, and nowhere higher. That is not timidity. It is where the arithmetic lands."},"check":{"question":"A florist has run an AI-drafted \"your order is ready\" message at rung two for six weeks with no corrections, and moves it to rung four with an hourly notification. The same week, she adds a new delivery-surcharge policy to her facts sheet. What is the most likely consequence?","options":[{"text":"The model's error rate will rise because the facts sheet is now longer, and longer instructions reliably degrade the quality of generated messages.","correct":false,"tempting":"Treats prompt length as the mechanism, when nothing about six clean weeks was ever evidence about a changed input."},{"text":"The clean run says nothing about messages involving the new policy, and any systematic mistake with it will go out repeatedly before the hourly notice is read.","correct":true},{"text":"Rung four is unsuitable for customer messages in general, because sending a message is irreversible and no amount of evidence can justify automating an irreversible action.","correct":false,"tempting":"Correctly spots that sending is irreversible, but turns a reason for caution into an absolute rule the ladder does not claim."},{"text":"The notification will catch the problem within the hour, so the exposure is bounded at one hour of messages, which for a florist is a small and acceptable number of affected customers.","correct":false,"tempting":"Assumes the notification is read and understood on arrival, which is the assumption rung four exists to make people examine."}]},"prev":{"slug":"working-with-your-accountant","title":"Arriving prepared, and what never to outsource","url":"https://addaly.com/learn/ai-for-small-business/working-with-your-accountant"},"next":{"slug":"fields-not-prose","title":"Getting fields out instead of paragraphs","url":"https://addaly.com/learn/ai-for-small-business/fields-not-prose"},"authorship":"Founder and series editor: Dr Mustafa Mun. Written by AI models to a written standard. Exam keys checked blind by a second model: 3,665 of 3,666 agreed. Academic review invited.","review":{"status":"Draft, open for academic review"},"cite":"Addaly (2026) 'Five rungs of letting go, and what licenses each one', in AI for a Small Business. Edited by M. Mun. Addaly. Available at: https://addaly.com/learn/ai-for-small-business/the-permission-ladder (Accessed: 13 October 2026).","licence":{"name":"CC BY-NC-SA 4.0","url":"https://creativecommons.org/licenses/by-nc-sa/4.0/"},"reader":{"schema":1,"title":"Five rungs of letting go, and what licenses each one","courseTitle":"AI for a Small Business","module":{"index":7,"slug":"letting-it-act","title":"Letting it act: automation you can trust with the keys"},"answer":null,"keyIdea":"The rung you are on decides how long an error lives before anybody sees it, so climb only where the action is cheaply reversible and you have evidence about errors of the kind sampling can actually detect.","hero":{"kind":"figure","figure":{"t":"figure","label":"Figure 7.1","figure":{"type":"stack","title":"Five rungs, read as time to detection","layers":[{"label":"One, suggest","detail":"The model offers phrasing and you type what goes out. An error costs seconds."},{"label":"Two, draft","detail":"It writes the whole item, you read every one, you press send. Caught in the ten seconds before sending, and where the great majority of useful small-business AI lives."},{"label":"Three, queue","detail":"The action is prepared but not taken: the reply sits in drafts, the invoice is created and unsent. Caught at the next approval sweep, so within a day."},{"label":"Four, act then tell you","detail":"It happens and a notification arrives with an undo window. Caught only if a human being genuinely reads that notification."},{"label":"Five, act silently","detail":"You find out from the monthly log, or from a customer. One error becomes one error multiplied by the number of times the job ran."}],"caption":"The error rate is identical on every rung; only the time to detection changes. Two per cent is unremarkable at rung two and is about six wrong messages a week at rung five in a business handling three hundred enquiries."},"alt":"Figure: Five rungs, read as time to detection. Layers, top to bottom: One, suggest (The model offers phrasing and you type what goes out. An error costs seconds.); Two, draft (It writes the whole item, you read every one, you press send. Caught in the ten seconds before sending, and where the great majority of useful small-business AI lives.); Three, queue (The action is prepared but not taken: the reply sits in drafts, the invoice is created and unsent. Caught at the next approval sweep, so within a day.); Four, act then tell you (It happens and a notification arrives with an undo window. Caught only if a human being genuinely reads that notification.); Five, act silently (You find out from the monthly log, or from a customer. One error becomes one error multiplied by the number of times the job ran.). The error rate is identical on every rung; only the time to detection changes. Two per cent is unremarkable at rung two and is about six wrong messages a week at rung five in a business handling three hundred enquiries."}},"blocks":[{"t":"h2","id":"autonomy-is-not-one-decision","runs":[{"text":"Autonomy is not one decision"}]},{"t":"p","runs":[{"text":"People talk about automation as though it were a switch: either a person does the job or the computer does. In practice there are five distinct positions, and most of the bad outcomes in small-business AI come from skipping two of them."}]},{"t":"p","runs":[{"text":"Rung one: suggest.","b":true},{"text":" You are writing; the model offers phrasing, an outline, three subject lines. You type the thing that goes out. Errors cost seconds."}]},{"t":"p","runs":[{"text":"Rung two: draft.","b":true},{"text":" The model produces the whole item. You read all of it, every time, and you press send. This is where the great majority of useful small-business AI lives, and a lot of businesses would be better off never leaving it."}]},{"t":"p","runs":[{"text":"Rung three: queue.","b":true},{"text":" The model produces the item and prepares the action: the reply sits in a drafts folder, the invoice is created but unsent, the stock order is filled in but not placed. You approve. The difference from rung two is that the work of "},{"text":"doing","i":true},{"text":" is already finished, so your job shrinks to judging."}]},{"t":"p","runs":[{"text":"Rung four: act, then tell you.","b":true},{"text":" It happens. You get a notification and a window in which you can undo it. Nothing is lost if you are quick."}]},{"t":"p","runs":[{"text":"Rung five: act silently.","b":true},{"text":" It happens. You find out from the monthly log, or from a customer."}]},{"t":"h2","id":"what-actually-changes-as-you-climb","runs":[{"text":"What actually changes as you climb"}]},{"t":"p","runs":[{"text":"Not the error rate. The model is the same model on rung two and rung five; it makes mistakes at the same rate either way."}]},{"t":"p","runs":[{"text":"What changes is "},{"text":"time to detection","b":true},{"text":". On rung two an error is caught in the ten seconds before you send. On rung three it is caught at the next approval sweep, so within a day. On rung four it is caught within your undo window. On rung five an error can run for weeks, which means it is no longer one error: it is one error multiplied by the number of times the job ran."}]},{"t":"p","runs":[{"text":"That multiplication is the whole risk. A 2% error rate is unremarkable at rung two and is roughly six wrong messages a week at rung five in a business handling three hundred enquiries."}]},{"t":"h2","id":"the-reversibility-test-done-honestly","runs":[{"text":"The reversibility test, done honestly"}]},{"t":"p","runs":[{"text":"Before any rung above three, ask what undoing this actually involves."}]},{"t":"p","runs":[{"text":"Cheaply reversible: creating a record, tagging something, filing a document, scheduling a task, drafting, adding a row, raising an unsent invoice."}]},{"t":"p","runs":[{"text":"Not reversible, whatever the software claims: sending an email or a message, taking a payment, placing an order with a supplier, publishing anything, deleting anything, making a promise to a customer, submitting a form to a government body."}]},{"t":"p","runs":[{"text":"“Unsend” features are a fourteen-second courtesy, not a control. Once a message has arrived in somebody’s inbox on their phone, the fact that you later retracted it changes nothing about what they read."}]},{"t":"h2","id":"what-licenses-a-climb","runs":[{"text":"What licenses a climb"}]},{"t":"p","runs":[{"text":"Not confidence. Evidence, and the evidence is boringly specific: a run of consecutive items where you corrected nothing."}]},{"t":"p","runs":[{"text":"Fifty consecutive clean items is a reasonable bar for moving from rung two to rung three on ordinary work. It is worth being clear-eyed about what that buys you: module five’s arithmetic applies unchanged here. Fifty clean items is consistent with a true error rate of about 6%. It is not proof of safety; it is the point at which watching every single one has stopped being the best use of your attention."}]},{"t":"p","runs":[{"text":"Moving to rung four needs two further things: an undo path you have actually tested, and a notification that a human being reads. A notification going to an address nobody opens turns rung four into rung five with extra steps."}]},{"t":"h2","id":"the-failure-the-ladder-does-not-protect-you-from","runs":[{"text":"The failure the ladder does not protect you from"}]},{"t":"p","runs":[{"text":"Sampling estimates the rate of "},{"text":"independent","i":true},{"text":" errors: the ones that happen one at a time for uncorrelated reasons."}]},{"t":"p","runs":[{"text":"It tells you nothing about correlated failure. If your facts sheet still says the old price, every single output is wrong, in the same way, at once. Fifty clean items last month is no evidence at all about the morning after the price changed, because the thing that changed was not the model."}]},{"t":"p","runs":[{"text":"This is why the climb is not permanent. Anything that changes the inputs (a price rise, a new product line, a model upgrade, a change in your terms) drops you back to rung two for a fortnight. Not as a punishment. As the only way to find out whether the thing still works."}]},{"t":"h2","id":"where-to-start","runs":[{"text":"Where to start"}]},{"t":"p","runs":[{"text":"Pick the job you already trust most at rung two, and move it to rung three. Not to rung four. Rung three costs you almost nothing, because approving a prepared action takes about four seconds, and it gives back nearly all the time that reviewing was costing."}]},{"t":"p","runs":[{"text":"Most small businesses that get value from AI over a year are sitting at rung three on two or three tasks, and nowhere higher. That is not timidity. It is where the arithmetic lands."}]}],"check":{"question":[{"text":"A florist has run an AI-drafted “your order is ready” message at rung two for six weeks with no corrections, and moves it to rung four with an hourly notification. The same week, she adds a new delivery-surcharge policy to her facts sheet. What is the most likely consequence?"}],"options":[{"text":[{"text":"The model’s error rate will rise because the facts sheet is now longer, and longer instructions reliably degrade the quality of generated messages."}],"correct":false,"tempting":[{"text":"Treats prompt length as the mechanism, when nothing about six clean weeks was ever evidence about a changed input."}]},{"text":[{"text":"The clean run says nothing about messages involving the new policy, and any systematic mistake with it will go out repeatedly before the hourly notice is read."}],"correct":true,"tempting":null},{"text":[{"text":"Rung four is unsuitable for customer messages in general, because sending a message is irreversible and no amount of evidence can justify automating an irreversible action."}],"correct":false,"tempting":[{"text":"Correctly spots that sending is irreversible, but turns a reason for caution into an absolute rule the ladder does not claim."}]},{"text":[{"text":"The notification will catch the problem within the hour, so the exposure is bounded at one hour of messages, which for a florist is a small and acceptable number of affected customers."}],"correct":false,"tempting":[{"text":"Assumes the notification is read and understood on arrival, which is the assumption rung four exists to make people examine."}]}],"order":[2,0,1,3],"why":null},"prev":{"slug":"working-with-your-accountant","title":"Arriving prepared, and what never to outsource","minutes":8},"next":{"slug":"fields-not-prose","title":"Getting fields out instead of paragraphs","minutes":9}}}