{"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-checking-tax","title":"The checking tax nobody counts","url":"https://addaly.com/learn/ai-for-small-business/the-checking-tax","position":3,"of":88,"minutes":9,"answer":"The saving from a drafting tool is the drafting time minus the checking time, and a task whose errors are silent rather than loud can cost more to verify than it ever cost to do.","keyIdea":"The saving from a drafting tool is the drafting time minus the checking time, and a task whose errors are silent rather than loud can cost more to verify than it ever cost to do."},"body":{"format":"markdown","text":"## The saving is not the drafting time\n\nA drafted reply appears in four seconds. It used to take you four minutes. The obvious conclusion is that you saved four minutes.\n\nYou did not. You saved four minutes minus the time you now spend reading, correcting and deciding whether to trust it. That second number is the checking tax, and almost nobody measures it, which is why so many businesses feel busier after adopting a tool that is working.\n\nMeasure it once, properly. Take ten real items. Time yourself doing three the old way. Then draft the other seven with help and time the whole loop: writing the request, waiting, reading, editing, deciding. Divide.\n\nOwners are routinely surprised. A typical honest result on customer replies: four minutes down to a minute and a half, so a real saving of about 60%. A typical honest result on a technical quote: eleven minutes down to nine, because the checking is nearly as hard as the writing. The first task is worth automating. The second is not, and no amount of better prompting changes that, because the cost is in the verification rather than the generation.\n\n> Figure: The whole loop, timed on ten real items. Bars (minutes per item, writing plus checking): Customer reply, done by hand 4; Customer reply, drafted then checked 1.5; Technical quote, done by hand 11; Technical quote, drafted then checked 9. The first task saves about sixty per cent and is worth automating. The second saves two minutes, because the checking is nearly as hard as the writing, and the cost sits in verification rather than in generation.\n\n## Loud errors and silent errors\n\nNot all checking is the same, and this is the distinction that decides whether a task is safe.\n\nA **loud error** announces itself. Clumsy phrasing, a paragraph in the wrong order, a greeting that does not sound like you. You catch it by reading normally, in seconds, and the worst case if you miss one is mild embarrassment.\n\nA **silent error** looks exactly like a correct answer. A delivery date that is wrong by a week. A tax rate that was right last year. A dimension transposed from 460mm to 640mm. A clause dropped from a contract summary. You cannot catch these by reading; you catch them only by checking each item against a source, deliberately, one at a time.\n\nSo the honest question about any task is not \"can it do this?\" but **\"how expensive is my checking, and what does a missed error cost?\"** A task with cheap loud errors is a gift. A task with expensive silent errors is a trap that pays you for weeks and then charges you all at once.\n\n## Read in passes, not in one go\n\nMost people read a draft once, for general quality, and their eye slides over the numbers because the sentences around them are smooth. Two separate passes catch far more in less total time.\n\n1. **The numbers-and-promises pass.** Ignore the prose entirely. Look only at figures, dates, percentages, quantities, guarantees, and anything phrased as a commitment. Every one must trace back to something you supplied. Anything that appeared from nowhere gets deleted rather than corrected. On a short reply this takes about twenty seconds.\n2. **The reading pass.** Now read it as a customer would, for tone and sense.\n\nThe order matters. If you read for tone first you have already accepted the text as a whole, and the numbers become part of the furniture.\n\n## Checking fatigue is real, and it is predictable\n\nHere is the failure that catches conscientious people, and it is worth naming because it feels like nothing is happening.\n\nThe first twenty drafts are good. You correct two small things. The next twenty are good. By the fiftieth you are skimming, because fifty consecutive fine outputs have taught you that the output is fine. The error rate has not changed at all; your attention has. Then one bad one goes out.\n\nThree defences, all cheap:\n\n- **Keep batches small.** Twenty items, then stop. Attention on a checking task falls off sharply after about twenty minutes.\n- **Make the check mechanical.** A three-item checklist you tick beats \"read it carefully\", because a checklist survives boredom and vigilance does not.\n- **Spot-audit on a schedule.** Once a week, pull five items that already went out and check them properly against source. This measures your actual error rate rather than your confidence, and it is the only way you will ever notice drift.\n\n## The rule of thumb\n\nIf checking a piece of output takes more than about half the time the task originally took, the task is probably the wrong one. Move it back to the list and pick another.\n\nThat is not a counsel of despair. It is the reason the two tasks you chose need to be the ones where you can glance at the result and know. Frequency multiplied by a real saving is where the money is, and a real saving is the one that survives the checking."},"check":{"question":"A workshop uses AI to draft cutting lists from customer measurements. The prose is always clean, the checking feels quick, and it has run well for six weeks. What is the strongest reason to distrust the current arrangement?","options":[{"text":"Six weeks is too short a period to draw any conclusion about a tool's reliability.","correct":false,"tempting":"Sounds cautious but names no mechanism; a well-chosen task can be judged in two weeks, and duration alone is neither the risk nor the reassurance here."},{"text":"Cutting lists are numeric, so a transposed dimension reads exactly like a correct one, and six weeks of clean output is precisely what erodes the attention needed to catch it.","correct":true},{"text":"The model may have been updated by the vendor, so the outputs could have changed without warning.","correct":false,"tempting":"A genuine risk covered later in the course, but it does not explain why this particular task is unsafe; the problem is silent numeric error, present from day one."},{"text":"The prose being clean shows the model is optimising for readability over accuracy, so the instruction should ask for a terse factual style that leaves any error more exposed to the reader.","correct":false,"tempting":"Misreads fluency as a tunable trade-off against accuracy, when clean prose and a wrong dimension are produced by the same process and neither causes the other."}]},"prev":{"slug":"what-the-machine-is-doing","title":"Fluent and wrong come from the same place","url":"https://addaly.com/learn/ai-for-small-business/what-the-machine-is-doing"},"next":{"slug":"where-ai-arrives","title":"Four doors AI comes through, and what each one costs","url":"https://addaly.com/learn/ai-for-small-business/where-ai-arrives"},"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) 'The checking tax nobody counts', in AI for a Small Business. Edited by M. Mun. Addaly. Available at: https://addaly.com/learn/ai-for-small-business/the-checking-tax (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":"The checking tax nobody counts","courseTitle":"AI for a Small Business","module":{"index":1,"slug":"where-it-belongs","title":"Working out where AI belongs in your business"},"answer":null,"keyIdea":"The saving from a drafting tool is the drafting time minus the checking time, and a task whose errors are silent rather than loud can cost more to verify than it ever cost to do.","hero":{"kind":"figure","figure":{"t":"figure","label":"Figure","figure":{"type":"bars","unit":"minutes per item, writing plus checking","title":"The whole loop, timed on ten real items","series":[{"label":"Customer reply, done by hand","value":4},{"label":"Customer reply, drafted then checked","value":1.5,"highlight":true},{"label":"Technical quote, done by hand","value":11},{"label":"Technical quote, drafted then checked","value":9}],"caption":"The first task saves about sixty per cent and is worth automating. The second saves two minutes, because the checking is nearly as hard as the writing, and the cost sits in verification rather than in generation."},"alt":"Figure: The whole loop, timed on ten real items. Bars (minutes per item, writing plus checking): Customer reply, done by hand 4; Customer reply, drafted then checked 1.5; Technical quote, done by hand 11; Technical quote, drafted then checked 9. The first task saves about sixty per cent and is worth automating. The second saves two minutes, because the checking is nearly as hard as the writing, and the cost sits in verification rather than in generation."}},"blocks":[{"t":"h2","id":"the-saving-is-not-the-drafting-time","runs":[{"text":"The saving is not the drafting time"}]},{"t":"p","runs":[{"text":"A drafted reply appears in four seconds. It used to take you four minutes. The obvious conclusion is that you saved four minutes."}]},{"t":"p","runs":[{"text":"You did not. You saved four minutes minus the time you now spend reading, correcting and deciding whether to trust it. That second number is the checking tax, and almost nobody measures it, which is why so many businesses feel busier after adopting a tool that is working."}]},{"t":"p","runs":[{"text":"Measure it once, properly. Take ten real items. Time yourself doing three the old way. Then draft the other seven with help and time the whole loop: writing the request, waiting, reading, editing, deciding. Divide."}]},{"t":"p","runs":[{"text":"Owners are routinely surprised. A typical honest result on customer replies: four minutes down to a minute and a half, so a real saving of about 60%. A typical honest result on a technical quote: eleven minutes down to nine, because the checking is nearly as hard as the writing. The first task is worth automating. The second is not, and no amount of better prompting changes that, because the cost is in the verification rather than the generation."}]},{"t":"h2","id":"loud-errors-and-silent-errors","runs":[{"text":"Loud errors and silent errors"}]},{"t":"p","runs":[{"text":"Not all checking is the same, and this is the distinction that decides whether a task is safe."}]},{"t":"p","runs":[{"text":"A "},{"text":"loud error","b":true},{"text":" announces itself. Clumsy phrasing, a paragraph in the wrong order, a greeting that does not sound like you. You catch it by reading normally, in seconds, and the worst case if you miss one is mild embarrassment."}]},{"t":"p","runs":[{"text":"A "},{"text":"silent error","b":true},{"text":" looks exactly like a correct answer. A delivery date that is wrong by a week. A tax rate that was right last year. A dimension transposed from 460mm to 640mm. A clause dropped from a contract summary. You cannot catch these by reading; you catch them only by checking each item against a source, deliberately, one at a time."}]},{"t":"p","runs":[{"text":"So the honest question about any task is not “can it do this?” but "},{"text":"”how expensive is my checking, and what does a missed error cost?”","b":true},{"text":" A task with cheap loud errors is a gift. A task with expensive silent errors is a trap that pays you for weeks and then charges you all at once."}]},{"t":"h2","id":"read-in-passes-not-in-one-go","runs":[{"text":"Read in passes, not in one go"}]},{"t":"p","runs":[{"text":"Most people read a draft once, for general quality, and their eye slides over the numbers because the sentences around them are smooth. Two separate passes catch far more in less total time."}]},{"t":"ol","start":1,"items":[[{"text":"The numbers-and-promises pass.","b":true},{"text":" Ignore the prose entirely. Look only at figures, dates, percentages, quantities, guarantees, and anything phrased as a commitment. Every one must trace back to something you supplied. Anything that appeared from nowhere gets deleted rather than corrected. On a short reply this takes about twenty seconds."}],[{"text":"The reading pass.","b":true},{"text":" Now read it as a customer would, for tone and sense."}]]},{"t":"p","runs":[{"text":"The order matters. If you read for tone first you have already accepted the text as a whole, and the numbers become part of the furniture."}]},{"t":"h2","id":"checking-fatigue-is-real-and-it-is-predictable","runs":[{"text":"Checking fatigue is real, and it is predictable"}]},{"t":"p","runs":[{"text":"Here is the failure that catches conscientious people, and it is worth naming because it feels like nothing is happening."}]},{"t":"p","runs":[{"text":"The first twenty drafts are good. You correct two small things. The next twenty are good. By the fiftieth you are skimming, because fifty consecutive fine outputs have taught you that the output is fine. The error rate has not changed at all; your attention has. Then one bad one goes out."}]},{"t":"p","runs":[{"text":"Three defences, all cheap:"}]},{"t":"ul","items":[[{"text":"Keep batches small.","b":true},{"text":" Twenty items, then stop. Attention on a checking task falls off sharply after about twenty minutes."}],[{"text":"Make the check mechanical.","b":true},{"text":" A three-item checklist you tick beats “read it carefully”, because a checklist survives boredom and vigilance does not."}],[{"text":"Spot-audit on a schedule.","b":true},{"text":" Once a week, pull five items that already went out and check them properly against source. This measures your actual error rate rather than your confidence, and it is the only way you will ever notice drift."}]]},{"t":"h2","id":"the-rule-of-thumb","runs":[{"text":"The rule of thumb"}]},{"t":"p","runs":[{"text":"If checking a piece of output takes more than about half the time the task originally took, the task is probably the wrong one. Move it back to the list and pick another."}]},{"t":"p","runs":[{"text":"That is not a counsel of despair. It is the reason the two tasks you chose need to be the ones where you can glance at the result and know. Frequency multiplied by a real saving is where the money is, and a real saving is the one that survives the checking."}]}],"check":{"question":[{"text":"A workshop uses AI to draft cutting lists from customer measurements. The prose is always clean, the checking feels quick, and it has run well for six weeks. What is the strongest reason to distrust the current arrangement?"}],"options":[{"text":[{"text":"Six weeks is too short a period to draw any conclusion about a tool’s reliability."}],"correct":false,"tempting":[{"text":"Sounds cautious but names no mechanism; a well-chosen task can be judged in two weeks, and duration alone is neither the risk nor the reassurance here."}]},{"text":[{"text":"Cutting lists are numeric, so a transposed dimension reads exactly like a correct one, and six weeks of clean output is precisely what erodes the attention needed to catch it."}],"correct":true,"tempting":null},{"text":[{"text":"The model may have been updated by the vendor, so the outputs could have changed without warning."}],"correct":false,"tempting":[{"text":"A genuine risk covered later in the course, but it does not explain why this particular task is unsafe; the problem is silent numeric error, present from day one."}]},{"text":[{"text":"The prose being clean shows the model is optimising for readability over accuracy, so the instruction should ask for a terse factual style that leaves any error more exposed to the reader."}],"correct":false,"tempting":[{"text":"Misreads fluency as a tunable trade-off against accuracy, when clean prose and a wrong dimension are produced by the same process and neither causes the other."}]}],"order":[0,2,3,1],"why":null},"prev":{"slug":"what-the-machine-is-doing","title":"Fluent and wrong come from the same place","minutes":10},"next":{"slug":"where-ai-arrives","title":"Four doors AI comes through, and what each one costs","minutes":9}}}