{"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-settings-that-matter","title":"The four settings that change the answer","url":"https://addaly.com/learn/ai-for-small-business/the-settings-that-matter","position":16,"of":88,"minutes":8,"answer":"Model choice, search on or off, reasoning mode and temperature are the four settings that change the answer.","keyIdea":"Model choice, search on or off, reasoning mode and temperature are the four settings that change the answer, and search state matters most because a model with search off answers from a stale memory with no visible sign it has done so."},"body":{"format":"markdown","text":"## Most controls do nothing useful. Four do.\n\nEvery tool buries some options. For business work, four of them materially change what you get, and knowing which is which saves both money and a class of confusing failure.\n\n### 1. Which model\n\nEvery vendor offers a range, roughly: a fast cheap one, a balanced one, and a slow expensive one. Marketing names change constantly; the shape does not.\n\nThe useful thing to know is that the difference is smaller than the pricing implies for the work you are doing. Sorting transaction descriptions, tidying a note, drafting a standard reply: the cheap model does these about as well, and it does them faster. The expensive model earns its price on long documents, multi-step reasoning and unusual languages.\n\nDefault to the cheap one. Escalate when a specific task is visibly failing, then check whether the escalation actually fixed it. On a pay-per-use plan this decision can be a factor of twenty in your bill for the same month's work.\n\n### 2. Web search on or off\n\nThis is the setting that produces the most confusion, because with search off the model answers from memory and gives no sign it has done so.\n\nOn: it fetches pages and can cite them, so it can tell you about last month. Off: it answers from training data with a cutoff, fluently, including about things that have changed.\n\nAsk any question about current prices, rules, availability or a named company, and know which mode you are in before you believe the answer. Some tools decide for themselves whether to search, which is convenient and means you have to check the reply for citations rather than assume.\n\n### 3. Reasoning or thinking mode\n\nSeveral tools now offer a mode that works through a problem step by step before answering. It is slower and, on pay-per-use, several times dearer, because those intermediate steps are tokens too.\n\nWorth it for: comparing three suppliers on five criteria, working out why the stock figures do not reconcile, anything with multiple constraints to hold at once.\n\nNot worth it for: drafting, rewriting, sorting, translating. The great majority of small-business work. Using it everywhere is a common and expensive habit.\n\n### 4. Temperature, where you can see it\n\nUsually exposed only in APIs and developer tools, sometimes as a \"creativity\" slider. It controls how much the model samples away from the most likely next token.\n\nLow, around 0.2, for anything where consistency matters: classification, extraction, standard replies. Higher, around 0.8, for brainstorming names or headlines, where you want range.\n\nIf you have no such control, you are near the middle. That is a reasonable default and not something to worry about.\n\n## Two more worth knowing\n\n**Custom instructions.** Where your facts sheet and standing rules belong. Free, permanent, applies to every conversation, and it survives the context problem from the previous lesson. If you set up one thing this week, set up this.\n\n**Data and training controls.** Whether your conversations may be used to improve the vendor's models. This one is not about output quality at all; it belongs to module seven, and the practical instruction there is to find the toggle, set it, screenshot it with the date.\n\n## The summary\n\nSettings are worth about twenty minutes of your attention, once. Then leave them alone.\n\nThe quality of what you get out is dominated by what you put in: the facts, the examples, the constraints. An hour spent on your facts sheet beats a week spent comparing models, and it is a much better use of an afternoon than reading comparison charts that will be out of date by the time you finish.\n\n## The reply that stops mid-sentence\n\nOne more control worth naming, because its failure looks like a bug. Every response has a maximum length, sometimes shown as a setting, sometimes fixed by the tool. Hit it and the reply stops, often in the middle of a word.\n\nNothing has crashed. Typing \"continue\" usually picks up where it left off. But it is a signal you asked for too much in one go (sixty product descriptions rather than fifteen at a time) and the fix is to batch, which also makes your checking pass manageable.\n\n## Where they hide\n\nModel choice is usually a dropdown near the message box. Search is a toggle or a globe icon, sometimes only appearing once you start typing. Reasoning mode is often a button beside the model name. Custom instructions live in settings under a name like personalisation, profile or projects.\n\nVendors move these constantly, which is mildly infuriating and worth knowing: if a control you relied on has vanished, it has probably been renamed rather than removed. Spend ten minutes after any major interface change confirming your standing instructions are still attached, because that is the one whose silent loss you will not notice for a week."},"check":{"question":"A business is on pay-per-use and its monthly bill has quadrupled with no increase in volume. Staff have started using the reasoning mode for everything. Why does this explain the rise?","options":[{"text":"Reasoning mode selects a larger model, which is priced higher per token than the default.","correct":false,"tempting":"Half right about the price and wrong about the cause; the additional charge comes mainly from the volume of intermediate steps produced, not from a change of model tier."},{"text":"Reasoning mode keeps the full thread in context for longer, so more is re-sent on each turn.","correct":false,"tempting":"Attributes the cost to context handling, which is unchanged; a reasoning reply on turn one of a fresh chat already costs several times a normal one."},{"text":"The intermediate steps are generated tokens and are billed, so a mode that thinks at length costs several times a direct answer even for simple drafting.","correct":true},{"text":"Reasoning mode disables caching of repeated content, so previously discounted input is charged at full rate on every turn, which multiplies the cost of any long facts sheet.","correct":false,"tempting":"Reaches for a plausible billing mechanism that is not what drives this; caching discounts apply to repeated input regardless of whether the reply is a reasoning one."}]},"prev":{"slug":"context-and-forgetting","title":"Why long conversations go strange","url":"https://addaly.com/learn/ai-for-small-business/context-and-forgetting"},"next":{"slug":"a-prompt-library","title":"A prompt library a two-person business can keep","url":"https://addaly.com/learn/ai-for-small-business/a-prompt-library"},"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 four settings that change the answer', in AI for a Small Business. Edited by M. Mun. Addaly. Available at: https://addaly.com/learn/ai-for-small-business/the-settings-that-matter (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 four settings that change the answer","courseTitle":"AI for a Small Business","module":{"index":2,"slug":"asking-properly","title":"Asking properly: prompting for business work"},"answer":null,"keyIdea":"Model choice, search on or off, reasoning mode and temperature are the four settings that change the answer, and search state matters most because a model with search off answers from a stale memory with no visible sign it has done so.","hero":{"kind":"plate","image":null,"alt":null,"caption":null},"blocks":[{"t":"h2","id":"most-controls-do-nothing-useful-four-do","runs":[{"text":"Most controls do nothing useful. Four do."}]},{"t":"p","runs":[{"text":"Every tool buries some options. For business work, four of them materially change what you get, and knowing which is which saves both money and a class of confusing failure."}]},{"t":"h3","id":"1-which-model","runs":[{"text":"1. Which model"}]},{"t":"p","runs":[{"text":"Every vendor offers a range, roughly: a fast cheap one, a balanced one, and a slow expensive one. Marketing names change constantly; the shape does not."}]},{"t":"p","runs":[{"text":"The useful thing to know is that the difference is smaller than the pricing implies for the work you are doing. Sorting transaction descriptions, tidying a note, drafting a standard reply: the cheap model does these about as well, and it does them faster. The expensive model earns its price on long documents, multi-step reasoning and unusual languages."}]},{"t":"p","runs":[{"text":"Default to the cheap one. Escalate when a specific task is visibly failing, then check whether the escalation actually fixed it. On a pay-per-use plan this decision can be a factor of twenty in your bill for the same month’s work."}]},{"t":"h3","id":"2-web-search-on-or-off","runs":[{"text":"2. Web search on or off"}]},{"t":"p","runs":[{"text":"This is the setting that produces the most confusion, because with search off the model answers from memory and gives no sign it has done so."}]},{"t":"p","runs":[{"text":"On: it fetches pages and can cite them, so it can tell you about last month. Off: it answers from training data with a cutoff, fluently, including about things that have changed."}]},{"t":"p","runs":[{"text":"Ask any question about current prices, rules, availability or a named company, and know which mode you are in before you believe the answer. Some tools decide for themselves whether to search, which is convenient and means you have to check the reply for citations rather than assume."}]},{"t":"h3","id":"3-reasoning-or-thinking-mode","runs":[{"text":"3. Reasoning or thinking mode"}]},{"t":"p","runs":[{"text":"Several tools now offer a mode that works through a problem step by step before answering. It is slower and, on pay-per-use, several times dearer, because those intermediate steps are tokens too."}]},{"t":"p","runs":[{"text":"Worth it for: comparing three suppliers on five criteria, working out why the stock figures do not reconcile, anything with multiple constraints to hold at once."}]},{"t":"p","runs":[{"text":"Not worth it for: drafting, rewriting, sorting, translating. The great majority of small-business work. Using it everywhere is a common and expensive habit."}]},{"t":"h3","id":"4-temperature-where-you-can-see-it","runs":[{"text":"4. Temperature, where you can see it"}]},{"t":"p","runs":[{"text":"Usually exposed only in APIs and developer tools, sometimes as a “creativity” slider. It controls how much the model samples away from the most likely next token."}]},{"t":"p","runs":[{"text":"Low, around 0.2, for anything where consistency matters: classification, extraction, standard replies. Higher, around 0.8, for brainstorming names or headlines, where you want range."}]},{"t":"p","runs":[{"text":"If you have no such control, you are near the middle. That is a reasonable default and not something to worry about."}]},{"t":"h2","id":"two-more-worth-knowing","runs":[{"text":"Two more worth knowing"}]},{"t":"p","runs":[{"text":"Custom instructions.","b":true},{"text":" Where your facts sheet and standing rules belong. Free, permanent, applies to every conversation, and it survives the context problem from the previous lesson. If you set up one thing this week, set up this."}]},{"t":"p","runs":[{"text":"Data and training controls.","b":true},{"text":" Whether your conversations may be used to improve the vendor’s models. This one is not about output quality at all; it belongs to module seven, and the practical instruction there is to find the toggle, set it, screenshot it with the date."}]},{"t":"h2","id":"the-summary","runs":[{"text":"The summary"}]},{"t":"p","runs":[{"text":"Settings are worth about twenty minutes of your attention, once. Then leave them alone."}]},{"t":"p","runs":[{"text":"The quality of what you get out is dominated by what you put in: the facts, the examples, the constraints. An hour spent on your facts sheet beats a week spent comparing models, and it is a much better use of an afternoon than reading comparison charts that will be out of date by the time you finish."}]},{"t":"h2","id":"the-reply-that-stops-mid-sentence","runs":[{"text":"The reply that stops mid-sentence"}]},{"t":"p","runs":[{"text":"One more control worth naming, because its failure looks like a bug. Every response has a maximum length, sometimes shown as a setting, sometimes fixed by the tool. Hit it and the reply stops, often in the middle of a word."}]},{"t":"p","runs":[{"text":"Nothing has crashed. Typing “continue” usually picks up where it left off. But it is a signal you asked for too much in one go (sixty product descriptions rather than fifteen at a time) and the fix is to batch, which also makes your checking pass manageable."}]},{"t":"h2","id":"where-they-hide","runs":[{"text":"Where they hide"}]},{"t":"p","runs":[{"text":"Model choice is usually a dropdown near the message box. Search is a toggle or a globe icon, sometimes only appearing once you start typing. Reasoning mode is often a button beside the model name. Custom instructions live in settings under a name like personalisation, profile or projects."}]},{"t":"p","runs":[{"text":"Vendors move these constantly, which is mildly infuriating and worth knowing: if a control you relied on has vanished, it has probably been renamed rather than removed. Spend ten minutes after any major interface change confirming your standing instructions are still attached, because that is the one whose silent loss you will not notice for a week."}]}],"check":{"question":[{"text":"A business is on pay-per-use and its monthly bill has quadrupled with no increase in volume. Staff have started using the reasoning mode for everything. Why does this explain the rise?"}],"options":[{"text":[{"text":"Reasoning mode selects a larger model, which is priced higher per token than the default."}],"correct":false,"tempting":[{"text":"Half right about the price and wrong about the cause; the additional charge comes mainly from the volume of intermediate steps produced, not from a change of model tier."}]},{"text":[{"text":"Reasoning mode keeps the full thread in context for longer, so more is re-sent on each turn."}],"correct":false,"tempting":[{"text":"Attributes the cost to context handling, which is unchanged; a reasoning reply on turn one of a fresh chat already costs several times a normal one."}]},{"text":[{"text":"The intermediate steps are generated tokens and are billed, so a mode that thinks at length costs several times a direct answer even for simple drafting."}],"correct":true,"tempting":null},{"text":[{"text":"Reasoning mode disables caching of repeated content, so previously discounted input is charged at full rate on every turn, which multiplies the cost of any long facts sheet."}],"correct":false,"tempting":[{"text":"Reaches for a plausible billing mechanism that is not what drives this; caching discounts apply to repeated input regardless of whether the reply is a reasoning one."}]}],"order":[0,3,1,2],"why":null},"prev":{"slug":"context-and-forgetting","title":"Why long conversations go strange","minutes":8},"next":{"slug":"a-prompt-library","title":"A prompt library a two-person business can keep","minutes":8}}}