{"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":"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","position":4,"of":88,"minutes":9,"answer":"Chat window, built-in feature, API and local model are four routes to similar capability that differ in setup cost, price per use and where your data ends up.","keyIdea":"Chat window, built-in feature, API and local model are four routes to similar capability that differ in setup cost, price per use and where your data ends up, and the built-in feature is usually cheapest because the data is already there."},"body":{"format":"markdown","text":"## The same model, four different decisions\n\nA great deal of confused advice comes from people comparing a chat window with an API and concluding one is better. They are not competing products. They are four delivery routes to broadly the same capability, and they differ in price, in where your data goes, and in how much work you have to do.\n\n### 1. The chat window\n\nChatGPT, Claude, Gemini, Copilot, DeepSeek, Le Chat, Qwen. A browser tab or a phone app. Free tiers are real and many small businesses never leave them; paid consumer tiers cluster around USD 20 per person per month, with lower regional pricing in some markets.\n\nWhat you get: everything, immediately, with no setup. What you give up: it is entirely manual. Somebody has to open the tab, paste the context, and copy the answer back. That is fine for six proposals a month and hopeless for four hundred product descriptions.\n\nThis is where every small business should start, and where most should stay.\n\n### 2. A feature inside software you already pay for\n\nYour accounting package, your booking system, your helpdesk, your online shop, your office suite. Zoho, QuickBooks, Xero, Tally, Shopify, HubSpot, Google Workspace, Microsoft 365: nearly all of them now have AI features, often included, sometimes as a per-seat add-on.\n\nThis route is systematically underrated, for a reason that has nothing to do with the model: **the data is already there.** No pasting, no stripping, no second vendor, no second contract, no second copy of your customer list in a system you have not read the terms of. It also tends to be the cheapest, because the vendor is bundling.\n\nBefore you buy anything new, spend an hour looking through what you already pay for. A surprising number of owners are paying twice.\n\n### 3. The API\n\nYou pay per token: input and output are usually priced separately. Order of magnitude for a page in and half a page out: about one US cent on a mid-priced model, closer to a tenth of a cent on a cheap one. Five hundred customer replies a month typically lands between USD 0.50 and USD 6.\n\nThat is startlingly cheap, and it is why people get excited. The catch is structural: **an API key is not an application.** Something has to call it: a script, a spreadsheet add-on, an automation tool, a developer. The model is the cheap part; the plumbing is the expensive part. Budget the plumbing, or use a no-code automation tool that provides it.\n\nThe API route is right when the volume is high, the task is identical every time, and a person is not in the loop for each item.\n\n### 4. On your own machine\n\nOpen-weight models (Llama, Mistral, Qwen, Gemma, Phi and others) run locally through Ollama or LM Studio, both free. A modern laptop with 16GB of memory runs a small model usefully; 8GB is tight but workable with a smaller one.\n\nMarginal cost is zero and nothing leaves the building, which dissolves a large part of the privacy problem in one move. Quality is a clear step below the best hosted models, and the gap is widest on reasoning and on languages other than English. It is good at repetitive sorting, classification, tidying and drafting, which is much of what a small business actually needs.\n\nSet-up is an afternoon. If your business handles health records, legal files or anything you would not want on somebody else's server, this is the route worth an afternoon.\n\n> Figure: Four doors, ordered by where your words end up. Layers, top to bottom: The chat window (Nothing to set up and everything available, but entirely manual. Free tiers are real; paid consumer tiers cluster near USD 20 a person a month. Your words go to a vendor's servers under terms you should read once.); A feature in software you already pay for (Your accounts package, booking system, helpdesk or shop. Usually the cheapest, because the vendor already holds the same data under a contract you already signed, and there is no second copy.); The API (About a cent for a page in and half a page out, a tenth of that on a cheap model. The model is cheap and the plumbing that calls it is not. Right when volume, rather than judgement, is the bottleneck.); A model on your own machine (Llama, Mistral, Qwen or Gemma through Ollama or LM Studio, free. A clear step below the best hosted models, an afternoon to set up, and nothing leaves the building.). The order of price is not the order of exposure. The second door is often both the cheapest and the safest, and it is the one nobody checks before signing up with a new vendor.\n\n## Choosing, in one paragraph\n\nStart in the chat window until you know the task. Look hard at what your existing software already offers before adding a vendor. Move to the API only when the volume is high enough that a human copying and pasting is the bottleneck. Go local when the data is the constraint rather than the cost.\n\nAnd notice the thing that stays constant across all four: the facts sheet, the checking pass, the two tasks. The door changes what you pay and who holds your data. It does not change the discipline.\n\n## One thing that is true of all four\n\nWhichever door you use, you are sending your business's words to something. On routes one and three that is a company's servers, under terms you should read once. On route two it is a company that already holds the same data. On route four it is your own machine and nothing leaves.\n\nThat ordering is worth keeping in your head, because it does not match the price ordering. The cheapest option per item is not the safest, and the safest is not the most convenient. Module seven works through what to check for each; for now, notice that the door you pick is a data decision as much as a cost one."},"check":{"question":"A furniture business needs 800 product descriptions generated from a spreadsheet of specifications, once, before a catalogue deadline. Which consideration should drive the route they pick?","options":[{"text":"Local models are free to run, so the only real cost is the afternoon of setup.","correct":false,"tempting":"Optimises for the wrong constraint; running cost is trivial at this volume either way, and a local model adds setup time and a quality step down for no benefit here."},{"text":"The API is cheapest per item, so it is the correct choice whenever volume exceeds a few hundred, and at eight hundred descriptions the per-item saving comfortably justifies whatever setup is needed.","correct":false,"tempting":"Prices the model and ignores the plumbing; per-item cost is pennies, but somebody still has to build and test the thing that calls the key before the deadline."},{"text":"Whether their existing shop or catalogue software already has a bulk description feature, since the specifications are already in it and no second vendor is involved.","correct":true},{"text":"Whichever model scores highest on published benchmarks for writing quality, since the output is customer-facing.","correct":false,"tempting":"Treats a benchmark as evidence about this task; general writing scores say nothing about following a specification sheet in one company's house style."}]},"prev":{"slug":"the-checking-tax","title":"The checking tax nobody counts","url":"https://addaly.com/learn/ai-for-small-business/the-checking-tax"},"next":{"slug":"draw-the-process-first","title":"Draw the process before you change it","url":"https://addaly.com/learn/ai-for-small-business/draw-the-process-first"},"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) 'Four doors AI comes through, and what each one costs', in AI for a Small Business. Edited by M. Mun. Addaly. Available at: https://addaly.com/learn/ai-for-small-business/where-ai-arrives (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":"Four doors AI comes through, and what each one costs","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":"Chat window, built-in feature, API and local model are four routes to similar capability that differ in setup cost, price per use and where your data ends up, and the built-in feature is usually cheapest because the data is already there.","hero":{"kind":"figure","figure":{"t":"figure","label":"Figure","figure":{"type":"stack","title":"Four doors, ordered by where your words end up","layers":[{"label":"The chat window","detail":"Nothing to set up and everything available, but entirely manual. Free tiers are real; paid consumer tiers cluster near USD 20 a person a month. Your words go to a vendor’s servers under terms you should read once."},{"label":"A feature in software you already pay for","detail":"Your accounts package, booking system, helpdesk or shop. Usually the cheapest, because the vendor already holds the same data under a contract you already signed, and there is no second copy."},{"label":"The API","detail":"About a cent for a page in and half a page out, a tenth of that on a cheap model. The model is cheap and the plumbing that calls it is not. Right when volume, rather than judgement, is the bottleneck."},{"label":"A model on your own machine","detail":"Llama, Mistral, Qwen or Gemma through Ollama or LM Studio, free. A clear step below the best hosted models, an afternoon to set up, and nothing leaves the building."}],"caption":"The order of price is not the order of exposure. The second door is often both the cheapest and the safest, and it is the one nobody checks before signing up with a new vendor."},"alt":"Figure: Four doors, ordered by where your words end up. Layers, top to bottom: The chat window (Nothing to set up and everything available, but entirely manual. Free tiers are real; paid consumer tiers cluster near USD 20 a person a month. Your words go to a vendor’s servers under terms you should read once.); A feature in software you already pay for (Your accounts package, booking system, helpdesk or shop. Usually the cheapest, because the vendor already holds the same data under a contract you already signed, and there is no second copy.); The API (About a cent for a page in and half a page out, a tenth of that on a cheap model. The model is cheap and the plumbing that calls it is not. Right when volume, rather than judgement, is the bottleneck.); A model on your own machine (Llama, Mistral, Qwen or Gemma through Ollama or LM Studio, free. A clear step below the best hosted models, an afternoon to set up, and nothing leaves the building.). The order of price is not the order of exposure. The second door is often both the cheapest and the safest, and it is the one nobody checks before signing up with a new vendor."}},"blocks":[{"t":"h2","id":"the-same-model-four-different-decisions","runs":[{"text":"The same model, four different decisions"}]},{"t":"p","runs":[{"text":"A great deal of confused advice comes from people comparing a chat window with an API and concluding one is better. They are not competing products. They are four delivery routes to broadly the same capability, and they differ in price, in where your data goes, and in how much work you have to do."}]},{"t":"h3","id":"1-the-chat-window","runs":[{"text":"1. The chat window"}]},{"t":"p","runs":[{"text":"ChatGPT, Claude, Gemini, Copilot, DeepSeek, Le Chat, Qwen. A browser tab or a phone app. Free tiers are real and many small businesses never leave them; paid consumer tiers cluster around USD 20 per person per month, with lower regional pricing in some markets."}]},{"t":"p","runs":[{"text":"What you get: everything, immediately, with no setup. What you give up: it is entirely manual. Somebody has to open the tab, paste the context, and copy the answer back. That is fine for six proposals a month and hopeless for four hundred product descriptions."}]},{"t":"p","runs":[{"text":"This is where every small business should start, and where most should stay."}]},{"t":"h3","id":"2-a-feature-inside-software-you-already-pay-for","runs":[{"text":"2. A feature inside software you already pay for"}]},{"t":"p","runs":[{"text":"Your accounting package, your booking system, your helpdesk, your online shop, your office suite. Zoho, QuickBooks, Xero, Tally, Shopify, HubSpot, Google Workspace, Microsoft 365: nearly all of them now have AI features, often included, sometimes as a per-seat add-on."}]},{"t":"p","runs":[{"text":"This route is systematically underrated, for a reason that has nothing to do with the model: "},{"text":"the data is already there.","b":true},{"text":" No pasting, no stripping, no second vendor, no second contract, no second copy of your customer list in a system you have not read the terms of. It also tends to be the cheapest, because the vendor is bundling."}]},{"t":"p","runs":[{"text":"Before you buy anything new, spend an hour looking through what you already pay for. A surprising number of owners are paying twice."}]},{"t":"h3","id":"3-the-api","runs":[{"text":"3. The API"}]},{"t":"p","runs":[{"text":"You pay per token: input and output are usually priced separately. Order of magnitude for a page in and half a page out: about one US cent on a mid-priced model, closer to a tenth of a cent on a cheap one. Five hundred customer replies a month typically lands between USD 0.50 and USD 6."}]},{"t":"p","runs":[{"text":"That is startlingly cheap, and it is why people get excited. The catch is structural: "},{"text":"an API key is not an application.","b":true},{"text":" Something has to call it: a script, a spreadsheet add-on, an automation tool, a developer. The model is the cheap part; the plumbing is the expensive part. Budget the plumbing, or use a no-code automation tool that provides it."}]},{"t":"p","runs":[{"text":"The API route is right when the volume is high, the task is identical every time, and a person is not in the loop for each item."}]},{"t":"h3","id":"4-on-your-own-machine","runs":[{"text":"4. On your own machine"}]},{"t":"p","runs":[{"text":"Open-weight models (Llama, Mistral, Qwen, Gemma, Phi and others) run locally through Ollama or LM Studio, both free. A modern laptop with 16GB of memory runs a small model usefully; 8GB is tight but workable with a smaller one."}]},{"t":"p","runs":[{"text":"Marginal cost is zero and nothing leaves the building, which dissolves a large part of the privacy problem in one move. Quality is a clear step below the best hosted models, and the gap is widest on reasoning and on languages other than English. It is good at repetitive sorting, classification, tidying and drafting, which is much of what a small business actually needs."}]},{"t":"p","runs":[{"text":"Set-up is an afternoon. If your business handles health records, legal files or anything you would not want on somebody else’s server, this is the route worth an afternoon."}]},{"t":"h2","id":"choosing-in-one-paragraph","runs":[{"text":"Choosing, in one paragraph"}]},{"t":"p","runs":[{"text":"Start in the chat window until you know the task. Look hard at what your existing software already offers before adding a vendor. Move to the API only when the volume is high enough that a human copying and pasting is the bottleneck. Go local when the data is the constraint rather than the cost."}]},{"t":"p","runs":[{"text":"And notice the thing that stays constant across all four: the facts sheet, the checking pass, the two tasks. The door changes what you pay and who holds your data. It does not change the discipline."}]},{"t":"h2","id":"one-thing-that-is-true-of-all-four","runs":[{"text":"One thing that is true of all four"}]},{"t":"p","runs":[{"text":"Whichever door you use, you are sending your business’s words to something. On routes one and three that is a company’s servers, under terms you should read once. On route two it is a company that already holds the same data. On route four it is your own machine and nothing leaves."}]},{"t":"p","runs":[{"text":"That ordering is worth keeping in your head, because it does not match the price ordering. The cheapest option per item is not the safest, and the safest is not the most convenient. Module seven works through what to check for each; for now, notice that the door you pick is a data decision as much as a cost one."}]}],"check":{"question":[{"text":"A furniture business needs 800 product descriptions generated from a spreadsheet of specifications, once, before a catalogue deadline. Which consideration should drive the route they pick?"}],"options":[{"text":[{"text":"Local models are free to run, so the only real cost is the afternoon of setup."}],"correct":false,"tempting":[{"text":"Optimises for the wrong constraint; running cost is trivial at this volume either way, and a local model adds setup time and a quality step down for no benefit here."}]},{"text":[{"text":"The API is cheapest per item, so it is the correct choice whenever volume exceeds a few hundred, and at eight hundred descriptions the per-item saving comfortably justifies whatever setup is needed."}],"correct":false,"tempting":[{"text":"Prices the model and ignores the plumbing; per-item cost is pennies, but somebody still has to build and test the thing that calls the key before the deadline."}]},{"text":[{"text":"Whether their existing shop or catalogue software already has a bulk description feature, since the specifications are already in it and no second vendor is involved."}],"correct":true,"tempting":null},{"text":[{"text":"Whichever model scores highest on published benchmarks for writing quality, since the output is customer-facing."}],"correct":false,"tempting":[{"text":"Treats a benchmark as evidence about this task; general writing scores say nothing about following a specification sheet in one company’s house style."}]}],"order":[0,2,3,1],"why":null},"prev":{"slug":"the-checking-tax","title":"The checking tax nobody counts","minutes":9},"next":{"slug":"draw-the-process-first","title":"Draw the process before you change it","minutes":8}}}