AI for a Small Business
Two or three tasks, honestly measured. Not a new department.
- Level
- Nothing assumed
- Lessons
- 88
- Reading time
- 767 min
- Price
- Free, no sign-up to read
A practical course for someone whose own money is at stake — a shop, a workshop, a clinic, a small agency, a nonprofit. How to find the two or three tasks where AI saves real hours, how to keep invented numbers and customer data out of trouble, what it should cost you a month, and how to run a two-week trial that gives you an honest answer instead of a hopeful one.
Opens after the AI at Work exam
Sign in, finish that course, and pass its exam. You can read this syllabus meanwhile.
Go to AI at WorkModule 1
Working out where AI belongs in your business
Before any tool, a diagnosis. This block starts from how the machine actually produces text — because that is what predicts which of your tasks it will handle and which it will quietly ruin — and ends with a one-page map of your own week: the two candidate tasks, the hours each takes now, the cost of checking the output, and the honest baseline that every later claim of improvement has to be measured against.
By the end you can
Produce a one-page map of your own week that names two candidate tasks, states the hours each takes today from a real five-day count, prices the checking those tasks will need, and explains from the model's mechanism which of your tasks it will fail at
- 1The Two Tasks Worth DoingStart from where your hours actually go — frequent, text-shaped, cheap-to-check work — not from what the tool can do.
- 2Fluent and wrong come from the same placeA language model predicts likely next words rather than retrieving facts, so its confidence is a property of the writing and never evidence about truth — which is why the facts must go in for the words to come out safely.
- 3The checking tax nobody countsThe 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.
- 4Four doors AI comes through, and what each one costsChat 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.
- 5Draw the process before you change itMap the job as steps, mark each one as text, rules or judgement, and ask why each step exists before automating any of them — because a step that can be deleted saves all of its minutes and needs no checking.
- 6The five-day tally, and why memory is not a baselineCount the task for five ordinary days before changing anything, because memory inflates hated work and erases frequent short interruptions, and without a written before there is no after — only the good feeling every new tool produces.
- 7Seven things it will fail at, and whyThe model fails wherever producing text that has the shape of a correct answer is not the same as computing one — arithmetic, rotas, exhaustive lists, counting — and the test is whether a wrong answer would look any different from a right one.
- 8One person owns it, for one monthCompetence with these tools is built by one person doing one task repeatedly until they can predict its failures, so adoption in a small business is one person, one tool, one month — with what is learned written into a file that survives that person leaving.
- 9Five ways a small-business pilot diesPilots die from missing baselines, impressive-but-rare tasks, uncounted checking, tools bought before problems, and knowledge held by one person — and even a successful one fails if the hours saved were never assigned to anything.
Module 2
Asking properly: prompting for business work
A prompt is not a magic phrase, it is a briefing. This block builds the briefing a small business actually needs — the one page of facts that travels with every request, the five parts a good request always has, the examples that carry your voice where adjectives cannot, and the instruction that turns invention into a visible gap — then tests it against ten real cases and hardens it against the customer message written to manipulate it.
By the end you can
Write, harden and test a reusable prompt for a real task in your business — with a dated facts sheet attached, a rule that makes missing information visible, and a ten-case test that says whether it is safe to use unsupervised
- 10The one page that travels with every requestOne dated page of prices, boundaries, terms and voice, stored where every conversation sees it, is what turns a plausibility machine into a useful assistant — and a stale sheet is more dangerous than no sheet, because wrong facts arrive with no hesitancy at all.
- 11Five parts a request should always haveA good request states situation, facts, one task, constraints and format — and the part most people skimp, the raw facts, is exactly the part whose absence produces the empty prose that reads as generated.
- 12Five examples teach what adjectives cannotFive real, varied, clearly labelled samples of your own writing transfer voice in a way no list of adjectives can, because the model continues the pattern in front of it — but examples carry manner rather than knowledge, so the facts sheet is still doing the other half of the job.
- 13Making the gaps visibleYou cannot stop a model inventing, but grounding it in a named source, licensing a NOT IN FACTS answer and requiring a quote for every claim converts invisible invention into visible gaps you can fix in seconds.
- 14Repairing a draft instead of starting againRepair a draft by naming the specific defect rather than asking for something better, stop after two failed turns, and promote any correction you make three times into the prompt itself so you never make it again.
- 15Why long conversations go strangeThe whole thread is re-read on every turn, so long conversations cost more, slow down, and silently drop their oldest material — which is why standing instructions beat a facts sheet pasted at the top of an hour-old chat.
- 16The four settings that change the answerModel 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.
- 17A prompt library a two-person business can keepKeep working prompts in one shared, dated file with a known-failures line for each, because that line is the accumulated experience of the person who used it three hundred times and is what actually transfers when they are away.
- 18The customer message written to manipulate your assistantA model cannot reliably distinguish instructions from the text it was asked to read, so hostile content in a message, file or web page can redirect any system that acts without a person — which is why capability, not cleverer filtering, is the thing to ration.
- 19Ten cases before you trust itTest a prompt against ten real cases — four ordinary, two with missing information, two awkward, one out of range, one hostile — scored pass or fail on behaviour, and re-run the whole set whenever the prompt, the facts or the model changes.
Module 3
Your own documents, and getting straight answers out of them
A model that has never seen your business is a guesser. The same model handed your price list, your contracts and your past emails is a reader, which is a different and far more reliable thing. This block explains what actually happens when you upload a file, why the passage it finds is chosen by meaning rather than by words, what OCR and transcription really cost you in errors, and how to tell whether an answer came from your document or from the model's memory.
By the end you can
Turn your business's own files into something you can ask questions of, and for any answer it gives, say whether that answer came from your document or from the model's memory and verify it against the source in under a minute
- 20What actually happens when you upload a fileUploading a document may put all of it in front of the model or only the few passages your question retrieved, which is why pinpoint questions work well, exhaustive ones return confident incomplete lists, and "is this clause absent?" is the least trustworthy question you can ask.
- 21Grounding: answering from your material rather than its memoryGrounding converts the task from recall, where models fabricate, to reading comprehension, where they are strong — but it inherits everything wrong with your documents, so a grounded answer from a stale price list is a confident error with a citation attached.
- 22Why it finds "refund" when the customer typed "money back"An embedding turns text into a position in a space of meaning, so search becomes a distance calculation rather than a word hunt — which is why it matches paraphrases and other languages, and also why "our refund policy" and "our no-refund policy" sit dangerously close together.
- 23Four ways to keep a small library, and what each costsNotebook products, projects in tools you already pay for, features inside software that already holds your data, and a local setup are four places a small document library can live — and the deciding questions are whether the material is confidential and how you would get it out again.
- 24Turning paper into rows you can checkOCR combined with a language model produces clean, plausible text rather than obvious garbage, so a misread digit arrives repaired and unmarked — which is why extracted figures get checked against the paper and against an external total, never against how sensible they look.
- 25Reading a contract you were going to skim anywayOn a long agreement a model reliably finds and explains a clause you name, and reliably fails to tell you a clause is absent — so use it to arrive at an adviser prepared, never to conclude that something is not there.
- 26Transcription, and the consent question nobody asksTranscription errors concentrate on names, numbers and unusual words — the very content you wanted — so a transcript is a memory aid rather than a record, and the short written confirmation you send afterwards is the thing that counts.
- 27The stale document problemA grounded system inherits the currency of its documents exactly, so an out-of-date source produces confident cited errors indefinitely — which is why every document carries a date and a named owner, superseded versions leave the library, and a change to a price is a change to the facts sheet on the same checklist.
- 28Checking a grounded answer in under a minuteCheck a grounded answer by confirming a citation exists, that the quoted text is really in the source, that it supports the claim rather than merely sitting near it, and that the source is current — and treat a completely confident answer with no gaps as the one most worth examining.
Module 4
Customers, from first message to resolved complaint
The front desk is where most small businesses spend their words: enquiries, bookings, reminders, complaints, reviews and the phone. This block works through each of them with one rule underneath — a customer must never receive a price, a promise or an apology the owner did not approve — and it treats the handover to a person as the part of the design that decides whether any of it helps.
By the end you can
Run the enquiry-to-resolution path with AI help without a customer ever receiving a price, promise or apology you did not approve, name the point at which the system must hand over to a person, and measure whether customers were served better rather than merely faster
- 29Answering Customers Without Losing Your VoiceGive it your facts on one page, ask it to flag what it does not know, and keep your hand on the send button.
- 30Sorting before draftingClassification is safer and often more valuable than drafting because the output is a label from your own list, and an explicit UNSURE bucket plus written tie-break rules are what stop awkward messages disappearing into the wrong pile.
- 31Putting a bot on your website, with your eyes openA bot on your site speaks as your business and may bind it, so ground it in your own documents, refuse it the ability to quote custom prices, discounts or dates, give it read access rather than write access, and read twenty real transcripts every week.
- 32Designing the moment it gives upThe handover is the part of the design that decides whether automation helps, so hand over after two failures, on any money or emotional signal, and instantly on request — and never measure success by deflection, which counts abandonment as a win.
- 33The angry message, and what your reply commits you toA drafted complaint reply defaults to accepting fault and offering remedies you did not authorise, so read every one for commitments alone and delete rather than soften — sympathy is free, fault and remedy are decisions you make.
- 34Bookings and no-shows: the case where AI is the wrong toolReminders and confirmations are rules work that a booking system does free and reliably, and a deposit beats every message; the model earns its place only on the varying parts — filling a cancelled slot, the awkward conversation, and finding the pattern in six months of bookings.
- 35Reviews: asking, answering, and the line you must not crossFabricating reviews is prohibited in a growing number of regimes and detectable by platforms, asking only your happy customers is gating and is banned on major platforms, and the highest-value use of AI here is reading two years of reviews in bulk to find what customers value that you never advertise.
- 36The phone: what a voice assistant can and cannot do yetVoice adds latency, interruption and irrecoverable mis-hearing to every text problem, so it suits narrow bounded jobs with no money in them — and for most small businesses transcribed voicemail with a summary solves the real problem, which is failing to call back.
- 37Serving customers in a language you do not readTranslation quality drops sharply outside the best-resourced languages and fails fluently rather than visibly, so back-translate to catch gross errors, have a real speaker read ten replies before you rely on a language, and never publish permanent text in a language nobody in the business can read.
- 38Did the customer actually get better service?Speed improves immediately and means little on its own, so measure first-contact resolution, repeat contact within 48 hours, enquiry-to-outcome and complaints about the service itself — and never report hours saved without the outcome rate beside it.
Module 5
Selling: listings, images, marketing and price
Everything a customer sees before they buy. This block covers the copy that carries facts only you have, the photograph that has to be accurate rather than flattering, two hundred listings produced in an afternoon with a stated error bound, what changed about being found when search engines started answering questions themselves, and the one thing in the whole block that must never come out of a model: the price.
By the end you can
Produce listings, images and marketing that carry facts only you have, comply with advertising and platform rules, and come with a stated bound on how many of them you actually checked
- 39Listings, Menus, Quotes and ProposalsThe model writes the sentences; every number, date and promise has to come from you.
- 40Posts That Do Not Sound GeneratedGenerated text reads generated because it is empty; fill it with what only you saw today.
- 41Editing a product photo, and where editing becomes a lieRetouching toward what the object actually looks like is fine and retouching toward something else is a misleading advertisement, and the test is whether the customer opening the parcel would call the photograph accurate — with colour the place honest businesses most often cross the line.
- 42Generated images: what is unsettled, and the three rules that are notWhether training on copyrighted images was lawful and whether you own the output are separate questions, both unresolved and answered differently by country — but never depicting a product you do not sell, never presenting a generated face as a real person, and never generating another brand's marks are settled everywhere.
- 43Doing two hundred at once, and knowing how many to checkBatch generation is safe when the facts come from a spreadsheet, missing fields are flagged rather than filled, and you know how many rows you checked — because zero errors in twenty samples still allows a true error rate near 15%, and in fifty it is about 6%.
- 44Being found, now that search engines answer questions themselvesSearch engines answering questions directly has cut informational traffic while transactional and local queries still end at a business, so the durable work is an accurate local listing, pages answering real customer questions with real numbers, and consistent descriptions elsewhere.
- 45The objection it cannot knowGenerated copy is polite and ineffective because the model cannot know which specific hesitation is stopping your customers, so mine the objections out of your own enquiries and feed them in — and remember that changing the offer moves conversion far more than any rewrite.
- 46Automated advertising on a small budgetAutomated bidding needs roughly fifty conversions a month to learn, so a small business never leaves the learning phase and pays for the guessing — which is why intent-based search advertising, negative keywords and a known value per lead beat any amount of platform automation at small scale.
- 47Price: the one thing that never comes out of the modelA model has no access to your costs, your market or your need for the work, so any price it produces is a plausible shape rather than a figure — use it for the spreadsheet formula, the tier structure and the increase letter, and never for the number itself.
Module 6
Money, paperwork and the back office
The unglamorous half of a business: spreadsheets, cash, stock, rotas, hiring, procedures and official forms. This block draws one line through all of it — the model writes the formula, the letter and the procedure, and never computes the number, decides the filing or chooses the person — and it names the places where a plain automation rule beats a model outright.
By the end you can
Use AI across the paperwork of the business — formulas, cash forecasts, reorder points, supplier comparisons, rotas, job ads and procedures — without it ever becoming the place a number is computed, a record is kept, or a decision about a person is made
- 48Bookkeeping Without Handing Over Your AccountsUse it to sort and to write; never to add up, and never as the place your records live.
- 49Ask for the formula, never for the totalA formula is structured text the model produces reliably and a total is arithmetic it produces plausibly, so ask for the formula and let the spreadsheet compute — then test it on three rows whose answers you already know, including one awkward boundary case.
- 50Thirteen weeks, and why profitable businesses run out of moneyProfit is a period and cash is a moment, and the gap between them is timing — so a thirteen-week rolling forecast, updated for fifteen minutes every week with your own figures, turns a future shortfall into a date you can act on.
- 51Reorder points, and comparing quotes that are not comparableA reorder point is lead-time demand plus safety stock and takes two minutes to calculate for the ten items that matter, and the real work in comparing supplier quotes is normalising them into the same units — where NOT STATED is often the most informative field.
- 52Rotas and routes: the tools that actually solve themRotas, routes and allocation are constraint problems that need a solver, not a generator — so use a spreadsheet solver or OR-Tools for the answer, use the model to elicit the constraints and explain the result, and always verify with a separate rule-by-rule check.
- 53Hiring: where automation is most regulatedHiring is the most regulated place an automated decision can sit, and bias arrives through proxies that no prompt removes — so use AI for the advert, the structured interview and the scoring guide, and never to sift, rank or score the candidates themselves.
- 54Getting the procedure out of your head and onto a pageSpeak the procedure aloud and have the transcript turned into numbered steps with GAP markers where something is missing, because the gaps are the unstated steps you no longer notice — and never let it fill them, or you get a plausible document describing a business that does not exist.
- 55Official letters, forms and the line at filingA model is good at explaining what an official letter says and what happens if you ignore it, and unreliable on tax because the rules are jurisdictional, annual and conditional — so use it to form the precise question, then check the answer against the authority's own guidance.
- 56When a plain rule beats a modelAnything that happens the same way every time is a rule, and rules are cheaper, instant, deterministic and impossible to talk into inventing something — so build the rules first and add a model step only where the input varies and a person checks the output.
- 57Arriving prepared, and what never to outsourceYou are buying a professional's accountability as much as their knowledge, so use AI on the preparation — the briefing, the questions, the organised papers, and the plain explanation afterwards — and take every answer back for confirmation rather than acting on it.
Module 7
Letting it act: automation you can trust with the keys
Module six drew the line between a rule and a model. This block is about what happens after that line is drawn and you start letting the thing do rather than suggest. It covers the ladder of autonomy and what evidence licenses each rung, how to make a model return fields a machine can check instead of prose a person must read, what a call actually costs and how a loop becomes a bill, what an agent is underneath the marketing, and the operational plumbing — permissions, duplicate protection, logs, approval queues — that decides whether an automated business is calm or occasionally catastrophic.
By the end you can
Take one repeating job from suggestion to supervised action: choose a rung of autonomy you can defend with evidence, make the model return checkable fields rather than prose, bound what a confused or hijacked run can do, guarantee that a retry cannot send the same thing twice, and reconstruct from your own log why any given action was taken three weeks after it happened
- 58Five rungs of letting go, and what licenses each oneThe 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.
- 59Getting fields out instead of paragraphsAsking for named fields turns an output a person must read into one a machine can check, because a missing field is an error your software can catch while a missing sentence is not.
- 60What a call costs, and how a loop becomes a billUsage pricing is charged per token in and out, so cost scales with how much context you resend rather than with how many tasks you do — which is why an automated loop is a financial exposure and a spending cap is not optional.
- 61What an agent is underneath, and why ten steps go wrongAn agent is a model in a loop that can call tools and decide when to stop, so its reliability is the per-step reliability raised to the number of steps — which is why long autonomous chains fail far more often than each step suggests.
- 62Blast radius: what one confused run can reachAutomations should be given the narrowest access that does the job, in their own account with their own revocable key, because the question is never whether something will go wrong but how much a single wrong run can touch.
- 63The duplicate problem, and the one line that fixes itA retry cannot tell a failed action from a slow one, so every automated action needs a key that makes doing it twice the same as doing it once.
- 64A log that answers "why did it do that?" three weeks laterA run is only defensible if you kept the input, the output, the decision and the version of the prompt and model that produced it, because a model's output is not reproducible from the prompt alone.
- 65Designing the four-second approvalAn approval step only protects you if approving is fast and rejecting is easy, because a queue that takes too long is rubber-stamped and a rubber-stamped queue is rung five wearing a costume.
- 66The AI employee pitch, and the five questions that deflate itA product sold as an autonomous member of staff is the same model and tools you have just learned about, priced against a salary, so the questions that matter are how many steps a run takes, who is liable, and how you get your data out.
- 67The thing you lose that nobody countsRoutine work carries incidental noticing — the odd invoice, the customer who has gone quiet — and automating the work removes the noticing unless you deliberately rebuild it somewhere else.
Module 8
Learning from your own numbers
Every business in this course is sitting on two or three years of transactions, bookings, invoices and messages, and almost none of it has ever been looked at properly. This block is about getting a defensible answer out of it: what is usable and what is not, why cleaning is most of the work, how to make a tool compute rather than guess, why a forecast that cannot beat last year's figure is not a forecast, what a prediction about people is really telling you when the thing you are predicting is rare, the mistake that makes a model look brilliant and useless, and the point at which your data is too small to answer the question and an experiment is the honest alternative.
By the end you can
Get a defensible answer out of a year of your own transactions: clean it so the counts mean something, make the tool show the code that produced every figure, beat a stated naive baseline before believing any forecast, read a classifier against its base rate rather than its accuracy, name the leak that would have flattered it, and say when the data is too small to support the question and an experiment is the honest alternative
- 68What you are already sitting on, and what of it is usableYour transaction records are far more useful than any dataset you could buy, but only for questions about things you recorded — and the row that matters is usually a customer over time, not a sale on a day.
- 69Why cleaning is most of the jobCounts are only as meaningful as the categories underneath them, and a category spelled four ways is four categories — so the work of making numbers comparable is the analysis, not preparation for it.
- 70Make it compute, do not let it estimateA tool that writes and runs code on your file produces arithmetic you can check and rerun, while a model reading the same file produces a plausible figure with no working — and the two look identical on screen.
- 71The number any forecast has to beatA forecast is only worth having if it beats the simplest possible guess on data it has never seen, so you state the naive baseline and the holdout period before you build anything.
- 72Patterns you can name, and the promotion that poisons the historyMost of the variation in small-business demand is explained by things you already know the names of, and an unrecorded promotion or closure turns a knowable pattern into noise for every analysis you do afterwards.
- 73Predicting something rare, and why accuracy liesWhen the thing you are predicting is rare, a model can be 95% accurate by never predicting it at all — so judge it by how many of its warnings were right and how many real cases it caught, at a threshold set by what the warning costs you.
- 74The mistake that makes a model look brilliantA model that scores far better than seems plausible is usually reading a column that would not exist at the moment of prediction, so the test is whether every input would have been available before the thing you are predicting happened.
- 75AutoML: what it genuinely does, and what it cannot do for youAutoML automates model selection and tuning, which was never the hard part, and leaves you the whole of the hard part — the question, the definition, the columns, and whether the score is real.
- 76When your data is too small, and what to do insteadSmall samples produce large swings by chance and a busy owner tests many ideas against the same data, so most patterns found this way are noise — which is why a deliberate change measured forwards beats a pattern discovered backwards.
Module 9
What it costs, what it risks, and what happens in a year
Everything so far has been about getting the work done. This block is about still being glad you did it twelve months later: pricing each tool against what it actually replaces, understanding what a free tier takes instead of money, writing the one page of rules your staff need, handling the fact that AI has made fraud against small businesses meaningfully better, keeping access straight when somebody leaves, telling customers the truth about where a machine is answering, knowing what you have warranted to your own clients, surviving a vendor's disappearance, and going back at six months to the figures you wrote down at the start.
By the end you can
Run the AI side of a business for a year without a surprise: price every tool against what it replaces rather than against zero, keep customer data inside terms you have actually read, verify a payment instruction in a way a cloned voice cannot defeat, keep working when a member of staff or a vendor disappears, say in one sentence where a customer is dealing with a machine, and re-measure at six months against the baseline you recorded at the start
- 77What It Costs, and Keeping It Near ZeroPrice every tool against the hours or the cheaper tool it replaces, never against zero.
- 78The bill that creeps, and the quarterly hour that stops itPer-seat subscriptions grow by addition and never by subtraction, so the cost of AI in a small business is decided by an audit nobody schedules rather than by any purchasing decision.
- 79Free tiers, and what they take instead of moneyA free tier is paid for in data rights, rate limits or a smaller model, and which of the three it is determines whether it is fine for your marketing copy and unacceptable for your client files.
- 80Your Customers' Data, and the Law Where You AreStrip the data down, check and date the settings, and tell people in one honest sentence — you stay responsible either way.
- 81The one page your staff actually needStaff will use AI whether or not you have a policy, so the useful document names the approved tool and the three things that must never be pasted, rather than attempting a ban that only removes your visibility.
- 82Why fraud against small businesses got betterGenerative tools removed the cues small businesses relied on to spot fraud — bad grammar, a wrong voice, a generic message — so verification has to move to a channel the attacker does not control.
- 83The week somebody leavesAI work concentrates in one person's personal accounts and chat history, so a departure removes the tooling, the accumulated knowledge and your access to the customer data in it unless the accounts were the business's from the start.
- 84Saying where the machine is, in one sentenceDisclosure is settled where a customer might otherwise believe they are talking to a person and genuinely unsettled elsewhere, so the practical rule is to disclose the interaction rather than to label every sentence.
- 85What you have already promised your clientsYour own contracts and professional rules may restrict AI use before any AI law does, and confidentiality clauses written years ago frequently forbid sending client material to a third-party service.
- 86When the tool goes awayAI products are discontinued, acquired and repriced at a rate that makes dependence on any single one a live operational risk, so the test is how long it would take you to work without it.
- 87The Cheap Tool, and an Honest Two WeeksMeasure the week before, name the kill line in advance, and count the time spent fixing the output.
- 88Going back to the numbers you wrote downThe trial measured a fortnight under attention, so the only honest verdict comes from re-measuring six months later against the same baseline, when the novelty has gone and the tool is being used the way it will actually be used.