Joanah Fermin

Tutorials / Disciplined AI-Assisted Development · Lesson 2 of 14

On this page

    Making AI Useful

    Many people hold back from using AI, and the worry is fair:

    Is it worth the time and effort to fit AI into my work, when it makes mistakes?

    Our real goal is not for AI to be accurate at everything. It is for AI to be very good, and efficient, at the specific jobs you give it, again and again.

    Here is the route, using one example throughout, the support desk of an online shop:

    1. Why an AI gets things wrong: a gap or a fork.
    2. A human analogy: a new support hire, and how they become reliable.
    3. The same story with AI: every new chat is a new hire.
    4. The handbook: guidance, how to grow it in a sandbox, and what it can and cannot enforce.
    5. Is it worth it? The handbook as an investment that pays back.

    We leave out how to word a request for one specific job (lesson 03) and which documents to write for a project (lesson 08).

    Why an AI gets things wrong

    Any worker makes mistakes, and an AI does too. When an answer is wrong, it is usually for one of two reasons.

    A gap: it never saw the fact

    Some questions have one right answer, and the AI does not have it: your shop's return window, say. Your shop's rules are like the private code in lesson 01: they were never in the training text. A database in that position would return nothing. An AI does something else.

    A model is built to always predict a next word (lesson 01), so there is no "nothing" result. To see the effect, picture a tiny predictor that learned only from five sentences about a bookshop, and is built to always answer: for a word it never saw, it picks any word it knows. Asked about words it had never seen, it answers like this:

    After "invoice": stock
After "refund":  shelf
After "invoice": the

    It always answers, the answers are nonsense, and asking twice gives two different answers. A real model is far better at this, because it has seen far more. But the shape is the same: when it lacks the fact, it can still produce a fluent answer. That is called a hallucination: text that reads as confident and correct and is not. The tone does not change.

    A fork: several answers fit, and only you know which

    Ask a computer "what is 5 × 5?" and there is exactly one answer. Programs and language models both get that right, almost every time. Now a customer writes: "I want a refund for my $120 order." That sounds like one job, but several answers are reasonable: refund now, offer a replacement, or hand over to a person. The AI picks one that is common in what it learned (and may pick a different one next time, as lesson 01 showed), which may or may not be yours. It cannot know, because the answer depends on your shop's rules and culture.

    Two reasons an answer goes wrong A gap "Return window?" it never saw your rule fluent guess: "45 days" the fix: add the missing fact A fork "Refund my $120 order" refund replace pass on picks the typical one the fix: you make the decision

    Mistakes like these are getting rarer. More and better training, and models that write out their reasoning before they answer (lesson 01), have made them rarer in many areas. That helps, with a catch: rarer mistakes are harder to notice, and your shop's rules were never in any training text.

    Now see how a person fixes both kinds of mistake.

    The human analogy: a new support hire

    Imagine a new person joins the support desk of an online shop on Monday. They are sharp, friendly and willing, and they know nothing about this shop.

    The first customers ask for refunds. You give them no instructions, so they use their own judgement. One refunds everything to keep customers happy, and it costs the shop money. Another refuses requests you would have approved. They are not careless. They just do not know how this shop does things. If nobody tells them, the next customer goes the same way. They stay at that level, however long they stay.

    No feedback: the same mistakes every time Case 1wrong Case 2wrong Case 3wrong Case 4wrong Case 5wrong Nobody says what was wrong, so every case starts from the same place.

    Now suppose you give feedback.

    • "That refund was too generous. For small orders we just refund, but not for big ones." Next time, they adjust.
    • "That reply was good, keep using it." Next time, they go straight to the good approach.

    Week after week, the same person improves. The mistakes you corrected stop coming back, and the good approach becomes habit. The number of corrections you have to make falls.

    Corrections fall as feedback builds up corrections you have to make never quite zero: a person still reviews day one weeks later, with feedback

    It never quite reaches zero, as with any good employee, who is still reviewed. But the work gets steadily more reliable and the review gets quicker.

    Write it down

    All that feedback lives in your head and in their memory. Write it down, in a handbook, and it becomes something more.

    Feedback goes into the handbook A case New hire A result Your feedback The handbook the next case starts from the handbook

    The handbook is an investment. It costs some effort to write now, and every future hire benefits. A new person who starts next month does not repeat the first one's mistakes. They begin where the first one ended up.

    One handbook, many new hires The handbook written once New hire A New hire B New hire C each is good from day one

    Nothing you wrote down is lost when someone leaves.

    An AI is a new hire every time

    Now picture an AI answering customer messages for the same shop. It can make choices: issue a refund, offer a replacement, apologise, pass the case to a person. (A real support desk is far more varied than this. We simplify it to show the idea.)

    Would it do a good job on day one, freshly set up?

    Probably not. It knows how customer support usually goes, from everything it learned (lesson 01). It does not know how your shop does it. So it does the typical thing, and the typical thing may be wrong for you. Companies differ: some refund quickly to keep customers happy, some are strict, some offer a replacement first, and each has its own tone and its own limits on what a front-line person may decide. All of these are reasonable. Only the company can choose.

    The typical thing is not your shop's thing Shop A refunds quickly Shop B strict about refunds Shop C replacement first A fresh AI does the typical thing It matches no shop exactly. The handbook says which shop it works for.

    Every new chat with an AI is a new hire on their first day. It is capable and well read, and it knows nothing about your shop. There are two differences from a person, and both shape the rest of the course:

    1. It does not remember your feedback. A hire you correct today does better tomorrow. The model does not change between chats (lesson 01: a new chat starts empty). So the memory has to live somewhere else.
    2. It does not flag its guesses. A careful person usually knows when they are guessing and says so. The model's tone stays the same whether it is right or wrong. So someone has to check.
    A person remembers. A new chat starts blank. A person Monday: corrected remembers Tuesday: does better An AI Chat 1: corrected nothing carried over Chat 2: starts blank So the memory has to live somewhere else: in the handbook.

    The "somewhere else" is the handbook: your documents. It is the same picture as before, with chats in place of hires. Each new chat is handed the handbook and starts in a good place, and whatever you learn goes back into it.

    One limit: the model sees only a fixed amount of text at once, and every word of the handbook is read again and paid for on every turn (lesson 01), so the handbook has to be short, organised and kept current.

    The same documents, every session The handbook Refund policy Scripts and tone Hand-over rules Chat today starts well Chat tomorrow starts well A teammate's chat starts well Different sessions, the same documents: each one starts in a good place.

    The handbook: guidance

    You cannot change what the model learned. You can change what it sees at the moment it answers. In this course we call everything you put in front of the model guidance: the message, the facts you attach, the rules you state, the documents you hand over.

    Here is a small piece of a support handbook:

    Refunds

Under $50:      Refund straight away. Do not argue.
$50 to $500:    Try a retention script first (below).
                If the customer still wants a refund, hand over to a person.
Over $500:      Do not decide. Transfer to a supervisor.

Script A:  "I am sorry about the trouble. I can send a replacement
            today, or add a credit to your next order. Would either help?"

Always: be polite, keep replies short, and never promise anything
that is not in this document.

    It contains decisions (the three tiers), scripts (what to say) and style (polite, short). None of that was in the AI's training. It is your shop. Each entry fixes one of the two kinds of mistake: a fact for a gap, a decision for a fork. Good guidance answers three things the AI cannot guess: what the goal is, what the limits are, and what a good answer looks like.

    Growing the handbook: practise in a sandbox

    You do not write a perfect handbook up front. You grow it by practising in a sandbox: a safe place where mistakes cost nothing. For the support desk, that means made-up customer cases and no real refunds. Include the awkward ones: a $49 order and a $51 order, a $500 order, an angry customer.

    Practise in a sandbox, then go live The sandbox nothing real is at stake Handbook so far Run the AI on made-up cases Compare with what you wanted Gap or fork? fix the handbook Go live limits enforced, still reviewed exit when corrections are low

    The exit is a count: for this job, how often does a person change the AI's reply or reverse its refund? When that number is low enough that you are happy to trust it, you go live. Even then, keep counting, because real customers will surprise you. Confidence here means "reliable enough that review is quick", not "never wrong".

    Walls, not a guarantee

    Think of the handbook as two walls along a road: what we do (our policies) and what we never do (our limits). They keep the AI on the way from the customer's message to a good outcome.

    The handbook as walls along the road Wall: what we do (our policies) Wall: what we never do (our limits) Customer message Good outcome the AI, kept on course a slip: reviews and checks catch it The walls are guidance. They steer, they do not seal. A limit that must never be crossed is enforced by the system, not only written down. A later lesson shows how.

    Three things to remember:

    • Hand it over each time. It only steers when it is in front of the AI.
    • It steers, it does not seal. "Never refund over $500" is advice to the AI. If it must truly never happen, the system should not give the AI that ability, and someone still reviews the refunds.
    • Keep it current. If the policy changes to "under $75" and the handbook still says $50, the AI follows the old rule with full confidence. When the plan changes, the handbook changes first. That is the course's first rule.

    This gives us the course's second rule:

    AI is only as accurate as the guidance it gets.

    Is it worth it? The investment

    Back to the opening question: is the time and effort worth it? Think of the handbook, and the sandbox practice that builds it, as an investment. It costs hours up front, and it pays back in three ways.

    It saves hours

    Take one person who spends 8 hours a day on a repeated job. Suppose you invest 400 hours guiding an AI, and afterwards the job takes 4 hours a day: checking the work, reviewing the costly cases and keeping the handbook current. The running total of hours looks like this:

    Payback: the investment crosses over at day 100 hours spent, running total By hand: 8 hours a day With AI: 400 hours up front, then 4 hours a day 400 h payback: day 100 working days shaded: time saved after payback, growing every day

    The two lines cross at day 100, because 8 hours a day by hand equals 400 plus 4 hours a day with AI at exactly that point. After that, every working day is time saved, for every year that follows. That is about five months of working days. The figures are an illustration, and yours will differ.

    Two things make it better:

    • More people. The 400 hours are spent once and each person saves 4 hours a day, so with five people payback comes at day 20.
    • Cheap checking. The 4 hours that remain are checking and review. The easier a wrong answer is to see, the smaller that number stays.

    Payback needs a job that repeats, with enough volume. A job done twice a month never repays 400 hours. And where a wrong answer is costly and hard to see, think twice, or build the checking first.

    It saves explaining over and over

    Without a handbook, every new chat starts the same way: you explain your shop again. You explain it differently each time, you forget a rule, and a teammate explains it their own way. With a handbook, you write it once and hand it over.

    Explain once, hand over every time Without a handbook Chat 1 explain it all again Chat 2 explain it all again Chat 3 explain it all again With a handbook Chat 1 hand it over Chat 2 hand it over Chat 3 hand it over Write it once. Hand it over every time.

    It gets quicker and better

    Each lesson you learn goes back into the handbook, so corrections fall (the curve from earlier) and a new chat starts further ahead. Results arrive in fewer rounds and match your shop's standard, not a typical one.

    That is the point of the course: good process makes AI useful. The handbook is its first piece.

    💡 Next: software development. The support desk was only an example. From here on, we apply the same ideas to building software.

    Recap

    • AI goes wrong from a gap (a missing fact) or a fork (a decision only you can make).
    • Every new chat is a new hire, so the handbook holds the memory.
    • Grow the handbook in a sandbox, and enforce hard limits in the system.
    • Rule 2: AI is only as accurate as the guidance it gets.
    • The handbook is an investment that pays back in hours saved and better results.

    Questions to think about

    The answers are left out on purpose. Try them in your own words first.

    1. Why is "is AI accurate?" the wrong question, and what should you ask instead?
    2. What are the two kinds of mistake an AI makes, and what does each need from you?
    3. Every new chat is a new hire. Why does that make the handbook an investment?
    4. Why does a handbook steer an AI but not guarantee its work?
    5. Why does an AI answer only the question you asked? (Lesson 03 answers this.)

    Next lesson: why one instruction can be interpreted many ways, and how choosing the hat makes it specific.

    ← How Language Models WorkGetting Better AI Results by Being Specific →