The Three Questions Your Data Should Answer Before You Spend a Penny

The question hiding behind every tool

A vendor lands in your inbox with a demo. The deck is gorgeous. There is an AI in it that predicts churn, recommends the next best action, and optimises your pricing while you sleep. It costs about as much as a junior hire and it promises to think for you. Before you sign a thing, I want you to ask one small question that will save you a fortune, and it is not about the price. It is this: which question is this actually answering, and is that the question I have?

Because there are only three questions that data can ever answer for a business, and they stack in a strict order. What happened. What will happen. What should I do. Descriptive, predictive, prescriptive. Every dashboard, every model, every shiny platform sits on exactly one of those three rungs, whatever the marketing calls it. Once you can name the rung, you can see what you are being sold, and more importantly you can see whether you are ready for it. Most businesses are not ready for the rung they are being sold, and that gap is where the money quietly leaks out.

I have watched this play out enough times that I now run every tool, every pitch, and every half formed idea an owner brings me through the same three question filter before we talk about budgets at all. It is astonishing how often a six figure ambition collapses back down to a modest, boring, deeply useful piece of reporting once you name the rung honestly. So let me walk you up the ladder properly, one rung at a time, with what each one answers, how it is actually built, what it costs, and a worked example you can hold in your head.

Three questions, one ladder

Think of it as a ladder with three rungs, because that is genuinely what it is. You do not get to skip rungs, and the whole point of this article is that skipping is exactly what everyone tries to do. The bottom rung tells you what happened. The middle rung tells you what will happen. The top rung tells you what to do about it. The value goes up as you climb, and so does the cost, and so does the fragility. Hold that last part in mind, because it is the bit the demos never mention.

Each rung is a genuinely different kind of work, built by different methods, priced in a different bracket, and useful for a different question. Treat them as interchangeable and you will overpay for the fancy one while starving the plain one that would have changed the most decisions. So let us take them one at a time.

Rung one, descriptive: what happened

What it actually answers

Descriptive analytics answers the past tense question. How many orders last month. Which three products carry the shop and which twenty are dead weight. Where customers drop out of the checkout. What a cohort of new customers did in its first ninety days. It is reporting, dashboards, and honest summary statistics, and it sounds boring because it is boring, and it is the single most valuable thing most small businesses are missing. You cannot manage what you cannot see, and an astonishing number of shops genuinely cannot see their own numbers cleanly.

The method behind it

The method is unglamorous and mostly plumbing. Get the data out of the systems it hides in, agree on definitions, join it together, and summarise it so a human will actually look. Most of descriptive analytics is counting, grouping, averaging, and ratios. A conversion rate is just conversion rate = orders / visitors. A retention number is just retention = customers still active / customers you started with. There is no magic in the maths. The hard part is that the orders live in one system, the refunds in another, the traffic in a third, and nobody has ever written down what counts as a customer, so the counting quietly disagrees with itself.

What it costs

Descriptive is the least expensive rung by a wide margin, and it is where almost every small business should spend first. There is no AI in the brochure sense, there is just the truth, finally visible. The cost is mostly a few weeks of careful plumbing and one uncomfortable meeting where everyone agrees on definitions. Nine times out of ten this alone changes decisions, because owners are usually steering on gut and half remembered numbers, and the first honest dashboard is a small shock.

A worked example: the espresso cart

Say you run a little espresso cart. You feel busy, so you assume you are winning. We pull the actual numbers into one place for the first time. Last month you served 1,800 cups. Your average ticket was 4.20. That is a tidy 7,560 in revenue. Then we group by hour and product, and the picture changes shape. Two thirds of the revenue arrives between seven and nine in the morning. The afternoon is dead. Flat whites and the one pastry you nearly stopped stocking carry the cart, and half your menu sells in single digits. Nothing here predicts anything. It just describes, honestly, and already you can see that hiring an afternoon shift is madness and that running out of pastries at eight is costing you the best hour of the day. That is the bottom rung earning its keep before a single clever algorithm shows up.

Rung two, predictive: what will happen

What it actually answers

Predictive analytics answers the future tense question. Which customers are about to churn. How much stock you will need in November. Which lead is likely to close, and which is politely wasting your time. This is forecasting, regression, and the machine learning everyone means when they say machine learning. It is real and it is genuinely useful, and it stands entirely on the rung below it. A prediction is only ever as good as the description it learned from.

The method behind it

At its simplest a forecast is a pattern pushed forward. The plainest version is a moving average, forecast for next month = average of the last three months, and you would be surprised how often that beats a fancier model. One step up you fit a trend line, value = base + ( slope times month number ), and let the numbers tell you the slope. Churn and lead scoring work the same way in spirit: you take the history of who left or who bought, find the features that came before it, and learn the pattern so you can score the next person before they act. None of it sees the future. It assumes tomorrow rhymes with yesterday, which is usually true and occasionally, expensively, not.

The number that keeps a forecast honest is its error. A simple, brutal one is the percentage miss, percentage miss = size of the miss / actual value, averaged across the months you got to check. If your average miss is five percent you can plan on it. If it is forty percent you do not have a forecast, you have a horoscope with a confidence interval, and you should say so out loud before anyone bets stock money on it.

What it costs

Predictive costs more than descriptive, because it needs clean history, a model, and someone who knows the difference between a forecast and a guess dressed up as one. It is worth it once you can describe, because a prediction built on a clean description really is like seeing round a corner. Built on a messy one, it inherits every crack in the foundation and hides them behind a decimal point.

A worked example: the espresso cart looks ahead

Back to the cart. Because the description is now clean, we have twelve honest months of daily cups. We notice cups climb through summer and sag in January, a steady seasonal rhythm. A plain forecast on that history says next July you will likely serve around 2,300 cups, give or take the error band we measured. Now you can order beans and milk against a number instead of a hope, and you can staff the morning rush without drowning in waste. Notice what made this possible: not the model, the clean twelve months underneath it. Feed the same model the messy version, where returns and comped drinks and a broken till day are all tangled in, and it would forecast confidently wrong, and you would over order against nonsense.

Rung three, prescriptive: what should I do

What it actually answers

Prescriptive analytics answers the imperative question. Not what will happen, but the move to make about it. Raise this price. Send this offer to these customers and not those. Reorder this exact quantity today. This is optimisation, simulation, and decision rules, the stuff that acts on your behalf. It is the most valuable rung and the furthest from the ground, and it only works when the two rungs below it are solid.

The method behind it

Prescriptive turns a prediction into a decision by weighing options against a goal. The workhorse idea is expected value: expected value = probability times payoff, added up across the options, and you pick the option with the best one. A reorder decision is a tidy example. You do not order when the shelf looks low, you order at a calculated trigger, reorder point = ( daily demand times lead time ) + safety stock, where the safety stock buys you cover against a bad week, safety stock = safety factor times demand wobble times the square root of the lead time. Feed those formulas a good forecast and honest costs and they will hand you a number to order, today, with the guesswork taken out. That is prescription: the maths does not just tell you the weather, it packs your umbrella.

What it costs

Prescriptive costs the most, in money and in trust. It is not just a model, it is a model plus the rules and the guardrails and the organisational nerve to let a system act. That last one is the real cost people forget. You have to trust it enough to do what it says, and trust is expensive to build and instant to lose. One confident wrong recommendation on shaky data and nobody in the building believes the machine again, and you are back to the gut you were trying to replace, only poorer.

A worked example: the espresso cart decides

Same cart, both lower rungs solid. The July forecast says around 2,300 cups, milk lasts a few days, and your supplier delivers in two. The reorder rule works out that once your milk stock drops below the level that covers demand across those two delivery days plus a safety buffer for a hot spell, you order a specific number of litres, automatically, that morning. You stop running out at eight and you stop pouring sour milk down the drain at close. The move is small and unglamorous, and it only works because the description was clean and the forecast was measured. Skip either and the same rule would confidently order the wrong amount, on a schedule, forever.

The value climbs, so does the fragility

Here is the first thing worth really sitting with, because it is the hinge the whole article turns on. As you climb, the value rises, but the whole structure gets more fragile, because each rung stands on the one below and inherits its cracks. A clean way to feel this is to multiply your trust at each level. confidence in the action = confidence in the description times confidence in the prediction times confidence in the rule. Suppose you are a reasonable eighty percent sure of each. Then your confidence in the automated move is 0.8 times 0.8 times 0.8, which is about 0.51. Three good enough rungs stack into a coin flip. Now imagine the description is only fifty percent trustworthy because nobody agreed what a customer is. The whole product collapses and the machine is acting, fast and automatically, on something barely better than a guess.

Read that diagram the wrong way, from the top, and prescription looks like the smart place to start. Read it the right way, from the left, and you see that every arrow of error flows upward. A wobble in the description becomes a wobble in the prediction becomes a confident wrong instruction in the prescription. The higher you build on a cracked foundation, the harder you fall, and the more it costs on the way down, because now the mistakes are automated and fast.

The mistake almost everyone makes

Here is the pattern I see again and again, and it is the reason I wanted to write this down. A small business gets sold the top rung when it has not built the bottom one.

The vendor demo is always prescriptive, because prescriptive demos beautifully. The pitch is always that the AI will tell you exactly what to do next. Who does not want that. So the owner signs up for an engine that recommends the next best action, and it plugs into a data layer that cannot yet reliably tell you how many customers you had last quarter, because orders live in one system, refunds in another, and nobody agrees on what counts as a customer. The AI dutifully optimises on top of numbers that are quietly wrong, and the owner gets recommendations that feel scientific and are built on sand.

You cannot optimise what you cannot yet reliably describe. That is the whole thing on one line. If your description is broken, your prediction inherits the break, and your prescription amplifies it and then acts on it automatically. I am not against prescriptive analytics. It is genuinely where the biggest wins live. I am against buying it before you can describe your own business, because in that order it is not a tool, it is an expensive way to be confidently wrong at scale.

The ladder as a decision

So use the ladder as a diagnostic rather than a wish list. Start with the question you actually have, work out which rung answers it, and be honest about what it costs to stand on that rung safely.

Read that from the bottom. Wherever you want to end up, the arrows point you back to the same place to begin, which is clean, honest description. If your numbers are messy, no amount of predictive or prescriptive cleverness bolted on top will fix them. It will only decorate them.

The map on one page

Let me put the three rungs side by side, because seeing them next to each other is what makes the pattern click.

Analytics type	Question it answers	Typical method	What it gives the business
Descriptive	What happened	Reporting, dashboards, clean summary stats	You finally see the truth about your own business
Predictive	What will happen	Forecasting, regression, machine learning	You see round the corner and plan ahead
Prescriptive	What should I do	Optimisation, simulation, decision rules	The system recommends or takes the move for you

The framing here is not something I invented over coffee. The descriptive, predictive, prescriptive ladder is the spine of how the field teaches analytics maturity, laid out in texts like Business Intelligence, Analytics and Data Science and mirrored in the process guides that firms like Demand Metric put in front of marketers. The academic version and the practitioner version agree on the order, which is rare enough that it is worth trusting. Describe, then predict, then prescribe. In that order, always.

What each rung actually costs, side by side

The reason knowing your rung saves money is that each rung has a very different price, and vendors are motivated to sell you the expensive one whether you need it or not. Here is the same idea as the map above, but from the money angle.

Rung	Main cost driver	Rough effort	When it pays off
Descriptive	Plumbing and agreeing definitions	Weeks	Almost always, and first
Predictive	Clean history plus modelling skill	Weeks to months	Once a future number changes a real decision
Prescriptive	Rules, guardrails, and organisational trust	Months, plus nerve	Only once the two below it are solid

Here is the second thing worth sitting with. The least expensive rung is usually the one that changes the most decisions, and the most expensive rung only pays off when the two below it are already solid. So the money saving move is almost never to buy higher. It is to build lower first, and climb only when the rung beneath you holds your weight.

A worked mini case, end to end

Let me thread one real shaped story through all three rungs, because the jumps between them are where the lesson lives. An online shop came to me certain they needed an AI that would set prices for them. That is a top rung wish. Before touching it, we asked the bottom question: what happened. Turns out they could not answer it. Revenue in the shop admin did not match the bank, because refunds and marketplace fees were counted in three different ways. So we spent the first fortnight on description, nothing clever, just one honest source of truth. That alone found a bundle of products sold at a loss once real fees were counted, and fixing that made more money than any pricing AI would have in a year.

Only then did the middle question earn its place. With clean history we could forecast demand per product and spot which lines were seasonal. And only then, on that base, did a modest pricing rule make sense, and even then we ran it as a suggestion a human approved for a month before letting it act. Here is the shape of it.

Question asked	Rung	What we actually did	Outcome
What happened	Descriptive	One honest revenue source, real fees included	Found loss making lines, fixed pricing by hand
What will happen	Predictive	Demand forecast per product from clean history	Ordered to a number, cut dead stock
What should I do	Prescriptive	A price rule, run as a suggestion first	Small steady margin gain, trusted because tested

The AI they walked in asking for was the last and smallest part, and it only worked because we refused to start there. That is the ladder doing its job.

Common mistakes I see

A short field guide to the ways this goes wrong, so you can spot yourself in it.

Buying the top rung first. The classic. A prescriptive purchase bolted onto a business that cannot describe its own last quarter. It automates a guess.

Confusing a dashboard with a decision. A wall of charts nobody acts on is descriptive theatre, not descriptive analytics. If no Monday behaviour changes because of it, it is decoration.

Trusting a forecast with no error number. A prediction without a measured miss is not a forecast, it is a vibe. Always ask how wrong it usually is before you bet on it.

Letting a system act before you have watched it. Prescription that goes live on day one, with no month of running as a suggestion first, is how one bad recommendation burns all the trust at once.

Never agreeing what a customer is. This one sits under all the others. If the definitions disagree across your systems, every rung above inherits the confusion, and the fancier the tool the more expensively it is confused.

How to apply this on Monday morning

You do not need a project or a budget to start. You need an afternoon and some honesty. Here is the sequence I would run.

First, write down the actual question you have, in plain words, as a sentence. Not a tool, a question. Then decide out loud whether it is a what happened, a what will happen, or a what should I do. That single act of naming does most of the work.

Second, sanity check the rung beneath it. If your question is predictive, can you already describe cleanly. If it is prescriptive, do you already trust a prediction. If the rung beneath is missing, that is your real project, not the one you came in wanting.

Third, build the smallest honest version of the rung you actually need and stop there. One trustworthy dashboard beats ten speculative ones. One forecast on one decision beats a model zoo. Prove the rung holds your weight before you climb, and let the next rung wait until this one is boring and reliable.

How to place yourself on the ladder

You do not need a consultant to work out where you are. Three honest questions will do it.

Can you, right now, pull last quarter's real numbers, the ones you would bet on, in an afternoon, from one place, with definitions everyone agrees on. If not, you are below the descriptive rung, and that is where every pound should go until you are standing on it. This is not a failure. It is the normal state of a business that grew faster than its reporting did.

If you can describe cleanly, do you have enough clean history to learn a pattern from, and a real decision that a prediction would actually change. A forecast nobody acts on is a hobby. Predictive earns its cost only when a number about the future changes what you do this week.

And only if both of those are solid does the prescriptive question even make sense. Do you trust your predictions enough to let a system act on them, with guardrails, on decisions that matter. If yes, that is where the largest gains live, and it is worth every penny. If you are not there yet, a prescriptive purchase is money spent to skip the queue, and the queue does not let you skip.

What this means when the vendor calls

Next time a beautiful deck lands promising an AI that will decide for you, put it on the ladder before you put it on the budget. Ask which question it answers. If it is prescriptive and you cannot yet describe your own last quarter cleanly, you have found not a solution but a very expensive way to automate a guess. Send it back and go build the bottom rung, which nobody will try to sell you because there is no margin in honesty, and which will change more of your decisions than anything above it.

This diagnosis, working out which rung you are actually on and what it would take to climb the next one safely, is a lot of what I do with owners before a single line of code gets written. Sometimes the answer is a modest dashboard project that pays for itself in a fortnight. Sometimes it genuinely is a predictive model, because the description is already solid and there is a real decision waiting on it. It is almost never the thing the last vendor tried to sell. If you want a straight answer about which rung your business is on, and what the next one is actually worth to you, that is exactly the conversation my data and analytics work is built around, and the fastest way to start it is to book a call.

The one idea to leave with

Three questions, one order. What happened, what will happen, what should I do. Descriptive, predictive, prescriptive, and never out of sequence. The value climbs as you go up, but so does the cost and the fragility, and you cannot optimise what you cannot yet reliably describe. So before you spend a penny, name the question you actually have, find its rung, and be honest about whether the rung beneath it will hold you. Do that and you will stop paying top rung prices for problems the bottom rung would have solved, which is most of them, most of the time.