What data science actually delivers for growth

Data science is a way of making better decisions, not a department

Ask ten people what data science is and you get ten answers, most involving "AI" and a slide with a brain on it. The definition I trust is plainer. The US National Institute of Standards and Technology describes data science as "the methodology for the synthesis of useful knowledge directly from data through a process of discovery or of hypothesis formulation and hypothesis testing".¹

Notice what is missing. No dashboards, no algorithms, no department. A method for turning data into knowledge you can use.

My own version, for a managing director, is shorter still. Data science is the habit of making a business decision a little better than you would have made it without the data, and then checking that it really was better. Everything else is tooling.

So what does it deliver? Better decisions. Not a prettier report, not a model that sits in a notebook. If no decision changes, nothing was delivered, however clever the maths.

In this article I show what that looks like in practice: the loop every useful project follows, a simple way to put a pound value on a better decision, and a worked example at an invented garden centre chain in Brighton that runs three small models. Where an idea has its own deep article, I point you to it rather than repeat it.

An example: Sam runs Downland Garden Co, three garden centres in Brighton, Hove and Shoreham. Sam is invented, the business is invented, and every number in this article is made up so I can show every step. Sam has a till system, a loyalty club and a spreadsheet of supplier orders, no data scientist, and no wish for "AI". Sam wants to stop throwing away bedding plants every June. That is a decision problem, and it is where data science starts.

What the evidence says about UK businesses and data

Before the method, a look at the field. How many UK businesses actually use data to decide things?

The best recent figures come from the Department for Science, Innovation and Technology, which surveyed 3,796 UK businesses between December 2024 and February 2025.² Of the businesses that handle digital data, 46% agreed that data processing or analysis supports or informs their decisions. That is less than half, among the businesses that hold data in the first place.

The same study found a clear gap in what businesses report. Among businesses that analyse their data, 47% said it had made their processes more efficient or saved costs, against 14% of those that do not analyse it.² The report is careful to say it does not establish causality, and so should I be. Businesses that analyse data may simply be better run in other ways too.

The government's National Data Strategy makes the same point at a higher level. It describes an association between efficiency, productivity and businesses that work from data, and it is careful to call it that.³ An association, not a proof. I find that honesty reassuring. Anyone who tells you data science guarantees growth is selling something.

Worked example: what 46% means in a high street

Picture 100 UK businesses on one long high street. The study found that 83% of businesses handle digital data, so 83 of them have data at all.²

Of those 83, 46% say analysis informs their decisions, and $83 \times 0.46 = 38.2$ . So roughly 38 of the 100 businesses use data to inform decisions, by their own account. The other 62 either hold no digital data or hold it and decide without it. Sam's garden centre is one of the 45 in the middle group: plenty of data in the till, none of it in the ordering decision.

That is the opportunity. Sam does not lack data. The data and the decision have simply never met.

The loop: question, data, model, decision, measure

Every useful data science project I have worked on follows the same five steps, in the same order. When a project fails, it is usually because one of the steps was skipped, most often the first or the last.

The question comes first, and it is always a decision someone has to make. The data comes second, chosen because it bears on that decision, not because it happens to exist. The model is third, and it is the smallest thing that answers the question. The decision is the step that earns the money, which is why it is highlighted. The measure closes the loop, so you know whether the new way beat the old way. Then you go round again.

A report describes the past. A loop changes the next decision and learns from it.

Worked example: one pass of the loop at Downland Garden Co

Sam's question: how many trays of bedding plants should each centre order for the four weeks around the May bank holidays?

The data: three years of till sales by product and by day, the order sheets, the write off log (plants thrown away unsold), and the daily weather for Brighton, which is free to download.

The model: a forecast of weekly demand per centre, using last year's sales, the dates of Easter and the bank holidays, and whether the weekend was dry.

The decision: Sam orders to the forecast, plus a small buffer, instead of "last year plus 10%, because it felt busy".

The measure: at the end of June, Sam compares the write off and the sold out weekends with the year before. I put numbers on this later in the article.

Start with the question, and make it a decision

The most common failure I see is a project that starts with the data. "There is all this loyalty card data, what can it do for us?" is a fishing trip, and the usual catch is a dashboard nobody opens after the second week.

A good question names a decision, a person who makes it, and a date by which it is made. "What are our customers like?" is not a question in this sense. "Which loyalty members should get the spring voucher, decided by 1 March?" is.

I wrote a whole article on this, start with the decision, not the dashboard, and a shorter one on the three questions your data should answer. If you read only one more thing after this, read the first.

An example: Sam's first instinct was to ask for "a loyalty dashboard". I asked what Sam would do differently on Monday morning if the dashboard existed. The honest answer was "look at it". So we rewrote the question: "Which of our 18,000 members are about to stop visiting, and is a £5 voucher worth sending them?" That question has a decision (send or do not send), an owner (Sam), a cost (£5 a voucher) and a date (before the spring rush). It can be answered, and the answer can be checked.

The data is usually already there, and messier than you think

Most growing businesses already hold the data for their first useful model. The till, the online shop, the email platform and the accounts package between them record nearly everything that matters. What they rarely do is agree with each other.

In my experience the first weeks of any project go on making the numbers trustworthy, not on modelling. Product codes change, a centre closes for a refit and its zero sales look like a collapse, a promotion was never recorded. I describe that stage in detail in the first data science month, and there is a quick check you can run yourself in is your dashboard lying to you.

The other data question is volume. How much history do you need before a model is worth building? Less than vendors suggest for some questions, more than founders hope for others. The numbers are in how much data do you actually need.

An example: Sam's till data says Hove sold no bedding plants for three weeks in April two years ago. A model would read that as a terrible April. In fact the Hove greenhouse was being reglazed and the plants were sold from Brighton instead. One line in a notes column fixes it: mark those three weeks as "closed", and move nothing. Without that line, the forecast would have under ordered for Hove every April since. The fix took ten minutes. Finding it took a conversation with the Hove manager, which is why the data step always includes talking to people.

The model is the smallest thing that answers the question

"Model" sounds grand. A model is any calculation that turns the data you have into an estimate you need. A seasonal average is a model. A rule that says "members who have not visited for 90 days are at risk" is a model. A gradient boosted forest with 400 features is also a model, and it is very rarely the first one you need.

I keep an article titled counting and summing beats machine learning, most of the time for exactly this reason. The simple model is quicker to build, easier to explain to the person who has to act on it, and easier to check when it goes wrong. You move to something heavier only when the simple one has been measured and found wanting. If the vocabulary is muddy, AI vs machine learning sorts out which word means what.

An example: for Sam's loyalty question, the first model is a table. Each member gets one row with three numbers: days since the last visit, visits in the last year, and spend in the last year. Members are sorted by days since the last visit. Anyone over 120 days who visited at least four times the year before is flagged. That is a model. It took an afternoon. Later, when it has proved its worth, it can be replaced by a proper survival model, but not before it has paid for that upgrade.

The value of a decision improvement

Here is the one formula in this article, and it is the one I use to decide whether a data science project is worth doing at all. The value of a model is not its accuracy. It is how much better the decision becomes, minus what the model cost.

\Delta V = \sum_{i} p_i \, v_i(a_{\text{new}}) - \sum_{i} p_i \, v_i(a_{\text{old}}) - C

Here $\Delta V$ is the value of the decision improvement, in pounds. The index $i$ runs over the possible outcomes, for example a quiet season, a normal one and a busy one. $p_i$ is the probability of outcome $i$ . $v_i(a)$ is the profit you make from action $a$ if outcome $i$ happens. $a_{\text{new}}$ is the decision you make with the model, and $a_{\text{old}}$ is the decision you would have made without it. Each sum, $\sum_i p_i \, v_i(a)$ , is the expected profit of that action, the long run average over all the outcomes weighted by how likely they are. $C$ is the cost of building and running the model. Expected value is explained from scratch in Foundations: expected value, and probability in Foundations: probability.

Two things follow from the formula, and they matter more than the formula itself. First, if the model does not change the decision, $a_{\text{new}}$ equals $a_{\text{old}}$ , the two sums cancel, and the value is minus $C$ . A model that confirms what you were going to do anyway is a cost. Second, the size of the prize depends on the size of the decision. A brilliant model on a small decision is worth less than a rough one on a big decision.

The deeper treatment of choosing under uncertainty is in decision theory for founders.

Worked example: Christmas trees in Shoreham

Every year Downland Garden Co orders Christmas trees for the Shoreham centre. Each tree costs Sam £20 and sells for £50. A tree unsold on Christmas Eve is worth nothing. The old rule is simple: order 2,000, "because running out looks bad".

A simple forecast from five years of sales gives three outcomes for demand this year:

Quiet: 1,400 trees, probability 0.3
Normal: 1,700 trees, probability 0.5
Busy: 2,000 trees, probability 0.2

Step one, the profit from the old decision, ordering 2,000, in each outcome. The order costs $2{,}000 \times 20 = 40{,}000$ pounds.

Quiet: $1{,}400 \times 50 - 40{,}000 = 30{,}000$ , so £30,000 profit
Normal: $1{,}700 \times 50 - 40{,}000 = 45{,}000$ , so £45,000
Busy: $2{,}000 \times 50 - 40{,}000 = 60{,}000$ , so £60,000

The expected profit of the old decision weights each by its probability: $0.3 \times 30{,}000 + 0.5 \times 45{,}000 + 0.2 \times 60{,}000 = 9{,}000 + 22{,}500 + 12{,}000 = 43{,}500$ , so £43,500.

Step two, the decision the forecast points to: order 1,700. The order costs $1{,}700 \times 20 = 34{,}000$ pounds. In a quiet year Sam sells 1,400, in a normal or busy year Sam sells all 1,700.

Quiet: $1{,}400 \times 50 - 34{,}000 = 36{,}000$ , so £36,000
Normal: $1{,}700 \times 50 - 34{,}000 = 51{,}000$ , so £51,000
Busy: $1{,}700 \times 50 - 34{,}000 = 51{,}000$ , so £51,000

The expected profit of the new decision is $0.3 \times 36{,}000 + 0.5 \times 51{,}000 + 0.2 \times 51{,}000 = 10{,}800 + 25{,}500 + 10{,}200 = 46{,}500$ , so £46,500.

Step three, subtract the cost. Say the forecast takes a day to build and check each autumn, and costs £1,000. Then $\Delta V = 46{,}500 - 43{,}500 - 1{,}000 = 2{,}000$ .

So the forecast is worth about £2,000 a year at one centre, on one product. Worth doing, and worth knowing before anyone buys a £20,000 forecasting platform for Christmas trees. Notice also what happens in a busy year: Sam runs out of trees and earns £9,000 less than under the old rule. Expected value is the long run average. It says nothing about the one year you cannot afford to get wrong, and Sam is entitled to keep a buffer if a sold out December would genuinely hurt the brand.

Measure, or the loop never closes

The last step is the one most often skipped, and it is the one that tells you whether any of this worked. Without it you have a model and a feeling.

Measuring properly means comparing the new way of deciding with the old way, on the same kind of customers or the same kind of weeks, and being honest about chance. The cleanest way is a test with a control group: some customers or some weeks get the new decision, the rest keep the old one, and you compare. When a clean test is impossible, there are other designs, each with its own catch. I cover them in five ways to prove a campaign actually worked and, with the sample size maths, in statistical testing: sample size, lift and full factorial designs.

An example: when Sam sends the £5 voucher to the members the model flags, a randomly chosen 400 of them get nothing. That is the control group. Twelve weeks later Sam compares how many in each group came back. Without the 400, Sam would credit the voucher for members who would have come back for the spring compost anyway. With them, the difference between the two groups is the voucher's real effect. The control costs nothing, it even saves 400 vouchers.

Worked example: three small models at a Brighton garden centre chain

Now the whole picture. Downland Garden Co turns over about £4.2 million a year across its three centres. Over one year Sam runs three small models, one at a time, each through the full loop. All numbers are invented.

Model one: a demand forecast for spring bedding plants

The question: how many trays of bedding plants to order per centre for the spring season.

The starting point: Sam buys £300,000 of bedding plants at cost each spring. Last spring 18% was written off unsold, and $300{,}000 \times 0.18 = 54{,}000$ , so £54,000 went in the skip.

On top of that, the most popular lines sold out on two busy weekends. Sam's managers estimate the lost margin at £20,000.

The model: a weekly forecast per centre and per product group, built from three years of till data, the Easter and bank holiday dates, and the weather (the full method, with honest error bands, is in forecasting demand: autocorrelation, seasonality and honest error bands).

The decision: order to the forecast, centre by centre, with a buffer on the bestsellers.

The measure: the next spring, write off falls to 11%, and $300{,}000 \times 0.11 = 33{,}000$ , so £33,000.

That is a saving of $54{,}000 - 33{,}000 = 21{,}000$ , so £21,000. Sold out weekends halve, recovering about half the lost margin, $20{,}000 \times 0.5 = 10{,}000$ , so £10,000. Together, $21{,}000 + 10{,}000 = 31{,}000$ : the forecast is worth £31,000 a year.

One caveat: a kind spring flatters any forecast. Sam keeps measuring for a second year before calling it settled.

Model two: which loyalty members are about to lapse

The question: which members are about to stop visiting, and is a £5 voucher worth sending them.

The starting point: the club has 18,000 members. Each year about 30% lapse, meaning they do not visit for 12 months, and $18{,}000 \times 0.30 = 5{,}400$ members.

An active member brings Sam about £60 of margin a year.

The model: the table from earlier, members ranked by days since the last visit and their visits the year before. The 2,400 at highest risk are flagged. (The proper tool for this, once the simple version has proved itself, is survival analysis, covered in survival analysis, churn and customer lifetime value.)

The decision: send the voucher to 2,000 of the flagged members, chosen at random, and hold back the other 400 as the control.

The measure: in the control group, 55% lapse. Among members who got the voucher, 45% lapse. The voucher saved 10 in every 100 from lapsing. Across the 2,000 who got it, $2{,}000 \times (0.55 - 0.45) = 200$ members kept.

At £60 of margin each, $200 \times 60 = 12{,}000$ , so £12,000.

Now the cost. Of the 2,000, 55% come back, which is 1,100 members. About 600 of them redeem the voucher, many of whom would have come back anyway, and $600 \times 5 = 3{,}000$ , so the vouchers cost £3,000.

The net effect is $12{,}000 - 3{,}000 = 9{,}000$ , so £9,000. Smaller than the forecast, and that is fine. It also teaches Sam something worth more than the £9,000: who the lapsing members are, which feeds the segmentation work later (see segmentation: strategy before algorithms).

Model three: a price test on compost

The question: is the 50 litre bag of compost underpriced?

The starting point: the bag sells at £7.99 and costs Sam £4.50, so each one earns $7.99 - 4.50 = 3.49$ , that is £3.49. Sam sells about 40,000 bags a year, and $40{,}000 \times 3.49 = 139{,}600$ , so the compost earns £139,600 of margin.

The model: a test, not a guess. For eight weeks the price alternates week by week between £7.99 and £8.49 at all three centres, so weather and season hit both prices alike. In the higher priced weeks, unit sales are 3% lower.

The price rose by $\frac{8.49 - 7.99}{7.99} = 6.3\%$ , and units fell by 3%, so the price elasticity is about

\varepsilon = \frac{-3\%}{6.3\%} \approx -0.48

Here $\varepsilon$ is the price elasticity of demand, the percentage change in units sold divided by the percentage change in price. A value between 0 and minus 1 means demand is inelastic: customers barely notice. The idea is built from scratch in Foundations: elasticity, and the regression method for real data is in modelling demand and price elasticity with regression.

The decision: move the bag to £8.49.

The measure: at the new price Sam sells about $40{,}000 \times 0.97 = 38{,}800$ bags, each earning $8.49 - 4.50 = 3.99$ , that is £3.99. And $38{,}800 \times 3.99 = 154{,}812$ , so £154,812 of margin.

The gain is $154{,}812 - 139{,}600 = 15{,}212$ , that is £15,212, call it £15,200. The caveat: an eight week test in spring says nothing about the autumn, and a competitor's price change could move the elasticity. Sam reruns the test the next year. For the plain English version of why one elasticity never fits every price, read price elasticity in plain English.

What the three models delivered together

The horizontal axis shows the three models. The vertical axis shows the money effect of each in pounds over one year, after the direct costs of acting on it (such as the vouchers) but before the cost of building the models. Every number is invented for this example. The takeaway is the order of the bars: the biggest win came from the biggest decision, the spring order worth £300,000, not from the cleverest model. The loyalty model was the most interesting to build and earned the least.

Together the three are worth $31{,}000 + 15{,}200 + 9{,}000 = 55{,}200$ , so £55,200 a year.

Suppose cleaning the data and building the three models costs £18,000 in the first year. Sam is left with $55{,}200 - 18{,}000 = 37{,}200$ , so £37,200, and most of the build cost does not recur. On £4.2 million of turnover, £55,200 is about 1.3%. That is what data science usually delivers for a business of this size: not a transformation, but a steady, measured improvement in decisions that were already being made, year after year. In my experience that is exactly why it compounds. Each pass of the loop makes the next one quicker.

What data science does not deliver

The hype is rarely plain about the limits, so here they are.

It does not deliver certainty. Every forecast has an error band, and an honest model says how wide it is. It does not tell you what caused what unless you design for that, with a test or a careful method (a correlation in last year's data is not a reason, see causal inference in marketing). It does not replace judgment. It gives judgment better material.

It also does not need a platform to start. Real time bid feeds, attribution across every channel and automated triggers are real and useful, but they are the second or third pass of the loop, not the first. Bidding on value is covered in bid on lifetime value, not cost per click, and the question of which channel earns the next pound in media mix modelling.

An example: after the compost result, Sam asks whether to raise every price by 6%. The data says nothing about that. The test covered one product, at one price step, in one season. Compost is a staple people buy by the boot load. A £40 ornamental pot is a different purchase with different rivals. The honest answer is "test the next category before you move it", which is less exciting than "the model says raise prices", and much less likely to empty the car park.

Who does the work, and when you need a data scientist

Does Sam now need to hire one? Probably not yet. The three models above need someone who can clean data, count carefully, run a fair test and explain the result. That can be a capable analyst, a part time specialist, or a curious owner with help.

I wrote you probably do not need a data scientist yet for exactly this moment, and data engineer, data scientist, data analyst: who does what if you want the roles untangled.

An example: Sam's rule of thumb, which I would sign up to, is to add a person only when the loop is turning faster than the current people can keep up. Three models in a year do not need a full time hire. In year three, with weekly forecasts, a survival model and price tests in six categories, the sums may differ. The same formula decides it: the value of the decision improvement, minus the cost.

Where to go deeper

This article is the map. Each deep article takes one step of the loop, or one kind of model, and builds it from scratch:

The question: start with the decision, not the dashboard
The data: the first data science month
Forecasting: forecasting demand: autocorrelation, seasonality and honest error bands
Churn and lifetime value: survival analysis, churn and customer lifetime value
Segments: segmentation: strategy before algorithms
Marketing budgets: media mix modelling: adstock, saturation and where the next pound goes
Pricing: modelling demand and price elasticity with regression
Hiring: you probably do not need a data scientist yet

And underneath all of them, the Foundations page, which explains every formula from the ground up.

NIST Big Data Public Working Group, NIST Big Data Interoperability Framework: Volume 1, Definitions, Version 3 (NIST Special Publication 1500-1r2), National Institute of Standards and Technology, October 2019, https://doi.org/10.6028/NIST.SP.1500-1r2 ↩
Department for Science, Innovation and Technology and Ipsos UK, Business data use and productivity study (wave 2): statistical report, GOV.UK, updated January 2026, https://www.gov.uk/government/publications/business-data-use-and-productivity-study-wave-2/business-data-use-and-productivity-study-wave-2-statistical-report ↩ ↩² ↩³
Department for Digital, Culture, Media and Sport, National Data Strategy, GOV.UK, updated December 2020, https://www.gov.uk/government/publications/uk-national-data-strategy/national-data-strategy ↩