Data engineer, data scientist, data analyst: who does what

TL;DR

Difficulty: Beginner

A data engineer builds the pipelines that collect, clean and join your data in one place. A data analyst uses that data to explain what happened and why. A data scientist predicts what happens next and designs experiments that show which action pays. The order matters: without clean data the analysis is wrong, and without analysis a model has nothing to stand on. In the UK the analyst is the least expensive hire and the scientist and engineer sit close together at the top. Most businesses need the engineer's work and an analyst first, and a scientist only once prediction and testing are what holds growth back.

Data engineer, data analyst and data scientist sound like three names for one job. They are three different jobs, and hiring the wrong one first is an expensive mistake. I follow one invented coffee subscription in Leeds through all three roles, with a worked experiment that shows why a free bag of coffee beat a discount, then set out what each role costs to hire in the UK and the order to bring them in.

Three job titles, three different jobs

Every business now records more than it reads. Each click on the website, each email opened, each abandoned basket, each support ticket and each payment lands in some system. A pile of records is about as useful as a pile of bricks. It becomes valuable only when someone turns it into a decision.

The confusion starts with the job titles. Data engineer, data analyst and data scientist sound like three names for one job. They are three different jobs, with different skills, different tools and a different question at the centre of each. Hire the wrong one first and you pay a senior salary for someone who spends months doing work they were not hired for.

This article follows one invented business through all three roles, so you can see exactly who touches what. It then shows what each role costs to hire in the UK, with sources, the order in which to bring them in, and what goes wrong when a business skips them. If you already know the roles and want fifteen concrete projects with figures, the companion piece is data roles in practice: how to grow SaaS and ecommerce with data.

The example: a coffee subscription in Leeds

Morning Ritual (morningritual.example) is an invented coffee subscription business in Leeds. Customers choose a blend and a rhythm, and a box of beans arrives every two weeks for £16. The business has about 4,000 subscribers, a website, an app, a newsletter and some paid social campaigns.

Growth has gone flat. Roughly as many subscribers leave each month as join. The founder, Hannah, is sure there is money hiding in the business, she just cannot see where. Her question is simple to ask and hard to answer: which customers are about to cancel, and what would keep them?

That one question travels through all three roles. Each person touches it in a completely different way.

Tap or click to open full screen and move around

The data engineer builds the plumbing

Right now Morning Ritual's data is scattered, and nobody did anything wrong. The customer list lives in the shop platform. Payment history sits with the payment provider. Opens and clicks are inside the newsletter tool. App activity is in another system, website behaviour in the analytics tool. None of them talk to each other.

So a question like "show me every subscriber who opened the last three emails but has not had a box in six weeks" has no answer. The information exists, but it is locked in five boxes that were never meant to be combined.

That is the data engineer's job. The engineer builds the pipelines that pull data out of each system, clean it and load it into one central store, usually called a data warehouse. When an order comes in at 3am, it flows into the warehouse without anyone touching it. The subscriber in the email tool and the subscriber in the payment system are matched to the same person. The dull but critical problems are the engineer's: duplicate records, missing fields, dates stored in three formats, a sync that fails at midnight. The engineer also decides what is collected at all, so the pipelines carry only the data the consent banner allows, which is covered in event tracking with consent mode.

Think of a restaurant. The engineer runs the plumbing and the electricity. They do not cook and they do not write the menu, but without clean water and working power nobody in the kitchen can do anything. When the engineer has done the job well you never think about them. When they have not, nothing else works.

For Morning Ritual, the engineer builds one place where every subscriber's whole story sits together: orders, payments, emails, app taps and delivery dates, refreshed automatically and trustworthy. Nothing clever has happened yet. But for the first time, the clever things are possible. How that warehouse is built, step by step, is in data warehousing for marketing automation.

The data analyst tells you what is going on

With the data clean and in one place, the analyst steps in. The analyst answers business questions with the data that already exists, and presents the answer so a person can act on it.

At Morning Ritual the analyst starts on the churn question, and within a few days comes back with a pattern. Of the subscribers whose second box arrived more than 30 days after sign up, 17% cancelled within 90 days. Of those whose second box arrived on time, 6% did. That is nearly three times the rate. Subscribers who ever contacted support about a late delivery were also far more likely to leave, even when the problem was fixed.

Notice what the analyst did not do. They did not predict anything. They looked at what already happened and found the story in it. A pattern like this is a correlation, not yet proof of cause: perhaps the late second box causes the cancellations, or perhaps the subscribers who signed up during a busy fortnight were less committed to begin with. The difference between the two is explained in the Foundations section on correlation. A good analyst says which it is, or says that nobody knows yet.

The analyst also builds the dashboard Hannah now opens every Monday: new subscribers, cancellations, revenue per subscriber, the most popular blends. Before, she ran the business on gut feeling. Now she has an honest picture, updated daily.

Back in the restaurant, the analyst reads the till receipts at the end of the night and tells you the truth: the specials sold out, nobody orders the fish, and Tuesdays are dead. Priceless, and all of it about the past.

For most small and medium businesses, a good analyst on top of clean data is the hire with the highest return. Many companies make large decisions without anyone turning their data into plain English, and an analyst closes that gap faster than any other role.

The data scientist predicts and experiments

The analyst has said that late second boxes and support complaints go with cancellations. Hannah wants the next step. She does not only want to know who cancelled last month. She wants to know who is about to cancel next month, while there is still time to act.

That is the data scientist's job. The scientist builds a model that learns from the subscribers who left in the past and gives every current subscriber a probability of leaving. Instead of reacting after a cancellation, Morning Ritual gets a list each morning of the subscribers most at risk this week, ranked. The methods behind such a model, from survival curves to lifetime value, are in survival analysis, churn and customer lifetime value.

Prediction is half the job. The other half is finding out what actually keeps people, and that needs an experiment. The scientist takes 600 at risk subscribers and splits them at random into three groups of 200. One group gets 20% off their next three boxes. One group gets a free bag of a new blend. One group gets nothing, which is the comparison. Eight weeks later the scientist counts who is still subscribed.

The only number that matters is the difference between a group that got an offer and the group that got nothing:

Δ=rt−rc\Delta = r_t - r_c

Δ\Delta is the incremental retention, rtr_t the share of the treated group still subscribed at the end, and rcr_c the share of the control group still subscribed. Whether a difference is real or luck is a hypothesis test, and how many customers you need for it is in sample size.

The offer is worth sending only if the extra subscribers it keeps earn more than the offer costs across everyone who received it:

V=n (Δ m−c)V = n\,(\Delta\,m - c)

VV is the net value of the offer, nn the number of subscribers who received it, mm the contribution margin a retained subscriber brings over the period you care about, and cc the cost of the offer per subscriber who received it. The margin logic is in the Foundations section on lifetime value.

A worked example: the discount that loses to a free bag

After eight weeks, 60% of the control group is still subscribed, 72% of the discount group and 70% of the free bag group. A retained subscriber brings £6 of contribution per box, 13 boxes over six months, so mm is £78. The discount costs £8.40 per subscriber who received it, the free bag £5 with postage.

Vdiscount=200×(0.12×78−8.40)=192V_{\text{discount}} = 200 \times (0.12 \times 78 - 8.40) = 192

For the discount, Δ\Delta is 0.72 minus 0.60, so 0.12, and the offer earns £192 over six months. For the free bag, Δ\Delta is 0.70 minus 0.60, so 0.10:

Vbag=200×(0.10×78−5)=560V_{\text{bag}} = 200 \times (0.10 \times 78 - 5) = 560

So VV is £560 for the bag against £192 for the discount. The discount kept two more people per hundred, yet the bag made almost three times as much. The reason is the 120 subscribers in each group of 200 who would have stayed anyway. The discount handed £8.40 to every one of them. The bag cost less to give to people who did not need it.

With 200 per group, both offers beat doing nothing by a clear margin. The two point gap between discount and bag is well inside the noise, so the scientist does not claim one kept more people than the other, only that the bag is the better use of money. The arithmetic of how much data a test like this needs is in how much data do you actually need.

The scientist might also build the recommendation that suggests a subscriber's next blend, or forecast how much green coffee to buy next quarter so the roastery is neither empty nor overflowing. In the restaurant, the scientist is the one who studies years of bookings, weather and fixtures and tells you to order 40% more steak for Saturday because there is a rugby match at Headingley. Not what happened, but what is likely to happen, and what to do about it.

The simple way to remember the difference

The data engineer builds and runs the systems that collect, clean and store the data. Their question is: can we get to this information reliably?

The data analyst uses that data to explain what happened and what is happening now. Their question is: what is going on, and why?

The data scientist uses that data to predict what happens next and to test what to do about it. Their question is: what is likely to happen, and which move pays?

The engineer builds the road. The analyst reads the map and tells you where you are. The scientist looks ahead and tells you where to drive. The map is no use without the road, and the forecast means nothing without the map. The order matters.

In a small company the lines blur. One person may build a pipeline in the morning and a dashboard in the afternoon, and a good data scientist can do some of both. The titles still matter, because they tell you which of the three questions a person is best at answering.

What each role costs to hire in the UK

Two public sources give a fair picture of UK pay, and they agree on the order. ITJobsWatch tracks permanent IT job adverts across the UK. In the six months to 10 November 2026 the median advertised salary was £80,000 for a data scientist, £70,000 for a data engineer and £57,500 for a data analyst.1 The National Careers Service, run for the Department for Education, gives £32,000 for a starting data scientist rising to £83,000 with experience, and £28,000 rising to £65,000 for a data analyst and statistician.23

Read the advertised medians with care. They skew towards London and towards senior roles, and they move with the mix of adverts: the analyst median was £42,500 in the same six months of 2025, and nobody's pay rose by a third in a year.

A bar chart of median advertised salaries for permanent UK jobs in the six months to 10 November 2026: data analyst 57,500 pounds, data engineer 70,000 pounds, data scientist 80,000 pounds. Source ITJobsWatch. Tap or click to open full screen and move around

The pattern is the same in every source: the analyst is the least expensive of the three, and the scientist and the engineer sit close together at the top, with the scientist usually a step ahead. Add employer's National Insurance and pension contributions on top of the salary before you compare a hire with outside help.

Why this matters for growth

Growth work is a loop. You have an idea, you test it quickly, you measure honestly, you keep what works and you drop what does not. That loop runs on data. Take the data away and you are guessing with confidence.

Most ideas do not survive the measuring. When Microsoft evaluated ideas with controlled experiments, only about a third improved the metric they were designed to improve. About a third made no significant difference, and about a third made things worse.4 A business that ships every idea without a test ships the harmful third along with the good one, and never learns which was which.

Each role keeps the loop honest. The engineer makes sure an experiment can be measured without a week of spreadsheet work. The analyst says whether it moved the number, and spots the side effects. The scientist designs the test so luck cannot fool you, as in the discount example above. Without them, a growth programme turns into opinion: the loudest person wins, and three months later nobody can say whether anything worked.

Why this matters for product development

When a developer ships a feature, the real question is whether anyone uses it and whether it helps the business. Without data roles that question never gets a proper answer, and a team keeps building things nobody wants.

The engineer makes sure features are instrumented, which means the product records what people do with them. The analyst reports whether the new checkout raised conversions or quietly lowered them. The scientist runs a controlled rollout, showing the new version to some users and the old one to the rest, so you know the change caused the improvement and did not simply coincide with a sunny week.

Why this matters for digital marketing

Marketing is where the cost of guessing shows up fastest, because the money goes out every day.

Picture Morning Ritual advertising on three platforms. Each platform's dashboard takes credit for the same sales, and Hannah cannot tell which channel earns money and which burns it. The engineer brings ad spend and revenue into one place, so the channels can be compared on the same terms. The analyst works out what a subscriber costs to win from each channel and what that subscriber is worth over their lifetime, not just their first box. The scientist predicts which new visitors are likely to become long term subscribers, so the budget chases the people who stay rather than the lowest cost clicks. The full method, channel by channel, is in bid on lifetime value, not cost per click.

What goes wrong without them

The costs are concrete. Decisions get made on numbers that are simply wrong, because nobody built the plumbing and orders were counted twice. Money keeps flowing to campaigns that do not work, because the true picture is split across five tools. Subscribers who could have been kept are gone before anyone notices. Features nobody uses keep getting built. A competitor who runs the measure and learn loop pulls ahead while meetings argue over opinions.

The subtler danger is hiring the wrong role. A business with no clean data and no habit of analysis hires a data scientist, the most expensive of the three. The scientist spends the first months doing the engineer's and the analyst's jobs, delivers little of what they were hired for, and leaves. The business concludes that data does not work for a company like theirs. It was never the data. It was the order. The ways data projects fail before they start, this one included, are listed in 16 ways data projects fail before they start.

The order to hire in

Most businesses do not need a team of three on day one. They need clean data they can trust and someone turning it into plain answers.

Start with the foundation. An engineer, or an engineer's work done once by someone outside, so the data lands in one place reliably. Then an analyst, so every important decision rests on a number you believe. Bring in the scientist when prediction and experiments are what stands between you and the next stage of growth, for example when you have enough subscribers that a churn model and a proper test like the one above pay for themselves.

What the first month of such work looks like in practice, before any model, is in the first data science month. If you would rather not hire three people to find out what your data can do, I do this as data science work: the plumbing, the analysis and the models, for as long as you need them. Your raw data stays yours, and the terms for code and models are in the analytics and data science terms.

Not sure which data role your business needs first? Bring your data to a free workshop and I will tell you what it can answer today and what it would take to answer the rest, and the first two weeks of work are on me.

Sources

  1. 1.

    ITJobsWatch, job trends for Data Scientist, Data Engineer and Data Analyst, UK, median annual salary of permanent job adverts, six months to 10 November 2026 with the same periods of 2025 and 2024, https://www.itjobswatch.co.uk/jobs/uk/data%20scientist.do, https://www.itjobswatch.co.uk/jobs/uk/data%20engineer.do and https://www.itjobswatch.co.uk/jobs/uk/data%20analyst.do

  2. 2.

    National Careers Service, Data scientist job profile, average salary a year, https://nationalcareers.service.gov.uk/job-profiles/data-scientist

  3. 3.

    National Careers Service, Data analyst statistician job profile, average salary a year, https://nationalcareers.service.gov.uk/job-profiles/data-analyst-statistician

  4. 4.

    Ronny Kohavi and others, Online Experimentation at Microsoft, Microsoft ThinkWeek paper, 2009, public version, https://exp-platform.com/Documents/ExPThinkWeek2009Public.pdf