You Probably Do Not Need a Data Scientist Yet

The most expensive two words in a small business

Every few months an owner tells me, with a slightly worried face, that they think they need a data scientist. Sales are a touch flat, or a touch up and they cannot say why, or the spreadsheet has grown into a monster nobody trusts, and somewhere on LinkedIn they read that the cure is data science. So they start looking at salaries, they see the number, they wince, and then they either overpay for a job title they do not need or freeze and do nothing at all.

I want to save you both the wince and the freeze. Most small and medium businesses that think they need a data scientist do not, at least not yet. What they actually need is someone who can take a fuzzy worry and turn it into one clear question, then pull the single number that answers it. That person is usually a lot less expensive than a data scientist, a lot easier to find, and worth more to you on a wet Tuesday morning than the fanciest modeller you could hire.

Let me help you work out which one you genuinely need, because getting this wrong is one of the pricier mistakes I watch owners make. The two words "data scientist" have a way of pulling fifty thousand euros a year out of a business that needed something much simpler. And the tragedy is not that the money is spent. It is that the money is spent and the original worry is still sitting there, unanswered, because nobody stopped to ask what the actual question was.

Three different jobs wearing the same hoodie

The first thing worth knowing is that "someone good with data" is not one job. It is at least three, and they solve different problems. The clearest way to carve them up comes from The Data Analytics Handbook, which drew the line between the analyst and the data scientist back in 2014, and it still holds. Let me take each of the three in turn, because the whole rest of this piece hangs on you being able to tell them apart.

The report builder

A report builder takes data that already exists and puts it in front of you in a form you can read. Last month's sales by product, this week's refunds, which supplier is slipping. They are not answering a hard question. They are removing the friction between you and a number that is already sitting in your systems, just buried where you cannot reach it. Most owners who say they are "flying blind" do not need analysis at all. They need a report that nobody has ever bothered to build.

The reason this role is underrated is that it feels too simple to matter. Surely if the number is already in the system, seeing it changes nothing? In practice the opposite is true. A number you have to hunt for is a number you never look at, and a number you never look at cannot change a single decision. The report builder converts a buried figure into a habit. That is a bigger deal than it sounds, and it is where a surprising share of "data problems" quietly dissolve.

The analyst

An analyst answers questions about what has already happened and why. Why did returns spike in March. Which two products are always bought together. Whether the discount actually lifted profit or just moved sales forward. The analyst lives in the past and the present, works mostly with the data you already own, and their core skill is not maths, it is asking the business the right question and then interrogating the data until it confesses.

The analyst is the role most owners actually need and least often hire, because the title sounds junior next to "data scientist" and the work sounds unglamorous next to "machine learning". But the analyst is the one who tells you the discount you love is quietly losing you money, or that your best channel is the one you nearly cut. That is decision changing work, and it runs on data you already have, which means it can start on Monday with no new systems at all.

The data scientist

A data scientist builds new things that did not exist before, usually to predict what has not happened yet. Which customers are about to churn, what this basket should recommend next, how demand will move if you change the price. This is where the models, the statistics and the real code come in. It is genuine skill and it is genuinely scarce, which is why it is priced the way it is. The catch is that a prediction is only worth building if you were going to act on it often enough to pay for the person who built it.

That last sentence is the whole game, so let me say it as a rule you can carry around. A prediction earns its keep only when the value of acting on it, times how often you will act, clears the cost of building and maintaining it. A model that would save you two hundred euros a month cannot justify a hundred thousand euro seat, no matter how clever it is. Cleverness is not the test. Repeated, valuable action is the test.

Here is the trap that ties the three together. All three of these people can call themselves a data scientist on LinkedIn, because the title is fashionable and nobody polices it. So owners hire for the title, pay the data scientist rate, and then hand that person a job that was really report building or analysis. You have bought a Formula One engine to do the school run, and then you are surprised it is expensive to run and bored in traffic.

To keep the three straight, here is the same distinction in a table you can glance at.

Role	The question it answers	Data it uses	What it costs you	Starts working
Report builder	What is the number, and can I see it every week	Data you already hold	Low, often a one off	Almost immediately
Analyst	What happened and why	Data you already hold	Moderate, a few days a month	Within days
Data scientist	What will happen next, predicted by a new model	New data pipelines plus what you hold	High, a full time seat	Weeks to months

The single most useful habit you can build is to read your worry, then read this table, and notice which row it actually points at before anyone mentions a salary.

The skill that is worth more than the maths

Now the part that took me a while to properly believe. The most valuable person in this whole story is often not the best modeller. It is the one who can take a vague, anxious sentence from a business owner, "I feel like we are busy but not making money," and turn it into a single question that a number can actually answer. "What is our gross margin per order this quarter versus last, split by channel." That translation, from worry to question to number, is the rarest skill of the lot, and it is the one that moves the business.

There is a nice way to think about the full set of skills, from Business Intelligence, Analytics and Data Science in 2018, which lays them out as a hexagon: domain knowledge, data handling, programming, communication, curiosity, and the statistical craft that ties them together. A data scientist is usually hired for the two hard technical points, programming and statistics. But for a small business, the points that pay the rent are the softer ones. Domain knowledge, because someone who understands your trade knows which question is worth asking. Communication, because a right answer nobody understands changes nothing. And curiosity above all, because the person who keeps asking why is the one who finds the answer that was never in the brief.

Let me draw the hexagon as a map, so you can see which corners you are really buying when you hire.

This is the first thing to sit with. A brilliant modeller who never asks why will faithfully build you the wrong thing. A curious person with modest technical skills who insists on understanding the business will point you at the number that actually matters. If you can only have one to start, take the curious one. The technical craft can be bought in later, or borrowed for a fortnight when a specific model finally earns its place. The judgement about which question is worth answering cannot be bought in later, because by then you have already spent the money on the wrong answer.

Why the softest corner is the hardest to hire

Here is the awkward truth about the soft corners. They do not show up on a CV. You cannot filter a stack of applications for "asks the awkward question that saves the quarter". So hiring processes reach for what they can measure, which is the technical craft, and quietly ignore the thing that actually determines whether the hire pays off. That is not a knock on anyone. It is just what happens when a skill is real but invisible on paper. It is also exactly why a short, low commitment arrangement beats a big permanent bet at the start: you get to see the judgement in action before you wed yourself to it.

So which do you actually need

Here is how to self diagnose, and it starts not with the data but with the question you cannot currently answer. Say it out loud. "I do not know which products actually make money." "I cannot tell which marketing is working." "I do not know why good customers stop coming back." Then trace where the answer lives.

Most owners are surprised where they land. They walk in convinced they need the data scientist at the bottom of the tree, and they walk out realising the answer to their question was sitting in an order table the whole time, waiting for someone to build the report. That is the second thing worth sitting with. The gap between what you think you need and what you need is usually one or two boxes, and every box you climb down the tree costs a great deal more money.

The only branch that truly justifies a full time data scientist is the bottom left one: a model that does not exist yet, that is central to how the business runs, and that you will lean on every single day. A pricing engine for a marketplace. A recommendation system for a shop doing serious volume. A demand forecast a logistics operation lives or dies by. If that is you, hire the data scientist and do not skimp. If it is not you, one of the other three boxes will serve you better and cost you less.

A quick test to place yourself on the tree

If you are not sure which box you are in, run your worry through three plain questions. First, is the number you want already sitting in a system somewhere, and you just cannot see it easily? If yes, you want a report builder. Second, would knowing why something happened change what you do next, using only the records you already keep? If yes, you want an analyst. Third, do you need to predict something that has not happened yet, and would you act on that prediction often, and is that prediction central to how you make money? Only if all three of those are yes do you want a full time data scientist. Notice how much has to be true before the most expensive box lights up. That is not an accident. It is the whole point.

The number that makes owners go quiet

Let me put the cost on the table, because this is where the decision usually gets made. These figures are illustrative and you should check them against your own market before you act, but the shape of them is what matters.

Arrangement	Roughly what you get	Indicative annual cost
Full time in house data scientist	21 days a month, one senior brain, plus recruitment, tools and management	EUR 90,000 to 110,000 fully loaded
Fractional analyst, a few days a month	3 days a month, senior thinking on your real questions, no on costs	EUR 30,000 to 36,000
One off report build	The report you have been missing, built once and yours to keep	Low thousands

What "fully loaded" actually means

The full time figure is not the salary. It is the salary plus employer on costs, recruitment fees, software, a laptop, and the management time it takes to keep a specialist busy and pointed in the right direction. In Austria that loading pushes a headline salary up by roughly a third before you have made a single decision with the data. A useful way to hold it is as a simple multiplier on the gross salary:

$\text{Loaded cost} = \text{Gross salary} \times 1.18 + \text{recruitment} + \text{tools} + \text{management time}$

So an eighty thousand euro salary is not an eighty thousand euro decision. Run it through the on cost factor alone and you are already at $80{,}000 \times 1.18 = 94{,}400$ before a single tool licence or a single hour of your own time managing the seat. Add recruitment and software and you are comfortably into the ninety to a hundred and ten thousand band in the table. The headline number on the job advert is the smallest part of what you actually commit to.

Cost per decision, the figure nobody prints

Here is the sum that changes minds. A full time hire only produces so many real decisions in a year, and you are paying for the whole seat regardless of how many that turns out to be. So the honest unit of cost is not the salary, it is the cost per decision the person actually helps you make. Spread a loaded cost of around ninety five thousand euros across, say, fifty real decisions in a year and you get:

$\frac{95{,}000}{50} \approx 1{,}900 \text{ euros per useful decision}$

Now do the same sum for the fractional arrangement. Three days a month at a day rate of nine hundred euros is $900 \times 3 \times 12 = 32{,}400$ euros a year. Suppose that senior brain, aimed only at your real questions, helps you make forty solid decisions across the year. Then:

$\frac{32{,}400}{40} \approx 810 \text{ euros per useful decision}$

The fractional route comes out at less than half the cost per decision, and that is before you count the decisions a full time seat wastes on report building it should never have been given. The lesson is not that data scientists are overpriced. They are priced exactly right for the scarce thing they do. The lesson is that most SME questions are not that scarce thing, so paying the scarce price for them is where the money leaks.

The break even nobody checks

There is a point at which a full time seat does become the better buy, and it is worth naming so you know what you are watching for. A full time seat pays off once you genuinely need more real decision support than a fractional arrangement can deliver. Roughly, if a fractional arrangement gives you three days a month, that is thirty six days a year of senior attention. A full time seat gives you around two hundred and thirty working days. So the full time seat only wins when you can honestly keep more than roughly one hundred days a year of a specialist filled with valuable, decision changing work, not busywork. Most small businesses cannot, and there is no shame in that. It just means the fractional arrangement is not a compromise for you. It is the correct answer.

A quick figure to hold onto. A full time hire costs you around three times a fractional arrangement, and a fractional arrangement at a few days a month is very often more than enough to answer the questions that are actually keeping you up. You are not buying a person full time. You are buying the thinking, and the thinking does not need to sit in your office five days a week to be worth having.

A worked mini case: the shop that thought it needed a model

Let me make this concrete with a composite example, the kind of thing I see play out. Picture an online shop turning over a couple of million a year. The owner is convinced they need a data scientist to build a recommendation engine, because the big players have one and the owner has read that recommendations lift revenue. They have a number in their head: a full time hire at around ninety five thousand loaded, and a plan to have the recommender live within the year.

We start instead with three days in the first month, and no modelling at all. The first job is to build the report that never existed: gross margin per order, split by product and by channel. It takes two of the three days. The moment it is on screen, the picture changes. Two of the shop's best selling products, the ones featured on the homepage, turn out to make almost no margin once returns and shipping are counted. A third product, quietly sitting three clicks deep, is carrying the whole business.

Now the analysis. Why are the low margin products the ones on the homepage? Because they were the first products the shop ever sold, and nobody ever revisited the layout. One afternoon of interrogating the order data confirms it. We move the high margin product to the front and demote the two that were losing money. No model. No prediction. Just a report that revealed the truth and a question that chased it down.

Here is the arithmetic on that single change. Say the shop does eight thousand orders a year and the reshuffle lifts average margin per order by four euros, a modest assumption:

$8{,}000 \times 4 = 32{,}000 \text{ euros a year}$

That is the entire cost of the fractional arrangement recovered by one decision, in the first month, before anyone has touched a predictive model. And here is the punchline. By month three, the owner no longer wants the recommendation engine. Not because it would not work, but because they can now see that the recommender was solving a problem they did not have. Their problem was never "we cannot predict what to recommend". It was "we cannot see which products make money". A report answered it. The ninety five thousand euro seat, had they hired it, would have spent its first three months building a clever model on top of a layout that was quietly bleeding margin the whole time.

That is the pattern, again and again. The expensive answer is exciting. The right answer is usually sitting in a table nobody has looked at.

Why fractional is usually the right first move

Here is what I have watched happen again and again. An owner who was about to spend a hundred grand on a full time data scientist instead brings someone in for a few days a month. In the first two sessions that person does almost no modelling at all. They sit with the business, they ask the awkward questions, they find the three numbers nobody was tracking, and they build the two reports that should have existed years ago. Suddenly the owner can see. And in seeing, they realise that the fancy predictive model they thought they needed was solving a problem they did not actually have.

That is the case for starting fractional, and it is the third thing worth sitting with. The first job is almost never modelling. It is framing. A few days a month buys you the senior brain that frames the question, and framing is the part that pays for itself immediately. If, months down the line, it turns out you genuinely do need a full time data scientist building something core every day, you will know, because a fractional arrangement will have shown you exactly what that person should build. You will hire from knowledge instead of from anxiety, and you will write a job description that actually describes the job.

Think of it as an escalation you earn your way up, not a leap you take on faith.

Start small, get the questions right, build the reports you are missing, and only scale the commitment when the work genuinely outgrows the arrangement. Precision and headcount come later. The direction is the win. Every step up that ladder is taken because the previous step proved it was needed, not because a salary advert made it sound impressive.

This is squarely the kind of thing I do, and it is what my data and analytics work is built around. A few days a month, on your real data, starting with the question you cannot currently answer. No new hire, no wince, no year long commitment to a title you were not sure you needed.

Common mistakes I watch owners make

Over the years the same handful of errors come up, and every one of them is avoidable once you can see it. Here are the ones that cost the most.

Mistake	What it looks like	The fix
Hiring the title, not the job	Paying the data scientist rate for work that was report building	Name the question first, then match it to the smallest role that answers it
Buying the tool before the question	A shiny dashboard licence nobody has framed a decision around	No tool until a real question is written down
Confusing more data with more insight	Collecting everything, deciding nothing	Collect the few numbers a decision actually needs
Building a model you will not act on	A churn score nobody has a plan to use	Only build a prediction you will act on repeatedly
Going full time to feel serious	A permanent seat that is idle half the year	Start fractional, escalate only when the work overflows

Let me pull out the two that do the most damage. The first is hiring the title rather than the job, which is the whole reason this piece exists. The second is subtler and just as expensive: building something you will never act on. A prediction that does not change a decision is a very costly ornament. Before anyone builds a model, the question to ask is not "can we predict this" but "when we can predict it, what will we do differently, and how often". If the honest answer is "not much, and not often", you have found a model that should not be built, and you have just saved yourself a small fortune.

There is a third, quieter mistake that deserves its own line: mistaking activity for progress. A team can be busy collecting data, wiring up pipelines, and admiring dashboards for a year without a single decision changing. Motion is not the same as movement. The test of any data work is always the same, and it is brutally simple: did a decision change because of it? If nothing changed, nothing was learned, no matter how much was built.

How to apply this in the next thirty days

You do not need a strategy deck for this. You need a page and a bit of honesty. Here is the plan I would run.

Week	Do this	You end up with
Week 1	Write down the one question you cannot currently answer, in a single plain sentence	A real question instead of a vague worry
Week 2	Run it down the decision tree and name the smallest role that answers it	Clarity on whether it is a report, an analyst or a model
Week 3	Find the data that holds the answer, or the report that would surface it	A concrete first task, not a hire
Week 4	Bring in a fractional brain for a day or two and answer the question	A decision made, and evidence of what to do next

The whole point of this plan is that it front loads the lowest cost, fastest, most reversible steps and defers the expensive, permanent one until you have proof you need it. By the end of the month you will either have answered your question for the price of a couple of days, or you will have discovered, with evidence in hand, that you genuinely need something bigger. Either outcome is a win, because both of them replace anxiety with knowledge.

If you want to shortcut the first three weeks, that is exactly what a call with me is for. We start with the question you cannot answer, place it on the tree together, and I tell you honestly whether it is a report, an analyst, or the rare case that really does justify a model. You can book that at contact#book, and you will leave the call knowing which of the three you need, which is worth having even if we never work together.

The one idea to leave with

If you take a single thing from this, make it this. Before you hire for a title, name the question you cannot answer, and follow it down the tree to the smallest role that can answer it. Nine times out of ten it is a report builder or an analyst, not a data scientist, and the right way to buy that thinking is a few days a month, not a full time seat. The person who turns your fuzzy worry into the one number that answers it is worth more than any modeller who never stops to ask why. Hire the question first. The title can wait.