Self-service BI means people answering their own data questions without raising a ticket and waiting a fortnight. That's the whole promise, and it's a good one. The reason it fails so often is that "self-service" gets interpreted as "everyone builds whatever they like from raw data", which produces dozens of private spreadsheet-grade models wearing Power BI's clothes. The version that works is narrower: one shared, governed model holding the definitions, and self-service happening on top of it. People explore, filter and build reports against numbers that have already been agreed. They don't each rebuild the numbers.
That distinction is worth spelling out properly, because it's the difference between the tool that transformed your reporting and the tool everyone quietly stopped trusting.
What does self-service BI actually mean?
The term was coined against a real enemy: the old world where every report request went into an IT queue, took weeks, and came back not quite answering the question. Self-service was the promise that the person with the question could get the answer themselves, while the question still mattered.
Notice what the promise actually covers: answering questions. It says nothing about who builds the data model, who defines what "revenue" means, or who owns the refresh that keeps the numbers current. Those jobs didn't disappear when the tools got friendlier. They just stopped being obviously assigned to anyone.
Power BI makes the confusion easy, because the same product does both jobs. It's a modelling tool for the person building the shared foundations and an exploration tool for the person consuming them. Hand it to twenty people with no agreement about who's doing which job, and all twenty do a bit of both. That's not self-service. That's twenty small BI departments who don't talk to each other.
Why does self-service BI usually fail?
Three failure modes account for nearly everything I see.
Everyone models, nobody agrees. Each department connects to its own export and defines its own measures. Sales counts an order when it's placed, finance when it's invoiced, operations when it ships. All three are defensible. All three appear in the same board meeting as "revenue". I've written up what that does to a business in six reports, six different numbers, and the short version is that the meeting stops being about the decision and becomes an argument about whose number is right.
The enthusiast leaves. Self-service estates are usually carried by one or two keen people who taught themselves. Their models are undocumented, their refreshes run under their personal credentials, and their leaving date is the estate's expiry date. Nothing about the tool caused this. The absence of anyone owning the foundations did.
Consumers get blamed for a builder problem. When the numbers disagree, the conclusion drawn is "our people aren't data literate enough", and the fix prescribed is more training on charts. But the consumers were never the problem. They were served ambiguity and asked to self-serve certainty.
Which version of self-service are you being sold?
When a supplier or an internal champion says "self-service BI", they mean one of three arrangements. They have very different outcomes, so it's worth knowing which one is on the table. This is a comparison of the three models as I see them play out, not a survey; there's no honest external statistic for this, so judge the rows against your own estate.
| Free-for-all | Locked-down | Shared model + self-service on top | |
|---|---|---|---|
| Who builds models | Everyone | A central team only | One owner (internal or hired) |
| Who defines the measures | Each author, privately | The central team | Agreed once, in the shared model |
| Speed of getting an answer | Fast, at first | Slow, ticket queue | Fast, and it stays fast |
| Do numbers agree across reports | No | Yes | Yes |
| What it feels like in year two | Nobody trusts anything | The queue is back | The useful one |
| Real cost driver | Rework and arguments | Central team headcount | Building the model well once |
The free-for-all is what most SMEs accidentally buy, because it's what happens by default when licences are handed out with no structure. The locked-down version is the old IT bottleneck rebuilt in a newer tool. The third column is the one the term "self-service" was always supposed to describe.
What's the smallest setup that actually works?
For a small or mid-sized business, the working version needs surprisingly little.
One shared semantic model. A single model holding the tables, relationships and measure definitions the business has agreed on. The model is the product; reports are windows onto it. When someone asks what gross margin means, the answer lives in exactly one place.
A named owner. Someone accountable for the model's definitions, its refresh and its changes. This can be an internal person or a fraction of an external one. It cannot be nobody.
A small number of report builders. Two or three people who build report pages against the shared model. They never redefine measures; they arrange and present what the model already agrees on.
Everyone else consumes. Opening reports, filtering, drilling, exporting, asking questions. This is where most of the organisation should sit, and there's no shame in it. A finance manager who reads a trusted report well is getting more value from BI than one who builds an untrusted model badly.
That's it. No governance committee, no centre of excellence, no platform team. If your organisation has forty people, the whole arrangement is one model, one owner and a couple of builders. I've made the same anti-overbuild argument about the reports themselves elsewhere; it applies just as much to the operating model around them.
Where does training fit, and what does it cost?
Training is how the builder tier gets competent without the enthusiast-roulette problem, and it's most effective when it's taught on your own data rather than a demo file, because your team goes back to work in the same model the next morning.
My published training rates as of September 2026, from my training page: a 1-2-1 day from £750 online or £850 in person, a small group of up to four from £950 online or £1,100 in person, and a team day of up to ten from £1,250 online or £1,500 in person, each including a half-day prep that sorts licences, installs and access before anyone sits down. For a stuck builder with one specific problem, a £100 unblocking session is the cheaper, faster tool.
The sequencing matters more than the spend. Training report builders before the shared model exists teaches them to build confidently on sand. Model first, then train people on top of it. If you're not sure whether your current estate is sand or foundation, that's what a £950 Health Check is for: it scores what's actually there, including whether your measures agree with each other, before you invest in scaling it.
Common questions
What is self-service BI in plain English?
It's the ability for people in the business to answer their own data questions, build their own views and explore the numbers without submitting a request to IT and waiting. Done properly, they do this on top of a shared, governed data model, so everyone's answers come from the same agreed definitions.
Why do self-service BI projects fail?
Mostly because "self-service" gets stretched to cover model building as well as question answering. When every department builds its own model from its own extracts, the definitions drift, reports disagree, and trust collapses. The tool then gets blamed for an operating-model problem. The other common cause is that the whole estate depends on one self-taught enthusiast, and it expires when they leave.
Does self-service BI mean everyone needs Power BI training?
No. Most people should consume reports, not build them, and consuming a well-built report needs an hour of walkthrough, not a course. Training pays off for the small group who'll actually build on the shared model. Published UK rates for that are on my training page: from £750 for a 1-2-1 day to £1,500 for an in-person team day of up to ten, as of September 2026.
What is a semantic model and why does it matter for self-service?
It's the shared layer holding your tables, relationships and measure definitions: the place where the business agrees, once, what revenue and margin mean. Self-service works when people explore on top of that layer, because their freedom is in the questions, not in the definitions. Without it, self-service means everyone quietly inventing their own definitions.
Can a small business do self-service BI without an IT department?
Yes, and arguably more easily than a large one. The minimum viable setup is one shared model, one named owner (who can be a part-time external), two or three trained report builders and everyone else consuming. No committee or platform team required at SME scale.
Is self-service BI cheaper than having reports built for you?
Not automatically, and this is the honest bit: the free-for-all version is usually the most expensive option once you price the rework, duplicated effort and decisions delayed by arguments over numbers. Self-service saves money when the shared foundations are built once, properly, and the business serves itself on top. It's a sequencing question, not a headcount question.
How do I know if our self-service estate has already gone wrong?
Three observable signs: two reports give different answers to the same question, key refreshes run under one person's personal login, and nobody can say where a given measure is defined without opening files to look. Any one of them means the foundations need attention before anyone builds more on top.
Training and Health Check prices are from my published training and Health Check pages, September 2026. All figures exclude VAT where applicable.