What a data platform actually costs: a budget breakdown across five European markets

Most published guidance on what analytics costs is either a vendor price list or a range so wide it is useless. This is the structure we use when we build a business case, with the figures we would quote and the ones we would not.

Everything here is an order of magnitude as at the date above. Vendor list prices change, salary markets move, and anything in this article older than a quarter should be sanity-checked rather than cited.

The five cost lines

A Microsoft analytics platform costs money in five places. Business cases that fail usually counted two of them.

1. Platform licensing, Power BI per-user licences or Fabric capacity.
2. Azure consumption, storage, compute, integration, networking, where you run engineering outside Fabric.
3. Build cost, internal time, external delivery, or both.
4. Run cost, operations, monitoring, incident response, small changes.
5. The cost of not doing it, the one nobody quantifies, and frequently the largest.

Licensing

Power BI Pro is around US$14 per user per month at list and Premium Per User around US$24, following the 2025 revision. Fabric capacity starts near US$260 a month for an F2 and reaches roughly US$8,400 for an F64 pay-as-you-go, with around 40 per cent off for a one-year reservation.

The number that matters is the break-even: F64 is the first SKU where report readers no longer need individual licences, and it becomes cheaper than per-user licensing somewhere between about 350 and 600 readers depending on your discount and whether you reserve. The full arithmetic, including the four common ways organisations overspend, is in Power BI licensing in 2026.

For a mid-sized organisation, say 25 authors and 150 readers, per-user licensing is clearly correct, and the annual licensing line is a five-figure sum rather than a six-figure one.

Azure consumption

Where engineering runs in Azure rather than Fabric, the consumption lines are storage, compute and integration. Rough shapes for a mid-sized estate:

The single most common avoidable cost we find is compute left running when nobody is using it. A dedicated pool paused outside working hours costs roughly 60 per cent less than one running continuously, and pausing it is a scheduled job rather than a project.

Build cost: the honest ranges

What a build costs depends on source-system count and data quality, not on how many reports come out of it. We price by project size, and the same figures apply across our markets in the local currency, a project that is £40,000 in the United Kingdom is €40,000 in Germany and CHF 40,000 in Switzerland.

Discovery sits in front of all three as a fixed fee in the low five figures, and the fixed price for the build is set after it rather than before.

What moves a project to the top of its band: more than four source systems, a source with no change-tracking capability, data quality requiring remediation, an organisation that cannot settle a KPI definition quickly, or a regulated environment needing validation evidence. What moves it to the bottom: two or three accessible sources, a sponsor who can resolve definitional disputes, and an internal team able to take over report building once the model is delivered.

Swiss engagements land at the upper end of each band rather than in a separate one. The work is the same; the cost of delivering it there is not, and the supplier onboarding alone adds weeks before anything starts.

Hiring versus engaging

This is the comparison most business cases get wrong, usually by comparing a day rate to a salary without loading the salary.

A salary is not the cost of an employee. Load it with employer contributions, equipment, software, management time, recruitment fees and the recruitment lead time, commonly 25 to 40 per cent on top depending on market. Then add the risk that one person holds all the platform knowledge, which is the specific failure we are called in to repair most often.

The honest comparison:

• Hire when demand is steady and ongoing, there is an existing platform to maintain, and a manager who can judge the work. An internal engineer is cheaper per day than any consultancy over a year.
• Engage for a first platform, a migration, a specialist problem, or a deadline that cannot absorb a three-to-six-month hiring cycle.
• Both is the most common right answer: internal ownership of business engagement and report building, external depth for architecture and peak delivery. This is usually cheaper than either extreme and considerably less fragile than a single internal hire.

Run cost

The line that is forgotten in nearly every business case. A platform needs monitoring, incident response, small changes, and someone who notices at 04:00 that the refresh failed.

Realistic shapes:

• Internal: a portion of an engineer's time, 20 to 40 per cent of an FTE for a mid-sized estate, more if there is no automation.
• Managed service: £3,000 to £8,000 a month for a small estate, £8,000 to £20,000 for a mid-sized one, and £20,000 to £50,000+ for a large multi-domain platform. Sized to the estate rather than sold as a standard package.
• Nothing: the default, and the reason platforms rot quietly. A source changes a column, a refresh starts failing on Sundays, workspaces fill with abandoned experiments, and the platform becomes less trusted each quarter without ever breaking loudly enough to trigger a project.

Budget 15 to 25 per cent of build cost annually for run. Less than that and you are planning to rebuild rather than to operate.

A worked example

A manufacturer with 600 staff, three plants, SAP plus an MES plus a quality system, 20 authors and 180 readers, in the UK:

Whether that is good value depends entirely on the fifth cost line.

The cost of not doing it

This is the number that decides the business case, and it is the one clients can usually produce fastest once asked the right question. In the engagements we have run, the quantified costs of the status quo have included:

• Four days a month of senior finance time assembling a pack that a model produces automatically.
• A month-end close extended by three days because reconciliation ran through individual spreadsheets.
• Markdown decisions taken on instinct rather than sell-through, with a margin cost the commercial director could estimate within an hour of being asked.
• A regulatory submission prepared by hand each quarter by two experienced staff, concentrating risk in two individuals.
• Loss-making customers invisible because cost was allocated at depot level rather than shipment level.

Each of those has a number attached to it inside your own organisation. Ask the four or five people who do that work how many hours it takes and what it prevents them from doing, and the business case usually writes itself, in either direction.

Three worked scenarios

Abstractions are less useful than numbers, so here are three shapes we quote regularly. All figures are indicative and would be fixed after a discovery.

Scenario A, a 120-person professional services firm in Ireland. Practice management system, finance system, CRM. 15 authors, 60 readers. One semantic model covering utilisation, realisation, WIP and pipeline. A medium project at the lower end.

• Licensing: Pro × 75, roughly €11,000 a year.
• Azure: minimal, the data fits comfortably, roughly €3,000 a year.
• Discovery plus build: around €55,000 one-off.
• Run: a small retainer, around €12,000 a year.
• Year one roughly €81,000; ongoing roughly €26,000.

Scenario B, a 600-person manufacturer in Germany with three plants. SAP, MES, quality system, historian. 20 authors, 180 readers, three years of event-grain retention required. A medium project at the upper end, pushed there by source count and the works council consultation German projects require.

• Licensing: Pro × 200, roughly €32,000 a year.
• Azure or Fabric: lakehouse storage and compute, roughly €18,000 a year.
• Discovery plus build: around €110,000 one-off.
• Run: managed service, around €30,000 a year.
• Year one roughly €190,000; ongoing roughly €80,000.

Scenario C, a 2,000-person financial services group in the UK. Core system, general ledger, CRM, several regional entities. 60 authors, 900 readers, regulatory reporting lineage required. A large platform, delivered in two phases.

• Licensing: reserved F64 plus Pro × 60, roughly £56,000 a year.
• Azure: engineering outside the capacity, roughly £30,000 a year.
• Discovery plus build: around £240,000 across two phases.
• Run: managed service with out-of-hours cover, around £70,000 a year.
• Year one roughly £396,000; ongoing roughly £156,000.

Scenario C is the one where the licensing decision flips, and it is worth noting why: at 900 readers the reserved capacity is cheaper than per-user licensing and the regulatory lineage requirement needs capabilities a per-user model does not provide. Both reasons point the same way, which is the comfortable case. Where they point in opposite directions, the licensing arithmetic should win.

The budget mistakes we see most often

Counting build and forgetting run. A platform with no operating budget degrades quietly and is rebuilt in three years at full cost. Budget 15 to 25 per cent of build annually.

Comparing a day rate to a salary. An unloaded salary understates the cost of an employee by 25 to 40 per cent, and omits the recruitment lead time entirely.

Assuming licensing is the big number. For most mid-sized organisations it is the smallest of the five lines. People optimise it because it is the easiest to see.

Budgeting by report count. Cost is driven by source-system count and data quality. Fifty reports on two clean sources is cheaper than eight reports on six messy ones.

No contingency for source access. The most common cause of delay is waiting for access to a system owned by someone with no stake in the project. Build in time, and name the person who will escalate it.

Treating discovery as optional. An estimate produced without discovery is an average with a confidence interval wide enough to be useless. Paying for two weeks of certainty before committing six figures is the cheapest insurance in the programme.

If it writes itself in the negative, that is a legitimate and useful conclusion. We have told clients that the honest answer was a licence change and a process fix rather than a platform, and we would rather be the people who said that. A technical review will produce the version of this breakdown that applies to your estate, and you keep it whatever you decide.