Analytics & BI Tools · Chapter 17 of 18
The economics, worked
The break-even that decides a licensing model, the four costs nobody budgets, and three organisations costed end to end.
BI cost conversations go wrong because people compare list prices, and list price is one of four costs. This chapter does the arithmetic properly — and the arithmetic is more useful than the numbers, because the numbers move and the method does not.
The four costs, in the order people forget them
| Cost | Typically budgeted? | Why it surprises |
|---|---|---|
| 1 · Licences | Always | The only one on the comparison spreadsheet |
| 2 · Warehouse compute | Rarely | Live-query dashboards generate queries per interaction. A popular dashboard on live query can cost more in warehouse credits than in BI licences |
| 3 · The people | Sometimes | One analyst maintaining models costs more than most licence bills. Two of them costs more than almost any of them |
| 4 · Migration and rework | Almost never | Rebuilding 200 reports is a year of someone's life, and it is the reason consolidation projects overrun |
The ordering matters because it inverts the attention. Teams spend weeks negotiating a licence discount and no time at all on the storage mode that quietly triples the warehouse bill. If your dashboards are on live query, the compute line is a BI cost — it just arrives on a different invoice.
The break-even that decides your licensing model
Every per-seat versus capacity decision reduces to one calculation. Take the monthly capacity cost, divide by the per-viewer licence price, and you have the number of read-only users above which capacity wins.
On current pricing, a reserved capacity at the threshold tier is around $5,000 a month and a viewer licence is $14, so the break-even is roughly 358 view-only users. On pay-as-you-go, where the same capacity costs about $8,400, it is closer to 601.
Three details that make the naive calculation wrong.
1 · It is a cliff, not a slope. One SKU below the threshold, every viewer still needs a paid licence. There is no partial benefit — buying a capacity half the size does not halve your licence count, it eliminates the benefit entirely. Any sizing exercise has to check that specific threshold rather than assuming capacity scales smoothly.
2 · Creators still need seats above the threshold. Free users can view with a viewer role. Anyone who builds or manages content in a shared workspace still needs a per-user licence. So the honest comparison is (creators × seat) + capacity against (everyone × seat), and forgetting the creators is the most common error in these business cases.
3 · The capacity buys more than viewing. It is a whole platform — data engineering, warehousing, pipelines — not just a licensing device. A pure licence-arbitrage comparison understates it, sometimes substantially, and that argument is the honest way to justify crossing the threshold below the break-even headcount.
And one detail worth knowing because it is counter-intuitive: on this vendor's capacity, the three-year reservation is the same annual rate as the one-year. Unlike compute reservations, longer lock-in buys no additional discount — so take the shorter term unless you have another reason.
Three organisations, costed
Same requirement — dashboards for the business — at three sizes. Figures are annual, at list, using August 2026 published prices, and rounded.
| A · 40 people | B · 800 people | C · 5,000 people | |
|---|---|---|---|
| Shape | 3 creators, 37 viewers | 20 creators, 780 viewers | 60 creators, 4,940 viewers |
| Per-seat (at $14) | ~$6,700 | ~$134,000 | ~$840,000 |
| Capacity + creator seats | ~$60,500 — absurd | ~$63,400 — wins | ~$70,100 — wins decisively |
| A premium-seat tool (Creator/Explorer/Viewer) | ~$9,400 | ~$170,000 | ~$1,000,000+ |
| Open source, self-hosted | ~$0 licence + ~0.2 FTE | ~$0 licence + ~1 FTE | ~$0 licence + 2–3 FTE + support |
| Warehouse compute if imported | negligible | modest, predictable | modest, predictable |
| Warehouse compute if live query | small | can exceed the licence cost | can dwarf the licence cost |
Three things fall out of that table that are worth more than the numbers:
- At 40 people, do not run an evaluation. The difference between the options is a few thousand a year, and a three-vendor selection process costs more than that in people's time. Pick the one your team already knows and move on. Recognising when you are in that regime is a real skill.
- At 800, the model flips and the capacity option becomes both cheaper and more capable. This is the size at which the licensing decision genuinely deserves a week of analysis.
- At 5,000, the licence choice is dwarfed by the people and the compute. Sixty creators is a real team; the warehouse bill on a badly-designed estate exceeds everything else. Optimising the seat price here is optimising the smallest term.
- And the open-source column is honest rather than free. Zero licence cost plus two or three engineers is not cheap; it is a different allocation. It is a genuinely good trade in an organisation where engineers are cheaper than software, and a bad one where nobody is available to own it.
Where the money actually leaks
Six leaks, in the order they are usually found
1 · Live query where import would do. The largest single cause of an unexpected warehouse bill in a BI estate. Every filter click is a query, multiplied by the audience. Ask what freshness was genuinely required before assuming it was.
2 · Refresh schedules that were shortened and never revisited. Moving a model from daily to hourly multiplies its refresh cost twenty-four-fold, and nobody notices because it happened one model at a time.
3 · Default filters set to 'all time'. A dashboard that scans three years on every open, for a user who wanted this month.
4 · Dashboards nobody opens. Usage telemetry typically shows a third unopened in ninety days — each still refreshing on a schedule and consuming capacity. Archiving with one-click restore is free and uncontroversial.
5 · Synchronised refresh schedules. Forty models at 06:00 creates a spike the capacity must be sized for. Staggering them is a configuration change that can let you buy a smaller tier.
6 · Over-provisioned seats. The authoring tier assigned to people who only read. On a premium-seat tool this is the largest single line, and an audit of who has actually published anything in six months is usually startling.
The new line item: metered AI
Something changed in the last year that is not yet on anyone's cost model. Vendors have diverged on whether natural-language analytics is metered.
- One major vendor now prices conversational analytics per data token — with a large included quota, and enforcement beginning in late 2026. That is the first dated point at which a BI bill acquires a token line.
- Another moved its equivalent feature to pay-as-you-go with a free monthly allowance per user.
- A third states explicitly that it does not meter or charge for LLM tokens — which is plainly a competitive position rather than an accident.
- And one went the other way on access entirely, dropping the minimum capacity for its assistant by roughly thirty-fold, which changed who can afford AI features at all.
The buying implication: if your organisation intends to use natural-language analytics at any volume, the token policy is a first-order cost question and it will not appear on the comparison spreadsheet you are shown. Ask for it explicitly, ask what the included quota is, and ask what happens on the day enforcement begins.
What to actually do
- Model the audience shape, not the headcount. Creators, explorers and viewers have different prices and different growth rates, and the viewer count is what determines your licensing model.
- Cost the storage mode as a licensing decision. Import versus live query changes the warehouse bill by more than most licence negotiations change the licence bill.
- Get quotes in parallel. The discount available is a function of whether the vendor believes you have an alternative — and with half the market no longer publishing prices, the sales process is the only way to find out what anything costs.
- Put a review date on the decision. Every situation in this chapter was created by a reasonable choice that nobody revisited when the audience, the freshness requirement or the pricing model changed. The tool matters less than whether someone will notice when it stops being right.
Check yourself