The Analytics Tier You Are Buying Is Probably the Wrong One
There is a predictable way that analytics budgets get wasted, and it has nothing to do with picking the wrong vendor or the wrong warehouse.
It is buying capability one tier above where you actually are.
The four tiers
Analytics builds in a fixed order, and the order is not negotiable no matter how much you spend.
Descriptive — what happened. Revenue by region. Customers by month. Churn by cohort. This sounds trivial and is not: most companies cannot produce these consistently across two different tools, because two departments define the terms differently and neither knows it.
Diagnostic — why it happened. This requires that your data be modelled well enough to slice along dimensions nobody planned for in advance. Someone asks whether refund rates differ by acquisition channel, and the answer should take hours rather than a week.
Predictive — what will happen. Forecasting, churn scoring, demand planning. Genuinely valuable when the foundation supports it.
Prescriptive — what to do, and increasingly, doing it. Systems that recommend or take actions within defined guardrails.
The arithmetic that catches people
A prediction model built on unreliable descriptive data does not fail. That is the whole problem.
Train a churn model on a customer table where thirty percent of records have a null contract start date. It will converge. It will produce probability scores with four decimal places. It will generate a beautiful lift curve on a validation set that shares exactly the same defects as the training set.
And it will be wrong in ways that are extremely difficult to detect, because there is no error, no exception, and no signal — just a number that looks like every other number.
Vendors selling machine learning engagements rarely lead with a data quality audit. The audit is unglamorous, it delays the interesting work, and it occasionally reveals that the interesting work should not start yet. Insist on it anyway.
Why AI made tier-skipping more tempting
This got worse recently, and for an understandable reason.
Natural-language querying demos extraordinarily well. Someone types a question in plain English, an answer appears, and the room is impressed. It takes about a day to wire up over any warehouse.
Over unmodelled data, it produces confident wrong answers at speed. The model guesses what columns mean from their names, picks a plausible join key, includes the test accounts nobody told it about, and returns a figure with no caveat attached.
So a tool that would have been a tier-three or tier-four capability now appears achievable in week one, on tier-one foundations. It is the most attractive shortcut in the current market and the most expensive one, because the damage is to trust rather than to a budget line.
Trust in an analytics platform behaves close to binary. One wrong number in a room that matters costs you confidence in every other number the platform has produced, and rebuilding that takes roughly eighteen months regardless of how much you spend.
What buying the right tier looks like
Spend the first month on definitions rather than pipelines. Get finance, sales operations and product in a room and agree, in writing, on the ten metrics that actually drive decisions. That document is the deliverable of month one.
Then build the descriptive layer properly — modelled, tested, documented — for exactly those ten metrics. Not fifty. Ten.
Add the semantic layer, which encodes those definitions so that anything querying the warehouse, human or model, gets the agreed answer rather than an inferred one.
Only then add prediction or a conversational interface, in the second quarter, on foundations that can carry the confidence they project.
A vendor promising a production AI analytics assistant in month one is promising you speed. What they are delivering is confident wrong answers, on schedule.
The full version — what data analytics services actually include, real cost ranges, engagement models, and the vendor questions worth asking — is here: Data Analytics Services: A 2026 Guide for Technical Buyers.
TechCirkle builds data platforms and does AI development work.
Frequently Asked Questions
What are the four analytics maturity tiers?
Descriptive (what happened), diagnostic (why it happened), predictive (what will happen), and prescriptive (what to do, and increasingly doing it). The order is fixed, and skipping a tier reliably produces expensive results nobody trusts.
Why do prediction models built on poor data fail silently?
They converge normally and produce well-formatted probability scores. The validation set shares the same defects as the training set, so the lift curve looks fine. There is no error to catch, only a number that looks like every other number.
Why is tier-skipping more tempting now?
Natural-language querying demos superbly and takes about a day to wire up over any warehouse. That makes an advanced capability appear achievable on foundational-tier data, which is the most attractive and most expensive shortcut available.
How many metrics should the first build cover?
Ten — specifically the ten that drive real decisions or appear in board reporting. Attempting fifty dilutes the work and exhausts the political capital needed for genuinely contested definitions.
What is a semantic layer for?
It encodes agreed metric definitions so that anything querying the warehouse — a person or a language model — receives the agreed answer rather than an inferred one. It is what makes AI querying safe.
How long does it take to recover from a wrong number?
Roughly eighteen months, and money does not accelerate it. Trust in an analytics platform behaves close to binary: one wrong figure in a consequential setting undermines confidence in every other figure produced.