/

Data

Critical data elements: how to find the data your business depends on

Critical data elements: how to find the data your business depends on

Data

Data

Mahesh Chandran

Mahesh Chandran

CEO

CEO

Guide cover: Which datasets does your business depend on? A three-signal method for ranking data by what breaks when it is wrong.

If the warehouse went read-only tomorrow morning, which tables would people notice by lunch? Most data teams can name three or four straight away. The rest of the list takes a meeting, and the meeting usually ends in disagreement.

That short list is what governance teams call critical data elements: the data the business cannot run, report or decide without. Writing it down is one of the most useful things a data team can do, because every later decision about tests, alerts, refresh schedules and spend depends on it.

This guide sets out a three-signal method for finding the datasets a business depends on, ranking them into tiers, and deciding what each tier gets. It expands on Step 4 of How to audit your data estate in seven steps.

The short answer

A business depends on a dataset when something stops working, or someone acts on a wrong number, as soon as that dataset is late or wrong. To find those datasets, score every table on three signals: who reads it, where its output lands, and what happens when it is late or wrong. Put anything that lands in billing, finance, filings, customer products or production models in Tier 1. Pass each tier upstream to the tables it is built from, and give each tier its own level of testing, monitoring and alerting.

Key terms

  • Critical data element (CDE): a single field the business needs to manage risk, make decisions and keep operating, such as a customer identifier or a booked revenue amount.

  • Dataset tier: a rank from 1 to 3 that records how much the business depends on a table or pipeline. Tier 1 is the highest.

  • Destination: the last place a dataset’s numbers appear, such as an invoice, a regulatory filing or a dashboard.

  • Upstream pass: giving every table that a dataset is built from the same tier as that dataset.

  • Dependency: what stops working, and for whom, when a dataset is late or wrong.

The method at a glance

  1. List every table and pipeline, with an owner for each.

  2. Score each one from 0 to 3 on readers, destination and consequence.

  3. Apply the tier rules. A score of 3 on destination or consequence means Tier 1. A total of 4 or more means Tier 2. Everything else is Tier 3.

  4. Pass each tier upstream through the lineage.

  5. Give each tier its level of care, and re-score every quarter.

Why don’t size and query counts show what matters?

The obvious ways to rank data are by how big it is, how often it is queried, or how long it has existed. Each one fails in a predictable way.

  • Size rewards raw event logs, which are often the largest tables in a warehouse and among the least read. Some of the smallest tables, such as currency rates or a customer-to-account mapping, sit under every revenue report.

  • Query count rewards machine traffic. A dashboard that refreshes every fifteen minutes generates thousands of reads a month whether or not anyone opens it.

  • Age rewards inertia. A table that has existed for six years has survived several reorganizations, which says more about how hard it is to remove than about how much it matters.

Dependency is the better measure: what stops working, and for whom, when this data is late or wrong. A table with three readers that feeds the invoice run matters more than a table with two hundred readers in an exploration sandbox.

Dependency also flows upstream. A raw table that nobody queries directly still matters if a Tier 1 table is built from it. Mikkel Dengsøe and Lindsay Murphy make the same point in their guide to identifying business-critical data, where anything upstream of a critical asset counts as part of the critical path.

What are the three signals of dependency?

Three questions, asked of every dataset, give a ranking that holds up when people argue about it.

The three signals and their 0 to 3 scales: who reads it, where it lands, and what happens when it is late or wrong.

1. Who reads it

Start with the query history and the BI tool’s view logs, and count people and systems separately. Then look at who the people are. Dengsøe and Murphy note that usage by seniority is highly correlated with importance. A table read every week by the finance lead and the head of sales carries more weight than one read daily by a single analyst exploring an idea.

Systems count too. A table read by the billing service or a production model has a reader that never takes a day off and never notices a problem on its own.

2. Where it lands

Follow each dataset downstream to the last place its numbers appear. The destination says more about importance than the reader count does. These are the usual destinations, roughly from highest stakes to lowest:

  • Invoices, payments and customer-facing products

  • Regulatory filings, the finance close and board reporting

  • Production models, including AI features that answer customers

  • Operational dashboards that teams act on every day

  • Ad hoc analysis and exploration

3. What happens when it is late or wrong

The consequence question settles most arguments about tiers. For each dataset, ask what happens if it stops refreshing tonight, and how long it would take someone to notice.

The second half matters as much as the first. A wrong number costs very little on the day it appears and a great deal by the time someone catches it. A broken invoice table is caught within hours because customers complain. A slowly drifting revenue figure in a board deck can go unnoticed for a quarter, and that delay is what makes it expensive. Datasets where errors are expensive and slow to surface belong at the top of the list.

How do you turn the signals into tiers?

A simple score keeps the ranking consistent across teams. Give each dataset 0 to 3 on each signal:

  • 0: nobody reads it, it lands nowhere, and nobody notices when it is wrong.

  • 1: a few people read it now and then, it feeds exploration and ad hoc analysis, and someone notices eventually.

  • 2: a team reads it every week, it feeds dashboards that teams act on, and a team is blocked for a day when it is wrong.

  • 3: senior leaders or a production system read it, it feeds billing, finance, filings, customer products or production models, and customers, money or regulators are affected when it is wrong.

Then apply three rules, in order:

  1. Tier 1 if the dataset scores 3 on where it lands or on what happens when it is wrong, whatever its other scores.

  2. Tier 2 if its total is 4 or more.

  3. Tier 3 for everything else. A Tier 3 dataset with a total of 0 is a candidate for retirement, covered in How to find and retire unused data pipelines.

The destination rule matters most. It stops a table with few readers but a high-stakes destination from sliding down the list because its reader count is low.

Last, pass each tier upstream. Every table that a Tier 1 dataset is built from takes Tier 1 as well, even if nobody reads it directly. Work back through the lineage from each Tier 1 and Tier 2 dataset until you reach the sources.

raw_payments and stg_payments have no readers but inherit Tier 1 from invoice_lines, which feeds customer invoices.

A worked example

The tables below are invented to show the method.

  • invoice_lines: readers 3 (billing service), destination 3 (invoices), consequence 3 (customers billed wrongly). Tier 1.

  • raw_payments: readers 0 (only the pipeline), destination 0, consequence 0. Tier 1, passed up from invoice_lines.

  • weekly_signups: readers 2 (marketing team), destination 2 (planning dashboard), consequence 1 (a plan slips). Tier 2.

  • experiment_q2_scratch: readers 1 (one analyst), destination 1 (exploration), consequence 0. Tier 3.

raw_payments scores zero on its own and still ends up in Tier 1, because invoice_lines is built from it. That one row is why the upstream pass matters.

A scoring template to copy

Paste these columns into a spreadsheet to run the method. The rows repeat the worked example.




What does each tier get?

A ranking only pays off when it changes how the data is run. Each tier should come with a different level of care.

What Tier 1, Tier 2 and Tier 3 datasets each get, from named owners and paging alerts down to quarterly retirement review.
  • Tier 1 gets a named owner, tests on every column that feeds its destination, monitoring for freshness and volume, alerts that page someone, first place in the refresh queue, and protected compute. Changes to its code need a second reviewer.

  • Tier 2 gets an owner, tests on the columns people use, monitoring, and alerts that go to the owning team’s channel during working hours.

  • Tier 3 gets basic freshness checks and no pages. It is reviewed every quarter for retirement.

The split also tells a data team where to spend new effort. Tests and monitoring belong on the Tier 1 and Tier 2 datasets the business depends on. An hour spent testing a Tier 3 table goes to data that nobody would notice breaking, and every alert it raises makes the important alerts easier to ignore.

How do you keep the ranking current?

Tiers drift as the business changes. A Tier 2 table becomes Tier 1 the week someone builds a pricing model on it, and a Tier 1 report drops when the business stops using it. Three triggers keep the ranking honest:

  • A new destination. When a dataset starts feeding billing, a filing or a customer-facing feature, re-score it that week.

  • A lineage change. When a pipeline is rebuilt or a source is replaced, pass the tiers upstream again.

  • The calendar. Re-score every dataset each quarter, using the same query history that produced the first ranking.

How do critical data elements relate to dataset tiers?

The term critical data element comes from data governance, and especially from banking. Under BCBS 239, the Basel Committee’s principles for risk data, banks identify the data elements they need to manage risk and report on it. Outside banking, the data governance practitioner Irina Steenbeek uses a broader working definition of critical data elements, covering the data a company needs to manage business risk, make decisions and keep operating.

Critical data elements are usually defined at the level of a single field, such as a customer identifier or a booked revenue amount. Dataset tiers work at the level of the tables and pipelines that carry those fields. The two fit together. The governance team names the fields that matter, and the tiers show which tables carry them, who reads those tables, and what the data team has to protect.

Some catalogs build tiers in. OpenMetadata, for example, ships a five-level tier classification in which Tier 1 is the most important data and Tier 5 marks unused datasets that can be deleted periodically. Three tiers are enough to start with. More levels can come once the first ranking has been through a quarter.

Questions people ask

How many datasets should be Tier 1?

Fewer than most teams expect. If half the warehouse ends up in Tier 1, the tiers are no longer telling anyone where to look first. Tighten the rule for where data lands until Tier 1 is a list people can name from memory.

Who decides the tiers?

The data team proposes them from usage and lineage, because that evidence lives in the platform. The business owners of each destination confirm them, because they know what happens when a number is wrong. Finance should review anything that touches revenue or filings.

Can tiering be automated?

The evidence can be. Readers, destinations and lineage all come from records the platform already keeps, and the upstream pass is mechanical. The judgment about what happens when a dataset is wrong needs a person, at least the first time.

How is tiering different from data classification for privacy?

Privacy classification asks what kind of data a table holds, such as personal or payment data. Tiering asks how much the business depends on it. A table can be low-tier and sensitive, or Tier 1 and contain nothing personal at all, so most teams need both labels.

Where Dataring fits

We built Dataring to keep this ranking current without a quarterly spreadsheet. It ranks every dataset by how much the business depends on it, then acts on that ranking. The datasets the business relies on get the resources to stay reliable, and the pipelines nobody depends on are retired within limits each company sets, under its supervision.

Book a free call to see it on your data estate.

Newsletter

Product news by email. We use your address for nothing else.

© 2026 ShareData Inc.Last updated October 29, 2026