A data stack is the group of tools a company uses to move information from the systems where it is created into a place where it can be reported on and analyzed.

It is not a single product. It is a sequence of five steps, and each step can be handled by software, by a person, or by a combination of both. Every company that has ever built a report combining figures from two different systems has completed all five steps, whether or not anyone called them a data stack.

The term usually comes up in one of a few situations: a company is choosing analytics software, a consultant or accountant has recommended a change, a report has produced a figure that nobody can explain, or a new tool cannot be connected to an existing one. In each case, the question underneath is the same — where does the data go between the system that records it and the chart that displays it.

‍

The basic sequence

A restaurant kitchen follows a similar order. Ingredients arrive from several suppliers. They are stored in a walk-in refrigerator. A cook prepares them according to standard recipes. A server brings the finished plate to the table.

A data stack works the same way. Business systems supply the raw material. A central database stores it. A set of rules prepares it. A reporting tool presents it. In both cases the recipes carry as much weight as the ingredients, and the recipes are the part most often left unwritten.

‍

The five layers

LayerWhat it doesCommon names for itExamples of tools1. SourcesCreate and record the dataSource systems, systems of recordQuickBooks, Xero, Salesforce, HubSpot, Stripe, Shopify, payroll and advertising platforms2. IngestionCopy data out of each source on a schedulePipelines, connectors, ETL, ELT, data movementFivetran, Airbyte, Stitch, Meltano, or custom scripts3. StorageHold all of the data together in one queryable placeData warehouse, data lake, lakehouseSnowflake, BigQuery, Redshift, Databricks, PostgreSQL4. TransformationConvert raw source tables into figures that match how the business defines themModeling, transformation, semantic layerdbt, SQLMesh, SQL scripts, stored procedures5. PresentationDisplay the results so a person can read themBI, business intelligence, reporting, visualizationPower BI, Tableau, Looker, Superset, Metabase, Excel

Layer 1: Sources

These are the everyday systems that record what happens in a business: the accounting ledger, the CRM, the payment processor, the e-commerce platform, the payroll system. Each holds a partial view. None is designed to answer questions that require the others.

Layer 2: Ingestion

Ingestion is the work of copying data out of each source system on a regular schedule. Done manually, this is the act of exporting a spreadsheet. Done with software, a connector maintains the copy automatically and adjusts when a source system changes how it stores information.

Layer 3: Storage

Storage is a single database that holds copies of the data from every source, in a form that can be queried. A data warehouse is the most common type. This is the layer that makes cross-system questions possible, because a query can reach two sources at once only if both are in the same place.

Layer 4: Transformation

Transformation converts raw source tables into figures that match the business's own definitions. This is where a company decides what its words mean. Whether revenue includes refunds. Whether a customer is a billing account or a parent company. Whether the sales team's "closed" and the finance team's "booked" describe the same event.

When these definitions are written down once, in this layer, every report built afterward inherits them. When they are not, each report carries its own version, and figures diverge. This layer requires the least technical skill and produces the most disagreement.

Layer 5: Presentation

Presentation is the reporting tool: the dashboard, the chart, the scheduled PDF, the spreadsheet. It is the only layer most people in a company ever see, which is one reason the other four are easy to overlook.

‍

Key terms

Pipeline — An automated process that moves data from one place to another on a schedule.

ETL and ELT — Extract, Transform, Load, and Extract, Load, Transform. The difference is whether data is reshaped before or after it is stored. ELT is now more common, because cloud storage is inexpensive enough to keep a raw copy, which allows a mistake in the rules to be corrected without re-extracting historical data.

Data warehouse — A database designed for analysis rather than for running daily operations.

Data lake — Storage that holds data in its original format, including files and documents. A lakehouse combines features of both.

Semantic layer — The record of what each business term means, stored alongside the calculations, so that reports share one definition.

Orchestration — Software that runs the steps in the correct order and reports failures.

Reverse ETL — Sending a figure calculated in the warehouse back into an operational system, so that it appears where staff already work.

Modern data stack — A marketing term, not a technical one. It describes cloud-based tools with one vendor per layer, as distinguished from older systems installed on a company's own servers.

‍

Identifying what a company already has

Most small and mid-sized companies have layers 1 and 5 and nothing formal in between. They own many business systems and at least one reporting tool. Layers 2, 3 and 4 are performed by a person who exports files, combines them in a spreadsheet, and applies the definitions from memory.

Four signs indicate that a manual arrangement has reached its limit:

1. Preparing a recurring report takes longer than reviewing it.
2. Two reports give different figures for the same measure.
3. The process depends on one person's knowledge and is not written down.
4. A question cannot be answered because it would require joining three systems.

A separate situation is a company that has a warehouse and a reporting tool but no transformation layer. Its data is centralized, but its definitions are not, so each analyst applies their own. Survey data suggests this is widespread: in dbt Labs' 2025 State of Analytics Engineering report, 56% of respondents identified data quality as a problem. Those respondents work at organizations that already own the tooling, which indicates the difficulty lies in definitions and process rather than in software.

‍

A commonly cited figure

A claim that appears throughout material on this subject is that data professionals spend 80% of their time cleaning data. Its origin is a single survey, reached through several intermediate citations.

In 2020, the data specialist Leigh Dodds traced that chain: a vendor product page cited a 2018 Harvard Business Review article, which cited a 2017 blog post, which cited a 2016 CrowdFlower survey. That survey reported 60% of time spent cleaning and organizing data, with a further 19% spent collecting it. The 80% figure is reached by combining the two categories. A 2018 Kaggle survey, which asked about cleaning separately, recorded approximately 15% for cleaning and 11% for gathering.

The figures differ because the surveys count different activities. A company's own measured hours — time spent on exports, reconciliation and correction — is a more reliable basis for an estimate than any of the published percentages.

‍

Recent developments

Two changes have affected this category in 2026.

The middle layers are consolidating. Fivetran and dbt Labs, widely used tools in layers 2 and 4 respectively, completed a merger on 1 June 2026, having announced it in October 2025. For most of the past decade, the prevailing approach was to select a specialist tool for each layer separately. Two of the most prominent specialist tools now belong to one company, so the boundaries between layers are becoming less distinct.

AI tools are becoming consumers of the data stack. The same dbt Labs report found 80% of data practitioners using AI somewhere in their workflow. Where a tool answers questions in plain language, it draws on the definitions in layer 4. If those definitions are absent, the tool has no way to know which figure was intended, so the transformation layer affects the reliability of AI-generated answers directly.

‍

Common questions around data stacks

Is a data stack the same as a data warehouse? No. The warehouse is one layer, the storage layer. A data stack includes the tools that fill the warehouse and the tools that report from it.

Does a small company need one? It depends on whether routine questions require more than one source system. If all necessary information is held in a single platform, that platform's own reporting is usually sufficient. A data stack becomes relevant when answers require combining systems, or when manual preparation consumes significant time.

Where should a company start? Where the difficulty currently is. If figures disagree, the definitions in layer 4 are the constraint, and agreeing on them requires discussion rather than purchase. If preparation is slow but the figures are trusted, the constraint is in layers 2 and 3.

What does it cost? Costs vary widely, because most storage and ingestion tools charge according to volume rather than a flat fee. Estimating a budget requires knowing how many source systems will be connected and how often data will be refreshed. The staff time needed to configure and maintain the tools is often the larger figure and is easier to overlook.

‍