Single source of truth
A single source of truth (SSOT) is the designated place where a given piece of information is maintained and whose version wins when copies in other systems disagree.
For every piece of data, it is settled where it is valid and who maintains it.
A single source of truth is a company decision: for every piece of data, it is settled where it is valid and who maintains it. The customer address then comes from the CRM, the list price from the ERP, the travel expense rule from the handbook. Other systems read from there or receive copies that update automatically. The authoritative place for a data domain is also called the system of record.
Software development has a related principle. In a 2003 interview, Andy Hunt and Dave Thomas, authors of “The Pragmatic Programmer,” put it this way: every piece of system knowledge should have one authoritative, unambiguous representation. They explicitly included database schemas and documentation.
An AI assistant that searches company documents through retrieval-augmented generation (RAG) answers from whatever it finds. If three versions of a price list sit on the shared drive, it can quote any of them. Contradictory copies lead to contradictory answers.
Chapter 1 of 7
One authoritative place per item is not one system for everything.
A single source of truth does not require pulling all data into one big system. It requires that each piece of information be maintained in exactly one place. A company then has several authoritative places: the customer master in the CRM, items and prices in the ERP, product copy in the PIM, work instructions in the handbook.
One all-in system also hits the limits of business language. Martin Fowler describes an electricity utility where “meter” meant the grid connection, the customer relationship, or the device, depending on the department. He quotes Eric Evans: total unification of the domain model for a large system will not be feasible or cost-effective. Terms like customer or product behave the same way.
Terms therefore come first: what a data domain covers and which place is authoritative for it. The tool comes after.
Chapter 2 of 7
Maintained in one place, read everywhere.
For each data domain, it is settled which system is authoritative and who makes changes there. Every record carries a unique identifier, such as a customer or item number, which other systems use instead of creating their own version. They read through interfaces or receive copies that flow one way, from the authoritative place outward. Changing a price directly in the store bypasses that path.
Copies are still allowed. Reporting warehouses, search indexes, and an AI assistant’s knowledge base are all copies. They feed from the source, know how current they are, and are never written back without the source knowing.
Each record also shows when and by whom it was last changed, so a conflicting figure in a report is quickly traced to the version that counts.
Chapter 3 of 7
The business function that knows the data owns it.
Technology alone does not create a single source of truth. Someone has to own each data domain: define valid entries, fix errors, approve changes. The natural owner is the function that produces the data day to day, such as purchasing for supplier prices or HR for job descriptions. IT provides the systems, interfaces, and permissions.
Zhamak Dehghani describes the same idea for analytical data in her data mesh concept: accountability for data quality shifts upstream, as close to the source of the data as possible. The same logic applies to master data and process knowledge: errors get fixed at the source, not overwritten in a copy.
Without named ownership, the official version goes stale and teams return to their own spreadsheets. Ownership belongs in data governance, with one role per domain.
Chapter 4 of 7
For AI, the source caps the quality of the answer.
Language models store training knowledge in their parameters. For company knowledge, many applications use retrieval-augmented generation, described by Patrick Lewis and colleagues in 2020: a retrieval step fetches relevant passages, and the model writes its answer from them. The answer rarely gets better than the passages it finds.
If the same information exists in several versions, the search picks by similarity to the question. It has no idea whether a version is still valid. Two employees ask nearly the same question and get different prices or deadlines. The model invented nothing; it quoted existing copies.
A single source of truth fixes the cause. The knowledge base is filled from the authoritative places, old versions are removed, and every passage carries its source and date.
Chapter 5 of 7
For personal data, the law already demands accuracy.
Under Article 5(1)(d) GDPR, personal data must be accurate and, where necessary, kept up to date; inaccurate data must be erased or rectified without delay. Article 16 gives data subjects the right to demand rectification. If the company has disclosed the data to recipients, Article 19 requires it to tell them about every rectification, unless that proves impossible or involves disproportionate effort.
If a customer address sits in several systems and spreadsheets, a correction has to reach every one of them. With one authoritative place and automatic distribution, it is a single step. An SSOT is not a legal requirement and does not replace review by data protection counsel, but it makes these obligations easier to meet and to prove.
Chapter 6 of 7
What good implementation looks like.
It starts with a list of all important data domains, each with its authoritative place, its owner, and the systems that read from it. That list alone exposes double maintenance and gaps.
In a good implementation, each data domain has exactly one place where changes are allowed. Reading systems pull their data automatically and show how current it is. A regular reconciliation flags differences between source and copy before they reach reports or AI answers. New tools, AI included, connect to the authoritative places, not to exports.
There is no standard price. Effort goes mainly into the inventory, cleaning and merging old data, interfaces, and ongoing upkeep. With a single system and few users, a formal designation adds little. Once several systems hold the same data and discrepancies keep triggering questions, the payoff grows.
Chapter 7 of 7
Where the term stops working
A single source of truth settles where a piece of information is valid and maintained. Whether it is correct is a matter of data quality, since wrong data stays wrong in one place too. Data governance sets the roles and rules for designating these places. A knowledge graph links terms but does not replace deciding which place is authoritative for a value.
Frequently asked questions
What is a single source of truth?
A single source of truth is the designated place where a piece of information is maintained and whose version wins in case of doubt. Other systems read from it instead of keeping their own version.
What does single source of truth mean?
It means one agreed source per piece of information that everyone treats as correct. System of record names the same idea for a specific data domain.
Is a single source of truth one system for all data?
No. It designates one authoritative place per data domain, so a company usually has several, such as the CRM for customers and the ERP for prices.
Is a data warehouse a single source of truth?
For master data, usually not: it is a reporting copy, and the data is maintained elsewhere. For metrics, it can be the authoritative place if the company designates it that way.
What is the difference between a single source of truth and master data management?
Master data management (MDM) is the discipline companies use to reconcile and clean master data from several systems into a valid record, often with dedicated software. The single source of truth is the principle behind it; MDM is one way to get there.
Why does a single source of truth matter for AI?
An AI assistant answers from the documents and data it finds. If contradictory versions exist, it can quote any of them, and different users get different answers. A designated, dated source makes answers traceable.
Sources
- Bill Venners (Artima): Orthogonality and the DRY Principle, interview with Andy Hunt and Dave Thomas, March 10, 2003
- Martin Fowler: Bounded Context, martinfowler.com, January 15, 2014
- Zhamak Dehghani: Data Mesh Principles and Logical Architecture, martinfowler.com, December 3, 2020
- Lewis et al.: Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks, arXiv:2005.11401, 2020
- European Union: Regulation (EU) 2016/679 (General Data Protection Regulation), Articles 5, 16, and 19, 2016