The Systems of Provenance
AI now generates the analysis, the numbers, and the decisions your business runs on. Trace one back and most companies cannot say where it came from, whether it holds up, what it is worth, or what it has already changed. The problem is no longer what AI produces. It is whether you can trust it.
Study the price of bitcoin over the years and the climb can be explained a dozen ways. Some of it is hype. Some of it is supply and demand. Some of it is the technology itself. But under all of those explanations sits a quieter one that holds the rest up: Provenance.
prov·e·nance (noun)
- The place of origin or earliest known history of something.
- A record of ownership of a work of art or an antique, used as a guide to authenticity or quality. From the Latin provenire*, "to come forth." — Oxford Languages*
Every bitcoin carries its whole history, every transaction traceable back to the block it was mined in, on a ledger nobody can quietly edit. That history is its provenance, the unbroken record of where a thing came from and every hand it passed through, complete enough to prove it is real. With a bitcoin, provenance is not attached to the asset. It is the asset. Break the chain and there is nothing left to own.
NFTs were sold as the same promise, and for a while people paid millions on it. But the token rarely had authority over what it supposedly proved. Some pointed to an image on a server that could go dark. Others came wrapped in vague promises of access, royalties, or membership that no code actually enforced. The token proved you owned the token. It could not make the world honor what the token was said to represent. When the servers went dark and the promises went unkept, the tokens did not get cheaper. They became worthless.
Here is the part people missed. Provenance only holds value when the record has authority over the thing it records. Bitcoin's ledger does not point at the coin. It is the coin. A record that only points at something it cannot control is not provenance. It is a claim.
Companies are about to learn that their data works the same way.
Someone generates a long report with AI and never fully reads it. It gets forwarded, quoted in a meeting, pasted into a strategy deck, and everyone downstream accepts it because it looks right, or close enough. Now the company is making real decisions on data pulled from sources no one checked, shaped by logic no one can see. The model could have been quietly compromised by a competitor months ago and the output would look exactly the same. You could not tell, and almost no one checks.
Before AI, a person would write most of what mattered, and a person would read it before it counted. Now the machine writes it and no one reads it. The gap that used to be a nuisance is now a security risk dressed up as productivity.
As companies wake up to this, the Systems of Provenance will be how they prove the integrity of what AI does for them.
Systems of ...
The phrase "system of" has a lineage, and the lineage tells a story about where value kept moving.
It starts with Geoffrey Moore and the system of record. The ERP, the ledger, the CRM. The job was control and integrity: one accurate, auditable version of the truth. Nobody was supposed to enjoy using it. It was supposed to be right.
Next came the system of engagement. The interesting work was no longer storing the truth but using it: the collaboration tools, the mobile apps, the surfaces where people actually did their work. Success stopped meaning "is it correct" and started meaning "will anyone use it."
Then the system of insight, Forrester's 2015 name for turning data into a decision through a closed loop: find the insight, put it where people act, measure, repeat. Later, the system of action, on the argument that insight was never the point, action was. Stop showing the human, do the thing.
Around the same time, the system of intelligence, Jerry Chen's term at Greylock for the AI layer sitting on proprietary data. His point was strategic: as the record and engagement layers got commoditized, the advantage moved to the intelligence you could build on data nobody else had.
Read those in order and a pattern falls out. Record, engagement, insight, intelligence, action. Each new "system of" appeared at the exact moment the previous layer became a commodity and the frontier moved somewhere harder.
So the Jeopardy question of the day is this: if intelligence is now getting commoditized, where does the frontier move next?
What are the Systems of Provenance
The Systems of Provenance is the set of systems and practices that establish whether a fact can be trusted, from the moment it is born to the moment something acts on it. It answers four questions about any piece of data. Where did it come from. Who owns it. Which version is canonical when systems disagree. And can you prove it was never quietly altered. Origin, ownership, reconciliation, integrity.
Provenance is a property that either survives the whole path or does not exist. A fact travels a long way inside a company. It is captured somewhere, committed to a system of record, reshaped into metrics, read on a screen, and eventually acted on. Provenance is the thread that has to run the entire length of that path. Break it at any point and everything downstream inherits a number it cannot vouch for.
You can watch provenance break in something as ordinary as an email. A price gets agreed in a thread, someone pastes it into a slide, the slide feeds a board deck, and months later the number is quoted as settled fact with no one able to say whether it was ever countersigned. The email felt authoritative because it was in writing and it was where people first saw the number. But an email records what was said, not what is true or owned. The deal lives in the contract. Treating the thread as the source is the small, daily version of the mistake AI is about to make at machine scale.
It helps to separate provenance from the words it gets confused with. Data quality asks whether a value is accurate. Lineage asks what path it took. Governance asks what the rules are. Those are not the same thing. Provenance is the standard all three serve: the ability to stand behind a fact when someone who assumes you are wrong asks you to prove it.
Governance is the policy. Provenance is the evidence.
When responsibility moves to AI
For most of computing history, provenance took care of itself. A human stood at every important step, entering the number, reading the report, making the call. Slow, but accountable. Every layer we built assumed that person was there to vouch for what passed through, and that assumption is still baked in.
AI removes the human in both directions at once. It reads the raw material and produces an answer with no one turning it into a governed record, and agents are starting to act on those answers, writing back into the systems that run the business. A model gives you a confident answer whether or not the data underneath was ever true, and it cannot tell you which.
Speed got cheap. Belief did not.
The failure mode is no longer a wrong answer. It is a confident one nobody can trace, executed before anyone thought to check. The frontier is not a better answer. It is a trustable one.
What the market is already telling us
Three signals say the market is already moving toward provenance without having named the shape.
Regulation is pricing it in. The EU AI Act's high-risk rules land in August 2026, requiring proof that datasets are sound and system logs kept for at least six months, which is lineage by another name. Content provenance is going hardware-deep too: Leica, Nikon, Canon and Sony now ship cameras that cryptographically sign images to the C2PA standard, OpenAI and Google are tagging generated media, and CISA recommended C2PA in early 2025. When the camera and the regulator both demand a chain of custody, provenance is no longer optional.
Enterprise AI is stalling on trust, not models. MIT's 2025 research found roughly 95 percent of enterprise AI pilots delivered no measurable impact, and the share of companies abandoning most of their AI initiatives reportedly jumped to around 42 percent from 17 percent a year earlier. The models are fine. Informatica's 2025 survey put data quality and readiness at the top of the obstacle list. Pilots die between demo and production because nobody can vouch for the data the system stands on, so nobody will let it act. That is a provenance problem wearing an AI costume.
The market is already buying the pieces. Data observability is a roughly three billion dollar market growing in the mid-teens a year, the data governance market is set to multiply this decade, and lineage tooling is booming, all of it tied openly to AI readiness. The tell is convergence: catalogs adding lineage, observability adding quality checks, governance suites reaching for both, as customers try to assemble one capability out of parts. What they are buying, under different names, is provenance. The category exists. It just does not know what it is yet.
Put the signals together and the direction is clear. AI output will have to prove it can be trusted, and the Systems of Provenance are built to do exactly that.
References
Geoffrey Moore. Systems of Record and Systems of Engagement. The original distinction between systems built for control and integrity and systems built for interaction and use.
Forrester. Systems of Insight Will Power Digital Business and Insight Was Never the Point: Arise, Systems of Action. The naming of the insight and action layers, and the argument that action, not insight, was always the goal.
Jerry Chen, Greylock. The New New Moats. The case that systems of intelligence, AI built on proprietary data, became the durable competitive moat.
European Union. EU AI Act, Article 26: Obligations of Deployers of High-Risk AI Systems, read alongside the Alation compliance guide on Article 10 data governance, six-month log retention, and the August 2026 timeline.
C2PA Content Credentials adoption status, 2026, and CISA, Content Credentials: Strengthening Multimedia Integrity in the Generative AI Era. On cryptographic content provenance moving into cameras, AI providers, and government guidance.
Fortune, MIT report: 95% of generative AI pilots at companies are failing, and Congruity360, Why 95% of generative AI pilots are failing, citing Informatica's 2025 CDO Insights on data quality as the top obstacle to AI success.
Data Observability Global Market Report 2026, Data Catalog Market Size, Share & Growth, and Decube, Data Lineage: the foundation of AI-ready data. On the growth of observability, catalog, governance, and lineage tooling, tied explicitly to AI readiness.