Forgery has become cheap
For years a company's IT security was measured against one question: can someone get in? Firewalls, passwords, two-factor authentication. They are still necessary, and we wrote about them when discussing login attempts on Odoo.
Artificial intelligence adds a second, more uncomfortable question: how do I prove that what I am looking at is genuine? A supplier's email announcing a new IBAN, an invoice, a voice recording of the owner, a log file after an incident. Producing a credible fake used to take time and skill; today it takes a prompt. And inside business systems, software agents are starting to read, write and record on behalf of people.
The problem moves from the perimeter to integrity: keeping strangers out is no longer enough, you need to be able to prove what existed, when, and that nobody touched it afterwards.
This is where blockchain is back in the conversation. After the wave of projects that promised to revolutionise supply chains and payments, many of which did not survive the market (the best-known case is TradeLens, the Maersk and IBM logistics platform, shut down in 2022), there is one job the technology is genuinely suited for. But it is a precise job, smaller than the story we were told, and it is often done without any blockchain at all. Let's see where the line falls.
What a blockchain guarantees, and what it doesn't
Jargon aside, a blockchain is a register you can only append to. Each block contains the cryptographic fingerprint (hash) of the previous one, so changing a past entry breaks all the ones that follow. And the register does not live in one place: there are many copies, held by different parties. In public blockchains, such as Bitcoin or Ethereum, anyone can hold one. In permissioned ones, designed for businesses, such as Hyperledger Fabric, copies are held only by identified and authorised organisations.
The chain of fingerprints on its own is nothing new. It has been used for years by systems that have nothing to do with blockchain, from public logs of web certificates to business software. The difference lies entirely in the second part: who holds the copies. If the register is mine alone, whoever takes control of my system can rewrite it from start to finish and recompute every fingerprint. If it is replicated on thousands of nodes I do not control, nobody can: not the attacker, not a dishonest employee, not me.
That is the property security cares about. It does not keep anyone out; it proves, after the fact, that a piece of data existed in that form at that moment, without asking anyone to take my word for it.
The rest of the promise holds up far less well.
It does not tell you whether data is true. It records what it receives. A fake invoice written to a blockchain is still fake, just with a certain date. The problem shifts to who writes and how.
Confidentiality comes at a price. Public blockchains can be read by anyone and forget nothing, which clashes with personal data protection and the right to erasure. The European Data Protection Board, in its guidelines on blockchain adopted in July 2026, is clear: personal data should be kept off the chain, and even the fingerprint of personal data is still personal data. If it really has to be recorded, it must be built with a secret key or a random value stored elsewhere, so that deleting them makes the trace on the chain useless.
Permissioned networks tackle the confidentiality problem, and do it well: Fabric shares private data only among authorised organisations, leaves their fingerprint on the common ledger and even allows them to be purged, keeping the hash as evidence. But confidentiality is obtained by restricting who holds the copies, so the guarantee that nobody rewrites history is only as strong as the independence of the participants. A consortium of companies that keep an eye on each other is a guarantee; a lead company with two suppliers in tow much less so.
It is only worth as much as those who share it. The TradeLens case says it better than any analysis: in its closing announcement Maersk writes that it had developed a working platform, but that the full industry collaboration it needed was not achieved . The technology was there; the participants were not. And if all the nodes are run by a single party, permissioned or not, the blockchain is just a database with a few extra complications.
This leads to the form in which blockchain is useful to security today, and it is the most modest one: anchoring. The data stays where it is; only its fingerprint is written to the public blockchain. Anyone who wants to verify recomputes the fingerprint from the document and checks that it matches the recorded one. There are open standards that do this at no cost to the user: OpenTimestamps, for example, relies on the Bitcoin blockchain and offers free servers that record fingerprints for us. As its page puts it, a timestamp proves that some data existed prior to some point in time. Nothing more, but that is not nothing.
Where the focus on security gives it a role again
If the useful property is "nobody can rewrite history, not even whoever holds the keys", the interesting cases are those where the risk comes precisely from whoever holds the keys: an administrator, an attacker who has taken their place, a software agent acting with someone else's permissions. These are the cases AI makes more frequent. Let's look at them one by one, asking each time whether blockchain is really needed or whether something simpler will do.
Content and evidence: when was this file created?
A photo, a recording, a signed contract, an expert report. With generative AI the question "is it authentic?" applies to any digital content, and the first useful answer is knowing since when it has existed in that form. A fake built after an event cannot have a fingerprint recorded before it.
The main tool here is not blockchain. The industry standard for content provenance, C2PA, relies on digital signatures with certificates and on timestamps issued by time-stamping authorities. Blockchain comes in as an alternative when you do not want to depend on an authority, or as a second independent proof.
In Italy it also has legal recognition: since 2019, storing a document on distributed ledgers has the legal effects of an electronic time stamp under the European eIDAS regulation. But it is a simple time stamp: under Article 41 of eIDAS it cannot be denied evidential value merely because it is electronic, while the presumption that date, time and integrity are accurate applies only to a qualified one, issued by a recognised provider. The Italian law also ties the effect to technical standards the technologies must meet. Anyone who needs a certain date to rely on in a dispute would do well to consult a lawyer before relying on blockchain alone: a qualified time stamp remains the safer route.
Which blockchain: public, for the fingerprint only.
Logs and audit trails: the black box after an incident
After an attack, logs are the only way to reconstruct what happened. They are also the first thing a skilled attacker alters, because that is where their traces are. And an AI that helps find and exploit a vulnerability also helps clean up afterwards.
European regulation is clear on this. For cloud service providers, managed service providers and the other categories it applies to, the NIS2 implementing regulation requires logging all activities of administrative accounts too, protecting logs "against unauthorised access or changes" and monitoring logging systems "independently of the systems they are monitoring" (Annex, points 3.2.3, 3.2.5 and 3.2.6). The delicate point is precisely the administrator: logs must withstand even whoever has the highest privileges on the system that produces them.
The regulation does not prescribe a technology, and a copy on an external, immutable archive is enough in many cases. Anchoring on a public blockchain adds something the external archive lacks: the proof can be verified by a third party, an auditor or a judge, without having to trust whoever runs the archive. Every day you compute the fingerprint of the logs and record it; if a file turns out different tomorrow, it shows.
Which blockchain: public, for the fingerprint only. It is the most concrete and least expensive case.
AI agents: who did what
In business systems, mailboxes and ticketing systems, AI agents are starting to read, write and approve. When something goes wrong the question is always the same: was it a person, the agent following a legitimate instruction, or the agent manipulated by malicious content?
TheAI Act requires high-risk systems to allow "the automatic recording of events" over their lifetime (Art. 12), and providers and deployers to keep logs for at least six months (Art. 19 and Art. 26). It says nothing about how to make them tamper-proof. And the problem exists outside high risk too: an agent that places orders or edits master data is not "high risk" for the law, but it is for the company.
Here the register of the agent's actions is a natural candidate for a chain of fingerprints with external anchoring: every action recorded with its context, none deletable without leaving a trace.
Which blockchain: public for anchoring; permissioned only if the register has to be shared with other organisations.
Supply chains and sensors: data shared among parties who do not trust each other
This is the ground on which blockchain promised the most, and where TradeLens stopped. It is topical again because of the digital product passport, which the European ecodesign regulation wants based on "a decentralised data system" run by economic operators, with "data authentication, reliability and integrity" ensured (recital 41 and Art. 10). The word blockchain does not appear in the regulation: the legislator asks for the result, not the technology.
Where several companies must write to the same register without any one of them controlling it, a permissioned network makes sense. But the guarantee is only as good as the source: if the temperature sensor is tampered with, the blockchain carefully preserves false data. Security is decided on the device, which must sign the data at the moment it measures it.
Which blockchain: permissioned among the companies in the chain, provided there really are several companies willing to take part.
Where it is not needed
The same European data protection guidelines include, among their final recommendations, a question that applies to any project: why is blockchain a necessary and proportionate means, and what were the alternatives? Often the honest answer is that a simpler alternative existed.
Documents and invoices: the problem is upstream
The typical fraud of the AI era is the perfect email from a supplier changing the IBAN, or an invoice identical to the real one. Recording it on a blockchain achieves nothing: the fake would enter the register with a very certain date. The point is to verify who produced the document, and for that, tools with full legal value have existed for years: digital signatures and electronic seals. A qualified seal enjoys by law the presumption of integrity of the data and of correctness of their origin (eIDAS, Art. 35); a qualified time stamp, of accuracy of date and time (Art. 41). And for a changed IBAN the oldest countermeasure still works: a phone call to the number you already knew.
AI data and models: where do they come from?
An AI model is as reliable as the data it was trained on, and data can be poisoned. TheOWASP, a non-profit foundation working on software security through open and free projects and resources, lists data poisoning among the top ten risks of systems based on language models. The countermeasures it recommends are tracking the origin and transformations of data with bills of materials such as CycloneDX, and versioning them: no blockchain.
Transparency logs: the same guarantee, without blockchain
The software world, facing the same provenance problem, chose another path. Website certificates have for years been recorded in append-only public logs , and open source software signed with Sigstore in a transparency log that anyone can monitor to check that it stays append-only and that no entry is changed or removed. These are chains of fingerprints run by a known operator but watched by independent auditors: no distributed consensus, no cryptocurrency. When there is a recognised party willing to keep the register and others willing to watch it, this is the simplest solution.
Private chains among a few
We saw it in section 2: if the nodes are controlled by a single party, or by a group that depends on it, blockchain adds cost and complexity without adding guarantees. A database with a good change log and off-site backups does the same job.
An example we know well: Odoo
We work on Odoo every day, and Odoo already contains a small blockchain without a blockchain. On accounting journals you can enable the option "Secure Posted Entries with Hash": from then on every posted entry receives a SHA-256 fingerprint computed on its data (number, date, journal, accounts, amounts, partner) and on the fingerprint of the previous entry. It is the same principle as in section 2: changing a past entry breaks all the ones that follow. You can see it in the source code, which is public.
It is a good feature, with two things to know. The first: it is off by default, and once enabled on a journal with secured entries it cannot be turned off. It should be a decision, not a discovery. The second is subtler. The computation uses no secret key, and the chain lives in the same database it protects. Anyone with write access to the database, an administrator or an attacker who has taken their place, can change an entry and recompute all the following fingerprints; Odoo's integrity check will then come out clean.
Odoo itself provides the missing piece, though it does not call it that: the "Data Inalterability Check" produces a report with the first and last fingerprint of each journal. If that report is kept outside the system, the chain protects for real again: a later recomputation would produce a last fingerprint different from the stored one. The question is where to keep it so that nobody can replace it.
And that is exactly the job a public blockchain anchor is suited for. Periodically you take the last fingerprint of each journal and record it with a tool such as OpenTimestamps: only a fingerprint of a few bytes leaves the company, no accounting entry, and from then on anyone can verify that on that date the accounts were exactly those. The same scheme extends to server logs and to the register of actions of AI agents working in the business system.
You do not need a blockchain platform, a token or a consortium: you need a chain of fingerprints that the business system partly already has, and a place outside your own control to deposit its tip.
In short: fewer promises, a precise job
Artificial intelligence has not made blockchain necessary. It has made it necessary to prove the integrity of data, documents and records in a world where forging them is cheap and where autonomous software acts with people's permissions. For part of this work blockchain is the right tool: when you need proof that anyone can verify, that does not depend on the word of whoever holds the data, not even ours.
For the rest there are simpler tools, often with more legal weight: qualified signatures and seals, qualified time stamps, transparency logs, external immutable archives. The right question to ask is the one in the European guidelines: why a blockchain specifically, and what were the alternatives?
If there is a convincing answer, the renewed interest is deserved. If there is not, it is the same enthusiasm as a few years ago with a new label.
Blockchain and artificial intelligence: where it really helps (and where a signature is enough)