Introduction
Following the rise in popularity of LLMs and generative AI, it didn’t take long for people to start asking what this means for XBRL.
- Would there be exciting new opportunities making better use of AI with XBRL?
- Would AI reduce the cost of XBRL adoption?
- Would AI eliminate the need for XBRL entirely?
If you read the opinion pieces written to answer these questions, there are different views and strong voices on all sides. Have a look at the quotes below and then see our view of where AI sits as a complementary technology to the digital reporting process powered by XBRL technology:
“[We recommend r]emoval of the electronic tagging requirement (iXBRL tagging) of the ESEF Regulation for financial and non-financial reporting. Tagging is complex, causes a great deal of effort for companies and creates legal uncertainty without any additional benefit for the public. Sophisticated AI tools can now also extract, interpret, and evaluate the reported information from PDFs without having to rely on electronic labelling in accordance with iXBRL.”
Deutsches Aktieninstitut: Better competitiveness through lean sustainability reporting
“Following [ESEF] formatting and tagging requirements is especially burdensome for undertakings which have not been subject to electronic reporting requirements so far.
ii. Suggested Solution
Given the rapid development of text analyzing tools based on AI, there should be an evaluation as to whether the requirements are still necessary and reasonable in view of their intended goal. If this evaluation concludes that there already are sufficient text analyzing tools to exploit information more efficiently, the requirements of Article 29d of the Accounting Directive should be eliminated.”
Institut Der Wirtschaftsprüfer: Reduction of bureaucratic burdens in sustainability reporting (pages 12-13)
“It’s also time to eliminate outdated ESEF-tagging.”
Deutsche Post DHL: Takeaways from “Simplification Kick-off Event”
“XBRL isn’t a burden; it’s the fuel for AI-powered analysis, ensuring accuracy, comparability, and decision-usefulness. Ditch structured data in favour of AI scraping PDFs? Sure—if you want your investment models built on guesses and markets running on vibes.”
XBRL International: XBRL is outdated? That’s news to AI
“This presentation … explores why structured data and AI are not competing forces but rather complementary tools in financial reporting. While AI can analyze vast amounts of unstructured data, structured formats like XBRL ensure accuracy, comparability, and regulatory compliance—things that AI alone still struggles with.”
XBRL Europe: Do We Really Need XBRL and Structured Data in the Age of AI?
Using AI to replace iXBRL tagging
Tagging is undoubtedly one of the more time-consuming activities in an iXBRL programme so the desire to reduce this effort is understandable. Advanced tagging solutions such as Seahorse have already been using AI — usually with what is now known as “traditional Machine Learning” — for years along with other features to greatly reduce tagging effort compared to more naïve solutions.
But it’s important to understand that there’s a natural complexity to the tagging problem which isn’t a consequence of the technology at all. Tagging isn’t just a technical means to identify data for extraction; it is an explicit declaration which reflects a set of judgements about what the data means. There is value in knowing that the preparer has explicitly made those judgements and is willing to put their name to them, and they have to be communicated in a way that is unambiguous.
If you remove this requirement in favour of extraction by data consumers, then these consumers assume all the risk. Misrepresentations become misinterpretations and you introduce opportunities for obfuscation.
That is not to say that a wider range of AI tools can’t be used to improve the overall user experience and efficiency. They can, and this is an area of intense focus for Seahorse development. These features work best when they are designed to work as part of the tagging experience as a whole, not as part of an AI tick-box exercise which simply moves the problem elsewhere (increasing review effort or reducing data quality).
Using AI for analysis
AI has many suitable applications related to XBRL, but replacing preparer declarations is not one of them. It is prone to hallucinations and vulnerable to manipulation which is even more of a concern if the responsibility has been shifted to the consumer.
We see great potential in using AI to aid in the analysis of reports; not in place of preparer tagging, but to derive the most value from it. AI is at its best when it has well-defined input data. iXBRL’s distinguishing feature — that it encapsulates the human-readable text, machine-readable data and mapping to a precise definition, all while retaining the full document context — makes it an ideal technology to support AI-enabled analysis.
The impact of AI on ROI
Regulation and transparency legislation are increasing and the additional data collection required does come at a cost.
So, it is tempting to try to reduce the burden of these growing demands by saying the data needs to be reported, but not tagged. This is a false economy.
This approach keeps all the cost of capturing and preparing the data, but without a standard and transparent mechanism for relating that data to the regulator-defined model. Without this clarity, the chance of miscommunication sky-rockets. Deprived of the validation mechanisms built into the XBRL standards, preparers, auditors, regulators and analysts will each produce their own independent (and therefore inconsistent) processes to perform validation.
In other words if the data is worth producing and collecting, it’s worth tagging.
Complexity
The complexity of ESEF is cited as a reason to turn to AI but most of the complexity in ESEF tagging comes from the rules in the Filing Manual — which are not defined by the XBRL standards but are layers of bespoke complexity piled on top. The ESEF ones are particularly arduous and have been quite ambiguous at times.
It is notable that those who complain loudest about this complexity are also those who took a rigid position on block tagging (counter to many others) and therefore chose to increase the burden rather than reduce it.
For us, the lesson here is unrelated to AI, instead data collectors should use the standards to greatest effect and minimise non-standard requirements.
Volume
There’s certainly an argument for saying that in some cases the cost of collecting the data exceeds the potential value you might derive from it. It is a mistake to mis-attribute this cost to tagging and analysis of the costs of reporting show a tiny proportion of these are attributable to the format of the report.
In fact, especially where the justification for data collection is under fire, there will be a high incentive to minimise the effort reporting that data fully and correctly. If the data is needed, then tagging provides important data quality protections, in terms of ensuring an explicit declaration, and providing robust validation.
Again, we take the view that if the data is worth collecting, it’s worth tagging.
New, high-quality data is required to avoid hallucinations
If the tagging AIs of today — which learn from experts’ tagging decisions — were replaced by AIs which learn from their own prior decisions, then you would start to see generation loss.
Without the feedback loop from human experts, the AI will produce something further and further away from a reasonable result.
This also means we can’t judge the output of such a system on the quality of results it produces today where all of the training data is produced by humans, if the whole value proposition depends on being able to remove those humans from the process.
XBRL is “outdated”
As others have pointed out, a technology with a long history might be outdated, but might equally be mature, stable and reliable. One would not suggest we stop using the internet or toilets or fire simply because it’s yesterday’s tech.
There is a long trail of “next big things” which failed to deliver — or, being slightly more generous, proved to be a solution to a range of well-fitting problems, not a panacea.
AI is neither a silver bullet nor a dead end. If applied with good judgement, it has potential to be an excellent tool to improve quality and efficiency. If applied indiscriminately, it represents a significant risk to the quality of data used for regulation and decision making.
So, what do we do?
AI has great potential to move XBRL forward, if it is applied with good judgement. In summary:
- Don’t defer all judgement to AI
- Protect the role of user feedback
- Use AI as part of a suite of cooperating features to streamline tagging
- Use AI for analysis, in a way which capitalises on the structured data (i)XBRL provides