On 30 December 2025, Denmark’s postal service delivered the last letter in its 401-year history. The 1,500 remaining postboxes came off the streets during that year. Meanwhile, in July 2026 the German federal cabinet approved GeDIG, the healthcare data and digital innovation act, which phases the fax out of doctors’ offices by autumn 2029 and makes electronic referrals mandatory.
So, how is it that one country can implement such changes while another is still phasing out its reliance on fax machines in this century? Factoring in each country’s use of AI in business, 42% of Danish firms used at least one AI technology in 2025, the highest share in Europe. Germany stood at 26%, above the EU average but sixteen points behind Denmark.
To me, the contrast nudges us to revisit our beliefs around AI adoption. Particularly the belief that model choice is the starting point of AI adoption, ignoring the condition of the systems around the model. Both countries have access to the same models and cloud providers, but the difference is whether business and public processes already produce reliable digital records. Since an AI agent does not work with reality directly, but with the version of reality available in a system, naturally this causes some constraints.
The First Constraint: The Information Was Never Digitised
An agent cannot retrieve a document that never entered the system. The shape of the organisation’s information flow determines what an agent can even perceive as a task.
At the country level, Eurostat data shows that business AI adoption correlates with the digitalisation of administrative processes. But the key mechanism is that countries where citizens already receive official documents digitally have fewer off-system interactions for firms to reconcile, reducing the amount of interpretation work required before AI can be safely used in operations.
On the firm level, a 2026 IZA study on 1,907 Danish companies finds that a one-standard-deviation increase in digitalisation is associated with an 8 to 10 percentage point higher probability of adopting core AI systems. This suggests that AI adoption depends on whether the firm’s internal operations are already organised in a digitally structured, machine-readable way. If processes are fragmented, there is little stable information for AI systems or agents to act on.
This may also help explain why policy-driven digitisation does not automatically translate into AI readiness. Italy’s mandatory B2B e-invoicing created a fully digital channel, but overall AI adoption remains below the EU average. Digitising a single step does not eliminate the surrounding analogue system.
The Second Constraint: The Information Exists Only in Someone’s Head
Humans run operations, which means most of the time they know details that the AI does not. A finance administrator knows that a tenant paid in cash, but the payment was never entered into the accounting system. As a result, the bank records and the system do not reconcile. The administrator understands why, whereas AI only sees inconsistent data. A property manager remembers that a landlord asked not to be contacted for two weeks. A team knows the workaround for one building or client, but nobody has documented it.
ClarEval tested tasks with missing goals, missing premises, and ambiguous terminology. Model performance fell from about 89% after clarification to roughly 9% when ambiguity remained unresolved. Meanwhile, Ambig-SWE found that allowing a model to question a user who possessed the missing information recovered up to 74% of the lost performance. This implies that engineers should be building a system that’s capable of asking for clarification from the get-go.
The Third Constraint: The Information Exists but Cannot Be Passed
Some data is available inside the company and still cannot reach the model because of privacy rules, security controls, contracts or internal policy.
A study based on seven years of cloud data found that GDPR—the EU’s core data privacy law—reduced data storage by 26% and processing by 15% among EU firms relative to comparable US firms. In other words, privacy regulation made firms less data-intensive by limiting the amount of information they could retain and process for analytics and AI. Generative AI adoption was almost equal, at 37% of EU firms and 36% of US firms. But the gap in the depth of AI adoption is vast, with 81% of US firms using AI in at least two internal processes, compared with 55% in the EU. One possible reason is that information which already exists inside organisations is often harder to reuse for AI in Europe. Customer support conversations, for example, often contain personal data collected to resolve customer issues. Under the GDPR’s purpose-limitation principle (Article 5(1)(b)), this data cannot simply be repurposed for AI training without a valid legal basis. US firms generally face fewer equivalent federal restrictions on reusing such data.
How to Build the Correct System
Retrieval only works when the answer exists in the system. Agents should ask when information is missing, never guess, and turn calls or informal decisions into structured records. Companies should also remove parallel paper workflows because if a task can happen outside the system, part of the context will remain invisible.
Each company must reckon with its own gaps. Measure whether a fact needed to act correctly was within the system or not. Test the workflows with and without the excluded privacy review fields. Lastly, for the workflow that can be completed on paper, that’s the agent to build last.
Agents may improve human output, but they do not repair what’s missing.
Credit: Source link
