Memory(One)
Get in touch
Back to insights

Applied AI

AI adoption is not AI capability

Many companies now use AI, but fewer have built real AI capability. The difference is whether AI is connected to useful data, workflows, systems, access control and operation.

Published
  • AI Readiness

Denmark is one of Europe’s strongest AI adopters on paper. Eurostat reported that 42.0% of Danish enterprises used AI technologies in 2025, the highest share in the EU. Statistics Denmark also shows that AI use rises sharply with company size: 75% of companies with more than 250 employees used at least one type of AI-based technology in 2025, compared with 37% of companies with 10-49 employees.

Those numbers matter. They show that AI is no longer a niche technology discussion. But adoption is not the same as capability.

A company can have many employees using ChatGPT, Copilot or embedded AI features inside vendor systems without having built anything that changes how the business actually operates. The more important question is not whether AI is being used. It is whether AI is connected to the company’s data, systems, workflows, permissions and operating model.

That distinction matters because the Nordic AI value gap is already visible. BCG’s 2026 Nordic AI research reports that only 4% of Nordic companies see meaningful ROI, defined as returns of at least five times their AI investment. The same research warns that Nordic companies allocate a disproportionate share of AI investment to off-the-shelf productivity tools rather than transformative, end-to-end use cases.

Three levels of AI use

A practical way to separate AI adoption from AI capability is to look at three levels.

1. Personal productivity AI

This is the easiest form of AI adoption. Employees use tools to draft text, summarise documents, translate, create ideas, analyse snippets of information or support everyday tasks.

This can be useful. It can save time and improve individual output. But it usually does not change the company’s systems. It does not automatically improve the quality of internal data, remove duplicate data entry, connect the CRM to the ERP or make a manual approval flow reliable.

The business benefit is often fragmented because the work happens at the edge of the organisation, one user at a time.

2. Embedded vendor AI

Many business platforms now include AI features. CRM systems, HR tools, ERP platforms, support tools, development tools and marketing systems increasingly contain summarisation, recommendations, classification or assistant features.

This can also create value, especially where the vendor system already contains structured, useful data. The limitation is that embedded AI usually works inside the boundaries of one product. It may not understand the company’s full workflow, internal exceptions, data from other systems or practical operating constraints.

Embedded AI can improve a tool without improving the overall system.

3. System-connected AI capability

This is where AI becomes more strategically useful. System-connected AI works with the company’s own data, business rules, integrations and workflows. It may support document processing, reporting, internal search, customer support triage, approval preparation, knowledge retrieval, operational routing or recurring workflow automation.

This level is harder because it depends on foundations that many companies have postponed for years: data quality, integration, access control, logging, system ownership, deployment discipline and process clarity.

It is also where the real value usually is.

Why high-value AI is harder

Useful AI needs useful context. If information lives across spreadsheets, email threads, ageing databases and disconnected tools, AI has little stable ground to work from. A model can generate fluent text, but it cannot safely automate a workflow when the source data is inconsistent, access rights are unclear and no one owns the process.

The same pattern appears in many organisations:

  • reports are assembled manually from exports;
  • customer information is split across CRM, ERP and inboxes;
  • approvals depend on informal knowledge;
  • business rules live in people’s heads;
  • documentation is incomplete;
  • integrations are fragile;
  • data quality is good enough for humans to interpret but not good enough for automation.

In that environment, AI can still assist individuals. But it will struggle to become a reliable part of business operation.

CBS research also points to the organisational side of the problem. In its 2026 article on AI at work, CBS reports large differences between leader and employee AI use, low formal AI training and a lack of shared direction. The issue is not only whether employees have access to AI tools. It is whether the company has made AI concrete and relevant to daily work.

Signs that AI adoption has not become capability

A company may be adopting AI without building capability if:

  • AI use is mostly private, informal or unmanaged;
  • employees use open tools with unclear data boundaries;
  • AI experiments do not connect to systems of record;
  • use cases are selected because they sound innovative, not because they solve recurring work;
  • no one can explain what data AI may access;
  • there is no owner for AI workflows after launch;
  • the business cannot measure whether AI saves time, improves quality or reduces risk;
  • the same manual exports, copy-paste routines and approval bottlenecks remain in place.

These are not reasons to avoid AI. They are reasons to approach AI as software and workflow work, not as a disconnected tool purchase.

What AI readiness should include

AI readiness is often treated as a data question. Data matters, but it is not enough. A practical AI readiness review should look across both technical and organisational conditions.

A useful review should ask:

  • Which recurring workflows create measurable friction?
  • Which data sources are needed for those workflows?
  • Are those sources reliable, accessible and permissioned correctly?
  • Which systems need to be integrated?
  • What decisions require human approval?
  • What logs and monitoring are needed?
  • What cost limits apply if AI calls models or tools repeatedly?
  • Who owns the workflow after launch?
  • How will the business know whether the AI workflow is working?

This turns AI from a broad ambition into a concrete software and operating question.

A practical sequence

The safest path is usually not to start with the model. Start with the workflow.

  1. Identify recurring business work. Find processes where teams repeatedly search, classify, summarise, route, reconcile, report or prepare decisions.
  2. Map the systems and data. Understand which tools, databases, documents and people are involved.
  3. Review access and risk. Decide what AI may read, prepare, recommend or trigger.
  4. Improve the foundations. Fix the data flow, API, documentation or deployment issue that blocks automation.
  5. Prototype narrowly. Test one workflow with clear value and human review.
  6. Operationalise. Add monitoring, logging, cost control, ownership and maintenance.

This sequence is slower than buying another AI subscription, but it is much more likely to create durable value.

The practical takeaway

AI adoption shows that a company is experimenting. AI capability means the company can apply AI inside real workflows with useful data, clear access boundaries, human review and long-term operation.

That capability is built through software architecture, integration, data foundations, application modernisation and practical workflow design. It is not created by a chatbot alone.

Sources and inspiration

Next step

Need a practical route from article topic to working software?

Memory(One) helps organisations review, modernise and build the systems their teams depend on.

Get in touchView servicesRequest a Technical Review