From 'we've tried ChatGPT' to AI in operation: the gap in Danish companies
Many Danish companies experiment with AI, far fewer have AI in actual operation. Here is the difference between the two states, why the gap arises, and the staircase that leads from one to the other.
Table of contents
What is the difference between having tried ChatGPT and having AI in operation?
The problem is that "tried" is AI in a browser tab with no access to your data, while "in operation" is AI embedded in a fixed workflow with controlled access to your systems and a named owner.
The gap between having tried ChatGPT and having AI in operation isn't about better models, it's about access, ownership and measurement. "Tried" is AI in a browser tab with no knowledge of your data. "In operation" is AI embedded in a fixed workflow, with controlled access to your systems, a named owner and a measured effect. That is the gap Agent360 helps companies close through AI implementation.
Most Danish companies have by now tried AI: a manager has tested ChatGPT for a draft, an employee summarises meeting minutes, marketing has experimented with copy suggestions. But it's a long way from having tried AI to having AI in operation, meaning AI that is embedded in fixed workflows, has access to the company's own data and delivers a measurable effect week after week. That gap is the most important explanation for why many companies experience enthusiasm without results: the experiments create aha moments, but never a change in operations.
AI in operation is when artificial intelligence is part of a recurring workflow with access to the company's own systems and data, with defined responsibility, an owner and a measured effect, as opposed to one-off experiments in a browser tab.
The figures confirm that AI use is growing fast in Denmark: according to Statistics Denmark, the share of companies with at least 10 employees that use artificial intelligence rose from 15 per cent in 2023 to 28 per cent in 2024. But the same survey shows what the AI is used for: the most widespread uses are text production (18 per cent) and text analysis (17 per cent), while only around one in ten companies uses AI to automate workflows. In other words: adoption is growing, but for many it is still about text tasks on the edge of the business, not about AI as part of the machinery itself.
This article is about the gap between the two states: why it arises, what "operation" concretely means, and what staircase leads from one to the other.
Where Danish companies actually stand
Let's stay with the official figures for a moment, because they paint a precise picture of the situation. Statistics Denmark's survey shows a doubling in AI use in one year, from 15 to 28 per cent of companies. That is remarkable growth, and it means AI is no longer a pioneer phenomenon: more than one in four companies is under way.
But if you look at what companies use AI for, the gap emerges. Text production and text analysis (the classic ChatGPT tasks) dominate, while automation of workflows sits down at around one in ten companies. That is the difference between using AI as a personal writing tool and using AI as infrastructure.
There's nothing wrong with the writing tool. But the value of a draft here and there is limited and hard to measure. The big value lies in the other category: AI that works in the business's systems: updating data, answering enquiries, compiling reports, keeping track of follow-ups, every day, without anyone having to remember to ask.
Why don't experiments become operation?
When so many companies are under way, why do relatively few get beyond the experiment? Four barriers recur.
1. The AI has no access to the company's data
A chatbot in a browser tab knows nothing about your customers, orders, agreements or history. It can write well in a generic way, but nothing company-specific. As soon as the task requires knowledge of your reality ("write a follow-up proposal for the customers who haven't replied in 14 days"), the experiment grinds to a halt, because the AI can't see into the CRM system, the finance system or the inbox. Without data access, AI remains a writing tool. We have previously described how data quality and CRM hygiene decide what AI can achieve at all with the company's data.
2. Nobody owns the experiment
Most AI trials start with a single curious employee. That's a fine start, but without an owner the trial never becomes a workflow. Nobody is responsible for the AI solution running tomorrow, nobody measures the effect, nobody maintains it. When the enthusiast gets busy or changes job, the experiment quietly dies.
3. Dependence on individuals instead of process
Related to ownership, but more insidious: even successful experiments often live only in one person's head, as a personal collection of prompts or a private workflow. The company hasn't gained a capability; a single employee has gained a superpower tool. It doesn't scale, and it disappears with the person.
4. Uncertainty about compliance and data security
Many managers put their foot on the brake out of a sensible instinct: are we even allowed to do this? What happens to data entered into external AI services? What does GDPR require? And what does the EU AI Act, which we have reviewed for Danish companies, mean? The European Commission's rulebook is risk-based: the requirements depend on what the AI is used for, as described on the Commission's official page on the AI Act. The uncertainty is understandable, but the consequence is often that AI is banished to harmless (and therefore valueless) tasks, instead of being implemented properly with data handling under control.
What "AI in operation" concretely means
It's worth being precise, because "we use AI" covers vastly different realities. AI in operation has five hallmarks:
- Recurring: the AI solves the task every day/week, not when someone happens to remember.
- Data-connected: the AI has controlled access to the systems the task requires: CRM, mail, calendar, finance.
- Owned: a named person is responsible for the workflow, its quality and its further development.
- Delimited: it is defined what the AI may do itself, and what requires human approval.
- Measured: the effect is quantified (time saved, faster response times, fewer errors) so the decision to continue is data-driven.
If one of the five is missing, it is an experiment. That's not an insult, experiments are the right start. But they must be recognised for what they are: step one on a staircase.
The difference between an AI experiment and AI in operation, dimension by dimension
| Dimension | Experiment ("we've tried ChatGPT") | AI in operation |
|---|---|---|
| Frequency | Ad hoc, when someone remembers | A fixed part of a recurring workflow |
| Data access | None, only what is typed in manually | Controlled access to the company's systems |
| Ownership | An enthusiast, informally | Named owner responsible for quality and operation |
| Knowledge | Lives in one person's prompts | Embedded in shared, documented workflows |
| Approvals | Undefined | Clear limits on what the AI may do itself |
| Effect | Anecdotal ("it works very well") | Measured and quantified on an ongoing basis |
| Survival | Dies when the enthusiast gets busy | Keeps running, it's part of operations |
The staircase: from browser tab to fixed workflows
The road from experiment to operation isn't one big leap, but a staircase with three steps. Most companies can recognise which step they're on.
Step 1: a prompt in the browser
AI is used as a personal tool: drafts, summaries, translations, idea development. Data is typed in manually, results are copied out manually. This is where the vast majority start, and it's a fine place to learn what language models can and can't do. The limitation is structural: without access to the company's systems, the AI can never take work off the operation itself.
Step 2: an agent with access to your systems
The leap from step 1 to 2 is the decisive one: the AI gets controlled access to the company's own data, typically via connectors to the systems where the work actually takes place. Now the tasks can become company-specific: "find the quotes that haven't been followed up", "compile last week's customer enquiries by topic", "update the contact with the notes from the meeting". Here the AI changes character from a writing tool to a colleague with system access, and here questions of access control, approval gates and data handling become really important. How workflows are concretely automated at this step we have covered in the guide to automating sales processes.
Step 3: fixed workflows with ownership and measurement
At the last step the AI tasks are no longer something somebody "runs", they're part of operations. The workflows have owners, limits and measurement points. New tasks are added as the organisation learns what works. This is where the effect turns into bottom line instead of enthusiasm: the time that used to go on routine tasks is genuinely freed up, week after week.
What the leap to operation requires
Three building blocks recur at companies that get from step 1 to step 3 in one piece.
Data access via connectors, not copy-paste
If AI is to work with the company's data, the connection has to be systematic: connectors that give the AI controlled, logged access to CRM, mail, calendar and the other systems. Manual copy-paste doesn't scale and is in practice also a bigger data security risk than a managed integration, because data flows uncontrolled between systems and clipboards.
A place to live: your own instance and control over data
The compliance uncertainty from earlier has a structural solution: run the AI somewhere you know where data lives. If every company gets its own instance with its own database, the answer to "where is our data?" becomes short and verifiable. It doesn't remove the responsibility for good data handling, but it makes it possible to take responsibility concretely instead of postponing the whole AI effort out of uncertainty.
Governance: limits and approvals from day one
Trust in AI in operation is built with limits: define what the agent may do itself (read, compile, suggest), and what requires a human yes (send, change, delete). With clear approval gates you can start carefully and widen the agent's room for manoeuvre as it proves itself, instead of choosing between blind trust and total restraint. We have written more about the organisational foundation in the guide to revenue operations, where the same principle recurs: clear owners, clear processes, measured effect.
Three recognisable examples of the staircase in practice
The staircase becomes more concrete with everyday examples. The three here are deliberately undramatic: that is precisely the point: AI in operation is about the tasks that eat the hours, not about the spectacular.
Example 1: customer enquiries that need sorting and answering
The step 1 version: an employee occasionally copies a difficult customer email into ChatGPT and asks for a draft reply. Helpful, but random, and the AI knows neither the customer's history nor the company's reply practice.
The operation version: the agent reads the inbox continuously via a connector, categorises the enquiries, looks up the customer's history in the CRM system and has finished draft replies ready for approval: every morning, for a named owner, with clear limits (the agent never sends anything itself). The effect is measured in response time and sorting time saved.
Example 2: meeting follow-up that is never forgotten
The step 1 version: the minutes are written with AI help, if the meeting leader remembers. The follow-up points end up in the minutes, and stay there.
The operation version: after every meeting the agent gathers notes, creates the follow-up tasks in the right systems, updates contacts and agreements in the CRM and reminds the owners of deadlines. No decision drowns in a document. That kind of systematic follow-up is related to what we have described in the guide to automating sales processes.
Example 3: the weekly report that writes itself
The step 1 version: a manager spends Sunday evening gathering figures from three systems and asking ChatGPT to help with the wording.
The operation version: the agent pulls the figures directly from the source systems via connectors, compiles the week's development and has the draft report ready on Friday afternoon. The manager spends their time assessing the content, not moving figures between tabs.
The common denominator in all three: the leap from step 1 to operation isn't "better AI", it's data access, ownership and fixed frameworks around exactly the same kind of task.
What about the tasks AI shouldn't have?
A credible AI-in-operation strategy is also about knowing what should NOT be automated. Three categories belong with humans, however good the technology gets:
- Decisions with significant consequences for individuals: hirings, dismissals, credit assessments. Here the EU AI Act also sets special requirements for high-risk uses with human oversight, cf. the Commission's review of the rulebook.
- Relationships where authenticity is the very value: the difficult customer conversation, the negotiation, the personal thank-you. AI can prepare; the human has to deliver.
- Tasks without repetition, one-off tasks with a low time cost are seldom worth automating; here the browser tab (step 1) is actually the right tool.
Being able to point to these limits isn't a weakness in the AI strategy, it's proof that it has been thought through. It also creates the employee trust that all operational AI ultimately depends on: people cooperate with the agent when they can see where its mandate ends.
Getting started: a checklist for the leap
If you want to go from experiment to operation, here is the short checklist:
- Choose one recurring task with a clear time cost, not ten tasks, one.
- Give the task an owner with the mandate and time to see it through.
- Establish the data access via connectors to the systems the task requires, never via manual data transfer.
- Define the limits: what may the AI do itself, what requires approval?
- Measure before and after: note what the task costs in time today, and quantify the effect after four weeks of operation.
- Decide on data: if it works, extend to the next task; if it doesn't, adjust or close, but don't let it fade away.
The most important rule of thumb
If you only take one thing from the checklist, take this: choose small, measure honestly, expand on evidence. Most failed AI ventures share the same story: they started too broadly ("we need AI across the whole organisation"), with no owner and no measurement point, and ended up as an expensive report in a drawer. The successful ones share a story too: one concrete workflow, one owner, four weeks of measurement, and from there a staircase of expansions where each step was justified by the measured effect of the previous one. It's less impressive to talk about at the management meeting. But it is how AI ends up in operation instead of in the drawer.
Frequently asked questions
Q: Isn't it fine just to use ChatGPT for drafts and texts?
Yes: as a personal tool it is both useful and a good learning step. But the value is limited to text tasks on the edge of the business. The big, measurable effect only comes when AI gets access to the company's own data and becomes part of fixed workflows, and that requires a deliberate leap, not just more use of the browser tab.
Q: How many Danish companies use AI today?
According to Statistics Denmark, 28 per cent of companies with at least 10 employees used artificial intelligence in 2024, almost a doubling from 15 per cent the year before. The most widespread uses are text production and text analysis, while only around one in ten companies uses AI to automate workflows.
Q: What is the first step from experiment to operation?
Choose one recurring task, give it an owner, and establish controlled data access to the systems the task requires. It's the data access that separates an experiment from an operational solution; without it AI remains a writing tool.
Q: Is it legal to give AI access to our customer data?
Yes, when it is done with data handling under control: you need to know where data is processed, have data processing agreements in place and manage access. The EU AI Act also sets risk-based requirements depending on the use, see the European Commission's official review. An architecture with your own instance and your own database makes the documentation considerably simpler.
Q: Do we need to build AI skills in-house before we can get AI into operation?
You don't need to be able to build language models, but you do need an owner who understands the workflows and can assess the AI's output. The technical heavy lifting (data access, integration, governance) can sit in the platform; the business understanding can't.
Q: How do we measure whether AI in operation pays off?
Measure the concrete workflow before and after: time spent per week, response times, error rates or throughput time. Quantify the effect after a fixed period (e.g. four weeks) and let the figures decide whether the solution should be extended, adjusted or closed. Anecdotes ("it feels faster") are not a basis for decisions, and that very discipline is the difference between an AI effort that grows on evidence and one that runs out of momentum once the enthusiasm has died down.
#ai-strategy#implementation#automation
Contents
Which of your manual steps could disappear?
We map where data is moved by hand between your systems, and what can realistically be automated.
Book a mapping sessionRelated articles

Own or rent software? A decision guide for Danish businesses
Own or rent software? Cost over time, data lock-in, GDPR and exit, self-host and fair source, plus a concrete decision framework for Danish businesses.

Claude Code in a business: what an AI developer agent can do, and what it requires
What Claude Code, Anthropic's agentic developer tool, can do in a business, and what it requires in security, governance and organisation.

AI strategy for businesses: a step-by-step guide
A practical guide to an AI strategy: map workflows, choose use cases, fix your data, run a pilot, bring staff along and know when not to automate.