There's a divide running through Norwegian working life right now, and it follows neither industry, age nor job title.
On one side are those who use Claude or ChatGPT as a better search engine. They ask a question, get an answer, copy what they can use, and go back to the system where the work actually happens. The tool is useful, but it sits beside the job.
On the other side is a smaller group that has done the opposite. They've moved the work itself into the assistant. They don't leave it to carry out the task. The systems come to them.
It looks like a skills gap. It isn't. It's two different ideas about what the tool is for, and the two groups see completely different futures because they already live in different presents.
The numbers say most people are in the first group
This isn't a gut feeling. It's been measured.
The report from NHO, the Confederation of Norwegian Enterprise, and Samfunnsøkonomisk Analyse, published in January and covering around 4,000 Norwegian businesses, shows that 55 percent use AI. That's more than a doubling since 2023. But the report also says this outright: AI is only to a small degree an integrated part of work processes.
An American study from September this year, covering nearly 14,000 workers, finds the same pattern even more clearly. Generative AI is used in over 80 percent of occupations, but fewer than 3 percent of work tasks have above 50 percent adoption. No single task passes 70 percent.
And the UK Department for Business and Trade, which measured actual Copilot use against a control group, ended at an average of 1.14 actions per user per day.
Broad use, shallow use. Almost everyone has picked the tool up. Very few have put their working day inside it.
What the second group actually does
What's happening in the smaller group is something other than faster answers.
A finance manager asks the assistant to pull the general ledger, find the discrepancies, compare against last year and write the memo. The accounting system is never opened. A project manager has the assistant read the tender documents, pick out the relevant items and set up the supplier enquiries. A quality manager has the entire documentation pile read and checked without opening a single PDF.
This isn't a prediction. It's described by a Norwegian practice system vendor, about their own product. Tripletex launched an official MCP server in May, meaning a connection that lets Claude and ChatGPT work directly inside the accounting system, and writes the following on their own pages:
"You no longer need to click your way through the system to register hours, create invoices or check trial balances."
Charlotte Thjømøe at Skya Regnskap put it to Økonomi24 better than any vendor has managed: "That way Tripletex suddenly becomes a colleague rather than a system."
That's a Norwegian vendor with 160,000 customers saying you don't need to use their interface. Worth noting: PowerOffice, Fiken and Xledger have nothing equivalent officially, for now. The bridges to them are built by third parties, and most can only read, not act.
How we think the work divides
Looking ahead, we think business processes split into two tracks, and that the two need completely different things from us.
One track is the processes that can run without people at all. Simple customer enquiries, document processing, reporting, reconciliations, data sorting. Fast, identical every time, and nobody sits watching while it happens. The value is volume and speed.
The other track is the processes that require professional judgment. Here the human doesn't disappear, but the workplace moves. We think more and more people will do this kind of work from Claude or ChatGPT, connected to the practice systems, rather than inside them. The value is quality, and how much ground one professional can cover.
It's the second track that changes most, and it's the one that gets misunderstood.
The big players have already taken a position
If this sounds like a marginal debate, look at what the platform companies have done over the past twelve months.
Marc Benioff wrote on X in April, when Salesforce launched Headless 360:
"No browser required. Our API is the UI. Entire Salesforce & Agentforce & Slack platforms are now exposed as APIs, MCP, & CLI."
Co-founder Parker Harris put it even more plainly the same week: "Why should you ever log into Salesforce again?" That's the world's largest CRM vendor telling its customers they don't need to log into the product.
Satya Nadella had said something similar back in December 2024:
"SaaS applications or biz apps, the notion that business applications exist, that will probably collapse in the agent era. Because if you think about it, they are essentially CRUD databases with a bunch of business logic."
And the connection layer has become infrastructure. MCP, the standard that lets the assistant talk to the practice systems, was built by Anthropic, who gave it away in December 2025. It's now governed by the Agentic AI Foundation, a new foundation that Anthropic, OpenAI and Block set up together under the Linux Foundation. That's the same umbrella that governs Linux and Kubernetes, which is where infrastructure ends up when no single vendor is meant to own it.
Among the members you'll find Salesforce, SAP, Oracle and IBM. The practice system vendors are themselves helping govern the protocol that lets the work move out of their own interface. It's worth considering why they do that.
The nuance that makes the simple picture wrong
Here we have to correct ourselves, because the claim that the interface disappears is too simple.
The interface doesn't disappear. It changes host.
In January came MCP Apps, the first official extension to the protocol, which lets tools return interactive components rendered inside the conversation. The vendor's interface moves into the assistant instead of going away. At the same time the movement runs the other way: Claude now runs inside Excel, Word, PowerPoint and Outlook.
So what's actually happening isn't that screens die. It's that the question of who owns the surface has been reopened, for the first time since the browser won.
What the real story is
But look at what the same companies built while they argued about the interface.
Microsoft launched Entra Agent ID and Agent 365, which they describe as "the control plane for agents". Google launched the Gemini Enterprise Agent Platform with Agent Identity, Agent Registry and Agent Gateway. ServiceNow launched AI Control Tower and an AI Gateway. Workday has Agent System of Record and Agent Passport. OpenAI launched Frontier and Presence. Anthropic launched Managed Agents with what they describe as "scoped permissions, identity management, and execution tracing built in".
And the protocol itself followed. In June, Enterprise-Managed Authorization became stable in MCP, so businesses can govern agent access centrally through their own identity provider. Agent identity sits at the top of the protocol's roadmap.
Six competitors built essentially the same product within twelve months. That's no coincidence. That's an industry that has concluded where the value is moving.
And it isn't moving to the interface. It's moving to the layer underneath: identity, access, rules, logging and approval.
Where the human remains
This is the point we think is under-communicated, and one we've been too poor at stating clearly ourselves.
When we say we don't build user interfaces, we don't mean there is no surface. We mean the surface moves from being where the work is done, to being where the work is governed and verified.
That isn't only our claim. Jared Spataro at Microsoft wrote it in the Work Trend Index in May:
"As agent use increases, human involvement doesn't disappear, it changes shape. What rises is the need for humans to set direction, define standards and evaluate outcomes."
Anthropic writes about Cowork that "nothing ships until you've reviewed and approved it". OpenAI writes about Presence that "the company sets the policies: what the agent can do, when it needs approval, and when a person should take over". And Tripletex, in their own technical documentation, says it straight to the user: "Always review what the AI does before confirming."
Three platform companies and one Norwegian practice system vendor saying the same thing. The human moves from carrying out to setting direction and approving.
That's also why the governance surface and governance itself are the same thing. They aren't two layers stacked on each other. They're one layer with two names.
Why this is hardest in regulated industries
In accounting and audit this gets very concrete.
The auditing standard for documentation has a test called the experienced auditor test. The working papers must enable an experienced auditor with no prior knowledge of the engagement to understand what was done. A chat thread does not pass that test. It isn't deterministic, it can't be reproduced, and it isn't a working paper.
For these industries the governance surface therefore isn't a bonus. It's the deliverable itself. The requirement isn't that the work gets done. The requirement is that someone else can reconstruct that it was done correctly.
The counterarguments, which are better than we'd like
We don't want to write a post that only confirms what we already believe, so here's what argues against it.
Gartner measured in April that 17 percent of businesses have actually deployed AI agents, while over 60 percent plan to. The gap between intention and production is large, and it's easy to mistake one for the other.
Software spending is growing. Gartner estimated in July that the global software market is growing by 15.5 percent this year, to around 1,470 billion dollars. If the value really were leaving the practice systems, that number would point the other way.
And Aaron Levie at Box has formulated the best counterargument we've read:
"The power of software is we've codified our workflows and business processes into a system that will run the same way every single time. I don't think that reduces in value."
He's right. Determinism doesn't become less valuable as the models get better. It becomes more valuable, precisely because the models don't have it.
What we're left with
The thesis holds, but in a narrower form than it's usually stated.
The interface for carrying out the work is moving, and it's happening now. The interface for governing and verifying it isn't moving away. It's becoming more important, and it's where the large companies have put their money over the past year.
But it's happening slowly, and unevenly. The connections are in production in Norway already, while the Norwegian AI act won't reach parliament before spring 2027 and the EU's high-risk rules are deferred to December that same year. That gap isn't an argument for waiting. It's an argument for building the governance yourself, before someone demands it of you.
And if you read this and recognised yourself in the first group, the one using AI as a better search engine, that isn't a problem. It's where almost everyone starts. But it's worth knowing that the difference between the two groups isn't primarily about the tool. It's about how much structure sits underneath it.
AI doesn't start with technology. It starts with structure.
Get in touch for a free AI Ready assessment →
Sources
NHO and Samfunnsøkonomisk Analyse, report 1-2026, January 2026. Bick, Blandin, Deming and Schumacher, NBER Working Paper 35677, 1 September 2026. UK Department for Business and Trade, evaluation of Microsoft 365 Copilot. Tripletex, ai-og-automatisering and developer.tripletex.no, May 2026. Økonomi24, 14 April and 12 May 2026. Benioff on X and Salesforce Headless 360, 15 April 2026. Nadella on the BG2 podcast, December 2024. MCP donated to the Linux Foundation, 9 December 2025. MCP Apps, 26 January 2026. Enterprise-Managed Authorization, 18 June 2026. Microsoft Work Trend Index 2026, 5 May 2026. Anthropic on Cowork, July 2026. OpenAI Presence, 22 July 2026. Gartner Hype Cycle for Agentic AI, 15 April 2026. Gartner IT forecast, 27 July 2026. Aaron Levie in diginomica, 9 February 2026. Regjeringen.no on the Norwegian AI act, 4 August 2026.
Read also: AI Agents and Secure Access: Entra Access Packages · The System That Reads the Quality Documentation Before the Inspector Does