It's the explanations AI takes, not the advisory work

The escape route was mapped out long ago. Bookkeeping gets automated, margins on transactions get squeezed, and the way forward is up the value chain, toward advisory work. The message has been so unanimous it has almost stopped being debated. The problem is that part of the room you're fleeing to can now be filled by the client.

What Norges Bank is actually hearing

In the latest round from Norges Bank's regional network, published September 17, some businesses report that artificial intelligence is already affecting their need for labor. The report describes AI affecting consulting firms unevenly. For some, the technology creates new business opportunities in advisory work and AI implementation for clients. For others, it reduces demand for certain services, because clients increasingly solve the tasks themselves. At the same time, pressure in the labor market has eased, with fewer businesses reporting full capacity utilization and shortages of qualified labor (Norges Bank, Regional Network 3/2026, DN).

An important caveat: Norges Bank's network is talking about IT and consulting businesses, not accounting firms, and the effect is uneven. There's no reason to read this as a measurement of the accounting industry. But it's an early signal of what happens once the client gets access to the same tools as the consultant, and it's worth asking the question before the measurement arrives.

This isn't an isolated finding. Already in Regional Network 3/2025, some reported that clients were using AI for tasks they previously paid for help with, while several also reported that clients needed assistance adopting the technology.

Regional Network 2/2026 makes it clearer: clients are more hesitant, assignments are being postponed or stopped in areas including technology and strategic advisory, and some consulting firms report that AI itself is reducing the need for their services.

The pattern, then, is two forces that can pull the same way: the business cycle and technology. The first passes. The second doesn't.

It's the explanations that are disappearing, not the insight

Capassa reads the signal as a possible price correction on one specific commodity: the explanation. A large part of what gets sold as advisory work is, in practice, knowledge asymmetry turned into a system. The client doesn't know what a cash flow forecast is, how dividend rules work, what makes a good budget model, or how to structure a downsizing process. Someone who knows communicates it, and bills hours for the communication. That marginal cost is heading toward zero. A language model explains just as well, around the clock, with no meeting to book. And it does so without sending an invoice.

What doesn't head toward zero is access to the client's own numbers, structured, reconciled, up to date, and with someone professionally accountable for them. There, a general model has nothing to work with, unless someone has already built the data foundation and the pipeline.

Three layers of advisory work with very different shelf lives

It's useful to split advisory hours into three, because they carry very different AI risk:

Generic knowledge. Regulations, methodology, templates, definitions, industry norms. This is the layer AI delivers for free. It's also the layer most naturally read into Norges Bank's description of clients solving tasks themselves.

Interpreting the client's own numbers. Why did the contribution margin fall in one segment, which clients are dragging down liquidity, what happens to profitability if the largest contract gets renegotiated. This requires the numbers to exist in machine-readable, consistent form. If they don't, the client gets an answer that sounds right, based on an exported spreadsheet nobody has quality-checked. But this layer isn't protected either. Once accounting systems get AI built in, the client can get an interpretation of their own numbers straight from there. Its shelf life depends on the firm owning data quality and moving first.

Accountability and follow-up. Someone who signs, someone who flags an issue before the quarter closes, someone who can be held accountable for the judgment being sound. Here, AI is structurally unable to compete, because accountability can't be automated away, it can only be shifted to someone.

Firms that have packaged advisory work as meetings, presentations, and reviews are often selling layer one. Firms that have packaged it as ongoing decision support built on the client's own numbers are selling layers two and three.

The detail in 1/2026 that's easy to miss

Regional Network 1/2026 contains something more interesting than the fact that some are losing assignments: within IT and other consulting businesses, the companies that help enable clients to adopt AI themselves are seeing by far the strongest growth. Others are seeing an increasing share of tasks performed by clients themselves, at the expense of hired labor and consulting services. The value, in other words, is shifting from doing on behalf of to enabling.

Our reading, translated to the accounting industry: it isn't the advisor who explains the numbers who wins, it's whoever makes the client's numbers usable, including for the client's own AI tools. The firm that owns the chart of accounts, the dimensions, the accrual periods, and the data quality owns the precondition for all analysis that happens afterward, regardless of who presses the button.

The same report notes that AI had a fairly neutral effect on employment needs in 2025, but gets a dampening effect in 2026, particularly at large companies, and most for employees with little work experience. This applies across all the industries in the network, not accounting specifically.

That last point should still give accounting leaders pause. The industry's staffing model is a pyramid: juniors produce the volume hours, and the pyramid is simultaneously the career ladder that produces tomorrow's senior advisors. Our hypothesis is that if AI eats the entry level first, it isn't just some hours that disappear, the recruitment logic itself changes.

At the same time, labor scarcity has long been the industry's main excuse for not industrializing. Pressure in the labor market has eased somewhat in the latest round, but that's a cyclical movement. Samfunnsøkonomisk Analyse estimates that demand for labor will grow by roughly 14 percent toward 2045. The point, then, isn't that people become easier to find, it's that technology, not the labor market, is what changes the equation.

The client is reading the AI numbers too

The macro picture points the same way. NHO and Samfunnsøkonomisk Analyse estimate that generative AI has the greatest potential to streamline tasks in office occupations, and that AI could make it possible to produce the same goods and services as today with around 15 percent fewer hours worked. SØA sees this as a way to meet growing labor scarcity, not as job loss in itself. The point here isn't the percentage, it's that the client is reading it too. A business owner who has just cut their own hours with AI asks different questions of an invoice based on hours.

But accounting advisory competes in the same budget at the client, and much of it sits closer to the client's data than strategic advisory does. Proximity to the data is the industry's structural advantage, but it's only real if the data is actually structured.

What the 2027 plan has to be able to withstand being measured against

The questions that follow from this are uncomfortably concrete.

What share of the firm's advisory revenue comes from hours where the deliverable is, in reality, an explanation the client could get for free? How much comes from deliverables that require the client's own, reconciled numbers, and therefore can't be reproduced without the firm?

How many advisory hours are triggered by the client asking, and how many are triggered by the firm flagging something first? Reactive advisory competes directly with a chatbot that's always available. Proactive alerts on actual numbers don't, because they require someone continuously monitoring something.

And: what's the actual pricing unit? As long as decision support is priced as time, it's priced in a unit where the market price can fall. When it's priced as an ongoing product, deviations, liquidity, profitability per project, an alert before quarter-end, backed by professional accountability, it's priced in a unit AI doesn't deliver on its own.

Advisory work was never the rescue in itself. It was a direction. Norges Bank's network businesses describe AI affecting consultants unevenly: some lose assignments because clients do the work themselves, others win by helping clients get started. Our reading is that the divide runs between those who sell explanations and those who sell decision support anchored in the client's own numbers.

The difference between the two long looked like a nuance. It's becoming a business model.