Subscribe now

For a decade, European tech has been defined by what’s missing: deep public markets, trillion-dollar champions, exits that never quite arrive. Klarna’s New York listing in September 2025 confirmed the familiar verdict — another champion lost to Wall Street.

The problems are real: thin liquidity, valuations that don’t hold, an EU Listing Act that helps at the margins but not the core. Only 19 venture-backed European startups went public in all of 2025, at a median age of 29 (PitchBook).

In the US, a strong IPO market gives a winning startup a real choice: list, or sell to a strategic. In Europe that choice is more narrow - with the public-market door barely open, the exit that counts is the strategic one. And the timing could not be better, because AI is forcing Europe’s incumbents to buy. When they can’t build fast enough alone, they reach outward.

The data on where European companies actually exit is clear - M&A accounted for 98% of EMEA venture-backed exits in H1 2025, up from 90% a decade earlier; IPOs fell to 2% (JPMorgan). That M&A dominates is not new; what is new is what drives it. Buyout exit value for VC-backed companies hit €19 billion last year, up more than 35%, the highest share of exits this decade (PitchBook). And the most telling number: without AI, European deal value in 2025 would have fallen by 5.7% (PitchBook). With AI, it grew. The exit market is not leaning on M&A by default; it is being rebuilt around AI capability — and that is where I see Europe hold an asset.

Europe’s industrial depth offers a unique opportunity

The EU contains roughly 33.5 million enterprises — around 251,000 mid-sized, 55,000 large — and SMEs alone produce nearly half of EU value added (Eurostat, 2024). Add the large incumbents — SAP, ASML, Siemens, Schneider, Bosch — and you get a density of mature, cash-generative buyers that does not exist at this scale anywhere else.

For thirty years that density was a curiosity, sometimes a liability — the incumbents founders learned to route around. Now almost all of them are under pressure they haven’t felt in a generation. KPMG’s 2025 work puts it starkly, and the pattern repeats across European peers: 91% of large enterprises now consider AI business-critical, while 43% of mid-sized companies still have no concrete plan (KPMG; DMB Mittelstand AI Index). The technology moves faster than any internal roadmap can absorb — a capability that didn’t exist 18 months ago is now table stakes. No incumbent, however large, can build the whole stack fast enough alone, so the build-versus-buy reflex has swung towards an evaluation of where to build on your own and where to buy-and-partner. Europe’s industrial depth now meets a capability gap — and on the other side sits a generation of European AI startups with exactly what those incumbents need.

The European-on-European deals show what this looks like

ASML — the Dutch lithography maker at the center of the global chip supply chain — put €1.3 billion into Paris-based Mistral in September 2025 for an 11% stake, becoming its largest shareholder, to embed Mistral’s models in its own systems. Siemens joined the €117 million Series B of London’s PhysicsX, a UK deeptech building AI-driven physics simulation for industry — a sign that Europe’s industrial giants now want strategic positions in European AI startups, not a view from the sidelines.

SAP shows how broad this can go. For one, it agreed to acquire Prior Labs — an 18-month-old Freiburg lab building tabular foundation models — committing over €1 billion to run it as an independent frontier lab, in a deal that handed the founders well over half a billion dollars in cash, one of Germany’s largest venture outcomes. Alongside it, SAP took a sub-1.3% stake in Berlin’s n8n that doubled its valuation to $5.2 billion; partnered with and invested in Berlin’s Parloa, whose voice agents now plug into SAP Service Cloud; and named London’s Conduct a strategic AI partner for Cloud ERP transformation. Four European AI companies, four structures — acquisition, minority stake, partnership, endorsed app — all the same move: combining an incumbents reach with startup capability.

The mid-market is starting to move too — though more as strategic investor than outright acquirer. The Schwarz Group behind Lidl and Kaufland became a core backer of Heidelberg’s Aleph Alpha; family offices such as Peugeot’s are writing checks into industrial-AI startups like Munich’s Augmented Industries. The largest acquisitions still come from the corporates, but the buyer base widening to include the Mittelstand is how an exit market deepens.

And the effects are multiplying down the line - fresh off ASML’s capital, Mistral acquired the Austrian startup Emmi AI in May 2026 — reportedly one of the largest exits in Austrian history. The company that was an acquisition target in September became an acquirer by spring. This is what a healthy exit market looks like: not a single ladder to a public listing, but a recombining web in which capability and reach keep finding each other inside Europe, each exit seeding the next.

AI has moved these conversations to the board

When capital is scarce, an incumbent buys to consolidate a market, and the logic is financial. When capability is scarce, it buys or partners to stay technologically alive, and the logic is existential. That shift explains how these deals now look: they close faster, lean on minority stakes and partnership as readily as outright purchase, and are decided at board level — AI is the number-one agenda item for almost every large European incumbent right now. That is what makes it a real chance for startups: when a deal sits with the board, timelines compress and strategic conviction does much of the work diligence used to do. Of course access is still the challenge ;)

VCs could play a much more strategic role here

One underused entry point for an incumbent looking to combine with a startup is the VC — not because incumbents need selling, but because a VC has a constant view across the market. A good investor can flag the startups closing a capability gap and — even with no portfolio match — point to comparable deals or other founders worth a call. I've experienced this firsthand: the handful of investors who took the time to understand what the corporate was actually trying to build shaped the conversation far more than those who simply pitched their portfolio. That overview is genuine value, and costs the corporate almost nothing. Of course, the most direct path still runs through the founder, who can open the door alone with the right workflow fit and governance story — but the investor can get there earlier and de-risk it. This is especially true for the AI pre-seed and seed companies that don't yet clear a corporate's usual bar on revenue or enterprise-readiness, and would otherwise never make it onto the radar.

What’s worth watching next

Europe will not fix its IPO market this cycle — the companies that truly need a deep listing will probably still keep going to New York. The more interesting story is underneath, and it benefits a far larger share of European founders, investors, and corporate leaders.

There are early signs the ecosystem is organizing around exactly this, on example stands out. In May 2026 the European Commission selected EQT to lead the new Scaleup Europe Fund — a flagship initiative first announced in von der Leyen’s 2025 State of the Union, targeting €5 billion to back European companies in AI, robotics, semiconductors, energy, biotech and adjacent deep-tech fields (EQT). Its stated purpose goes beyond writing cheques. The fund is explicitly designed to close the late-stage funding gap so European companies can scale internationally while staying anchored in Europe, and — through the Commission’s parallel European Corporate Network, which builds on the EIC’s corporate days — to connect those companies with industrial partners and customers across the continent. That is the matchmaking role described above, now with institutional weight behind it. It is worth being clear-eyed: a €5 billion fund and a corporate-networking programme do not, by themselves, build a single combination.

The actual deals will still be made one at a time, by founders who show up with the right fit and incumbents whose boards decide they can no longer wait. But a public vehicle whose whole reason for existing is to keep European capability scaling inside Europe — and to put it in front of the incumbents that can deploy it — is exactly the infrastructure a strategic exit market needs to deepen. If it delivers, the European innovation ecosystem has a genuine chance to convert this moment into a durable advantage rather than a one-cycle spike.

What I’ll be watching through 2026 and 2027 is not the deals announced, but what happens after: how well the capability gets built into the product, how fast it reaches the market, how much value is generated for the customer.

What will you be looking out for? :)

Subscribe now

Share

Sources

• PitchBook — 2025 European Venture Report

• JPMorgan — EMEA venture exit data

• Eurostat — Structural Business Statistics, 2024

• KPMG — AI in German business, 2025

• DMB / Salesforce — Mittelstand AI Index

• EQT — Scaleup Europe Fund

• Company announcements & public filings (ASML/Mistral, SAP/Prior Labs, n8n, Parloa, Conduct, Mistral/Emmi, Siemens/PhysicsX)

The cause wasn’t a failed contract or an infrastructure overrun. It was a single coding tool — Anthropic’s Claude Code — rolled out to Uber’s engineers in December. Usage nearly doubled by February. By April, roughly 95% of Uber engineers were using AI tools monthly, and most of the company’s coding work was running through them. The rollout worked exactly as intended, yet the budget didn’t survive it.

Subscribe now

If you run finance, strategy, or engineering at a company with more than a few hundred technical staff, you’re about to have a version of this conversation with your CFO — or are already having it. And the uncomfortable thing I want to state is that the number you submitted six months ago was never going to hold, no reasonable forecasting process would have caught it. The question worth asking now isn’t how do we spend less - it’s how do we spend better when the total is going to keep moving anyway

Why the number keeps moving

Capability creep. Every model release resets consumption at a higher floor. When Claude Opus 4.7 shipped on April 17, engineering teams found workflows within 48 hours that weren’t viable on 4.6 — longer-context reviews, multi-hour agent runs, vision tasks that used to need a human in the loop. Every time that happens, consumption steps up and doesn’t step back down.

Agent-to-agent traffic. Teams are running setups where one agent supervises another, where a coding agent hands off to a testing agent, where a research agent coordinates three sub-agents. Your bill includes tokens generated by “AIs” having conversations with each other. This line item didn’t exist a year ago and is now a meaningful share of spend in any team that’s moved past single-shot prompting. Most FP&A teams haven’t modelled it yet because there’s no historical pattern to extrapolate from.

Tool sprawl. Most organisations are paying for three to five AI coding tools in parallel — Claude Code, Cursor, Copilot, ChatGPT Enterprise, sometimes Perplexity or Codex on top. Each team picks a preference, none of them are mutually exclusive, and total token consumption compounds across the stack. At the same time, employees are getting pressure to adopt the tools.

The macro signal sits on top of this. Around the same time Uber was publishing the overrun, Anthropic signed an expanded agreement with Google and Broadcom for roughly 3.5 gigawatts of TPU capacity, coming online from 2027. That’s a provider telling the market it expects demand to keep compounding, and pricing accordingly. Inference has overtaken training as the dominant driver of AI compute demand. The running cost of the AI you already decided to use is going up, and it’s going up for reasons outside your contract.

The optimisation trap

Let’s talk about real levers that optimise spend but don’t solve the problem at hand

a) Model routing: sending cheap tasks to cheap models, reaching for frontier capability only when the work warrants it
b) Agent loop pruning: removing redundant tool calls inside long runs
c) Prompt caching: dropping repeated reads to a fraction of input price
d) Waiting: token prices have fallen roughly tenfold every years, and there’s no reason to think the trend breaks this year

Yet - pull every lever. Your total AI spend will almost certainly grow anyway.

The reason is - as inference gets cheaper per unit, teams find three, five, ten times more use cases that now pencil out at the new price. Optimisation at the task level expands the surface at the portfolio level. The cheaper the unit, the more units you consume, and the bigger the line becomes. Once the unit economics improve, workflows you previously shelved suddenly become worth shipping.

So the question changes. It stops being how do we spend less and becomes how do we spend better, which is an entirely different discipline.

Spending better, in three parts

1. Get consumption visibility, then tier it

Most companies treat all tokens as equal. The highest-value work and the lowest-value work are drawing from the same undifferentiated pool at the same model price. That’s the cheapest possible mistake to fix, and almost nobody has fixed it.

Fixing it requires two things finance teams usually don’t have yet. The first is consumption visibility by team, workflow, and feature — not by vendor. “We spent €2.1M on Anthropic last quarter” is useless for decisions. “Shipping the new checkout flow cost us €34K in tokens” is a sentence you can act on. The second is unit economics per workflow: cost per resolved ticket, cost per pull request, cost per qualified lead, cost per contract reviewed. Tomasz Tunguz (Theory Ventures) has been calling this “productive work per dollar of inference,” and the companies that track it end up with bigger, better-defended budgets — because they can prove which parts are working and cut what isn’t.

This will help you set-up a “Token Tiering”, which Aaron Levie (Box CEO) recently described in the 20VC podcast.

Tier 1 — revenue-critical. Core product engineering, customer-facing AI features, top-of-funnel growth. Best available models, no caps, explicit permission to spend more next quarter if the output justifies it.

Tier 2 — operational leverage. Finance, customer success, marketing analytics, ops workflows. Mid-tier models, soft caps, quarterly usage review.

Tier 3 — convenience. Internal docs, meeting summaries, scheduling, routine admin. Cheap models, hard caps, no exceptions.

The business case defines the tier, not HR grade. If a senior VP is using a frontier model to summarise Monday’s staff meeting, that is a tier-3 job at a tier-1 price, and someone has to be allowed to say so out loud. The tiering falls apart the first time it’s overridden for optics, which it will be, which is why you need the next part.

2. Put one person on the horizontal line

AI procurement, model selection, governance, and spend approval are distributed across the CIO, CTO, CFO, and individual function heads in most companies. Which means nobody owns any of it. Every company I know with AI spend actually under control has a single accountable lead, usually reporting into finance or strategy rather than IT, often titled something along the lines of Head of AI FinOps. Shared responsibility is how this problem got here in the first place.

Tiering requires a referee, routing decisions need an owner, vendor consolidation needs someone with the authority to say no to a function head who wants their own tool. Without that person, every optimisation decision becomes a committee conversation, and committee conversations don’t keep up with a budget that re-baselines every quarter.

3. Stop forecasting AI budgets annually

Annual numbers stop being meaningful by July because every model release resets the floor. If you’re still running a 2026 AI budget as a single fixed line, you’re managing last quarter’s problem.

Reforecast quarterly, and build in a 20–30% capability cliff reserve for the next model release you know is coming but can’t yet size. Treat it as a normal part of the budget, not a contingency you have to justify releasing. The alternative is that every frontier release triggers an emergency ask, which is both expensive politically and signals to the board that you don’t understand your own cost structure.

Pull the optimisation levers — routing, caching, loop pruning — but primarily to make the remaining spend legible, not to save money. Optimisation is a reporting tool disguised as a cost-cutting tool, and that’s its real value.

And plan for compute as compensation. An engineer with a generous inference budget out-produces one without it by a wide margin, and that gap is widening every quarter. Tomasz Tunguz at Theory Ventures flagged this first in February, pointing out that at a 75th-percentile software engineer salary of $375K, adding $100K in annual inference effectively makes the fully loaded cost $475K — roughly one dollar in five going to compute That’s the new shape of a fully loaded engineering seat, and the companies that accept it earliest will out-ship the ones that don’t.

The companies that will spend the most on AI in 2027 aren’t the ones running the tightest cost controls. They’re the ones who can tell you, on any given Tuesday, what last Tuesday’s AI spend actually produced. That sentence is the whole discipline.

If you’re working through this in your own organisation and want to compare notes, I’d love to hear what’s working and what isn’t. The hardest part of this transition is that everyone is figuring it out in parallel.

Thanks for reading NoelleRoemer’s Substack! Subscribe for free :)

Sources

• https://www.theinformation.com/newsletters/applied-ai/uber-cto-shows-claude-code-can-blow-ai-budgets

• https://www.anthropic.com/news/google-broadcom-partnership-compute

• https://a16z.com/llmflation-llm-inference-cost/

• https://tomtunguz.com/inference-as-compensation/

• 20VC podcast - 20th April with Box CEO Aaron Levie: https://www.thetwentyminutevc.com/aaron-levie-3

Share

Meanwhile, Nvidia CEO Jensen Huang told the All-In Podcast that he’d be “deeply alarmed” if a $500,000 engineer wasn’t consuming at least $250,000 in AI tokens per year. Shopify CEO Tobi Lütke made “reflexive AI usage” a baseline performance expectation and added AI consumption questions to peer reviews. Meta has also begun integrating AI usage into its performance culture — its internal “Claudeonomics” leaderboard gamifies consumption, and leadership has publicly encouraged engineers to maximise their token spend.

These leaders are responding to something real: most companies are still dramatically underusing AI, and a push toward adoption makes strategic sense. But the way we measure that adoption is about to matter enormously — and I don’t think we’ve figured it out yet.

When usage becomes a scoreboard

The term has already entered the lexicon: tokenmaxxing — the competitive consumption of AI compute as a signal of productivity. Meta’s “Claudeonomics” leaderboard tracks usage across 85,000 employees and ranks the top 250 power users. In a recent 30-day window, employees collectively consumed 60 trillion tokens — roughly $90 million in compute at standard API rates, or $1.08 billion annualised.

The gamification is deliberate. Employees earn badges from bronze to emerald. Titles include “Session Immortal” and “Model Connoisseur.” An OpenAI engineer reportedly consumed 210 billion tokens in a single week — 33 complete Wikipedias. Gizmodo reported that multiple tech companies are now evaluating employees in part by how quickly they burn through LLM tokens.

On April 2, 2026, OpenAI formalised the economics by moving Codex to token-based pricing, replacing per-seat licenses with pay-as-you-go billing. Anthropic’s enterprise plans already operate on per-token economics. The infrastructure now exists to track every interaction, every query, every dollar — down to the individual.

For the first time, knowledge work has a variable cost that’s attributable at the person level. But it means companies need to decide: what are we actually optimising for?

More tokens don’t mean more output — up to a point

Here’s the hypothesis I think matters most, and the one I’d want to test if I were running an AI-forward engineering org. The relationship between token consumption and productivity isn’t linear. It’s an inverted U.

At the left end of the curve, you have zero or minimal usage — people who haven’t adopted AI at all, either from friction, scepticism, or lack of training. They’re leaving real value on the table. Research from MIT Sloan found that the least experienced workers gained up to 35% more output per hour with AI access, while highly skilled consultants given structured AI guidance achieved gains exceeding 40%. The average boost was 14%, but the distribution is what matters: the gap between a skilled AI user and an unskilled one isn’t marginal - it’s a multiple.

In the middle, you have the sweet spot: people who use AI selectively and precisely. They get useful output on the first or second try. They apply AI to problems where it genuinely saves time — drafting, analysis, code generation, research synthesis — and do the rest themselves. Their token consumption is moderate. Their output is disproportionately high.

At the right end, you have the tokenmaxxers — people who use AI as a default for everything, including tasks they’d complete faster on their own. They re-prompt the same request twelve times because they can’t articulate what they want. They generate content nobody asked for. Some keep sessions running just to climb the rankings. Their marginal productivity may actually be negative once you account for the cost of reviewing, correcting, and discarding output that wasn’t needed.

If this curve is real — and I suspect any honest internal analysis would confirm it — then the optimisation problem is more interesting than “encourage more usage” or “control costs.” It’s about finding the peak for each function, and building systems that push people toward it.

Why $15 gets more scrutiny than $5,000

There’s a second-order effect here that I haven’t seen discussed, and it may matter more than the measurement question.

When AI usage becomes a visible, per-person cost line, the cost of using AI becomes salient in a way that the cost of not using AI never does.

A senior analyst spends three days manually synthesising data from multiple sources — a task that a well-constructed AI workflow could accomplish in two hours for perhaps $15 in token costs. The three days of analyst time cost the company roughly $3,000–5,000 in loaded salary. But that cost is invisible - it’s buried in headcount — a fixed line item nobody examines at the task level.

Now imagine the analyst uses AI instead. The $15 shows up on a dashboard. It’s variable and attributable. If the culture is one of cost scrutiny, the analyst learns that AI usage attracts attention in a way that slow manual work never does.

This asymmetry is the real risk. The cost of AI is measured in dollars on a dashboard. The cost of not using AI is measured in invisible hours nobody tracks. Any organisation that manages token spend without simultaneously measuring time-to-completion will systematically underinvest in the high-value applications — which is the opposite of what you want.

What organizations should put on the dashboard"

So how do you manage AI as a cost line without suppressing the behaviour you’re trying to encourage? Here’s what I think would actually work.

Give teams a budget, not individuals a score. Set a monthly compute allocation at the team level and let them self-optimise. Teams will figure out which applications are worth the tokens because they’re closest to the work. Nobody earns a badge, but every team has an incentive to spend wisely.

Track what came out, not just what went in. The number you want isn’t “how many tokens did this team use.” It’s “what changed because they used them.” Deals closed, features shipped, reports delivered, time-to-completion on recurring tasks — track those alongside consumption. A team whose output rises faster than its spend is using AI well, regardless of the absolute number.

Talk to both ends of the spectrum. Zero-usage teams need a conversation about barriers — tooling, training, use-case identification. The research consistently shows the least experienced workers gain the most from AI, so non-adoption is costliest where it’s most common. High-usage teams need a different conversation: what did you build? If a team burned $50,000 in tokens and automated a workflow that saves $500,000 annually, celebrate it. If they burned $50,000 and produced nothing measurable, that’s worth understanding too.

Put the hidden cost on the dashboard too. If you track AI spend per team, also track estimated time savings - force the comparison. A team that spends $10,000 a month but saves 200 analyst-hours is generating a 5–10x return. Without that context, the spend number means nothing.

Treat compute like R&D, not like office supplies. Jensen Huang’s instinct — that engineers should be spending heavily on AI — is right in principle. The companies that extract the most value will treat compute as investment in capability, not consumption to be minimised. The question is whether that investment is directed toward outcomes or toward leaderboard positions.

—

Full disclosure: I’ve drifted off and token-maximised plenty of times myself ;)

Sources:

• OpenAI Codex token-based pricing shift

• Exponential View / Azeem Azhar: The AI capacity trap

• OpenAI internal token leaderboard discussion

• https://www.businessinsider.com/tokenmaxxing-ai-token-leaderboards-debate-2026-4

Leave a comment

"What's your personal agent called?" has been a hot topic at dinner with friends lately. Mine’s called Nio. My friend calls his Q. My friend then admitted that Q had gone on a solo WhatsApp tour a few days ago — pinging colleagues, old friends, people he hadn’t spoken to in months. Fourteen confused replies before he managed to kill the process.

Everyone laughed.

But the joke stuck with me — because underneath the chaos, something real is happening. People are building AI agents that run 24/7, manage their workflows, handle their communication, and act on their behalf while they’re asleep. OpenClaw became the fastest-growing open-source project in history for a reason: always-on agents can provide real benefits.

Now imagine deploying that inside a company. The upside is enormous — but so is the risk if your agent starts freelancing with customer data or your CRM. An AI strategy isn’t enough anymore. Every company now needs a strategy for agents that run autonomously, around the clock, making decisions on their behalf — and that strategy better include how to keep them in check. This week, NVIDIA and Anthropic both shipped their answer to that problem — NVIDIA with NemoClaw, a security wrapper that makes OpenClaw enterprise-safe, Anthropic with Claude Code Channels, which rebuilds the whole idea from scratch inside their own product.

Subscribe now

It started at GTC on March 16th. Jensen Huang walked onto the stage in San Jose — 10,000 people, the leather jacket, $3 trillion company — and said something that made every enterprise exec in the room shift in their seat: “For the CEOs, the question is — what’s your OpenClaw strategy?” (TechCrunch)

Not your GPU strategy, not your AI model strategy — your strategy for a free, open-source tool that one developer built as a side project and that just became the fastest-growing open-source project in history, growing faster than Linux, in weeks rather than years.

He announced $1 trillion in chip orders, unveiled a next-gen supercomputer, and brought a walking Olaf robot from Frozen on stage. But the thing he kept coming back to was OpenClaw — the tool behind both Q and Nio. The tool that lets anyone build an AI agent on their laptop that actually does things: manages your calendar, writes code, handles workflows, runs 24/7.

“Mac and Windows are the operating systems for the personal computer,” Jensen said. “OpenClaw is the operating system for personal AI.” (NVIDIA Newsroom)

He’s not wrong. When agents run 24/7, manage workflows, and act on your behalf while you sleep — that’s not a productivity hack, that’s a fundamentally different way of running a company. But everything that makes always-on agents powerful also makes them dangerous without the right guardrails — which is exactly what NVIDIA and Anthropic each shipped an answer to, within the same week.

NVIDIA’s answer: make the wild thing safe.

Nvidia built NemoClaw: a security wrapper that installs on top of OpenClaw with a single command. Privacy controls, sandboxing, guardrails — so your agent can’t go on a WhatsApp adventure or, more importantly, wander through your company’s financial data unsupervised. Every file access, every network request, every external call — governed by rules you set.

The world’s biggest chip company built something that doesn’t need its chips. Why? Because Jensen is playing a different game. The models are commoditizing fast, and the chips will keep selling regardless. But the thing that’s actually stopping companies from deploying AI agents at scale isn’t capability — it’s trust.

Can I give an agent access to our CRM without a Q situation hitting a client? Can legal sign off on any of this? Will my board even let me try?

NVIDIA just positioned itself as the answer to those questions — not by building agents or models, but by building the layer that makes everybody else’s agents safe enough to actually deploy. Look at who’s already plugging in: SAP, Salesforce, ServiceNow, Siemens, Adobe, Atlassian.

If you’re the company every enterprise trusts to make agents safe, you’re in the room every time an agent gets deployed — and that’s not a chip business anymore, that’s a platform business.

Anthropic’s answer: skip OpenClaw entirely.

Four days after Jensen’s keynote, on March 20th, Anthropic shipped Claude Code Channels — and the approach couldn’t be more different. Where NVIDIA wrapped OpenClaw to make it safe, Anthropic looked at what made OpenClaw popular (an always-on agent you can message from Telegram or Discord that actually does work for you) and rebuilt that functionality natively inside Claude Code, with their own security model from the ground up. In practice, you message Claude from your phone via Telegram or Discord, and it executes the work on your laptop — building apps, running scripts, processing files — then replies on the same channel. Sender allowlists, explicit opt-in per session, everything running on the developer's own machine.

No wrapping, no open-source dependency. They saw what we all loved about OpenClaw, and built it themselves — safer, simpler, and inside the product you already pay for.

Two trillion-dollar-scale companies, two completely opposite strategies, but the same underlying conviction: every company will need a secure, always-on agent strategy. If you haven’t started thinking about yours, this was your week to start.

My friend has — he’s turned Q’s WhatsApp permissions off. At least for now.

Got your own agent story? Comment below!

This is part of an ongoing series on where enterprise AI is actually heading. Previous pieces: The week the orchestration layer became the product in big tech; Business users are becoming builders. What does that mean for product management at SaaS companies?

Leave a comment

Three very different strategies, one shared conviction: the chase in enterprise AI has moved to the orchestration layer — the governance, the process logic, the operational context that determines whether an agent can actually do anything useful inside a business.

Agents getting email addresses before the interns did

Agent 365 ships May 1 — and Microsoft is already running it internally. Over 500,000 agents across the company, generating more than 65,000 responses a day. The key move: agents get the same identity infrastructure as employees — IDs, security policies, e-mails adresses — which means they can be governed through the same control plane. That's how you orchestrate a workforce of agents at scale: you treat them like the workforce you already manage. Somewhere in New York, an agent is about to get its own Outlook inbox before this quarter's interns do.

NVIDIA’s playbook is the one that made Kubernetes the default container scheduler: offer the orchestration layer for free, collect the value and monetize as the ecosystem grows on top of it. NemoClaw is hardware-agnostic, set for reveal at GTC on March 16, and designed to be the runtime layer that makes the models underneath interchangeable.

Moltbook gives Meta something nobody else has — a live dataset of what happens when over a million agents interact without human oversight. Where they coordinate, where they fail, what guardrails they need. That's the empirical foundation for building the orchestration layer everyone else is designing from theory.

The orchestration layer is the bottleneck. It’s also the opportunity

Gartner predicts over 40% of agentic AI projects will be scrapped by 2027. Not because the models fail, but because organisations can’t operationalise them. Microsoft’s own Cyber Pulse report found that 29% of employees have already turned to unsanctioned AI agents for work tasks — agents running on production data with no governance in sight. Think about it for a moment: nearly a third of the workforce is deploying agents that IT and security teams can’t even see, let alone govern.

The pattern looks familiar — cloud and mobile both had their version of this, where capital flooded in before the operational layer caught up. The difference this time is that agents don’t just process data — they act. An agent can reason. But reasoning is not the same as doing. It cannot act inside a procurement workflow, optimise a cross-border supply chain, or flag a quality deviation three tiers deep — not because the model isn’t good enough, but because there’s no orchestration layer connecting it to the governed data, process logic, and compliance constraints it needs. The institutional memory of how an organisation actually works doesn’t live in the model. It lives in the systems the model can’t reach yet.

Incumbents didn’t build a moat. They’re sitting on one.

There’s a quiet realisation spreading through the companies that run enterprise infrastructure. The large platforms sitting on decades of process logic — the ones that know how a purchase order routes through approvals, how a supply chain reroutes when a port shuts down — are starting to understand that this institutional knowledge isn’t a legacy burden. It’s the defensible asset.

The model layer is commoditising fast. The orchestration and context layer is not. The companies that have spent thirty years encoding how businesses actually operate are the ones best positioned to connect agents to reality — not because they'll build the best agents, but because they own the operational infrastructure agents need to function.

The smartest incumbents aren’t trying to build everything. They’re mapping where to engage, pulling startups in to fill the gaps, and focusing their energy on the one thing that can’t be replicated from a standing start: operational depth at enterprise scale.

The orchestration race is on

The companies that win this next phase won’t be the ones with the best agents. They’ll be the ones whose orchestration layer becomes so embedded in how work gets done that it turns invisible.

The question is who gets there first — and whether the answer turns out to be an incumbent, a startup, or the combination of both that nobody saw coming.

This is part of an ongoing series on where enterprise AI is actually heading. Previous pieces: When models converge, context becomes the moat; Business users are becoming builders. What does that mean for product management at SaaS companies?

Sources: Microsoft Official Blog, “Introducing the First Frontier Suite” (Mar 9, 2026) · Wired, “Nvidia plans open-source AI agent platform NemoClaw” (Mar 9, 2026) · CNBC, “Nvidia plans open-source AI agent platform” (Mar 10, 2026) · Axios, “Meta acquires Moltbook” (Mar 10, 2026) · TechCrunch, “Meta acquired Moltbook” (Mar 10, 2026) · Gartner, “Over 40% of Agentic AI Projects Will Be Canceled by End of 2027” (Jun 2025) · Menlo Ventures, “2025 State of Generative AI in the Enterprise” (Dec 2025) · Microsoft Security Blog, “Cyber Pulse Report” (Feb 2026)

In brief: The people closest to the problem can now build the first version of the solution themselves. This changes SaaS product management. The Forward Deployed Engineering (FDE) model proved that proximity to the problem beats specification from a distance. But FDE don’t scale when the builders are your customers. Instead, the future will require Distributed Product Operations (DPO): not a deployment strategy, but a product operating model for a world where building happens everywhere — shifting Product Management (PM) from specifying features to curating, governing, and connecting what citizen builders (i.e., business employees with little / no formal coding training who create applications / automated workflows using low-code/no-code platforms) create.

Your customers are already building

A customer success lead prototypes a churn-risk dashboard in Lovable over lunch. A finance director builds a cash-flow tool with Claude — no ticket filed, no sprint planned. A sales ops manager ships a CRM workflow in Replit between meetings. No product requirements document (PRD). No PM involved.

These aren’t engineers. They’re your customers. And they’re no longer waiting for your next release.

Billions are flowing into citizen building tools

a16z’s analysis of spending across 200,000+ companies found vibe-coding platforms surging into top-tier enterprise AI expenditure. Replit ranked third overall, Cursor sixth, Lovable eighteenth. Consumer AI tools are being “yanked into enterprise faster and faster.”

Anthropic’s Boris Cherny told Fortune his team built Claude Cowork in roughly ten days. Non-technical users now deploy Claude for everything from tax prep to event booking. Microsoft — which sells the competing GitHub Copilot — has adopted Claude Code internally for non-developers.

Gartner forecasts that citizen developers (business users who build applications using AI or low-code tools without formal engineering training) will outnumber professional developers 4:1 by 2026.

Your roadmap now competes with your own customers

Previous low-code waves promised similar things. The difference: AI collapses the feedback loop entirely. When a finance director prototypes her own model, she’s not saving engineering time — she’s eliminating the translation layer. No PRD losing context. No sprint diluting intent. The person with the problem is building the solution — and they’re not on your payroll.

For software companies, this is a fundamental shift. Your users are no longer just requesting features and waiting. They’re building workarounds or extensions.

ICONIQ’s recent analysis calls this the “Great Sorting”: when anyone can produce a tool in an afternoon, the bar for what users will tolerate rises sharply. Mediocre features don’t just get ignored — they get replaced by something a customer built over lunch. From a PM perspective, the roadmap is no longer competing only with other vendors. It’s competing with your own users’ ability to build.

Product Management doesn’t die. It transforms.

Shopify CEO Tobi Lütke declared that teams must “demonstrate why they cannot get what they want done using AI” before requesting budget. Not cost-cutting — distributing the act of building across the organisation.

When building becomes abundant, the scarce resource shifts. The SaaS PM of the next era looks less like a feature factory manager and more like a profit-and-loss (P&L) owner with technical literacy: strategic frame, governance, and engineering capacity for what users should not build on their own — security, compliance, scale, integration.

Palantir had the right idea — but given recent developments in AI, it doesn’t scale

There’s a precedent. Palantir pioneered the Forward Deployed Engineer — software engineers embedded directly in client environments to build alongside users. OpenAI, Anthropic and other SaaS vendors have all adopted the model. FDE postings rose fivefold in 2025. a16z called it the hottest role in startups.

The FDE insight: proximity to the problem beats specification from a distance.

But FDE has a scaling problem. You can deploy ten engineers to ten clients. You can’t deploy ten thousand. Citizen development takes the FDE logic and inverts it: instead of sending your people out, you recognise the builders are already there — your customers. They don’t need deploying. They need enabling.

This is why the response can’t only be “hire more FDEs.” What software companies need is a distributed product operations (DPO) model — an operating model designed for a world where building happens everywhere, by everyone, on top of your product.

Three structural moves:

1. First, make your platform extensible by design — APIs, templates, integration patterns that let citizen builders create safely without breaking what’s underneath

2. Second, build a governance layer: security standards, data policies, a registry of what’s being built on top of your product. Not to control, but to learn

3. Third, create feedback loops from the edge back to your roadmap — what customers build themselves reveals exactly what your product is missing

What Product Management actually owns in this model

ICONIQ frames the emerging software landscape as two interdependent layers: a data context layer (structured data, workflows, domain expertise) and an agent layer (AI systems that act on top).

A PM’s job is to own the context layer — the governance, the data integrity, the integration patterns — while enabling citizen builders and agents to extend it. Not controlling what gets built, but making it safe and coherent. I.e., A product operating system for distributed building.

The new Product Manager job description

The PM role reorients:

• Curate, don’t specify. Watch what customers are building. Identify which patterns should be absorbed into the core product and which should stay local. Be the editor, not the author

• Make governance the product. Design the guardrails that make it safe for customers to extend your product

• Hold the map. Track where the gaps are, where integration risks live, and which customer-built experiments signal where your product should go next

• Build feedback loops. What customers build tells you what your product is missing. Route those insights into the roadmap systematically

Where this leaves Product Management

If the people who understand the problem best can now build the first version themselves - where does this leave PMs?

Creating the conditions in which the right things get built, by the right people, safely and coherently — whether those people are inside your company or your customers. Harder than writing PRDs. But it’s the job the world is asking for.

Sources:

• https://a16z.com/the-ai-application-spending-report-where-startup-dollars-really-go/

• https://fortune.com/2026/01/24/anthropic-boris-cherny-claude-code-non-coders-software-engineers/

• ICONIQ, “A Coming Age of Reason”

• Gartner Low-Code Forecast / BetaNews citizen developer predictions

• https://researchmoneyinc.com/article/the-insights-report-december-3-2025

• https://a16z.com/services-led-growth/

Executive Summary
Everyone is watching the Big Tech $700B race like it’s the main event. Racks of GPUs, new data centers, capex charts that look like rockets. But if you zoom out, the real story is stranger: as those investments make powerful models widely rentable and models start to converge, the separation moves off the chip and into the plumbing — who has process‑native, decision‑ready data that AI can actually push through the system.

Most enterprise AI pilots still live in what MIT calls “pilot purgatory”: they launch, they demo well, and then they quietly stall out on the way to the P&L. In a world where most companies can rent similarly strong models, the next AI advantage won’t show up in model cards. It will show up in gross margin, cycle time, and who escapes that purgatory first.

The $700B buildout and the economics gap

By 2026, Big Tech is on track to pour close to $700 billion per year into AI data centers and compute — Amazon, Microsoft, Google, Meta, and a new cast of AI‑native infra players like CoreWeave. This isn’t a normal product cycle; it’s the buildout of a new utility layer for intelligence.

As that spend flows into the major platforms, most companies will end up renting access to broadly similar frontier models through clouds, SaaS, and APIs. The more the model layer converges, the less room there is to differentiate there — and the more the edge shifts into context: process‑native data, governed workflows, and decision traces that actually make it into production.​

Inside enterprises, that looks like two worlds sharing one building. On one floor, teams talk about what the models could do. On another, operators talk about what already breaks on a Tuesday afternoon.

The investment race is optimised for the first floor. Durable advantage is decided on the second. That distance between shared models at the platform layer and uneven context inside enterprises is the AI economics gap.​

Enterprise context as the real moat

Jensen Huang’s line is starting to sound less like a quote and more like a law of motion: every company, every industry will need AI grounded in its own data. You can see the gravitational pull of that idea in how the big players are moving. Google keeps stretching context windows so models can ingest entire codebases and operating histories, not just a prompt of the day. Anthropic is optimising for controllability over fireworks — what models can be trusted to do in real workflows, not just in benchmark PDFs. NVIDIA is selling “AI factories” that are designed to sit right next to proprietary data under enterprise‑grade governance.

Different instruments. Same direction of travel. Models are flattening. Data environments are getting more idiosyncratic, more governed over time.

Why enterprise AI still feels like pilot purgatory

Across McKinsey, BCG, and MIT work, the pattern is consistent: most companies experiment with AI, a small minority get material value, and an even smaller slice manage to scale it across the firm. The core blocker isn’t that models are too weak. It’s that enterprise data and workflows are shaped wrong for them.

Most corporate data is still a scrapbook: snapshots, reconciliations, PDF reports created long after the fact. AI sees still photos. Businesses run on video — live, stateful, exception‑heavy flows where context changes minute by minute.

In the few places where companies have rebuilt around process‑native data — shared semantics, stable identifiers, clean handoffs — the story flips. Same base models, but double‑digit reductions in cycle times and cost start to show up, because the system is finally wired for continuous decisions instead of occasional reports.

ICONIQ’s latest snapshot is a good tell: around 40% of companies with more than $500M in revenue are already deploying AI agents, versus under 30% of sub‑$100M companies. The divide isn’t who can access GPT‑class models. It’s who already has workflows, governance, and data structures that agents can plug into without rewiring the whole building first.

Where compounding will actually happen

Over the next 18–24 months, a lot of today’s AI layer — especially copilot features bolted onto unchanged systems — will quietly face the gravity of churn as models converge and prices compress. When everyone can rent roughly the same intelligence, the question shifts from “What can your model do?” to “What can your organisation do with it that others can’t?”

In that world, the scarce resource isn’t parameters, it’s context: how deeply your systems capture and reuse the data, workflows, and decision histories that cheaper models can act on at scale.

ICONIQ sees builders adjusting accordingly: nearly 70% now focus on vertical, application‑layer AI, and about half say their differentiation comes from workflows, UX, integrations, and how they apply data, not from owning a special model. As base models become interchangeable, the defensibility flows to whoever best understands customer workflows and can embed AI into the boring, mission‑critical parts of the business.

The winners are unlikely to look like science fiction. They’ll look like “systems of record for how work actually gets done” — places where process‑native data lives, agents have somewhere to run, and switching costs quietly ratchet up quarter after quarter.

Satya Nadella put it in one line: “the hard part isn’t building AI; it’s making it work with the systems that run the world”.

Sources:

• https://www.reuters.com/technology/nvidia-invests-2-billion-coreweave-expands-partnership-2026-01-26/

• https://www.venturecapitaljournal.com/ai-reality-is-sinking-in/

• https://www.linkedin.com/posts/iconiq-capital-llc_ai-reality-is-sinking-in-activity-7426670865431310337-uORl

• https://mitsloan.mit.edu/ideas-made-to-matter/why-most-ai-projects-fail-scale

• https://aiindex.stanford.edu/report/
  

Davos has a way of cutting through the noise

This year it wasn’t really about AI. It was about power, dependency, and pressure — and AI sat quietly inside all three. Between geopolitics, renewed questions about U.S. reliability, and the pace at which the U.S. and China are deploying technology, Europe is entering a familiar phase: external pressure is turning into a deadline.

We’ve seen this before.

Defence showed what happens when pressure for Europe becomes real

In defence, once the pressure hardened, Europe moved. At the 2025 NATO Summit in The Hague, allies committed to spending 5% of GDP by 2035. The effect was practical: budgets unlocked, procurement accelerated, industrial capacity began scaling before a crisis forced the tempo. Funding rounds like Helsing (€600m Series D) and Quantum Systems (160€M Series C) show what happens when incentives stop being abstract.

The benefit for Europe was tangible. Pressure didn’t just increase spending — it pulled capability, control, and value creation back into Europe, enabling the emergence of domestic champions in a sector that had long depended on external players.

AI is now under similar pressure — but a competitive one

There’s no formal commitment forcing action on AI. But the competitive pressure is unmistakable. The U.S. and China are setting global baselines for cost, speed, and performance across markets.

Stanford’s AI Index 2025 estimates U.S. private AI investment reached $109.1B in 2024. China’s push is even more explicitly industrial: Morgan Stanley frames an AI-related market that could reach $1.4T by 2030, with investments potentially breaking even by 2028 and delivering ~52% ROIC by 2030.

The market has already shifted into execution mode. Menlo Ventures estimates companies spent $37B on generative AI in 2025, up from $11.5B in 2024, with more than half going into applications — the “productivity now” layer. Even more telling: in 2024, 47% of AI solutions were built internally. By 2025, 76% of use cases were purchased rather than built. Execution has overtaken experimentation.

Europe’s instinctive response is often to look to policy. Policy matters — but it won’t execute for anyone. On AI, companies will have to move.

Europe’s constraint isn’t ambition. It’s execution.

Europe’s advantage isn’t frontier labs. It’s its industrial base and SMEs: 99.8% of enterprises, roughly two-thirds of employment, and over half of EU value added. If these firms don’t execute, Europe doesn’t execute.

Recent Handelsblatt reporting makes the gap visible — and it’s uncomfortable reading. Only a small minority of German firms see measurable revenue or cost gains from AI so far. Many are experimenting, piloting, talking — but not shipping at scale.

What this really comes down to: committing to AI

At this point, the exact percentage matters less than the decision to commit. In defence, Europe didn’t wait for perfect ROI models. It set a readiness threshold — a fixed share of capacity — and let execution follow. AI is reaching a similar moment, just without a treaty.

Most European companies still treat AI as something to optimise after the fact: a pilot here, a tool there. Under real pressure, that approach breaks. What matters is whether AI carries enough weight inside the organisation to force trade-offs. That’s why a revenue-based commitment works — it makes AI unavoidable.

Across global surveys, firms that are clearly ahead didn’t jump straight to large AI budgets. They moved in steps: from experimentation, to commitment, to deployment at scale. In practice, that shift tends to happen once AI absorbs around 1–2% of organisational capacity. Below that level, AI stays optional - above it, execution compounds.

This is also where a common misunderstanding needs clearing up. A 1–2% commitment is not the final AI bill. SaaS and cloud didn’t stop at their early budgets either. Once adoption became unavoidable, spend expanded naturally across software, infrastructure, and operating costs. AI will follow the same path. The commitment is a floor, not a ceiling.

For European firms, the implication is simple: treat AI adoption like readiness, not like an innovation programme. A practical 2026 commitment looks like:

• Industrial and large SMEs: ~1–1.5% of revenue-equivalent capacity

• Large corporates: 1–2%, enough to push AI into core operations

This isn’t about spending for its own sake. It’s about crossing the threshold where AI moves from experiments to economics.

After 12 months, that commitment should be visible in numbers, not narratives:

• Cost impact: clear, measured reductions in labour, error, or rework costs in at least one core function

• Revenue or capacity impact: faster quoting, onboarding, planning, or service that increases throughput or conversion

• Adoption: a meaningful share of employees using AI-enabled workflows weekly, not occasionally

• Spend mix: the majority of AI spend tied to deployed systems, not pilots or tools

In absolute terms, this means tens of millions per year for large industrial firms — and that’s precisely the point. Readiness is never cheap, and it’s never optional once pressure sets in.

The AI baseline is already being set in the U.S. and China. Europe’s chance to win looks the same as it did in defence: commit early, execute across the industrial backbone, and pull capability in before dependency hardens.

Sources

• Stanford AI Index 2025
  https://aiindex.stanford.edu/report/

• Morgan Stanley — China AI market & ROI
  https://www.morganstanley.com/ideas/china-ai-sleeping-giant-awakens

• Menlo Ventures — 2025 State of Generative AI in the Enterprise
  https://menlovc.com/perspective/2025-the-state-of-generative-ai-in-the-enterprise/

• BCG — As AI Investments Surge, CEOs Take the Lead
  https://www.bcg.com/publications/2026/as-ai-investments-surge-ceos-take-the-lead

• Handelsblatt — AI adoption & SME execution gap https://www.handelsblatt.com/unternehmen/davos-die-ki-enttaeuschung-der-unternehmen-nur-wenige-profitieren/100190164.html

• NATO — 2025 Summit, The Hague (defence spending commitment) https://www.nato.int

Germany is under pressure. 23,900 corporate insolvencies are expected for 2025 (+8% from 2024), with losses of ~€57 billion and 285,000 jobs affected. And there is no relief in sight: insolvency numbers are not expected to fall in 2026. At the same time, the Mittelstand is pulling back on AI. A Horváth study shows AI spending in the Mittelstand fell from 0.41% to 0.35% of revenue, while the average across German companies rose to 0.5%. That gap matters.

This post argues that AI is not a technology bet in 2026, but a practical survival tool. The companies that get through the downturn won’t be those with the biggest AI budgets, but those that systematically reduce time, errors, and rework in everyday processes. I lay out concrete steps to do exactly that: how to test AI on one costly task, prove value in weeks, decide when to invest, and then repeat the process. Done in this phased way, AI becomes a way to increase productivity under pressure — not an expensive experiment. In 2026, the winners won’t be companies that “do AI,” but those that use phased AI adoption - buying time, stabilizing cash flow, improving resilience, and lowering the risk that operational stress turns into insolvency.

A downturn changes what matters. When cash gets tight, the winners don’t magically find budget — they get more output from the same input

That’s why I keep coming back to two numbers that don’t belong in separate conversations.

First: insolvencies. Creditreform projects 23,900 corporate insolvencies in Germany for 2025, and the associated damage is not abstract: ~€57B in losses and ~285,000 jobs affected. And they’re not calling for relief in 2026 - Creditreform expects insolvency figures won’t stagnate or fall.

Second: adoption. A Horváth survey of 200 Mittelstand companies shows AI spend falling to 0.35% of revenue in 2025 (from 0.41% in 2024), while the overall average increases to 0.50%. The study warning is basically: if the AI transformation isn’t pushed now, the gap becomes an existential strategic risk.

Before anyone writes this off as “macro doom,” there’s an important correction from inside the insolvency world. VID chair Christoph Niering warned against blaming every failure on tariffs or energy costs—calling that a “dangerous misjudgment” because it delays restructuring until it’s too late.

The real divide is mentality: winners stop asking why this happened and start asking where they can still create impact today. That mindset only matters if it leads to action. In AI, the risk isn’t missing the newest tools — it’s failing to change how work gets done, step by step, with clear decisions to stop or scale.

Why the Mittelstand hesitates (& why it’s understandable)

To act, it helps to understand why many Mittelstand companies hesitate in the first place. The hesitation isn’t irrational — but it becomes a problem when it turns into paralysis. Horváth-linked reporting points to three very German brakes on AI adoption:

• Data protection and sovereignty come first. Around three in four companies prioritize European providers to meet EU rules — while a similar share criticizes the lack of innovative European technologies. The result is a stalemate: many firms end up working with global vendors and trying to “do privacy inside” those solutions.

• Bureaucracy acts as a direct adoption tax. Horváth’s study highlights show that 68% of companies see documentation and reporting duties (for example in supply chains) as strongly hindering their development.

• Unclear ROI. Too many pilots, too little ownership, and too few hard decisions to stop or scale

None of this is irrational. But pausing adoption entirely is still a strategic mistake — especially in a stressed economy.

The move that works in 2026: accelerated, phased AI adoption

The mistake many companies make is starting too big. The right approach is step-by-step adoption, with clear stop points.

More details on the steps shown can be found below:

Step 1: Pick one problem that costs you money every week (3-5 days)

Choose one recurring task where:

• people spend a lot of time,

• errors or rework happens,

• volume is high.

Typical examples:

• customer support emails and ticket routing

• invoice matching and payment preparation

• procurement requests and supplier questions

• preparing quotes or sales documentation

If you can’t clearly say “this costs us X hours or Y euros per month”, don’t start there.

Step 2: Test whether AI helps — before buying anything (4–6 weeks)

Do not sign long contracts or commission custom systems yet.

For 4–6 weeks, test whether existing AI tools can improve that one task:

• measure how long it takes today,

• introduce AI with clear rules (human checks, access limits),

• measure again.

You are looking for one simple signal: Is there an improvement on KPIs by at least 10–20%?

If not, stop. Nothing is lost.

Step 3: Only then decide how to implement properly (6–12 weeks)

If the test clearly helps, then decide:

• which vendor or provider to use long term,

• how to integrate it into your systems,

• how to document it for compliance.

At this point, investing money makes sense — because you already know where the value is.

Step 4: Repeat — that’s where the advantage comes from

Once you’ve done this once, the second and third workflows are much faster:

• the rules are clearer,

• approvals take days, not months,

• employees trust the process.

The goal isn’t one AI project. The goal is the ability to repeat this cycle calmly, even under pressure.

Who needs to be involved (no new department required)

You don’t need AI experts. You need:

• one business owner of the task,

• one person coordinating tools and IT,

• one privacy/security contact defining what’s allowed,

• one experienced employee who does the work daily.

The positive picture

Germany doesn’t lack awareness. It lacks an ability to focus and diffuse under pressure.

In 2026, the winners won’t be companies that “do AI,” but those that use phased AI adoption to turn uncertainty into operational advantage — step by step, before pressure becomes irreversible.

Phased AI programs won’t prevent every failure. But applied early and deliberately, they can buy time, stabilize cash flow, improve resilience, and lower the risk that operational stress turns into insolvency.

Sources:

• https://www.reuters.com/business/german-corporate-bankruptcies-surge-decade-high-2025-2025-12-08/

• https://www.handelsblatt.com/politik/konjunktur/firmenpleiten-keine-trendwende-in-deutschland-in-sicht/100181933.html

• https://www.handelsblatt.com/unternehmen/mittelstand/studie-mittelstand-investiert-weniger-in-ki-kostenoptimierung-im-fokus/100189594.html

• https://za.investing.com/news/economy-news/german-corporate-insolvencies-to-reach-decadehigh-in-2025-4017235

• https://www.reuters.com/business/germanys-mittelstand-cuts-ai-investments-2025-study-shows-2026-01-08/

• https://m.economictimes.com/tech/artificial-intelligence/germanys-sme-firms-mittelstand-cuts-ai-investments-in-2025-study-shows/amp_articleshow/126409251.cms