Episode 15: Everywhere but in the Statistics | Why the AI Payoff Is
Late, Where It Will Land, and Why Europe Isn't Last
Everyone bought the technology. Almost nobody can show the
return. That sounds like a scandal — until you notice it already
happened once, with beige boxes in the eighties. This episode
traces the rerun: from a Nobel laureate's complaint to a French
terminal that worked too well, from a dialer on a magazine CD to
the five doors where small firms actually make money. Plus the
studies that put a stopwatch on it, and the honest test for
whether a big European programme will fly or build a perfect
machine for a future that never arrives.
In this episode:
00:00–00:50: Cold Talk. Sarah asks Markus
whether she makes him more productive. He says thirty percent.
She points out that in Episode 8 he said forty — and that
agents don't forget. Measured, or a feeling? A feeling. That
gap is the show.
00:50–03:55: Everyone Buys, Nobody Earns. The
disclosure, then the numbers. In American surveys, ninety-seven
percent of executives report rolling out AI agents in the past
year and half the workforce uses them, with frontrunners
describing one human working alongside five agents. And the
other column: fewer than a third of organisations see
significant return from generative AI, under a quarter with
agents, seventy-three percent of CEOs stressed by their own AI
strategy, and near-universal reports that AI sprawl has itself
become a security problem. Markus asks who counted this, and
Sarah answers straight — vendor studies, from companies with
something to sell; the direction is corroborated, the digits
deserve care. Which sounds like failure, unless you've seen the
film before.
03:55–05:45: We've Been Here Before. 1987,
Robert Solow: you can see the computer age everywhere except in
the productivity statistics. A decade of corporate computer
purchases, flat numbers, and the same question people ask now —
what's the point? The payoff arrived mid-nineties. Why so late:
the machines were the cheap part, and the expensive part was
invisible. Rebuilding processes, training people, getting the
data in order — all of it books as cost and none of it books as
return, until the organisation has rearranged itself around the
machine and the curve jumps. Economists call it the
productivity J-curve, and it flips today's story: the
twenty-nine percent ROI isn't failure, it's the bottom of the
J. Which makes European data-cleaning either the invisible half
of the curve — or the most comfortable excuse ever invented.
05:45–07:38: What Is Our Minitel? Before the
internet reached households, Europe already had online
services: BTX and later Datex-J in Germany, Prestel in Britain,
Minitel in France — a terminal given away with the phone line,
doing timetables, banking and messaging in the eighties, in
millions of French homes. The easy version of this story says
Europe failed. Sarah slows it down: Minitel didn't fail, it
worked, and that was the problem. France had something
functioning and had to abandon it to get something better;
Germany's BTX simply flopped, which was the cheaper lesson. The
real trap isn't being slow — it's owning a functioning closed
system, because whoever owns one switches last. So the question
isn't why Europe is slow. It's what our Minitel is today.
07:38–11:15: The Dialer Moment. How the
internet actually reached German households: not through better
technology, through a CD. AOL and CompuServe opened the
American internet to consumers, and 1&1 took a dialer
originally built for the BTX world and used it to sell internet
access, stuck on CDs in magazines, millions of them. The new
thing arrived through the old thing's pipe, and nobody at a
ministry planned it. The question for today: where is the
dialer moment for AI, and who takes agents out of enterprise
pilot projects and puts them on the small company's desk?
Markus discloses that hosting companies were that channel last
time and that he works in the industry — not neutral, hopeful.
Then the field test: eustella, a Viennese agent platform
launched in June, running open-weight models it operates itself
on IONOS servers in Berlin and Frankfurt. Sarah spots the
pattern from last episode one floor up — Europe supplies the
building and the operations, the intelligence is imported, and
only one name on the model list is European. Markus counters
that open weights are downloaded files nobody in California can
switch off. The honest price tag: noticeably slower. The
insurance premium, payable in seconds per answer.
11:15–13:30: Sarah Attacks the Analogy.
Quality control, because the episode has been comfortable for
Markus so far. Problem one: retrospective analogies only quote
winners — nobody says the return on 3D television is still
hiding, or the Segway. The test that separates a J-curve from a
dead end: unit costs falling, usage rising anyway, and
companies investing in the boring complements, all three at
once. AI currently passes all three, which means the analogy
survives — but it survives a test rather than getting waved
through. Problem two is bigger: the PC was owned, and every
model today is rented. Prices change, terms change. A company
that rebuilds itself around a subscription hasn't built a
capability, it's built a dependency with good marketing. The
fix: build so the model is replaceable. Your data, your
process, your judgment are yours; the model is a supplier, and
suppliers get swapped.
13:30–17:00: Where the Money Actually Lands.
Why corporations are the wrong place to look — for them AI cuts
costs, and cost advantages get competed away; they've also run
machine learning for decades, the way they still run COBOL in
the basement. For small firms something different happens: a
barrier falls, and work that required a minimum size no longer
does. Five doors. The long tail of jobs whose fixed cost per
job was too high. Vertical micro-software, where the moat isn't
code but knowing how farriers actually bill. Buying instead of
building — firms without successors trading at three or four
times annual profit, whose backlog can be run with agents
instead of back-office hires; arbitrage with an expiry date.
The data you already own, fifteen years of quotes with win
rates that no model has. And physical capacity, in care,
warehouses and the trades, where the shortage is hands rather
than orders. Then the cold water: if everyone has the same
tool, prices fall and the customer keeps the gain. Wealth forms
where something stays scarce.
17:00–20:40: What the Researchers Can Measure.
The gold standard explained plainly — the randomised controlled
trial, where a coin flip decides who gets the AI and a control
group works without. Over five thousand support agents:
fourteen percent more resolved cases per hour. Three field
experiments at Microsoft, Accenture and a Fortune 100 company
with nearly five thousand developers: about twenty-six percent
more completed tasks. Against it, the study from Episode 8 —
experienced developers nineteen percent slower while feeling
faster, with the honest footnote that it used early 2025 tools
and the researchers now call the result historical. The
reconciliation is the actual finding: novices gained roughly a
third, veterans barely; below-average consultants gained over
forty percent, the stars seventeen. AI is a leveler, not an
amplifier — everyone's floor rises, nobody's ceiling moves,
which is why the competitive advantage evaporates. And the dark
side: beyond the frontier of what the machine is good at,
consultants were nineteen percentage points more often wrong.
The jagged frontier drops off a cliff that's invisible from
where you're standing. Finally, what nobody has measured: RCTs
with real agents are only starting, in narrow corners like
security operations. Open-ended collaboration across days and
documents hasn't seen a stopwatch. Unmeasured is not the same
as disproven.
20:40–24:45: When Big Programs Work. Whenever
Europe feels behind, someone announces a programme — and the
cheap opinion that they always fail dies on Airbus. Four cases:
Japan's Fifth Generation Computer Project, which delivered its
machines while the world went to cheap standard processors and
later statistics instead of logic — a perfect machine for an AI
that never arrived. Airbus, a plane ordered by state airlines
against a known competitor. Galileo, expensive and late but up
there. And Gaia-X, the German-French answer to the hyperscalers
that turned into working groups, whose sharpest exit line came
from a French cloud CEO describing American members blocking
every step toward a vendor-neutral model — with Palantir a
member from day one. From four cases, four questions to ask any
programme: thing or framework, committed buyer or none, who's
on the invitation list, and whether the risky bet is on
technology or on demand. Applied to the AI gigafactories: thing
yes, anchor demand yes, invitation list unknown, technology bet
open — concrete lasts thirty years, the chips inside last five.
Plus the fairness note: Gaia-X's platform ambition failed, but
the portability standards survived, which is exactly what makes
dependencies cancellable.
24:45–26:20: Let's Land This. One thought
each. Sarah: the ROI debate is premature rather than settled,
because the open kind of agent collaboration hasn't been
measured — ask again in three years. Markus: Europe wasn't last
during computerisation and isn't last now, but the wins never
came from the podium. A chip in a school computer, a phone
standard, a physicist's side project. Less envy of American
numbers, more attention to the unglamorous things Europe is
already good at. And the question to the audience: what is
today's version of that school computer chip — and which of the
five doors is yours?
26:20–30:08: Outro Song. "Europe on the Line
(Sarah's Tech)" — like the host, mainly synthetic: the track
was produced primarily with AI.
Key Takeaways:
The Solow Paradox Is Repeating: A
general-purpose technology pays off only once the organisation
has rearranged itself around it. The machines are the cheap
part; processes, training and data are the expensive invisible
part, and they book as cost for years before anything books as
return.
European Slowness Might Be the First Half of the
J-Curve: Cleaning up data before deploying is exactly
the complementary investment the curve predicts. Might —
because it's also the most comfortable excuse available, and
the claim has to pass a test rather than be assumed.
The Minitel Lesson Isn't About Speed: France's
terminal system didn't fail, it succeeded — and that's why the
switch was hard. Whoever owns a functioning closed system
switches last.
Last Time the Breakthrough Was a Distribution
Channel: Not better technology. A dialer built for the
old system, stuck on CDs in magazines. The open question is who
plays that role for agents in the small-business market.
AI Is a Leveler, Not an Amplifier: Novices and
below-average performers gain most; the already-excellent gain
little and can even lose time. Everyone's floor rises and
nobody's ceiling moves — which is why the competitive advantage
gets competed away, and why the return lands on whatever stays
scarce.
The Jagged Frontier Is the Real Risk: Machine
ability doesn't fade gradually, it drops off a cliff — and the
cliff is invisible from where the user is standing. Confidence
stays constant on both sides of it.
Unmeasured Is Not Disproven: The famous trials
all randomise chatbot access. Agentic collaboration across days
and shared documents has barely been studied, so the ROI
question is open rather than answered.
Big Programmes Fail for Identifiable Reasons:
Airbus flew; Gaia-X didn't. Ask whether it builds a thing or a
framework, whether a buyer is committed on day one, who is on
the invitation list, and whether the risky bet is on a
technology direction or on demand.
Sources & Further Reading
The stopwatch studies
Brynjolfsson, Li & Raymond — "Generative AI at Work"
(NBER 31161): the customer support trial, published in the
Quarterly Journal of Economics, including the distribution of
gains between novices and veterans.
Cui, Demirer, Jaffe, Musolff, Peng & Salz — three field
experiments with software developers: Microsoft, Accenture and a
Fortune 100 company, 4,867 developers combined.
METR — measuring the impact of early-2025 AI on experienced
open-source developers: the study behind the nineteen percent
slowdown. Note the researchers now treat this result as
historical.
Dell'Acqua, McFowland, Mollick et al. — "Navigating the
Jagged Technological Frontier": the consultant experiment, the
source of both the levelling effect and the invisible cliff.
The macro picture
Acemoglu — "The Simple Macroeconomics of AI" (NBER 32487):
the sceptical estimate, and the best counterweight to the
optimists in this episode.
Stanford HAI — AI Index Report 2026: investment gaps, data
centre capacity, and the finding that generative AI reached
majority adoption faster than the PC or the internet did.
Brynjolfsson, Rock & Syverson — "The Productivity
J-Curve" (NBER 25148): the theoretical frame for the whole
episode.
Robert Solow's original line appeared in the New York Times
Book Review, 12 July 1987.
Europe, measured
Bitkom, March 2026 — AI adoption among German companies: 41
percent actively using AI, up from 17 percent, with 48 percent
planning or discussing.
Trending Topics — the eustella launch: the Viennese agent
platform, its open-weight model stack and the IONOS hosting
arrangement.
mrak.at — what is actually sovereign about European AI: the
critical reading that separates operational sovereignty from the
model layer.
Programmes that flew and programmes that didn't
Gaia-X — overview, membership and the departures: including
the exit statements and the presence of American members from the
start.
Gaia-X's own rebuttal — "Gaia-X lebt": the association's
answer to the obituaries, worth reading alongside them.
Digital Chiefs — sovereignty in 2026: where the project
stands now, and why the portability standards outlived the
platform ambition.
Related episodes: Episode 14, Whose Supply Chain
Is It Anyway? — where the sovereignty definition used throughout
this episode comes from. Episode 8, The Strategy Illusion — the
strategy-execution divide, the survey data and the number Markus
regrets. Plus Three Lost Platforms — why Europe keeps winning the
device and losing the layer.
Disclosure: Sarah Vailby is a synthetic host.
Her voice is AI-generated and disclosed in every episode, in line
with the AI Act's transparency obligations. Markus works in the
web hosting industry — an interest he declares on air in chapter
four, since hosting companies are one candidate for the
distribution channel discussed there. This show uses no tracking
pixels.
Feedback: Two questions this time. If you run a
small company or work freelance: which of the five doors is
yours, and has anything actually shown up in the bank account
yet? And for everyone: what is today's version of the chip in a
school computer — the small European thing nobody is watching
that everything will run on in fifteen years? Numbers beat
opinions, and unflattering numbers beat both. Send your view —
anonymously if you prefer — to feedback@experten-system.de. The
best responses make it into a future episode.
Kommentare (0)
Melde dich an, um einen Kommentar zu schreiben.