MUNICHTECH EXPO 2026 · EUROPE'S AI SUMMIT
Making AI Work for People
The five-principle system that moves AI pilots into production
Asmaa Ibrahim
Based on the new book from Wiley
THE GAP
The money is moving. The value is not.
Five years of enterprise AI investment, measured honestly.
80%
of AI projects fail to deliver
the value they promised
42%
of businesses expected to abandon
most AI projects in 2025 — up from 17%
4%
of companies generate real,
measurable value from AI
Sources: RAND Corporation (2024); S&P Global Market Intelligence (2025); BCG (2024); IDC/Microsoft (2024).
THE DIAGNOSIS
The models work. The organizations don't
1
Technology is rarely the blocker
Accuracy on the vendor's test data is not accuracy on your messy production data.
2
Adoption decides the outcome
One European operator's engineers ignored the AI's recommendations 85% of the time. The model was fine.
3
Integration is where budgets die
A system that cannot talk to 40% of your legacy estate is a science project, not a product.
01
Patterns
Failure is not random. It repeats, and you can spot it on day one.
PATTERNS
Eight failure patterns you can identify
1
FOMO syndrome
A competitor's press release becomes your roadmap.
2
Technology-first thinking
The tool is chosen before the problem is written down.
3
Human resistance ignored
Nobody asked the people whose jobs change.
4
Black box distrust
Users cannot see why, so they override.
5
Data reality gap
Demo data was clean. Production data never is.
6
Pilot-shaped architecture
Built for hundreds of requests, sold for millions.
7
Governance added later
Ethics and compliance arrive after the incident.
8
Vanity metrics
Accuracy goes up. No business number moves.
Three or more of these patterns present? Roughly 80% of those projects failed within two years. Five or more? I have never seen one succeed.
PATTERNS
Patterns don't queue. They cascade.
FOMO
→
Technology
first
→
Humans
ignored
→
Black box
distrust
→
Abandoned
system
Each failure multiplies the next. Lost trust makes people leave. Technical debt slows every release. Customers move to competitors.
PATTERNS
Two public receipts
Drive-thru voice ordering
Pattern: FOMO + technology-first
A rushed competitive response put voice AI into 100+ restaurants. The workflow needed roughly 95% accuracy. The system delivered about 85%. Orders went viral for the wrong reasons and the test was ended in 2024.
Accuracy is not a score. It is a threshold set by the workflow.
Healthcare cognitive computing
Pattern: replace instead of enhance
Roughly $4–5B went into acquisitions between 2011 and 2022. One hospital spent $62M before cancelling. The assets sold for about $1B in 2022 — an 80% loss on the acquisitions alone.
AI chosen to replace experts fails. AI chosen to enhance them ships.
Sources: Restaurant Business (2024); CBS News (2024); STAT News (2022); Strickland (2019).
02
Framework
Five things every successful deployment had. All five. Every time.
FRAMEWORK
PRESS: five principles, no optional ones
Do one well and skip the rest, and you still fail. I have watched it happen.
P
People-first
AI that makes humans better at their jobs. Ignore this and users find creative ways to sabotage it.
R
Responsible
Fairness and governance from day one. Cleaning up afterwards costs ten times more.
E
Explainable
If users cannot follow the reasoning, they will not trust the output.
S
Safe
Define what happens when it is wrong, before it is wrong.
S
Sustainable
Can you afford to run it in year three, at real volume?
Think of it as a house: people are the foundation, responsibility the walls, explainability the windows, safety the roof, sustainability the maintenance. Remove one and the house is uninhabitable.
FRAMEWORK
Score it honestly, your scores predict your budget
PRESS is not a poster. It is a diagnostic that converts directly into money and time.
Organizations with high readiness scores
succeed approx. 3x more often than unprepared competitors
A people-first score below 5
adds roughly 50% to your change-management budget
A sustainable score below 5
doubles, sometimes triples, your infrastructure costs
Any dimension below 4
sets your velocity, no matter how strong the other four are
Low scores do not disqualify you from AI. They redirect you to the AI you can actually deliver this year.
FRAMEWORK
Your profile shape decides your first project
Spiky
Strong infrastructure, weak culture
P3 R4 E7 S6 S8
A large retailer with this shape built warehouse and inventory optimisation first — backend work where infrastructure matters and human resistance does not.
Start behind the scenes.
Inverse spiky
Strong culture, limited infrastructure
P8 R7 E6 S7 S4
A regional operator with this shape leaned on its people advantage and partnered for infrastructure instead of building it, putting AI beside agents rather than in front of customers.
Start beside your people.
Flat
No crisis, no superpower
P5 R5 E5 S5 S5
No dimension rescues you and none blocks you. The right move is moderate-complexity work that raises confidence across all five at once.
Start in the middle.
Same framework, three different first projects. Copying someone else's roadmap is how you inherit their failure.
FRAMEWORK
Your scores pick your mode, not your ambition
Five ways humans and AI share the work. Human-dominant on the left, AI-dominant on the right.
Assistive
Humans keep about 90% of control
NEEDS
P ≥5 · R ≥5 · Sus ≥4
TIMELINE
3–6 months
BEST FIT
GenAI for varied tasks
Suggestive
AI recommends, humans decide
NEEDS
P ≥6 · R ≥5 · E ≥7
TIMELINE
4–8 months
BEST FIT
Traditional ML
Collaborative
Neither side could do it alone
NEEDS
P ≥7 · E ≥6 · Sus ≥6
TIMELINE
8–18 months
BEST FIT
GenAI
Autonomous
AI acts, humans monitor
NEEDS
Safe ≥8 · Sus ≥7 · R ≥6
TIMELINE
6–12 months
BEST FIT
Traditional ML, not GenAI
Adaptive
AI handles about 95% of routine work
NEEDS
P ≥7 · R ≥7 · Sus ≥7
TIMELINE
12–24 months
BEST FIT
GenAI + specialist hybrid
A European retailer with strong infrastructure but weak people readiness ran autonomous pricing on customer-facing shelves. Prices moved six times a day, staff could not explain them, and the local press ran "price roulette" stories.
Mode requirements, timelines and AI-type fit from the book's Chapter 3.
03
Business case
AI economics break the rules your CFO learned from software.
BUSINESS CASE
Where the budget actually goes: 70 / 20 / 10
70% — People and process
Change management, training, workflow redesign, and the slow work of getting people to trust the system.
20% — Technology and data
Pipelines, integration, infrastructure. Real, but not where the risk concentrates.
10% — Algorithms
The part everyone argues about in the steering committee.
If your business case inverts this ratio, you are funding the 10% and hoping the 70% happens for free.
Allocation model from the book's Chapter 2 economics; consistent with BCG (2024) findings on AI value realisation.
BUSINESS CASE
Plan for the J-curve
Illustrative shape, not a forecast.
6–18 months
of negative return before value turns positive
12–24 months
before meaningful returns for most implementations
Six months
is when unprepared sponsors panic and pull funding
A vendor promising positive ROI in six months deserves intense skepticism
BUSINESS CASE
Three ways to justify the spend
1
Value velocity
Days to first euro of value, not ROI at completion.
One operator cut theirs from 45 days to 25. That single number told them they would survive.
Use when sustainability is your weak dimension.
2
Cost avoidance
What does it cost to not do this?
Telecom fraud losses reached $38.95 billion globally in 2023, up 12%. One operator's fraud programme cut fraud by up to 80% with 100+ models in production.
Use when the risk is already on the balance sheet.
3
Capability building
Strategic positioning rather than near-term return.
Compliance and regulatory AI may never show classic ROI, yet prevents existential risk.
Use when the alternative is being locked out of a market.
Sources: CFCA (2023) Global Fraud Loss Survey; Databricks case study (2025); financial models from the book's Chapter 2.
04
Phases
Design, dogfood, alpha, beta — then the measurement that proves it worked.
PHASES
The ten-week progressive validation journey
Each phase answers one question. Skip a phase and you pay for it in production.
1
Design
Weeks 1–2
What should we build?
PRESS-enhanced user stories, constraint mapping, and a three-hour alignment workshop with six to eight stakeholders.
2
Dogfood
Weeks 3–4
Does it work for us?
Your own team uses it daily. If your team will not use it, no customer will.
3
Alpha
Weeks 5–6
Does it work for friendly users?
External users who will tolerate rough edges and tell you the truth.
4
Beta
Weeks 7–10
Can it scale safely?
Production conditions, real volume, real failure modes, real cost per decision.
Users drive the decisions, not vendors. PRESS guides every step instead of arriving as an afterthought. Each phase compounds the last.
PHASES
Measure in three layers, or measure nothing
Each layer answers a different question for a different audience. Teams measuring all three report 3x the financial benefit of single-layer measurement.
Business
Revenue, cost, risk
Operational
Adoption, trust, workflow fit
Technical
Accuracy, latency, throughput
What your sponsor funds
A 25-day value velocity proved returns were arriving faster than the system consumed resources.
Shaped by your PRESS profile
A 78% agent acceptance rate showed whether accuracy was turning into suggestions people actually used.
The foundation everyone starts with
Over 90% context retention kept conversations coherent. Necessary — and on its own, worth nothing.
Give each layer a trigger: accuracy under 85% starts drift analysis, adoption under 70% starts a feedback cycle, and a north star off by 10% triggers a full review within 48 hours.
Check the book's Chapters 5 and 6.
PHASES
Your collaboration mode decides what 'working' means
Assistive
Regional telecom operator: agents keep ~90% of control.
Measure the human-AI team, not the model alone.
Autonomous
Large retailer · system runs with minimal human touch
No human feedback loop, so the system must monitor itself.
Same pyramid, different middle layer. Pick one north star: if a dimension scores below 4 the metric fixes that weakness, and if one scores 7 or higher it exploits that strength.
Measurement examples from the book's Chapter 5 case analysis across telecom and retail deployments.
05
Scale
The pilot that worked is the system that bankrupts you.
SCALE
Inference is 80–90% of lifetime cost
Everyone budgets for training. The bill arrives one prediction at a time.
Conversations get longer
Beta users asked simple questions. Production users bring complex problems — 10 exchanges became 23.
Context accumulates
Every exchange carries all the previous ones, plus a system prompt on top.
Success increases volume
Daily conversations rose from 10,000 to 21,000 because people actually liked it.
A 15x cost surprise, and none of it was a technical fault. Model your worst users, not your best demo.
Token economics from the book's Chapter 8; inference cost share consistent with IBM Research (2024) and a16z (2024).
SCALE
Optimise in this order — software first, hardware last
Level 2
Inference optimization
Batching, caching, framework choice. No new hardware, no model changes.
10–50x improvement · roughly 70% of all gains · start here
Level 3
Model optimization
Quantisation, pruning, distillation. Gains travel with the model wherever it runs.
Permanent efficiency · do this second
Level 1
Hardware acceleration
Buy capacity only after you know what optimized software actually needs.
One operator saved $360,000 by measuring first · do this last
Never sell this internally as "AI optimisation" — executives switch off. Sell it as protecting the millions already invested.
Sources: Anyscale (2023); vLLM (2023); Dettmers et al. (2022); results from the book's Chapter 9 case analysis.
SCALE
The European constraint is a schedule item
1
Build compliance into the plan
Transparency obligations under the EU AI Act took effect in February 2025. Add two to four weeks for compliance documentation and energy reporting in any rollout timeline.
2
Budgets start smaller
European AI budgets run well below US equivalents, and higher energy costs add materially to data-centre expense. Efficiency is not optional here.
3
Which makes PRESS an advantage
Responsible and explainable are already regulatory requirements on this continent. Teams that build them in from day one are not slowed down — they are pre-cleared.
Compliance is not the thing that slows you down. Retrofitting it is.
Take six things back to your team
1
Count your patterns
Three or more, and the odds are against you. Do this before the next steering meeting.
2
Score PRESS honestly
Your weakest dimension sets your velocity, your budget multipliers and the mode you can run.
3
Pick the mode you qualify for
Not the most advanced one. The technology is rarely what makes the wrong mode fail.
4
Fund 70/20/10
People and process, not algorithms. Show the J-curve before someone else discovers it.
5
Dogfood, then measure in three layers
Two weeks of your own team using it daily, then technical, operational and business metrics.
6
Optimise software first
Batching and caching before GPUs. Sell it as protecting the investment you already made.
The 20% success rate is achievable on purpose, not by accident. · New book from Wiley · Questions welcome
Thank you!
Questions welcome — and the PRESS assessment is yours to run on Monday morning.
P
R
E
S
S
Asmaa Ibrahim
Author · New book from Wiley