AI as a Catalyst, Not a Replacement - Erik Slooten | Craft 2025

June 04, 2026

AI as a Catalyst, Not a Replacement - Erik Slooten (Talk Outline)

A Craft 2025 opening morning talk by Erik Slooten (introduced as “Eric Slotton”) — a seasoned executive in strategic transformation via AI and data science, active as CIO and senior executive for automation and delivery excellence (at EY), and a 1990s-trained network engineer who lived through the internet’s arrival. Explicitly framed as a motivational speech, not a technology talk: “I want to talk about what we do as a person now that this wave is on its way.” Structure: AI is already reality → craft becomes a niche → data science is the new foundation → a set of forward-looking thoughts → a personal reinvention story → the “be a master, not a hamster” pitch → an extended Q&A.

1. The Internet Parallel

1.1 A 90s engineer who had to pivot

Educated in 1990s network engineering; computers got smaller/bigger/faster, the internet and mobile telephony arrived, and fresh graduates “had to pivot really fast and learn to adjust really fast.”

1.2 The internet was net-good despite its costs

It came with threats, risks, and demands (like energy demand), but “a world without the internet would probably be less good than a world with it.”

1.3 The talk’s real subject

Not AI technology, but what engineers do as people now that AI can do primary coding, software design, and UX — “not without help and not without learning, but neither can you.”

2. This Is Already Reality

2.1 American and Chinese dev environments are half agents

In American and Chinese software teams, ”half of which are now agents.”
Agents open meetings, interact with people, and answer budgetary questions, look up code, or check the current lifecycle of development.

2.2 His EY team all use AI

The team has developers, data scientists, architects, and business analysts — ”none of them work without AI.”
Code is partially developed, reviewed, or debugged by AI — but it’s not “lights out with no one working.”

2.3 Idea-to-implementation went from months to weeks

Throughput from idea → concept → client implementation dropped from months to weeks.
A high-profile security case delivered in three weeks with two people and five agents.

2.4 Modern tooling is table stakes

Like testers who don’t use GitHub to run unit tests are “wasting your time and everyone else’s” — you could do something faster or more interesting.

2.5 The existential question

“What did I work for all these years if I now have agents that can code?” — a question he’s asked himself, and invites the audience to answer.

3. Disruption Everywhere

3.1 Saudi Arabia’s fully-AI clinic

Launched ”last month” — fully AI avatars, deep learning on patient records, “the full works,” built within 1.5 years.
Doctors are still there but behind the scenes: managing agents, supervising, correcting, retraining, keeping data clean, handling collisions and complaints.

3.2 AI-run hotels and apartments

Real-estate companies rent out fully-AI hotels; only a person to move suitcases.
His Madrid stay two months earlier: “I didn’t see a single person” — beautiful apartment, everything controlled by phone, chat the AI (“make sure I have coffee tomorrow”) and it just happened.

3.3 People already ask Perplexity medical questions

He sent his blood results to Perplexity to double-check the doctor.
People trust AI for “I have a weird spot, should I go to the doctor?” instead of “sitting for two hours looking at a grumpy nurse.”

3.4 The bottom line on jobs

AI won’t take all our jobs, “but it replaces the need for certain things to be done by people.”

4. Craft Becomes a Niche

4.1 The handmade-furniture analogy

~150 years ago all tables were made by hand.
Today “handmade” is premium — guests say “somebody actually made this table, you didn’t go to IKEA — that’s amazing.”

4.2 Handcrafted code as a premium niche

Handcrafted code becomes a premium niche; most lower-level coding done by agents, supervised by people who understand it.
Those people still need to know collisions, memory leaks, and that quantum computers run things differently than binary systems — but “you don’t need to do these things yourself anymore.”

4.3 The carpenter with laser-guided tools

A skilled carpenter with laser-guided, even AI-infused tools is still master of their craft.
“You’ll be the master of your craft, but surrounded by a whole room full of software at your service” — stop thinking you write everything yourself; manage your world differently.

5. Data Science Is the New Foundation

5.1 Data exploded, then AI exploded it more

The internet forced a data explosion; AI needs ”clean and huge amounts of data.”
In the last three years: from 79 zettabytes to 181 zettabytes by the end of this year — “data has exploded.”

5.2 Learn data science deeply

“If you haven’t studied it already, data science is your new foundational key to anything you do.”
Without deep data-science understanding “you won’t understand what happens to your world” — knowing what’s wrong “starts with the data underneath.”

5.3 Master of the toy shop and master constructor

You’ll have more tools (Copilot, GitHub, Cursor, whatever coding assistant) and mountains of construction material — zettabytes of it.
You can artificially generate more synthetic data, customized exactly for what your model needs to learn.

6. Forward-Looking Thoughts

6.1 There is no OSI layer for AI

He’s talked to international standards institutes: “we need an OSI layer for AI.”
You can’t treat AI as just another application — its function in the architecture and its whole logical function differ from normal software.

6.2 Agents are your new friends

“If you don’t start making friends with your agents now, you’ll be very lonely very soon.”
Bring agents into your life as colleagues; learn to trust them and how to talk to them.
Prompt engineering = your new communication skill: not every AI/agent reacts the same, so you talk to a finance AI differently than a software-development AI.

6.3 The smartphone will be disrupted

OpenAI is investing billions in new devices with ”the designer of the iPhone,” having raised billions in starting capital.
Prediction: purpose-built AI-carrier devices — more ergonomic, less disruptive, an assistant/buddy ”on your shoulder or on your ear.”
In 10 years, everyone staring at phones in the elevator will look like “a thing of the past.”

6.4 Ethical code matters

If you ”bitch at your AI co-pilot all day,” it learns behaviors and generates output based on your inputs.
”Mistreating AI should be a crime just as much as mistreating your colleagues.”
Corporate conduct for AI use should be codified in company law.
AI can profile you and create biases (gender, age, racial) — these should be filtered, blocked, and retrained; frameworks can reset AI bias fast, “much harder with people” (who all see “a stamp on your forehead”).

6.5 Data science, again, as the foundation

Without deep data-science understanding, “you’ll be polishing the building but won’t understand what it’s made of” or what will make it fall down.

7. Reinvent Yourself (Personal Story)

7.1 The heart scare at 41

At 41, with young kids, he suddenly started falling over; the doctor diagnosed a heart problem. “Something big and powerful had taken over my life.”

7.2 The choice

He could have slowed down, switched off, or changed to a less stressful career; the doctor said take it easy, stop smoking.

7.3 Learning to surf

Instead he took six months off to do what he’d always said he wanted — learn to surf (hard at 41: “slower, fatter, not as strong as at 25”).
He’s proud of the photo of one of his first waves; now the family surfs every year.

7.4 The lesson — reinvent yourself anytime

At 35 he’d have said “no time, no money, won’t happen” — but the shock let him pivot, and he’s “way healthier now.”
His own reinventions: network engineering → software development → cybersecurity → AI.
”You can reinvent yourself anytime you want” — stop clinging to what you thought you’d do forever; accept the world has changed.

7.5 “Be a master, not a hamster”

Don’t be a master only at writing code — be a master of your software environment, your client’s expectations, and the workshop of AI.
Become an orchestrator: producing creative outcomes by combining different AIs while warranting quality, trust, testing, and ethics.
“I can teach someone to code zero-to-hero in a few months. Being a master of your own destiny and an orchestrator, that’s a real art. Writing C is just a trick — this is mastery.”

8. Q&A

8.1 Q1 — The “dead internet theory”: if real-people data is shrinking, won’t AI train only on AI data and hallucinate?

Hallucinations aren’t caused by the source (human vs. synthetic); they come from discrepancies/conflicts where the model must “improvise.”
There’s a huge craft in cleaning up / recycling trash data — “data science will become the absolute cement of everything.”
EY spends huge time cleaning client environments before building agents; nothing wrong with synthetic data if it’s fit for purpose and trains the model to a higher-quality outcome.

8.2 Q2 — Where are the biggest mid/long-term limitations and boundaries of AI?

He teaches AI at university; a key challenge is awareness — AI isn’t conscious of the world and lacks common sense (“something a lot of people also lack”).
New models may be retrained to be self-aware/world-aware before relearning everything else.
Conflicting requirements cause hallucination — an agent designing software for 2–4 clients with different perspectives struggles to resolve conflicts.
Analogy: self-driving cars facing “hit the child or the dog” — the AI just stops or hallucinates.
AGI timelines are contested (Silicon Valley says “now / few years,” others “10, 20, if ever”); if AGI arrives, ethical/philosophical questions must be answered before the (easier) technical ones.

8.3 Q2b — (Aside) The five agents his company uses

Document ingestion — a RAG agent pulling historical client work, depersonalizing it into reusable EY intellectual property (“Lego blocks”).
Copilot — “everyone uses Copilot.”
GitHub and Cursor — for coding.
A ServiceNow agent — for service management.
An HR agent — running 24/7 for policy/holiday/foreigner questions, loaded with Hungarian labor laws.
(He can think of “at least 10 more.“)

8.4 Q3 — Is there a zero-to-hero path for data science?

Learn statistics first — you need to be good at math.
Then good online courses (Udemy and similar), including interactive code-writing courses — easy to get into as an engineer, ”strong understanding within six months.”
Find an internship / hands-on environment — EY takes Corvinus, ELTE, BME students for half a year, half training + real practice with data scientists.
“Don’t just read the book” — you only “feel it” by doing.

8.5 Q4 — How do you handle compliance of customer data fed into AI agents?

A framework called Responsible AI covering compliance (e.g., an agent offering a non-compliant offer or saying gender-biased/ageist things).
(Joke: “they started shooting developers, saying you don’t need them anymore… it’s a joke, guys.“)
Combines a technical framework with monitored repositories where all AI interactions are checked for bias, cyber threats, policy/legal non-compliance (European laws), and unethical practices (emotion/facial-expression profiling).
Governance standards read by the quality department, risk manager, even the CEO — and the framework is continuously refreshed.

8.6 Q5 — What about the negatives of LLM-generated code (speed up, but quality/maintainability/defects/security/critical-thinking down)?

“That’s just a temporary thing” — lots of experimenting.
LLMs will move back to a foundational role; agents will do the actual coding — “I don’t think LLMs long-term will be the thing that does the coding.”
Cautionary tale: a Sweden company replaced all customer-care staff with 700 agents in 3 months, saved money and pleased shareholders — then care quality collapsed, customers left, the company nearly went bankrupt, and they pulled all the agents at once.
LLMs don’t run the constant self-questioning feedback loop (“is this right?”) that humans do — but agents can be trained to be critical, ask questions, and rethink assumptions.
Mechanisms are coming — ”agents checking agents on code quality”; it’s early days but this will be fixed.

8.7 Q6 — (Re: Simon Wardley’s opening talk on contextual tooling) What day-to-day AI tools fulfill “contextual tooling” for a front-end developer?

For some clients they use natural language plus Cursor.
But for the trendy real contextual coding, “we are not using anything in practice just yet” — it’s early days.
Advice: start playing, do some vibe coding, learn prompting instead of actual coding.

8.8 Q7 — How can we trust AI to avoid bias when so much training data comes from white men and carries historical bias?

You can’t fix all the data produced by white males — instead detect the bias, fix it, and retrain.
Synthetic data can remove some bias, but “retraining the whole internet is unthinkable.”
White-male bias isn’t the only problem — there are pro-Chinese, pro-American, and political biases being generated now, plus cyber data steering opinions.
Clean data matters, but guardrails / a protective harness against biased output are as important, if not more — you’ll never fully clean your company’s data.

8.9 Q8 — What’s more economically viable: engineering/using AI agents, or moving manual data work to shared service centers (SSCs)?

EY’s concept is ”fast shoring” (also used elsewhere) — ”it’s both, not one or the other.”
Combine agents with people to go faster; moving work to India/Hungary is “never the only answer.”
Shoring shouldn’t be purely economic; shared service centers have evolved into delivery centers / innovation hubs improving products and processes.
Smart companies build concepts including both agents and people wherever economically viable.

People & References Cited

Erik Slooten (introduced phonetically as “Eric Slotton”) — speaker; EY executive, ex-CIO; 1990s network engineer; teaches AI at university.
EY — his employer; “Responsible AI” framework and “fast shoring” concept.
Simon Wardley — gave the opening talk (referenced for “contextual tooling for coding”).
OpenAI — investing billions in new AI devices with “the designer of the iPhone.”
Perplexity — used for a medical/blood-results question.
Companies / tools: Copilot, GitHub, Cursor, ServiceNow, Udemy, IKEA (analogy).
Universities: Corvinus, ELTE, BME (Hungarian students interning at EY).
Case examples: Saudi Arabia’s fully-AI clinic; AI-run hotels; a Madrid AI apartment; a Sweden company that replaced customer care with 700 agents and nearly went bankrupt.
Concepts: AI as catalyst vs. replacement; “age of infinite input”; craft as a premium niche; data science as “the new cement/foundation”; zettabyte data growth (79 → 181); “no OSI layer for AI”; agents as colleagues; prompt engineering as a communication skill; ethical/responsible AI and bias reset; AGI and its ethical questions; the self-driving trolley problem; the dead internet theory; “agents checking agents”; fast shoring; “be a master, not a hamster.”

Video: https://www.youtube.com/watch?v=GFztkIhesO4 — Transcript via yt-transcript.sh; outline generated from the transcript.