---
id: software-ceo-dinner
type: deck
title: 'The AI Reality Check: talk at the Software CEO Dinner Vienna'
summary: >-
  AI models now complete software tasks that take a person over 17 hours, yet
  only 37% of companies report any EBIT impact from AI. The talk shows why, and
  how one weekly task moves in five stages from no AI to an agent workflow.
language: en
date: '2026-09-23'
speakers:
  - Nik Redl
publisher: Specialty Tokens
audience: Software company CEOs and founders deciding where AI belongs in their company.
topics:
  - ai agents
  - ai adoption
  - ai strategy
  - ai spend
  - agent workflows
  - ai for software companies
page: 'https://specialtytokens.com/resources'
markdown: 'https://specialtytokens.com/resources/software-ceo-dinner.md'
slides: 'https://specialtytokens.com/slides/s/software-ceo-dinner?p=1'
slide_count: 29
access: The slides open after a short form on the page.
---

# The AI Reality Check: talk at the Software CEO Dinner Vienna

> AI models now complete software tasks that take a person over 17 hours, yet only 37% of companies report any EBIT impact from AI. The talk shows why, and how one weekly task moves in five stages from no AI to an agent workflow.

## Key points

- AI models can now complete software tasks that take a person 17 hours 25 minutes with 50% success (Claude Mythos Preview), up from 4 hours 53 minutes for Claude Opus 4.5 in November 2025. (slide 7)
- Only 37% of companies report any EBIT impact from AI, while 80% report higher personal productivity. (slide 10)
- Monthly AI spend per employee varies by more than 600 times: the median is $11.95, the top 10% spend $650 and the top 1% spend more than $7,400. (slide 11)
- Ramp's AI agents build customer integrations in minutes while engineers review them; 75 integrations have shipped this way, and AI use at Ramp grew 6,300% in one year. (slide 13)
- Allianz built Project Nemo, seven AI agents that check storm claims for cover, weather and fraud, in under 100 days; the time to process and pay fell by 80%, and a person approves every payout. (slide 15)
- A recurring task moves through five stages on the way to an agent: no AI, chat, skills (stage 2.5), connected systems, and an agent workflow where a trigger starts the work and people take the exceptions. (slide 19)
- An AI agent has four parts: a trigger (time or event), a model (the LLM), a skill (the team's know-how) and a connection (access to the system it works in); each part alone is only a tool. (slide 24)
- Small AI systems compound: in the talk's example, hands-on work falls from 40 hours today to 10 hours by month 6, because two hours of setup once gives five hours back every week. (slide 25)
- Nik Redl recommends ChatGPT or Claude, says tools like Copilot or Langdock fall behind, and advises Opus or Sol as the minimum model. (slide 26)
- Nik Redl bets that AI can already do more than 90% of office work today, given enough time and money. (slide 27)

## What's inside

1. Models crossed a threshold
2. Spend is up, results are not
3. Four companies that turned AI into results
4. One weekly task, five stages
5. The four parts of an agent
6. Small systems compound
7. Five points, and what comes next

## The written edition

AI models can now complete software tasks that would take a person more than 17 hours, yet only 37% of companies report any EBIT impact from AI. Nik Redl's talk at the Software CEO Dinner in Vienna argues that the gap lies in how the work is set up, not in the models. It follows one weekly task from no AI to an agent workflow.

### 1. Models crossed a threshold

The talk opens with Peter Steinberger, a founder and engineer from Vienna who made OpenClaw. His model bill for 30 days came to $1.3M: 603B tokens, 7.6M requests and about 100 Codex instances.

By December 2025, with Claude Opus 4.5, models had crossed a threshold. On software tasks, the human task length a model can complete with 50% success went from 4h 53 min (Claude Opus 4.5, November 2025) to about 12 hours (Claude Opus 4.6, February 2026) to 17h 25 min (Claude Mythos Preview).

AI moved from answers to actions: it reads, plans, uses tools and checks the work. If work has inputs, rules and a result, it can be codified. That started with coding; now it is knowledge work.

### 2. Spend is up, results are not

Only 37% of companies report any EBIT impact from AI, while 80% report higher personal productivity.

AI spend per employee varies by over 600×: the median is $11.95 a month, the top 10% spend $650 and the top 1% more than $7,400.

### 3. Four companies that turned AI into results

**Flexport** made freight cheaper and faster. Its planning AI fills cancelled bookings with later shipments. Ocean spend fell 2% and transit time 20%: $20K saved per $1M of ocean spend.

**Ramp's** agents build customer integrations in minutes. A customer describes what they need; an agent builds and tests it; engineers review it. 75 integrations have shipped this way. Non-engineers make 12% of human code changes, and AI use grew 6,300% in one year.

The third case makes AI part of every headcount request. Before a team asks for more people or resources, it must show why AI cannot do the work.

**Allianz** built Project Nemo in under 100 days. Seven AI agents handle storm claims: they check cover, weather and fraud, and a person approves every payout. The time to process and pay fell by 80%.

### 4. One weekly task, five stages

More AI spend does not become more output because companies do not jump straight to agents; they move through stages. The running example: once a week, someone checks the new Google Ads keywords. "Office chair meme" and "free office chair" sit next to "ergonomic office chair", and the list has to be cleaned.

- **01, no AI.** Every Monday the employee does the work. Every case starts from zero.
- **02, chat.** AI helps, but the employee still moves the data and carries the work.
- **2.5, skills.** The team writes its know-how down once: a skill tells the AI how the team checks keywords. The employee still moves the data.
- **03, systems.** Connected AI can read the real data. Data access removes copying, but it does not create a workflow.
- **04, agent.** The work runs as a system. The trigger starts the work; people take the exceptions.

### 5. The four parts of an agent

- **Trigger** (time or event): every Monday, or a new lead.
- **Model** (the LLM): reads, reasons and writes.
- **Skill** (know-how): how the team checks keywords.
- **Connection** (access): reads and changes Google Ads.

Each part alone is a tool. Together they make an agent that does the whole task on its own.

### 6. Small systems compound

The workload stays the same; the part you do by hand shrinks. In the deck's example, hands-on work falls from 40 hours today to 35 in month 1, 25 in month 3 and 10 in month 6. Two hours once gives five hours back, every week, and the freed time builds the next system.

### 7. Five points, and what comes next

Five things there was no time for:

1. ChatGPT or Claude: both work. Tools like Copilot or Langdock fall behind.
2. Use Opus or Sol at minimum. Fable and Astra are great, but expensive.
3. Computer use automates most of the rest. ChatGPT is especially strong here.
4. AI makes mistakes. People make more.
5. Test again, and often. "AI can't do that" is often wrong for the newest models.

Nik's bet: AI can already do 90%+ of office work today, with enough time and money. In 4 months, AI can do tasks twice as long as today, and the same performance gets 13× cheaper per year.

### Frequently asked questions

#### What is the main argument of the AI Reality Check talk?

Only 37% of companies report any EBIT impact from AI, although models now complete long, multi-step work. Results come from turning recurring tasks into small systems that run on their own.

#### What are the stages from no AI to an agent?

No AI, chat, skills, connected systems and an agent workflow, where a trigger starts the work and people take the exceptions.

#### What turns an AI tool into an agent?

A trigger, a model, a skill and a connection to the system it works in. Each part alone is a tool; together they do the whole task.

#### Which AI tools and models does the talk recommend?

ChatGPT or Claude; tools like Copilot or Langdock fall behind. Use Opus or Sol at minimum.

## The slides

### Slide 1: The AI Reality Check

SOFTWARE CEO DINNER
WIEN · 23.09.2026
The AI Reality Check
What changed this year, what agents can do now, and why most companies still see little of it.
NIK REDL · SPECIALTY TOKENS
HOSTED BY MAGUAR CAPITAL, I5INVEST AND HERBST KINSKY

### Slide 2: Engineer, operator and AI advisor

Nik Redl
Engineer, operator and AI advisor
Built Mokker past one million users before its acquisition by Soona
Former Chief of Staff at Soona
Co-founder of Specialty Tokens: we build AI systems inside companies
INTRODUCTION

### Slide 3: Where our learnings come from

Where our learnings come from
Gaslab
Lessons from real rollouts, not theory.
INTRODUCTION

### Slide 4: Peter Steinberger

Peter Steinberger
Founder and engineer from Vienna.
Made OpenClaw. Moved the frontier.

### Slide 5: 603B tokens · 7.6M requests · about 100 Codex instances

PETER'S MODEL BILL, 30 DAYS
$1.3M/Month
603B tokens · 7.6M requests · about 100 Codex instances
THE AI REALITY CHECK

### Slide 6: By December 2025, models had crossed a threshold.

CLAUDE OPUS 4.5
By December 2025, models had crossed a threshold.
THE AI REALITY CHECK

### Slide 7: The human task length an AI model can complete with 50% success.

CLAUDE MYTHOS PREVIEW · MEASURED ON SOFTWARE TASKS
17h 25 min
The human task length an AI model can complete with 50% success.
CLAUDE OPUS 4.5, NOV 2025: 4H 53 MIN · CLAUDE OPUS 4.6, FEB 2026: ABOUT 12 H
THE AI REALITY CHECK

### Slide 8: AI moved from answers to actions.

AI moved from answers to actions.
Read · plan · use tools · check the work
THE AI REALITY CHECK

### Slide 9: If work has inputs, rules and a result, it can be codified.

If work has inputs, rules and a result, it can be codified.
it started with coding
now knowledge work
THE AI REALITY CHECK

### Slide 10: of companies report any EBIT impact from AI.

37%
of companies report any EBIT impact from AI.
80% report higher personal productivity.
THE AI REALITY CHECK

### Slide 11: AI spend per employee varies by over 600×.

AI spend per employee varies by over 600×.
MONTHLY SPEND
MEDIAN
$11.95
TOP 10%
$650
TOP 1%
>$7,400
THE AI REALITY CHECK

### Slide 12: Flexport made freight cheaper and faster.

Flexport made freight cheaper and faster.
Its planning AI fills cancelled bookings with later shipments. Ocean spend fell 2%, while transit time fell 20%.
$20K
SAVED PER $1M OF OCEAN SPEND
THE AI REALITY CHECK

### Slide 13: Ramp's agents build customer integrations in minutes.

CASE 02 / 04
Ramp's agents build customer integrations in minutes.
A customer describes what they need. An agent builds and tests the integration, and engineers review it. 75 integrations have shipped this way.
12%
OF HUMAN CODE CHANGES BY NON-ENGINEERS
6,300%
MORE AI USE IN ONE YEAR
THE AI REALITY CHECK
Eric Glyman, co-founder and CEO

### Slide 14: AI is part of every headcount request.

AI is part of every headcount request.
Before a team asks for more people or resources, it must show why AI cannot do the work.
Required
FOR EVERY RESOURCE REQUEST
THE AI REALITY CHECK

### Slide 15: Seven AI agents handle storm claims. A person approves every payout.

CASE 04 / 04
Seven AI agents handle storm claims. A person approves every payout.
Allianz built Project Nemo in under 100 days. The agents check cover, weather and fraud. Then a person makes the final decision.
80%
LESS TIME TO PROCESS AND PAY
<100
DAYS TO BUILD AND LAUNCH
THE AI REALITY CHECK
Oliver Bäte, CEO · Photo: MsBlix, CC BY-SA 4.0

### Slide 16: Why does more AI spend not become more output?

Why does more AI spend not become more output?
THE AI REALITY CHECK

### Slide 17: Companies do not jump straight to agents.

Companies do not jump straight to agents.
One customer request, four ways of working.
01
No company AI
The employee does the work
02
Company-wide chat
AI helps, people move the data
03
Internal systems
AI can read company data
04
Agent workflow
An event starts the process
THE AI REALITY CHECK

### Slide 18: Someone checks the new Google Ads keywords.

THE RUNNING EXAMPLE · ONCE A WEEK
Someone checks the new Google Ads keywords.
ergonomic office chair
office chair meme
buy office chair vienna
free office chair
THE AI REALITY CHECK

### Slide 19: The employee does the work.

01
No AI
02
Chat
2.5
Skills
03
Systems
04
Agent
The employee does the work.
Every case starts from zero.
Every Monday
Employee
List cleaned
Google Ads
THE AI REALITY CHECK

### Slide 20: AI helps. The employee moves the data.

01
No AI
02
Chat
2.5
Skills
03
Systems
04
Agent
AI helps. The employee moves the data.
AI is available. The employee still carries the work.
Every Monday
Employee
List cleaned
Google Ads
AI chat
THE AI REALITY CHECK

### Slide 21: The team writes its know-how down once.

01
No AI
02
Chat
2.5
Skills
03
Systems
04
Agent
The team writes its know-how down once.
A skill tells the AI how the team checks keywords. The employee still moves the data.
Every Monday
Employee
List cleaned
Google Ads
AI chat
Skill
THE AI REALITY CHECK

### Slide 22: AI can read the real data.

01
No AI
02
Chat
2.5
Skills
03
Systems
04
Agent
AI can read the real data.
Data access removes copying. It does not create a workflow.
Every Monday
Employee
List cleaned
Google Ads
Connected AI
Skill
THE AI REALITY CHECK

### Slide 23: The work now runs as a system.

01
No AI
02
Chat
2.5
Skills
03
Systems
04
Agent
The work now runs as a system.
The trigger starts the work. People take the exceptions.
Every Monday
Agent
List cleaned
Google Ads
Exceptions
THE AI REALITY CHECK

### Slide 24: Four parts make an agent.

01
No AI
02
Chat
2.5
Skills
03
Systems
04
Agent
Four parts make an agent.
Each part alone is a tool. Together they do the whole task.
TIME OR EVENT
Trigger
Every Monday, or a new lead
+
LLM
Model
Reads, reasons and writes
+
KNOW-HOW
Skill
How the team checks keywords
+
ACCESS
Connection
Reads and changes Google Ads
=
RUNS ON ITS OWN
Agent
Does the whole task
THE AI REALITY CHECK

### Slide 25: Small systems compound.

01
No AI
02
Chat
2.5
Skills
03
Systems
04
Agent
Small systems compound.
The workload stays the same. The part you do by hand shrinks.
DONE BY HAND
RUNS AS SYSTEMS
HANDS-ON
TODAY
40 h
MONTH 1
35 h
MONTH 3
25 h
MONTH 6
10 h
Two hours once. Five hours back, every week. And the freed time builds the next system.
THE AI REALITY CHECK

### Slide 26: Five things we had no time for.

Five things we had no time for.
01
ChatGPT or Claude: both work.
Tools like Copilot or Langdock fall behind.
02
Use Opus or Sol at minimum.
Nothing below that. Fable and Astra are great, but expensive.
03
Computer use automates most of the rest.
ChatGPT is especially strong here.
04
AI makes mistakes.
People make more.
05
Test again, and often.
"AI can't do that" is often wrong for the newest models.
THE AI REALITY CHECK

### Slide 27: of office work AI can already do today, with enough time and money.

BONUS · MY BET
90%+
of office work AI can already do today, with enough time and money.
THE AI REALITY CHECK

### Slide 28: and AI can do tasks twice as long.

WHAT HAPPENS NEXT
4 months
and AI can do tasks twice as long.
TODAY
IN 4 MONTHS
13×
cheaper per year for the same performance.
TODAY
IN 1 YEAR
THE AI REALITY CHECK

### Slide 29: Get the slides

SLIDES
Get the slides
specialtytokens.com/go/ceo-dinner
LINKEDIN
Connect with me
linkedin.com/in/nredl
