Tobias Zwingmann

Author of AI-Powered BI (O'REILLY) | AI Coach & Mentor | I help organizations leverage AI and use data more intelligently to improve business performance. Sharing AI use cases & learnings.

My AI rule is simple. "If it doesn't cut costs, save time, increase throughput, remove risk or unlock revenue, it gets cut".In 5-continent wide enterprises and in the mid-sized companies finding $10k+ monthly savings in unexpected places I've helped, Profitable AI is about the same thing: Finding the overlap of what AI can do and what your bottom line needs it to do.Here's how I can help:→ Read the book: "The Profitable AI Advantage" – Frameworks for finding &amp; executing AI opportunities that pay off→ Use PAIGE: My AI strategy tool that helps you identify and evaluate opportunities for your specific business  → Join the Club: Direct access to me as you implement AI in your business (capped at 100 members)→ Work with me: When you're ready, my firm RAPYD.AI delivers AI strategy consulting and custom implementations that pay for themselves.I share what I'm learning every week in my Profitable AI newsletter – read by 10,000+ professionals who are tired of flashy demos and want real results instead.Follow along if you like. Or better: tell me what you're trying to figure out.

Managing Partner

We help you build amazing AI products faster.

Microsoft just launched Copilot Cowork – built with Anthropic. These are AI agents that run multi-step tasks across your entire M365 stack. What you see here is a shiny demo obviously. But it shows one thing crystal clear:

The technology is converging. Fast!

Every major player – OpenAI, Anthropic, Microsoft, Google – is racing to deliver the same capabilities: agents, automation, and multi-step workflows across your office work.

Which means AI technology is becoming commodity (since all of them pretty much rely on very similar models).

The differentiator for your business isn't whether you pick ChatGPT or Copilot or Claude. 

It's three things that no vendor can sell you:

A) Your data
B) Your processes 
C) Your organizational readiness to execute

That's it.

Otherwise, every "multi-step AI agent" will just live as a prototype and the only thing you'll have are expensive subscriptions.

The companies I work with that actually ship AI to production all started the same way: one specific problem, clear owner, clear roadmap, and an embarrassingly simple first step. Use whatever tool works.

I'm not saying this will be easy.

I'm saying there isn't an easier way.

undefined undefined

Author: The Profitable AI Advantage. Created profits with AI from enterprise giants to 20-person teams.

Some guy wanted to control his vacuum robot with a PS5 controller.

- Asks Claude Code (of course)
- Claude fetches an auth token for the Romo
- Turns out, the token works for every other Romo too
- Guy now has control over 7,000 vacuums (cams included)
- Could go "Breaking Bad", but calls the company instead
- Company ships a patch
- End of story

You could write this off as another geeky tech column.

But there's a bigger lesson here. 

A lot of the legacy systems I've seen in my career are sitting there with the same kind of exposure. (Looking at you, admin tool with hardcoded credentials and vendor API with last update in 2012!)

The only reason nobody's found the holes yet is because nobody took the effort to look.

But when the effort goes down to zero, we shouldn't be surprised to see more "incidents" like this.

Perhaps, you should take that effort first.

Pitching the flying car feels good for the ego.
 
You get to talk about fully autonomous agents and all. The bright AI future. 
 
But then reality sets in.

The first prototype kind of works. The production version doesn’t. 

You iterate. 

Again. 
 
Budget drains.
 
Timeline stretches.
 
Eventually you deliver the boring purple sedan just to show something. 
 
The sedan is actually a good car.
 
It does the work.
 
But it stops making sense if the road to get it cost more than the sedan itself.
 
When stakeholders get the sedan after paying for a flying car they ask questions.
 
"That's what I paid for?"
 
"Why did it take so long?"
 
Credibility gone.
 
No funding for the next project.
 
That’s why my preferred approach is the last row done in reverse.
 
I start with the sedan.
 
And build the flying car as we go.

We all know the famous Pareto Principle that says 80% of results come from 20% of effort. AI projects operate on Pareto Principle on steroids. Let me explain:

The first 80% of value comes fast. You build an Assistant that answers customer questions. It works, everybody gets excited. "This will be done next month!"

The what I call "AI Complexity Tax" kicks in after that.

A little improvement here.
Another iteration there.
“Make it more autonomous!”
Try to get rid of the “human in the loop” (because that’s the whole point, right?!)

Suddenly, you’ve spent 80% of your total budget on a minimal gain from 80% to 87%.

Complexity Tax.

The highest Complexity Tax is paid when you move a use case from one solution type to another.

From Assistant to Autopilot.
From Autopilot to Agent.
Or worst: From 0 to Agent.

And every transition has a dip. A period where you're spending more and getting less than what you had before.

Plus, every step up in autonomy or integration adds complexity that compounds quickly – I’m inclined to say "exponentially".

(See the visual)

That’s why the timing of when to say “stop” matters as much as the timing when you say “start".

The fix is simple: ship the 80% solution. The Assistant, the autopilot. Let humans catch exceptions. You're already profitable.

The autonomous multi-agent system can wait until the Autopilot has paid for itself three times over.

McKinsey's internal AI chatbot got hacked this week. Not by a nation state or a team of elite hackers.

It was a single AI agent that took two hours to exploit a security issue that's been known since the 1990s.

The chatbot called Lilli is used by 40,000+ McK consultants and sits on top of decades of proprietary research, client strategies, and M&A data.

Using this security bug from the 90s, the researchers were able to access 40+ million chats and 700,000+ files.

Plus, write access to the system prompts that control the AI behavior.

Security nightmare? For sure!

In the wrong hands, this could have turned into a disaster.

But luckily, the friendly researchers informed McK before anything too bad happened and they patched it within a day.

(Kudos to them.)

To be clear:

This is McKinsey making $15 billion a year and having world-class security teams.

Now think about your AI deployment.

Are you able to deploy AI agents securely? 

Who owns the security architecture for your AI layer specifically? 

If the answer is "we'll figure it out" or "Bob from IT", you probably have the same gap McKinsey had.

I'm not saying here "don't build AI".

But be aware what game you play before jumping head first into AI agents.

If the basics aren't locked down, adding more autonomy just multiplies the risk surface.

Secure what you've shipped before you ship more.

This is something a client recently told me that made me think: "If I don't know what AI can do, I can’t say which processes it could improve."

It’s true.

If you don’t understand something, how can you redesign your company’s workflows around it?

But if the conclusion is something like:

“I have to learn the fundamentals.”
“Learn prompt engineering.”
…then this is wrong.

Because the big question is:

WHEN?
(and once finished: “how do I know it’s all still relevant?”)

Thing is…

If you wait until you understand everything AI can do, you'll probably never start.

And honestly, nobody has complete knowledge.
Not even the cool folks at OpenAI.
(It’s actually shocking how little everyone knows but that’s a topic for another day)

So the fix isn't to "learn more about AI."
The fix is to narrow down the question.

You don't need to know everything AI can do.

You need to know which mode of AI fits the problem in front of you right now.

I introduce 5 AI Modes for Business when working with clients:

→ 1. Prediction Mode 
(AI that anticipates future outcomes based on historical data, e.g. forecasting)

→ 2. Perception Mode 
(AI that sees, hears and understands the unstructured world, e.g. analyzing images)

→ 3. Creation Mode 
(AI that produces original content – from funny stories to complex code)

→ 4. Thinking Mode 
(AI that "makes sense" of things and reasons through complex problems)

→ 5. Agentic Mode 
(AI that “does” things autonomously, like executing tasks or coordinating workflows without constant human supervision.)

When clients say:
"I don't know what AI can do."

I say:
"Which (combination) of these five modes would make your biggest pain points or bottlenecks disappear?"

Suddenly we have an answer.

And if not, we probably don’t have an AI problem at all.

Look at this! AI adoption is solved! McK reports that almost 90% of organizations say they're now using AI. But wait.. what's that number there on the right??

Only 20% say they use AI in 5 or more functions.

(I wonder how many have answered that based on the number of Copilot licenses shipped)

What this chart means is that a lot of companies get one AI win.
But then, many get stuck.

I see this frequently:

1) Someone on the team starts using ChatGPT.
2) It goes well. 
3) Leadership hears about it
4) Says "great, let's scale this AI thing!"

Except for there's nothing to scale. 

Because personal productivity doesn't have a clear scaling path. 
One person saving 2 hours a week is useful.
But it's not a system. 
It's a habit.

You don't get from "1 function" to "5 functions" by rolling out Copilot licenses.

You have to mature your AI cases from "this helps people" to "this scales with people."

From assistant to autopilot.
From Productivity AI to Engineered AI.

Which is a completely different project with a completely different business case.

And if you try to get there by telling more people to "just use AI more", you end up with 88% adoption and nothing to show for it in the P&L.

We’re used to govern software by code. But the code of an AI model in production is maybe 50 lines. This is why we need to look beyond and consider the data flowing in and out of the AI system, a.k.a. Data Observability.

What to watch:
→ Input quality – garbage in is still the main failure reason. 
→ Output consistency – is the model behaving the same way? 
→ Edge cases – what's happening at out of sample distribution?
→ Cost per action – are the economics still holding up at scale?

We discussed Data Observability in more detail during out last D2A2 webinar "Decoding the 2026 Report with Industry Leaders" where this video excerpt is taken from.

Link to the recording (and the 2026 Predictions report) in the comments.

"Should you use AI?" at this point isn't even a real question anymore.
Modern LLMs got 6x smarter over the last 2 years:

They can now run for hours autonomously.
There's no doubt there's progress.

But "Should you use AI?" does not translate to "You should use AI."

AI spending exceeded $1.5 trillion last year, and most businesses still can't show ROI.

Clients ask me:

"If the tech got 6x times better, then why aren't we seeing 6x better results?"

Because there's gap between what AI can do and what your business needs.

And there are two tracks to close that gap:

1) Productivity AI
ChatGPT, Copilot & co.
Tools that make your day easier.
Low cost, no business case needed.
Just try it, but don't expect big P&L impact

2) Engineered AI
Automated workflows that run with human checkpoints, not human babysitters.
Higher cost, higher stakes.
They require one question before you start:
"What does this problem cost us today?"

If you can answer that, you have a business case.
If you can't, you're not ready for Engineered AI yet.

A lot of that $1.5T went to fuel Productivity AI.
And that's fine – but it's why the ROI numbers look disappointing.

The companies showing real returns are the ones that figured out which problems are worth engineering a solution for. 
Not where to buy the most licenses.

The Limits of Productivity AI

It won’t transform your business.

Make sure it doesn’t break anything – and you’ve already done a great job.

New article: https://lnkd.in/dCjGD6Pe

"Can you give me a big cat?" – "Sure!" said the pet shop lady and returned with a tiger. Was she wrong? No, because the tiger is a big cat indeed. You, on the other hand, weren't precise. This is how most companies buy AI. 

They walk in and say "we want AI for customer service" or "we need AI for our documents." 

The vendor nods, delivers something that technically fits the brief, and three months later everyone's wondering why there's just cost but otherwise nothing to show for it.

The problem was never the vendor.
It was the ask.

"We want AI that reduces ticket volume by 40% within 6 months at a cost below €X/month" gets you a what you need.

"We want an AI chatbot" gets you what the vendor wants to sell.

When dealing with horizontal, general-purpose technology, then specificity isn't a nice-to-have.

It's the entire skill.

The more precisely you can describe what you need, the less you depend on anyone else to figure it out for you.

If you can't, you're in the pet shop asking for a big cat.

Give a team AI in Excel without a process.

Two things happen:

Group one ignores it.

They keep doing things manually.

They don't gain efficiency.

But they also do not break anything.

Harmless.

Group two embraces it.

The experimenters.

(These are usually the go-getters.)

And as harsh as it sounds, they might quickly turn into a major liability.

Because there's a massive difference between experimenting with marketing copy and experimenting with financial data.

If ChatGPT writes a bad post nobody cares.

If ChatGPT misreads Q3 revenue data catastrophic business decisions happen.

We used to have a safety net for this.

It was called the Excel Ninja.

People who knew how to do the "hard" stuff (Pivot tables, formulas, macros)

It required a high level of business understanding and technical knowledge.

That difficulty was a natural barrier to entry.

AI in Excel just crushed that barrier.

Anyone can generate any output now.

Someone with zero business context and zero technical understanding can ask a chatbot to analyze a massive spreadsheet.

Corporate data is usually what I call Ugly Data.

It is poorly formatted, full of merged cells and missing headers.

The AI tries to read it, gets confused, and silently skips the rows it does not understand.

It processes only a fraction of the actual numbers.

And then it confidently spits out a polished chart.

The user is thrilled.

They do not know what happened technically or have the business acumen to realize the numbers look slightly off.

They just copy the chart and paste it into a slide deck.

That deck gets emailed around.

It makes its way into a leadership meeting.

Executives end up making budget allocations based on ChatGPT missing everything down row 876.

The person who generated the number did not build it.

They just asked for it.

Productivity AI gives speed.

But speed without governance in data analytics is just a faster way to crash.

The solution is not banning AI and the efficiency gains are too good to ignore.

The deployment just has to change.

I stopped assuming that because a tool is easy to use it is safe to use.

I recommend clients to roll out these tools in batches.

We ship and learn what data is clean enough for an LLM to touch.

We build a process for validating the output before it leaves a desk.

When these constraints are in place AI really becomes an asset.

Get the speed of the experimenters without the blind spots.

It's the profitable way.

I normally don't post on new model updates but there's something going on. AI is coming for everyday office work.

GPT-5.4 launched yesterday with a ChatGPT-for-Excel add-in.
Anthropic did the same with Claude Cowork a few weeks back.
Microsoft's Copilot Agent in Excel suddenly became useful.

Why should you care?

Reason 1: Personal productivity.
Try these tools. If the last time was the GPT-4o era, you'll be surprised. This is Productivity AI – it makes your people faster.

Reason 2 (more important): Workflow redesign.
This is Engineered AI – it changes how work gets done.

I've lost count how many organizations are looking for "better" ways to get rid of their messy Excel spreadsheets. "Better" usually means buying a specialized tool. But that specialized tool (read: $100K/year Enterprise SaaS) is now competing with a $30/month ChatGPT subscription that says "Screw it. We're just working on the messy Excel. AI will figure it out."

Won't work in every case. But turns out, messy and accessible beats clean and locked away in 200 siloed tools.

Back in 2023, I wrote my first blog about the "AI make-vs-buy decision". Tomorrow it gets an update – because "buy" got a new dynamic.

You'll end up buying components for a shared AI platform. Question is which and how many. You'll end up building the things no AI can build for you: clean data, clear processes, and someone who owns the workflow end-to-end.

I see 3 different buying modes with different rules for each.
More in my newsletter tomorrow.

(Link below my name. Email goes out Saturday 5PM CET.)

Running another round of my popular ChatGPT for Data Analytics Bootcamp with O'Reilly next week. With the recent ChatGPT for Excel release, this couldn't come at a better time. Join me to learn:

- LLMs and ChatGPT fundamentals for data analytics
- Use prompt engineering to generate actionable insights
- Apply ChatGPT to everyday analytics problem-solving
- Improve data storytelling and communication
- Build custom GPTs for recurring analysis tasks
- Use ChatGPT with Excel and Google Sheets
- Automate workflows with simple no-code tools
- Integrate ChatGPT via API for scalable solutions
- Manage AI security, compliance, and governance
- Identify future AI opportunities in data analytics

Full agenda and registration in the comments.

Having shipped dozens of customer support chatbots to production with the best hitting 82% automated resolution rate, I thought it was time time for a quick reality check.

Because AI chatbots for customer support are among the lowest hanging fruits you could grab with Generative AI. 

But they can also turn quickly into a money pit, just moving resources from your customer support team to your IT department without contributing anything significant to the bottom line. 

Worst case I've seen: worse customer experience AND higher cost.

To avoid that situation, it's important to understand the decision criteria of when an AI customer support chatbot is really a low-hanging fruit (and when it's just a foul pea you wouldn't touch.)

If this sounds interesting, I published a new in-depth blog over the weekend. You can read it here: https://lnkd.in/d-Sv43mf

(Don't forget to subscribe while you're there.)

Is an AI Customer Support Chatbot worth it for your business?

Just wrapped up a new article that breaks down real costs, levels, and decision points.

If you want to get it straight to your inbox, sign up here before 5pm CET tomorrow (Saturday).

→ https://lnkd.in/eFzzQrMJ

taplio

Tobias Zwingmann's Linkedin Analytics