grep --tag * questions.txt

Questions I help answer

Questions companies are asking themselves right now, each with a short note on how I would approach it. If yours is not here, ask anyway. ~/contact

innovation

How do our innovations make use of AI, in the product and in the way we innovate?

We look at your innovation process as a whole, from where ideas come from to how they get tested, and find the steps where AI changes what is possible, not just what is faster.
How do we move our employees from the tail end of AI to the front end?

With practice on their own tasks, good examples and explicit permission to try. Tool training alone does not do it.
Are we using AI, or are we getting value from it? And how would we know?

We define what value means for one concrete process, measure where you are today and compare. Most companies skip the baseline and then cannot tell.
How do we run AI experiments without derailing the core business?

Small, time-boxed pilots with a clear stop criterion, run next to the core process rather than inside it until they have proven themselves.

How can we use AI for our specific business case?

I start with your processes and your data, not with the tools. Then we pick the one or two places where AI would make a measurable difference.
What is the difference between AI and machine learning, and which one do we need?

Machine learning is the part of AI that learns from data, and today's chatbots are one kind of it. Often a small, boring classifier beats a large language model. I will tell you when that is the case.
Where are we paying for hype, and where do we get measurable value?

An honest review of your current AI spending and projects: what is used, what is measured, and what could be stopped tomorrow.
What is a sensible first AI pilot that can fail cheaply?

One with real users, a narrow task, data you already have and a result you can check. Drafting replies to recurring requests is a classic.

Which data is sitting unused today and could give us a real advantage tomorrow?

Old documents, tickets, emails, reports and patents are often the most valuable data a company owns. Today's models can read them at scale. We find out which of yours are worth it.
How do we turn thousands of documents into structured, searchable knowledge?

This is my research: extracting structured data from large amounts of unstructured text, with a codebook, a checked sample and known error rates.
Has someone already solved our problem, in another industry, in a patent or in nature?

Often, yes. My PhD maps solution concepts from patents and links them to solutions found in biology. The same approach works for your problem.

How do we use AI without producing AI slop?

By keeping experts in charge of the result, giving the model your context instead of generic prompts, and having the discipline to throw bad output away.
How do we know the AI is right, before we rely on it?

With a test set your experts agree on, and numbers instead of impressions. In my research nothing is used until it holds up against such a set.
Which work should AI draft, and where must a human still sign off?

A useful rule: AI drafts, people decide. We go through your processes and mark where a wrong answer is cheap and where it is expensive.
How do we keep AI-generated code maintainable?

The same way as any code: reviews, tests, a clear architecture and developers who understand what they merge. AI makes these rules more important, not less.

How do we work with AI without compromising our customers' personal data?

Data minimisation, pseudonymisation, the right contracts and sometimes models that run on your own hardware. I have built a production platform under the GDPR and government privacy requirements.
Can we run AI models on our own servers instead of sending data abroad?

For many tasks, yes. Open-weight language models and local speech recognition are good enough today. My own transcription runs on my own server.
What do the GDPR, the EU AI Act and the European Accessibility Act mean for us in practice?

Less panic, more checklists. We map which of your systems fall under which rule and what that means for the engineering work. I am not a lawyer, but I know where the technical work is.
How do we avoid lock-in to a single AI vendor?

A thin layer between your systems and the model, your own evaluation data and contracts that let you leave. Then switching vendors becomes a configuration change.

How do we change our organisational structures for a world where AI plays a dominant role?

Decision rights, roles and team boundaries matter more than tools. We look at who decides what today, and what changes once routine work becomes cheap.
How do we bring our people along, so AI gets used well instead of feared or misused?

Hands-on workshops on their real tasks, clear rules for what is allowed and visible support from leadership.
How do junior developers still learn the craft when AI writes much of the code?

Juniors need to understand code, not just produce it. I have onboarded many people with little or no experience. The principles still hold, the exercises change.
How do we build a healthy development team from scratch?

Hire for curiosity and communication, set a few clear engineering standards early and give people ownership. I have formed several teams and helped startups do the same.

Which architecture should our startup start with, so we don't rewrite everything in two years?

Usually a boring, well-structured monolith with clean boundaries. It is easier to split later than to glue microservices back together.
How does AI fit into our SAP landscape? Can intercompany processes run on their own?

I have led SAP go-lives in SD, MM and WM and built autonomous intercompany processes. AI can take over reading, matching and drafting, while SAP stays the system of record.
What does a lean, modern data platform look like for a company our size?

Usually smaller than vendors suggest: a few well-chosen open-source building blocks, clear ownership and pipelines people can actually understand.

26 questions