Kris Peeters

The through-line

Kris Peeters keeps returning to one question: as agentic AI changes how organisations use data, what actually creates value, and what is just noise. Early posts frame this as an architecture problem, where should logic and semantic meaning live once agents start querying data directly . Later posts sharpen into a harder line against automation for its own sake: more tokens and more code do not equal more value, and much of the "automated by AI" tooling on the market solves nobody's real problem . The shift is from describing the technical opportunity (headless BI, lowered barriers to data work) toward prescribing discipline: pick models on cost and performance rather than loyalty , rethink the delivery process from scratch instead of bolting agents onto old roles , and keep paying for human expertise and accountability .

On the semantic layer

Peeters traces how agents change the BI stack. He notes that logic, metrics and definitions have historically lived "locked inside a BI tool," which "wasn't really problematic until agents became a thing" . Because agents also need to talk to data, he argues they "need a high quality semantic layer," which gave rise to the idea of Headless BI: strip logic out of BI tools, turn them into "a thin visualisation-only layer," and move logic to a semantic layer close to the data . But he is candid that this is aspirational rather than realistic in most organisations, since it forces analysts to learn dbt or Spark instead of their BI tool, and requires migrating "1000s of reports with built-in logic" to the platform .

On model commoditization

Watching foundational models proliferate, Peeters treats price and performance convergence as a structural fact to plan around, not a curiosity. He points out that "foundational models are becoming a commodity," that open-weight models "like Kimi and GLM are competitive on performance," and that there is "a HUGE variety in price" . Comparing GLM 5.2 to Claude Sonnet 5 or Opus 4.8, he observes it is "~0.95x the performance" but "~10x cheaper" . His conclusion is operational: "build your organisation to take advantage of all the models out there" and "don't let yourself lock into model providers that are too expensive" .

On AI slop and human accountability

Peeters is skeptical of startups packaging "<a-very-specific-human-experience>, Automated by AI," listing lead-gen, recruiting, strategic advisory and legal advice as examples, and asking bluntly, "who is buying this?" . His objection is that AI-generated output is now trivial to produce and to spam: "everyone can generate AI-slop lists and texts and slides now... and put it in a for loop to turn it into spam" . He draws a line between using AI internally and what he wants to buy: "I very much still want to pay for a human to understand my business, my needs and is accountable for delivering results," and he says he values "companies who have deep expertise" . The same instinct shows up socially: he calls a human-only meetup "a guaranteed AI-slop-free evening" .

On rethinking the SDLC

On engineering delivery, Peeters repeats an old line with new force: "'Code is a liability, experience is an asset.' We said it already 10 years ago. In the days of AI, it's even more true" . He argues that giving engineers token access "usually only results in more code. Not in more value," especially in large enterprises where "'typing in the code' was never the bottleneck in the first place" . His prescription is structural, not incremental: "throw out what you know. Start from scratch," questioning which handoff steps between business analyst, SCRUM master, engineer, QA, platform engineer, UX designer and architect an agent could absorb, since "each human is a hand-over, and loss" . This same experimental posture drives Dataminded's innovation days, where prototypes are benchmarked with real numbers, for example a knowledge-graph agent moving "from 37 to 83 out of 100" in accuracy, and to 100 with a human in the loop , and where his broader takeaway is that "we're still at the very early days with agents," comparable to the iPhone 1 or Netscape moment .

From the stage

On the podcast, Peeters frames the starting point for data work differently than in his written posts: teams should begin with a clear business question rather than a technology choice, because without one they "waste time and money exploring endless possibilities and creating unnecessary complexity" ¹². He also stresses that data's value is often counterintuitive, some use cases that look impossible become feasible once you understand data quality, while others that look obvious turn out impractical ¹². In conversation with Jonny Daenen, he explores a value chain that runs from data to applications and products, to insights, to decisions, and finally to actions that create real value ¹³. And discussing Dataminded's founding, he explains he built the company as engineering-first specifically because he had seen engineers treated as "second-class citizens" executing consultants' decisions, and describes his own move from hands-on engineering into delegating operational work like performance reviews and billing while staying close to clients and prospects ¹⁴.

Takeaways

• Before investing in a full semantic-layer migration, weigh the realism: moving logic out of BI tools into dbt or Spark and migrating thousands of existing reports is often harder than the "headless BI" pitch suggests .
• Architect for model portability rather than provider loyalty, since open-weight alternatives can deliver ~95% of a leading model's performance at roughly a tenth of the cost .
• Be wary of tools that just wrap "X, automated by AI" for legal, recruiting or advisory work; buyers still want a human who understands the business and is accountable for outcomes .
• When adopting agentic AI in delivery, don't just hand engineers more token access, redesign the workflow and question every human handoff in the SCRUM/QA/architect chain .
• Run structured, benchmarked experiments (like a four-tier accuracy benchmark or model-routing cost tests) before trusting agent-based systems in production .
• Start data initiatives from a specific business question, not from a technology or platform decision, to avoid open-ended exploration that burns time and money ¹².