Too many AI conversations start with platforms, data foundations and governance frameworks. I prefer to start with the work: what can we help people accomplish, where does AI add value, and what could go wrong?
These five principles bring together how I approach those questions.
1. AI for Data
Use AI to make existing data useful. It can help interpret unfamiliar structures, connect sources and expose gaps, so useful work need not wait for a comprehensive data transformation. Start with the question and improve the data where the answer requires it.
Prioritize expanding what people can understand, create and accomplish. Distributing AI capability puts experimentation in the hands of people who know the work, allowing useful applications and shared learning to emerge across the organization. That offers greater potential than a centrally selected portfolio of automation projects: the organization grows its capacity to discover and solve problems.
Concentrate AI where interpretation, exploration and judgment are needed. Its value comes from helping resolve uncertainty; applying it to already settled rules adds cost and variability without a corresponding benefit. Make the structure explicit as understanding improves, and keep reasoning focused on what remains uncertain.
Use AI to build dependable automation, then capture stable logic in software. Making a model repeatedly interpret the same rules creates variability that teams try to contain with ever more evaluations and guardrails. Explicit, testable code preserves what has already been learned and makes repeated execution easier to verify, operate and reuse.
Match oversight to what an application can affect and how far mistakes can spread. Model sophistication alone says little about the consequences of its use. Bounded, reversible experimentation should be easy; broader authority, sensitive data and consequential actions warrant stronger assurance. This allows capability to spread while concentrating scrutiny where it matters.
A powerful model helping someone draft a private note can have a small impact. A simple agent changing shared records or sending customer messages can have a large one. The model’s capabilities alone do not tell me how much governance its use requires.
I start with blast radius: the potential reach and consequences of an error, misuse, disclosure or unauthorized action. My proposed approach is to keep bounded personal experimentation simple and increase scrutiny as authority, exposure and consequences grow.
This is a governance proposal. It assumes existing platform, data, security and business controls remain in force. “No additional review” means no separate review for that particular AI solution within an already authorized scope.
Check the boundaries first
Before considering who will use a solution, I ask whether it stays within approved data access, authority, impact and safeguards. An expansion needs review. Unknown facts need clarification.
That first check matters because personal use is not necessarily low-impact use. An agent operated by one person can still send bulk messages, alter shared records or commit funds. Limiting access to its creator does not authorize those consequences.
Separate personal use, direct sharing and broad deployment
Within established boundaries, I would distinguish between solutions configured inside an approved platform’s existing capabilities and safeguards, and custom solutions that introduce capabilities or safeguards someone must separately design, operate and assure.
The distinction is responsibility. Writing code is not, by itself, the deciding factor. A separate runtime or an integration outside existing assurance can introduce obligations that configuration within an approved platform does not.
Who will use it?
Within approved platform capabilities
Custom capabilities or safeguards
Only the creator
No additional review
No additional review
Intended recipients, shared directly
No additional review if the conditions below hold
Review required
Broad or automatic distribution
Review required
Review required
Every row assumes the initial boundary check has passed. Where review is required, its depth follows the actual blast radius.
Direct sharing needs clear responsibilities
I think of direct sharing as borrowing a colleague’s tool. Recipients deliberately adopt it and remain responsible for how they use it. For that to work:
Access is limited to intended recipients who understand the purpose, limitations, permissions and actions.
Each recipient uses their own authorized access. Sharing a tool does not share its creator’s credentials or permissions.
The creator maintains the configuration and communicates limitations; recipients check outputs and actions in their own use.
Sharing does not imply organizational endorsement or bypass publication controls.
If the facts are unclear, clarify them. If these conditions are not met, treat the distribution as broad. The relevant question is who can actually access and adopt the tool, rather than which sharing button was clicked.
Assess what could go wrong
I use six questions to assess blast radius:
Reach: who could be affected or rely on the result?
Authority: can it read, draft, send, change records, spend or make consequential decisions? What gates execution?
Data: whose data is involved, how sensitive is it, and where can it go?
Consequences: what harm could follow from an error or misuse?
Containment: who can detect problems, stop execution, reverse actions and recover?
Propagation: how often can it run, and can its effects spread to other systems or agents?
Bounded, reversible use with limited exposure may need only a lightweight check of ownership, permissions and containment. Shared operational effects, sensitive data or repeated writes call for workflow assurance: failure testing, monitoring, recovery and maintenance ownership. Severe consequences or difficult reversal warrant formal review by the relevant business and specialist owners.
The most consequential exposure should guide the depth of review. A severe risk should not disappear inside an average score. These are guides for proportionate assurance, not automatic approval thresholds. Review may require tighter boundaries, stronger controls or rejection.
A human approval button is not an exemption
Human oversight helps when the person can understand, stop or correct the consequential action. An approval button that encourages routine acceptance provides little assurance.
I therefore assess the effectiveness of oversight alongside the system’s authority and impact. The presence of a human does not eliminate the need to understand what the system can do. This builds on my argument in Responsible AI: Enhance Human Judgment, Don’t Replace It.
Approve a scope that can be reused
Record the owner, capabilities, audience, data, authority, scale and safeguards covered by approval. Then allow reuse within those boundaries.
Review again when the scope materially expands or essential safeguards change. Routine prompts and executions should not restart the approval process. A model change matters when it changes the risk or falls outside existing assurance.
My aim is governance that allows capability and learning to spread while concentrating attention on consequential exposure. That supports the distribution of AI capability without confusing access to a tool with authority to affect others.
A language model can interpret a request, calculate an answer, format a report and decide what happens next. That makes it tempting to give the model the whole workflow. But capability alone is a poor guide to architecture.
My starting point is simpler: reserve AI for ambiguity, and use systems for structure. Apply reasoning where it resolves uncertainty or improves understanding. Capture the parts that are already understood in code, rules, schemas and templates.
The cost problem is a design problem
Token budgets, usage caps and model tiers make inference costs visible. Yet repeated spending often points to something more fundamental: a workflow is asking the model to solve the same problem again.
Regenerating a stable report layout, applying established calculations or rediscovering a known mapping consumes reasoning capacity without creating new understanding. Increasing the budget may keep that workflow running, but it leaves the design problem intact.
I want the value of AI to accumulate. A resolved question should become a reusable decision, specification or component, rather than another prompt that must be interpreted on every execution.
Ask whether the step needs AI
Before choosing a model, I ask what remains uncertain in the task.
AI is well placed when the work involves interpreting incomplete information, exploring alternatives, synthesizing evidence, explaining findings or making contextual judgments. Those activities benefit from flexibility because the answer cannot be fully specified in advance.
Established calculations, repeated formatting, fixed routing and stable classification rules have a different character. If the expected behavior can be stated explicitly and tested, it should usually be implemented directly.
A high-frequency step with little ambiguity is a particularly strong candidate for removal from the model path. The question is whether each invocation adds useful reasoning or merely repeats settled work.
Structure the work before selecting the model
I use five steps to make that distinction practical:
Define the structure. Clarify the inputs, outputs, constraints and boundaries.
Capture repeatable logic. Put established rules into code, schemas, workflows or templates.
Locate the ambiguity. Identify the interpretation, synthesis or judgment still required.
Apply AI there. Give the model a bounded role with evidence and a clear way to assess the result.
Choose the model last. Match its capability and cost to the reasoning that remains.
What this looks like in practice
Software engineering: use stronger reasoning to explore architecture, identify risks and clarify requirements. Once the design has been decomposed into clear implementation tasks, smaller coding models may be sufficient. A continuing need for frontier reasoning during routine implementation can indicate unresolved ambiguity upstream. I explore this further in Efficient Use of AI for Software Engineering.
Analytical reporting: compute facts, capture evidence, apply explicit rules and render the report through software. Use AI to explain findings and interpret their significance. The report should not depend on a model reinventing its calculations and layout each time.
Process automation: encode predictable steps in workflows and services. Use AI when the process encounters unfamiliar language, an exception or a decision that genuinely requires context. An agent can call dependable capabilities without having to reason through every operational detail.
Make placement a deliberate decision
This principle also changes how I assess AI adoption. More model calls do not necessarily mean more value. I look for useful reasoning, reusable outputs and less uncertainty in the work that follows.
Organizations need a shared understanding of where inference is justified, where explicit logic is preferable and how model tiers fit different tasks. Otherwise, scaling adoption can simply scale avoidable cost and variability.
The goal is to place AI where its flexibility matters—and preserve what it helps discover in forms that can be used reliably again.
We often talk about training large language models (LLMs) as if it’s one thing — but it really isn’t.
There are several distinct types of training, each with its own purpose, cost, and level of control.
Understanding the difference helps clarify what’s realistic to do in practice, and what should be left to the model labs with thousands of GPUs and power budgets larger than small towns.
Here’s a simple breakdown.
Base Training — Learning Language from Scratch
This is where it all begins.
The model learns to predict the next word in a sentence across trillions of examples. It’s how it develops a general understanding of language, reasoning, facts, and relationships between concepts.
Purpose: Build a general-purpose foundation.
Data: Huge, diverse datasets (Common Crawl, Wikipedia, code, books).
Cost: Astronomical — only done by major labs.
Analogy: Teaching a child how to speak and read.
Once complete, this produces what’s called a base model — capable, but not polite, safe, or even particularly helpful.
Post-Training — Teaching Behavior and Alignment
After base training, the model needs to learn how to behave.
This phase adjusts it to follow instructions, respond helpfully, and align with human preferences and safety policies.
It typically involves:
Supervised Fine-Tuning (SFT): The model learns from curated examples of correct input/output pairs.
Reinforcement Learning from Human Feedback (RLHF): Humans rank several model responses, and the model learns to prefer the higher-ranked ones.
Reinforcement Learning from AI Feedback (RLAIF): The same, but with AI systems acting as evaluators.
Purpose: Make the model cooperative and safe.
Analogy: Teaching manners, ethics, and social intelligence after language is learned.
All commercial models — GPT-4, Claude 3, Gemini, Llama 3 — go through this step before they ever reach users.
Fine-Tuning — Specializing for a Domain
Fine-tuning takes a general model and teaches it domain-specific knowledge: medicine, law, brewing, internal documentation — whatever your niche may be.
There are a few variants:
Full fine-tuning: retraining all model weights (rare and expensive).
Parameter-efficient fine-tuning (LoRA, QLoRA, PEFT): training small adapter layers on top of the frozen base model — vastly cheaper and reversible.
Purpose: Adapt to a domain or style.
Analogy: Sending a fluent speaker to medical school.
In practice, fine-tuning makes sense only if you have high-quality, well-structured data and a clear purpose — for example, improving factual recall in a specific knowledge domain or matching a company’s tone of voice.
Reinforcement Fine-Tuning — Teaching Preferences and Optimization
A newer development combines reinforcement learning with fine-tuning to optimize specific, measurable outcomes — such as factuality, brevity, or computational efficiency.
OpenAI’s Reinforcement Fine-Tuning (RFT) of the o1 model is one recent example: instead of relying on humans to rate outputs, the process automatically scores model responses using well-defined reward functions.
Purpose: Optimize behavior using measurable rewards.
Analogy: Practicing until performance metrics improve, rather than memorizing answers.
Retrieval-Augmented Generation (RAG) — Adding Knowledge Without Training
RAG isn’t training at all — it’s a retrieval technique.
The model stays frozen but is connected to a search index, database, or vector store.
When asked a question, it first retrieves relevant information and then generates an answer grounded in that content.
Purpose: Keep models current and connected to external knowledge.
Analogy: Looking something up rather than memorizing it.
RAG is ideal when data changes frequently, or when you can’t or shouldn’t embed sensitive data into the model itself.
Prompt Tuning — Lightweight Personality Shaping
The lightest-weight form of adaptation is prompt tuning, sometimes called soft prompting.
Here, a small vector (or a few tokens) is trained to steer the model’s behavior without modifying its core weights.
Purpose: Adjust tone or persona without retraining.
Analogy: Giving the same person a new job description — “today you’re the legal assistant.”
Prompt tuning is useful when you want to offer multiple personalities or roles from a single model.
What’s Reasonable to Do (and What Isn’t)
To make this practical, here’s how the different kinds of training align with the Four Categories of AI Solutions from simple Copilot-style automation to custom AI systems built from scratch.
Goal
Technique
Cost & Effort
Category (from “Four Categories of AI Solutions”)
Typical Use
General chatbot or Copilot
None (use aligned base model)
🟢 Low
Category 1 — Copilot / built-in AI features
Office copilots, internal Q&A bots
Domain expertise
LoRA / adapter fine-tuning
🟠 Medium
Category 2 — Configured / composable solutions
Industry copilots, internal assistants
Keep knowledge fresh
RAG or hybrid RAG + fine-tuning
🟠 Medium
Category 3 — Integrated or extended AI systems
Research assistants, customer-facing search
Optimize measurable output
Reinforcement fine-tuning
🔴 High
Category 4 — Custom AI / in-house LLMs
Scientific computing, advanced R&D
Create new model family
Base training
🚫 Extreme
Beyond Category 4
Reserved for foundation model labs
Summary
Base training — teaches language.
Post-training (SFT/RLHF) — teaches behavior.
Fine-tuning — teaches domain knowledge.
Reinforcement fine-tuning — teaches optimization.
RAG / Prompt tuning — extend without retraining.
LLMs aren’t trained once — they’re trained in layers.
Each layer shapes a different aspect of intelligence: from raw linguistic intuition to helpful conversation, domain expertise, and ongoing adaptability.
Knowing which layer you’re working with isn’t just a technical detail — it’s the difference between using AI and building with it.
They say “Data and AI,” as if AI is dessert—something you get to enjoy only after you’ve finished your vegetables. And by vegetables, they mean years of data modeling, integration work, and master‑data management. AI ends up bolted onto the side of a data office that’s already overwhelmed.
That mindset isn’t just outdated—it’s actively getting in the way.
It’s time to flip the script. It’s not Data and AI. It’s AI for Data.
AI as a Data Appendage: The Legacy View
In most org charts, AI still reports to the head of data. That tells you everything: AI is perceived as a tool to be used on top of clean data. The assumption is that AI becomes useful only after you’ve reached some mythical level of data maturity.
So what happens? You wait. You delay. You burn millions building taxonomies and canonical models that never quite deliver. When AI finally shows up, it generates dashboards or slide‑deck summaries. Waste of potential.
What If AI Is Your Integration Layer?
Here’s the mental flip: AI isn’t just a consumer of data—it’s a synthesizer. A translator. An integrator – an Enabler!
Instead of cleaning, mapping, and modeling everything up front, what if you simply exposed your data—as is—and let the AI figure it out?
That’s not fantasy. Today, you can feed an AI messy order tables, half‑finished invoice exports, inconsistent SKU lists—and it still works out the joins. Sales and finance data follow patterns the model has seen a million times.
The magic isn’t that AI understands perfect data. The magic is that it doesn’t need to.
MCP: OData for Agents
Remember OData? It promised introspectable, queryable APIs—you could ask the endpoint what it supported. Now meet MCP (Model Context Protocol). Think OData, but for AI agents.
With MCP, an agent can introspect a tool, learn what actions exist, what inputs it needs, what outputs to expect. No glue code. No brittle integrations. You expose a capability, and the AI takes it from there.
OData made APIs discoverable. MCP makes tools discoverable to AIs.
Expose your data with just enough structure, and let the agent reason. No mapping tables. No MDM. Just AI doing what it’s good at: figuring things out.
Why It Works in Science—And Why It’ll Work in Business
Need proof? Look at biology.
Scientific data is built on shared, Latin‑based taxonomies. Tools like Claude or ChatGPT navigate these datasets without manual schema work. At Carlsberg we’ve shown an AI connecting yeast strains ➜ genes ➜ flavor profiles in minutes.
Business data is easier. You don’t need to teach AI what an invoice is. Or a GL account. These concepts are textbook. Give the AI access and it infers relationships. If it can handle yeast genomics, it can handle your finance tables.
Stop treating AI like glass. It’s ready.
The Dream: MCP‑Compliant OData Servers
Imagine every system—ERP, CRM, LIMS, SharePoint—exposing itself via an AI‑readable surface. No ETLs, no integration middleware, no months of project time.
Combine OData’s self‑describing endpoints with MCP’s agent capabilities. You don’t write connectors. You don’t centralize everything first. The AI layer becomes the system‑of‑systems—a perpetual integrator, analyst, translator.
Integration disappears. Master data becomes a footnote.
When Do You Still Need Clean Data?
Let’s address the elephant in the room: there are still scenarios where data quality matters deeply.
Regulatory reporting. Financial reconciliation. Mission-critical operations where a mistake could be costly. In these domains, AI is a complement to—not a replacement for—rigorous data governance.
But here’s the key insight: you can pursue both paths simultaneously. Critical systems maintain their rigor, while the vast majority of your data landscape becomes accessible through AI-powered approaches.
AI for Data: The Flip That Changes Everything
You don’t need perfect data to start using AI. That’s Data and AI thinking.
AI for Data starts with intelligence and lets structure emerge. Let your AI discover, join, and reason across your real‑world mess—not just your sanitized warehouse.
It’s a shift from enforcing models to exposing capabilities. From building integrations to unleashing agents. From waitingto acting while you learn.
If your organization is still waiting to “get the data right,” here’s your wake‑up call: you’re waiting for something AI no longer needs.
AI is ready. Your data is ready enough.
The only question left: Are you ready to flip the model?
When driving value from generative AI (GenAI) it’s important to choose the right approach in order to be able to get a return on investment. This page attempts at explaining possible approaches and required resources.
Takers, Shapers and Makers
There seems to be 3 major categories of GenAI adopters according to McKinsey and Gartner:
McKinsey
Gartner
Description
Takers
Quick Wins
Focus on utilizing existing GenAI tools and models for productivity improvements with minimal customization.
These initiatives typically have short time to value and are task-specific, aiming for immediate efficiency gains in routine tasks.
Shapers
Differentiating Use Cases
Engage in integrating GenAI tools with proprietary data or adapting them for specific applications.
These initiatives aim to achieve competitive advantages, involving medium time to value with higher costs and risks than quick wins.
They leverage GenAI to extend current processes and create unique value propositions.
Makers
Transformative Initiatives
Concentrate on developing new GenAI models or tools for specialized applications, with the potential to transform business models and markets.
These are the most ambitious initiatives, characterized by high cost, complexity, and risk, and a long time to value.
They aim for strategic benefits that may be difficult to quantify initially.
TCO/ROI
The Total Cost of Ownership (TCO) and Return on Investment (ROI) for GenAI adoption across takers, shapers, and makers categories involve several considerations, including hidden costs, strategic implications, and potential benefits.
Gartner offers insights on measuring GenAI ROI, advocating for a business case approach that simulates potential cost and value realization across GenAI activities. This approach categorizes investments into quick wins, differentiating use cases, and transformational initiatives. Quick wins focus on immediate productivity improvements with short time to value, differentiating use cases aim at competitive advantage with medium time to value, and transformative initiatives have the potential to upend business models with longer time to value but higher costs and complexity. The guide emphasizes the importance of balancing financial returns with strategic benefits, which might be difficult to quantify initially.
Source: https://www.gartner.com/en/articles/take-this-view-to-assess-roi-for-generative-ai. Red box is added by me, see conclusion below.
Builders
I’m introducing an extra “Builders” category into the GenAI adoption landscape beyond merely adopting or adapting, Builders take a step further by crafting bespoke extensions and plugins for GenAI platforms. This initiative is driven by the ambition to tackle intricate, multi-step workflows that typically demand considerable human intervention. The essence of being a Builder lies in their ability to not just work with GenAI but to enhance its core capabilities, enabling solutions that seamlessly bridge various systems and processes. This approach demands a blend of creativity, technical prowess, and a deep understanding of both the technology and the problem domain.
Category
Description
Required People Resources/Skills
Tools
Takers
Utilize existing GenAI tools for productivity improvements with minimal customization.
Aimed at immediate efficiency gains in routine tasks with short time to value.
Basic understanding of AI/ML conceptsSkills in integrating and configuring APIs
Ability to adapt third-party GenAI tools to existing workflows
Microsoft Copilot
Microsoft Copilot Plugins
Enterprise “Chat”-GPTs
Shapers
Integrate GenAI tools with proprietary data or adapt them for specific applications to achieve competitive advantages, involving medium time to value with higher costs and risks.
Low/No-code developers
Domain experts for data interpretation
Project managers with a technical background
Retrieval Augmented Generation (RAG)
Microsoft Copilot Studio
Microsoft Azure AI Studio
Builders
Develop custom solutions or extensions to GenAI platforms to solve complex, multi-step processes that usually require significant human effort.
Advanced programming skills in relevant languages
Data scientists for model tuning
Experience with GenAI frameworks
Systems integration expertise
Creative problem-solving abilities
Microsoft Copilot Extensions
Microsoft PromptFlow
LangChain
LangGraph
LlamaIndex
AutoGen
CrewAI
(OpenAI Swarm)
LLM Function Calling
LLM Routing
LLM Threat Modelling
LLM Security
Makers
Develop new GenAI models or tools for specialized applications with the potential to transform business models and markets.
Characterized by high cost, complexity, and risk, with a long time to value.
Expertise in deep learning and neural networks
Experience in building and training large-scale AI modelsStrong research and development background
Ability to work with high-performance computing resources
LLM Models
LLM Frameworks
LLM Fine-Tuning
(LLM Creation and Training)
The “Builders” category fills the gap between “Shapers,” who mainly adapt existing models for their unique needs, and “Makers,” who create new GenAI models from scratch. Builders leverage powerful frameworks and platforms to create bespoke solutions that automate complex workflows, potentially revolutionizing how businesses approach process automation and efficiency. This distinction underscores the evolving landscape of GenAI adoption, highlighting the increasing sophistication and customization capabilities available to organizations.
Conclusion
The red box on the image above indicates that solutions made in the Takers and lower Shapers category are likely to be overtaken by standard solutions from vendors and the plethora of SaaS AI offerings appearing on a daily basis. Caution should be used when choosing to invest in solutions in this area unless quick wins are important.
Clearly it’s important to have a strategic, well-planned approach to integrating GenAI with emphasis on organizational readiness, skill development, and a focus on applications that offer a competitive advantage – otherwise GenAI just becomes a technology looking for a problem like Blockchain.
I finally got access to ChatGPT Vision and I uploaded a picture of some beers I had at home:
This is the response from ChatGPT:
Nice selection of beers you got there! I see Perlenbacher, which is often associated with Lidl supermarkets, and of course, the iconic Carlsberg. Carlsberg, founded in 1847, is one of the most well-known beer brands globally and is based out of Copenhagen, Denmark. “Birra Messina” seems to be a Sicilian beer. Each of these beers brings a different taste profile to the table. Enjoy responsibly! Would you like to know more about any specific beer?
I then prompted this using TypeChat:
Please attempt to recognize the cans in the picture and fill out the following response: interface Response { data: Array<{ name: string, brand: string, count: number}>; }
No training – no nothing. I know the internet is full of examples of people having ChatGPT turning Figma drawings into web sites and apps – but this is totally out of the comfort zone of ChatGPT and I have previous experience trying to teach image recognition of packaged goods to models and it’s not that easy.
Solutions based on this could destroy the incumbent vendors in image recognition in the CPG market.
Christian says that he doesn’t want a future where a Clippy has to talk to a Clippy which talks to another Clippy and so on. In this context Clippy is some sort of an “AI” which can provide some answer.
I don’t know if that is true or not or if it’s the future we want or not – but right now, that seems to be where we’re headed.
LangChain is about chaining prompts and answers together so that we can get better answers
Microsoft Copilot Plugins is about enabling Microsoft Copilot to extend your prompts to other systems and consolidate the answers and perform tasks on your behalf
Microsoft TypeChat is about providing types and structure to the data we want to pass between systems to make interoperability even easier and more precise
I’ve spent the first 20 years of my life as a developer and the last 20 or so as an “architect” in many different contexts and it almost always boils down to solving integration between business domains and figuring out how data can be accessed, translated and acted upon across the different domains either in real time or before some date and time.
I totally agree with the sentiment that most, if not all, business systems are really just a database with a UI on top and that actual business value comes from combining the data across all systems and extracting answers from it. However, as long as the “business” continues to lock data in silos and refuses to adapt business processes to standard systems and demands customization of systems to accommodate the last 20% of functionality in a way that almost makes integration impossible – then I think the Clippies are going to be around for some time to come.
I think short to medium term the scenarios with plugins to Microsoft Copilot (and similar) will prove valuable to most businesses. Longer term we still need to fix copyright (AIs will be stopped from scraping), privacy (business are scared of leaks) and authentication/authorization (who are you and what do you have access to across boundaries). For instance how can we safely solve problems where data is scattered across private and public domain.
It’s going to be interesting and some new rules/processes are going to have to be worked out.