Malcolm Chisholm

System Prompt Governance is the key to AI Governance. Data is passive, but AI is active, which makes governance much more important for AI than for data. Read more about what you can do in my latest Substack article: open.substack.com/pub/malcolmchi… #ai #aigovernance #llms #datagovernance

AI doesn't have a soul, or a conscience, or a personality. What it does have is a facsimile of all these things - and more - embedded in its System Prompt. System Prompts are the most consequential yet least governed artifacts in the AI stack. They control everything an AI system does - and almost no organization treats them as assets worth governing, or even managing. No owners. No versioning. No compliance review. No edge case testing. No lifecycle management. I've been working in data governance for over 30 years. I know what an ungoverned asset looks like - and I know what to do about it. I've developed the System Prompt Governance Framework (SPGF), built on the same of the same principles that transformed how organizations manage data: precise definitions, clear ownership, measurable quality, lifecycle discipline, and accountability. But I have extended the SPGF to account for what is genuinely new in AI. Read more in my latest Anchored Data article on Substack: anchoringdata.com/p/your-ai-has-… Ready to govern your AI's soul? Contact me at mchisholm@datamillennium.com #AI #AIGovernance #SystemPrompt #DataGovernance #ResponsibleAI #AIRisk #EnterpriseAI #PromptEngineering #AIStrategy #GenAI

Join me tomorrow for a discussion of Open Metadata and Semantic Intelligence in DM Radio’s Briefing Room! I will be discussing these very relevant and important topics with Suresh Srinivas, CEO of Collate (getcollate.io). Hosted by the great Eric Kavanaugh, this is going to be a very timely topic for the current singularity we are living through. Don’t miss it – Register now at: us02web.zoom.us/webinar/regist… Wednesday February 25 at 1:00pm ET #data #ai #metadata #standards

Are you proving that you are more productive as a result of using AI? This seems to be a vital question in US corporations now. You not only have to be using AI to be more productive, but you also have to prove it. I am starting to keep an AI journal for my work where I track what I am doing with AI and how it is making me more productive. It’s a personal kind of AI governance. What are you doing? #ai #aigovernance #data #datagovernance

AI is not a tool – it’s a symbiont. The response you get back from a prompt in part depends on your relationship with AI. This seems to be highly underappreciated by enterprises. I and my co-worker entered the same prompt into the same AI (Claude Opus 4.6). Her response was more aligned with an executive perspective, and mine more operational. So the way in which AI works is different for each of us. Which may mean – unfortunately – that if you are not a high functioning individual, then what you get back from AI may be less than it provides to someone else. Sounds a bit brutal, but it does seem the relationship matters. To what extent and how significant this is I am not sure, but it may limit AI literacy I am moving to ask teammates to run the same prompts I do and then I am combining the results. #ai #aigovernance #datagovernance #ailiteracy

With the advances this month in AI what matters is your ability to conceptualize and provide clear and precise intent to AI. Your inputs have to be close to exact to get the output you need. And definitions are important. You need to remove any ambiguity about the terms you are using. Are your definitions of the quality required? Especially for data - if you want AI to work with data! Shameless plug: amazon.com/Definitions-In… #ai #definitions #data #datagovernance

What’s in Your Moat? It’s the question of the day! As a participant in the data and software industry, if you cannot answer that question, then you probably do not have a moat and will be unable to face the onslaught of AI. As an AI/Data Governance professional, I am hoping AI & Data Governance will provide part of the moat – but it’s going to be a battle! #ai #moat #datagovernance #aigovernance

Pricing Power of Software Companies vs AI – Just Got Very Difficult. If a buyer thinks a software company is using AI to write code and not developers, then the buyer will demand lower prices to share in the efficiency gain. If a buyer thinks a software company is still using human developers to write code, and not AI, then they will find an alternative or demand lower pricing because the company just got a lot riskier. No wonder Wall Street is freaking out. x.com/TechCrunch/sta… #ai #SaaS #datagovernance #procurement

Semantic Intelligence as a driver of AI! What to learn more? You probably need to. Join me in DM Radio’s Briefing Room where I will be discussing semantic intelligence and open metadata with Suresh Srinivas, CEO of Collate (getcollate.io). Hosted by the great Eric Kavanaugh, this is going to be a very timely topic for the current singularity we are living through. Don’t miss it – Register now at: lp.dmradio.biz/semantic-intel… Tuesday February 25 at 1:00pm ET #data #ai #metadata #standards

This week’s news: No job is safe – starting with AI jobs. Unlike the SciFi AI plots it seems real AI does not respect its creators, and coders are not needed at Anthropic any more (well, not to actually code). It appears that February 2026 marks the crossing of a Rubicon. See the number of views of the below linked items. So what – if anything – will we be doing in the future. It seems there are a few things still needed: 1. Clarity of intent. The ability to conceptualize and provide specific guidelines to AI. It can’t build a C-compiler if it cannot be told what a C-compiler is. 2. Providing an understanding of enterprise data. In a way this maps back to intent as the data is used for something. 3. General semantics. Not sematic technology, which is often discussed here, but actual semantic content. Again, this overlaps with 1 and 2. Links: Matt Shumer: x.com/mattshumer_/st… Codie Sanchez: x.com/Codie_Sanchez/… Nate: youtube.com/watch?v=JKk77r… #ai #data #datagovernance #aigovernance

What do data policies mean to the business? In my latest Substack article I explore how businesspeople should understand data policies, and why data policies really matter: anchoringdata.com/p/a-business-u… #datagovernance #datamanagement #policies #data

SaaS Remediation Projects are Here – What Role Will Data Governance Play? Yes, companies are looking at strategies to get rid of SaaS software and replace it with AI-generated solutions. Unless the market rejects the narrative that AI can eliminate SaaS, or questions its fundamentals, this is going to be a force in 2026. Data Governance professionals would be well advised to prepared for SaaS remediation projects now. What are you doing? #ai #saas #datagovernance #data

Going forward, if you cannot prove you are NOT a victim of AI, then you will become one. This applies just as much to data professionals as to enterprise software experiencing the current stock market “SaaSpocalypse”. Remember that Data Administration was wiped off the face of the planet in the dot.com bust of 2000. Data Governance could suffer the same fate. Data professionals should show how effectively they are using AI in their daily work to boost their personal productivity. They should also find practical ways to get data and metadata prepared for AI. In the past few days I have tried to draw attention to MCP servers as a way to do this. There is more, but time may be short. Complaining that others in the organization are not fulfilling their “owner” responsibilities is not going to help going forward. Agree or disagree? #ai #datagovernance #saas #mcp

Have MCP Servers Realized the Original Dream of the Semantic Web? It’s probably a tall order, but I contend that XML and RDF did not achieve this goal. XML in practice rarely if ever had the semantic content envisioned for it, with its tags usually being nothing more than field names. Humans had to interpret them on the receiving end. By contrast,MCP Servers that expose databases to AI *must* have adequate semantics to work because AI agents have to interpret these semantics in order to process the data accurately. It is true that MCP Servers do not have the explicit goal of the Semantic Web, but they appear to be getting us closer to it. What do you think? Scientific American article: www-sop.inria.fr/acacia/cours/e… #ai #mcp #datagovernance #semanticweb #xml

Will Model Context Protocol servers (MCP) replace Data Products? MCP allows data to be interfaced with AI. Many Data Products I have seen are little more than individual database tables or simple joins. So, yes, MCP servers could replace this kind of Data Product. But MCP is only for AI, and there are many non-AI uses of Data Products. Fair enough, but AI agents could supply the data to these uses. So again, more pressure on traditional Data Products. I think MCP servers are more demanding of governance, and more rigorous in what has to be governed, than Data Products. So maybe a transition to MCP servers and away from Data Products may be a good thing. What do you think? #ai #mcp #dataproduct #datagovernance

Architecture Question: Do you bring data to AI or AI to data? Bringing the data to AI means building data pipelines. Bringing AI to the data means setting up Model Context Protocol (MCP) servers. The former requires teams of data engineers to build and maintain the pipelines every time there is a need. The latter requires one-time setup of MCP servers that then get called as needed. Of course, it is not quite as simple as this as the data is accessed in raw form and may still need reshaping for the use case in question. Do data engineers have a future building pipelines? There will always be some, but pipelines are no longer the only game in town, at least when it comes to AI. On the other hand, MCP servers open a world of AI and governance needs that have to be addressed. What do you think? #ai #mcp #datagoverance #aigovernance

An intersection of Data Governance and AI that we need to pay attention to is MCP – Model Context Protocol. Imagine you have a SQL query that returns a dataset from a Customer database. You can create an MCP server – easily, in minutes, with nocode tools – so the dataset is accessible from AI with an MCP client (which they all seem to have). But the MCP server has to describe in English (or whatever human language) exactly what it is about and what the data being provided is. If the same horrible track record of API documentation is followed, this will cause chaos. The MCP server essentially has microprompts in it. These have to be governed to ensure they are adequate, keep up with schema and reference data (yes! reference data) drift. The sloppy ungoverned technical documentation of the past few decades will not cut it with MCP and lead to real-world impactful failures. And this is before exposing sensitive data, not knowing where the data is leaking to, what AI’s are calling it, and MCP server owners (you can never have enough types of owner). But that’s all another governance story. Do you see this as urgent – let me know. #ai #mcp #datagovernance #database

In my latest Substack article I describe Data Disposition from a business point of view. Read more at: open.substack.com/pub/malcolmchi… #data #datamanagement #datagovernance #datadisposition

How long before every AI query you issue is metered and you have to pay for it? Or do you think it’s all free? See below where I ask Perplexity to estimate the Watt-Hours required to run a query I just asked it. The result is 60 Watt-Hours, which I find horrifying. Yes, I have no way of validating this and the Perplexity response included all kinds of disclaimers about uncertainties. However, I suspect it is somewhat right. It reminds me of the days when we were told moving to the Cloud would produce giant cost savings. Or the innocent days before Data Privacy. A new reality is dawning where energy is a scarce resource that is going to get priced in for AI usage. And I suspect it is not long before it happens. Let me know what you think. #ai #energy #datacenter #electricity

Hurricanes are your friend in AI cost! Why? Here in Florida it is riskier to put data centers, so they are built elsewhere, e.g. Northern Virginia (“Spy Country”). So when I run a giant AI query, I think some unfortunate Virginian is paying for the energy costs involved. I accept that the reality of the US electricity market is more detailed and complex than my simple thought process, and electricity rates are rising in Florida for reasons other than data centers. And data centers will come to Florida. Nevertheless, there is a definite geographical mismatch here, and AI costs are being disproportionately borne by citizens of certain mid-Atlantic and North-Eastern states - a kind of AI arbitrage. #ai #florida #datacenter #electricity