The Immutable Laws of Enterprise AI

Enterprise AI has had its share of growing pains. We started with hallucinations, moved through tokenmaxxing, and have now arrived at an abundance of AI-generated corporate slop that makes the author look productive while giving everyone else more to read and less time to do their actual jobs.

Yet the potential of AI applied to enterprise content is too compelling to ignore. And as organizations gain experience, some fundamental principles are beginning to emerge.

As a tribute to my marketing heroes, Al Ries and Jack Trout, and their famous 22 Immutable Laws of Marketing, I've decided to call these principles The Immutable Laws of Enterprise AI. I have three to offer. So far. 

Law 1: Always bring AI to content, not content to AI

Many enterprise AI projects start with a proof of concept. You copy a sample of corporate data into a data lake or lakehouse, connect an AI model, and start experimenting. So far, so good.

The problems often begin when you try to turn that successful proof of concept into a production system. Moving enterprise content out of its original environment introduces several challenges that are easy to overlook during a proof of concept.

First, content can lose its context. This is particularly problematic with unstructured data such as documents, presentations, and other business files. A document is rarely just a document. It comes with metadata, version history, container structure, links to other files, and inclusion in business processes. Simply copying the file into a lakehouse usually doesn’t preserve all that context. Rebuilding it later can be expensive, complicated, and error-prone.

Second, duplicating content compromises governance and compliance. Enterprise repositories enforce access permissions, retention policies, audit trails, and other controls. When content is copied elsewhere, those controls don't automatically follow it unless you put a lot of effort into rebuilding and synchronizing them. As a result, your AI applications may expose information to users who shouldn't have access to it and break other governance rules.

Similarly, the auditors might ask which copy is authoritative, or the attorneys may inquire whether all relevant documents have been made available for discovery. Suddenly, that simple AI project has created a whole new set of problems. Duplicating content is asking for governance trouble.

Third, copying content adds cost. Storage may seem cheap when you're dealing with structured data, but enterprise content is a different story. Documents, images, videos, engineering files, and other unstructured data can add up to enormous volumes. And it’s not just the cost of storage; it’s also the ingress and egress costs that add up rather quickly.

There's a useful parallel with enterprise search. Search engines have long indexed enterprise documents without requiring organizations to move their authoritative content into a separate repository. AI must follow a similar principle. There's nothing wrong with creating embeddings, indexes, or other derived data to support AI applications. The trouble begins when organizations create another copy of the actual content files and then struggle to keep their context, permissions, and governance synchronized with the original.

The principle is simple: Always bring AI to content, not content to AI. Keep the authoritative content in its governed environment and let AI access it securely.

Law 2: AI reliability rises with the depth of context

To deliver value in an enterprise, AI needs access to enterprise information. That's obvious. What's less obvious is that chunking and vectoring documents alone is rarely enough.

Every organization has its own language, business model, products, customers, suppliers, policies, and processes. New employees spend months learning these details. Experienced employees have spent years accumulating and internalizing that knowledge, and companies value that experience for good reason. They understand what information means in the context of the business.

AI needs that context, too.

Some of it can come from the content itself. Other context may come from metadata, company-specific dictionaries, industry ontologies, knowledge graphs, and structured data scattered across business systems. That last part is especially important: a lot of the context is contained in other data across various systems of record.

Consider a relatively simple example: asking AI to draft a new customer contract. The obvious inputs are the customer's order form and previous contract; you might use those documents as the input. But that’s hardly enough to produce a good contract.

The AI should also understand your current contractual terms, company policies, and applicable regulatory requirements. It should know which products the customer already owns, information typically stored in the CRM system. It may need to understand product compatibility and dependencies, which are documented in your CPQ or PLM system. And depending on the contract, the customer's payment history from the billing system might also be relevant.

The challenge is that this information lives in different systems, in different formats, and under different access controls. Some of it is structured data, some is unstructured content, and some exists only as relationships between the two.

A good enterprise AI architecture needs to connect these sources and make the relevant context available. That doesn't mean dumping every available piece of corporate data into a data lakehouse. The goal is to give AI the up-to-date, relevant information from authoritative sources.

And that leads to the second law: AI reliability rises with the depth of context. The better AI understands your business and the task at hand, the more useful and trustworthy its output can become.

Law 3: The value of AI increases with the degree of automation

AI can make individuals remarkably productive. Give someone a capable AI assistant, and they can draft documents, summarize meetings, analyze information, and produce work in minutes that previously took hours.

That's valuable, but it doesn't necessarily translate into equivalent productivity gains for an entire organization. What if you want an entire marketing team to produce content with consistent messaging, terminology, and tone of voice? To achieve that, you need to move beyond individual AI interactions and start thinking about automation.

You will need more than individual prompts. You need a shared prompt library, reusable skills, consistent data sources, and shared agents that perform specific tasks. As requirements become more sophisticated, so does the automation.

At one level, an agent might perform a simple task using a predefined workflow. More advanced implementations can orchestrate multiple agents, each responsible for a different part of a larger business process. Things get particularly interesting when AI agents become part of an automated business process.

Instead of asking an assistant to prepare a customer renewal proposal, imagine a workflow that automatically identifies upcoming renewals, collects relevant customer information, drafts proposals, checks them against company policies, and routes them for approval. Humans remain responsible for the important decisions, but much of the repetitive work happens automatically.

That's where AI starts delivering value beyond individual productivity. AI adopted across teams or the entire business becomes truly game-changing.

Which brings us to the third law: The value of AI increases with the degree of automation. The real opportunity isn't just helping individuals work faster. It's improving how work gets done across the organization.

Three laws, for now

So, there you have it. My three immutable laws of enterprise AI:

• First, bring AI to your content, not your content to AI. Keep authoritative information where it belongs, with its context and governance intact.

• Second, give AI the relevant business context. Documents alone rarely tell the whole story, and AI is only as reliable as the information it can access.

• Third, move beyond individual AI assistants to automation. The real value comes when AI improves how the entire organization works, not just how fast individuals complete tasks.

I'm sure there are more immutable laws of enterprise AI. After all, Ries and Trout came up with 22 laws of marketing, and I've only got three.

What other laws would you add?