Context is the new gold. Why AI agents are only as good as what your company knows about itself

Every company now has access to the same AI models. The same chat window, the same agents, often the same prompts. What it doesn't share with anyone is its context: what the brand stands for, which claims it can make, how its products work, who decides what and why. That context is turning into the most valuable thing a company owns.
From prompts to context
For two years, the industry talked about prompt engineering: finding the right words to ask a model. In mid-2025 the conversation moved on. Andrej Karpathy put it this way: context engineering is “the delicate art and science of filling the context window with just the right information for the next step.”
Anthropic followed with a practical guide for building agents. Its core advice is to find “the smallest possible set of high-signal tokens that maximize the likelihood of some desired outcome.” Models, the guide explains, have an “attention budget” they spend on everything you give them.
The shift sounds technical, but it isn't. A prompt is a question. Context is everything the model knows when it answers: your brand, your products, your rules, your history. You can write a better question in five minutes. You can't write a company's context in five minutes.
Agents are arriving faster than the context they need
Gartner expects 40% of enterprise applications to include task-specific AI agents by the end of 2026, up from less than 5% in 2025. The same firm also predicts that over 40% of agentic AI projects will be canceled by the end of 2027, because of “escalating costs, unclear business value or inadequate risk controls.”
Both can be true at once, and the reason is mostly the same. Gartner warns that through 2026, organizations will abandon 60% of AI projects that aren't supported by AI-ready data.
Industry surveys point the same way. They come from vendors who sell the cure, so read them as direction, not as precise measurement:
- In a Teradata survey of 1,000 technology and data leaders, 77% said that 20% or less of their enterprise data is described and contextualized well enough for agents.
- In the Contentstack Agentic Enterprise Report of 621 digital leaders, 78% of those already running agents in production ran into content or data readiness issues. 88% of them wish they had invested in content and data infrastructure earlier.
The pattern is hard to miss. The model is rarely the problem. The problem is that the agent doesn't know enough about the company it works for, or it knows the wrong things.
What company context actually is
When people hear “data”, they think of databases. For most companies, the context that matters to an agent is much broader, and much of it lives in places no one thinks of as data:
- Brand. Voice and tone, visual rules, the design system, the words you use and the words you avoid.
- Content. Product facts, approved claims, case studies, the texts on the website and in the app.
- Rules. What can be promised, what needs legal review, pricing policies, accessibility requirements.
- Decisions. Why things are the way they are. Which direction was tested and dropped, and why.
- Ownership. Which source is the source of truth, who owns it and who can change it.
In most companies, this knowledge is scattered across a brand manual in a PDF, a Figma file, a website CMS, an app, a shared drive full of presentations and the heads of a few senior people. A human colleague can navigate that mess. They know whom to ask. An agent can't.
More is not better
The obvious reaction is to give the agent everything. Connect every drive, every inbox, every document. That rarely works.
Anthropic's guide describes what researchers call context rot: as the amount of text in the context window grows, “the model's ability to accurately recall information from that context decreases.” An agent fed with ten outdated versions of the pricing deck will confidently quote the wrong one.
Good context is curated, not collected. It has one current version of each fact, clear priorities and clear limits. In design, we know this discipline well. It's the difference between a folder of nice screens and a design system: a set of rules, components and decisions that anyone, or anything, can build from without asking.
Why it's gold
Gold is valuable because it is scarce and keeps its value. Company context works the same way.
Models are becoming a commodity. When everyone uses the same models, the same tools and often the same prompts, the output drifts toward the same average. We wrote about this in Recognizable, or invisible. The only input your competitor can't copy is your context.
Context also compounds. Fortune's coverage of MIT NANDA's 2025 research on generative AI in business reported that most pilots stall with little to no measurable impact on profit and loss. The core issue, according to the researchers, is a learning gap: generic tools don't learn from or adapt to a company's workflows. A well-kept context is exactly that missing memory. Every decision written down, every rule clarified and every outdated fact removed makes the next automation cheaper and the next agent more reliable.
Standards make context portable
Connecting agents to company systems used to mean custom integrations for every tool. That is changing. The Model Context Protocol, an open standard for connecting AI applications to data and tools, was donated by Anthropic to the newly formed Agentic AI Foundation under the Linux Foundation in December 2025.
Standards solve the plumbing. They don't solve the content. A protocol can connect an agent to your brand guidelines in an afternoon. It can't make those guidelines current, consistent or clear enough for a machine to act on. That part is still work, and it is work that combines brand, design and technology.
Safe and sustainable, or not at all
An agent that can act on company context can also act wrongly on it. Two principles decide whether that context becomes an asset or a liability.
Safe. Every agent should see only what it needs and do only what it is allowed to. Permissions follow roles, sensitive data stays out, and anything that leaves the company, from a campaign to a price, passes human review. Gartner's “inadequate risk controls” are, in practice, missing rules about who and what may do which thing.
Sustainable. Context ages. Products change, the brand evolves, rules get updated. If nobody owns a piece of context and nobody reviews it, agents will keep working with last year's company. A sustainable context has owners, versions and a regular review, just like a good design system.
Where to start
You don't need a big program to begin. In our experience, a few steps take a company a long way:
- Map your sources of truth. For brand, content, product and rules, decide which source is the current one and retire the duplicates.
- Turn the brand into rules. A brand manual written for people becomes a design system, tone of voice guidelines and a list of approved claims that both people and machines can read.
- Name owners and a review cycle. Every part of the context has someone responsible for keeping it current.
- Connect it through standards, with permissions. Let agents read the context through open protocols, with access set per role.
- Start with one workflow. Pick one repeated task, for example adapting a campaign to new channels, let an agent run it on top of the context and measure the result before you scale.
Mine it, refine it, keep it
The companies that will get the most out of AI won't be the ones with access to the best model. Everyone will have that. They will be the ones that know themselves best and have written it down in a form their people and their agents can use.
Context is the new gold. Like gold, it has to be mined, refined and kept safe. That is the idea behind our Context system for AI agents: connecting marketing materials, the website and the app into one ecosystem that AI agents can run, safely and sustainably.