At a Glance
- Google, Microsoft, Alibaba, Snowflake and Databricks are all shipping a “context layer”: the part that tells AI agents what your business means and how it works.
- That’s good news. Agents fail more often from missing context than from weak models.
- The risk is that your business logic gets rewritten in five vendors’ formats. That’s a deeper lock-in than data lock-in ever was.
In our last take, we argued that AI isn’t killing the data platform; it’s feeding it. This week, the story moved one layer up.
The week everyone said “context”
This week, Microsoft starts rolling out a preview that connects Work IQ, the layer underneath Copilot and its agents, to business data in Dynamics 365 and Power Platform. The feature to watch is “business skills.” You define a process like renewal readiness once, and every connected agent can follow it.
A week earlier, Alibaba described its new agentic cloud in three layers. One of them is a dedicated Context Engine that holds real-time data and long-term memory.
Google got there first. At Cloud Next in April, it turned its data catalog into a Knowledge Catalog: a live map of your business that uses Gemini to tag data, define logic and connect relationships automatically. Google has also said this layer supplies the missing half of agent accuracy that clean data alone can’t. Notice the twist here. Google doesn’t only store your definitions. It writes many of them for you.
Snowflake announced its own version in June. Databricks is pitching a single live copy of data so agents don’t have to wait for pipelines. Salesforce said much the same thing at Dreamforce.
The vendors are different, but the bet is the same. The model is no longer the hard part. The hard part is knowing what “active customer” means at your company.
Why this matters
Agents rarely fail because they can’t reason. They fail because nobody told them that “revenue” in finance excludes returns, or that renewals above a certain size need legal sign-off. Context layers fix that, and that’s good.
But look at what each platform is asking of you. To make its agents smart, you have to encode your definitions, processes and playbooks inside that platform. Do that in Microsoft, Snowflake and Salesforce, and you now have three versions of “active customer.” Three teams maintain them, and they quietly drift apart.
Data lock-in was about where your tables lived. Context lock-in is about where your business meaning lives. That’s much harder to move, because it isn’t rows of data. It’s judgment.
Our take at Atgeir:-
- Write it down once, upstream. Keep business definitions and metrics in one place you control, usually your data platform’s semantic layer. Let every agent platform read from it instead of redefining it. If a platform auto-generates definitions for you, review them and bring them upstream too.
- Treat playbooks like code. Business skills and agent instructions are company IP. Version them, review them, and store them somewhere you could hand to a different vendor tomorrow. Protocols like MCP are the delivery pipe, not the vault.
- Ask the export question before you sign. For every context feature a vendor pitches, ask: if we leave, what comes with us? If the answer is “your data, but not your definitions,” you’re buying lock-in with a nicer interface.
The next platform war won’t be fought over your data. It’ll be fought over what your data means. Decide now who gets to own that.