Last time I argued that most enterprise AI is just expensive advice: it answers, and a human still does the work. The obvious next question is why. Why does the work always land back on a person? The answer is context, but not in the way the industry uses the word.
Ask any vendor what context is for and you'll get the same reply: it makes the answers better. Ground the model in your data, your history, your documents, and it hallucinates less and reasons more sharply. All true. And all beside the point.
Context doesn't exist to improve answers. It exists to eliminate handoffs.
Every handoff is a missing piece of context
Go back to the thirty minutes of human orchestration from part one, the verifying, the CRM update, the email to engineering, the Jira ticket, the roadmap edit, the Slack message, the follow-up. Look at why each of those steps needs a human at all.
- The human verifies because the AI couldn't see enough to be trusted.
- The human gathers context because the AI didn't have it.
- The human updates the CRM because the AI has no permission to.
- The human messages the team because the AI has no memory of who's involved.
Every single handoff exists to compensate for something the system didn't know or couldn't touch. The human is the patch. We've been hiring people to be the connective tissue between systems that were never designed to talk, and calling it a workflow.
The reframe changes what you're buying
Once you see context this way, the buying decision changes shape. You stop asking "how accurate is the answer?" and start asking "how much of the after-work disappears?"
Context as accuracy
More data in, smarter answer out. The human still executes. You've bought a better-informed advisor.
Context as execution
Enough context to act. The handoffs stop existing. You've bought completed work.
These sound similar and they are not remotely the same product. The first makes your people slightly faster at the same manual choreography. The second removes the choreography. One is a feature. The other is a category.
Why "more context" isn't the same as "enough context"
There's a trap here worth naming. Dumping more documents into a model gives it more to reason over, but it doesn't give it the ability to act. Accuracy context and execution context are different things:
- Accuracy context is retrieval: the right passages, the relevant history, the supporting facts.
- Execution context is state and permission: what's true right now across your systems, and what the system is allowed to change.
You can have a mountain of the first and still hand every task back to a human, because the model can read your world but can't move anything in it. Eliminating handoffs requires the second kind, and the second kind is much harder to build. That difficulty is exactly why so few tools cross the line from advice to action.
Count the handoffs, not the tokens.
If adding context made the answer better but the number of human steps after the answer stayed the same, you improved the advice. You didn't eliminate the work.
What it looks like when context does its real job
When a system has execution context, the shape of the output inverts. It doesn't return a recommendation for a human to go implement across six tools. It implements it, because it already knows the account, holds the permissions, and can see the current state of every system the task touches. The proposal isn't drafted for you to route; it's built and routed. The escalation isn't summarized for you to resolve; it's resolved.
That's not a smarter answer. It's the absence of the work that used to follow the answer. And it's only possible because context stopped being fuel for a response and became the foundation for an action.
The goal was never a better answer. It was less work.
Next in the series: if context is what lets a system act, what happens to the person who used to do all that acting? The uncomfortable truth is that in most "AI-powered" companies today, the human is still the operating system, and that's the real ceiling on ROI.
Part two of the Expensive Advice series, on why context is valuable for what it enables, not for how smart it makes the answer sound.