Think about onboarding a sharp new hire. On day one they're brilliant in the abstract and useless in the specific. They don't know your customers, your systems, how your month-end actually works, or why you do that one thing the weird way. Give them a few weeks of context and the same person becomes invaluable. The intelligence didn't change. The context did.
AI is in roughly the same spot. The biggest unlock isn't waiting for a smarter model. It's giving the model you already have the context it needs to be useful in your business.
The myth of the smarter model
There's a quiet assumption running through a lot of AI conversations: that the output is disappointing because the model isn't good enough yet, and the next version will fix it. Sometimes that's true. Usually it isn't.
<20%
accuracy when an AI queries a raw database with no context
AtScale, 2025
>95%
accuracy for the same model paired with a governed context layer
AtScale, 2025
+38%
lift in AI analyst accuracy from adding business context, in a 522-run test
Atlan AI Labs, 2025
Same model, wildly different results. The variable wasn't intelligence. It was whether the model understood the business it was working inside.
Two ideas worth separating
Here's the distinction we keep coming back to with leadership teams, because once it lands, a lot of the confusion clears up.
A project
Everything the AI needs to know about a workflow or a part of your business — your definitions, your systems, your rules, how the work actually flows. Write it down once and the AI stops guessing.
A skill
A specific, repeatable thing the AI does inside that context. The month-end export. The proposal in your format. The variance summary your CFO actually wants. Defined once, reused on demand.
Most teams skip straight to the skill — they ask for the output — and then wonder why it comes back generic. It comes back generic because the AI has no project to stand in. No context. It's the brilliant new hire on day one.
An example: month-end close
Say you want AI to help with the monthly close. The skill is obvious: produce the variance commentary. But the skill is worthless without the project around it — which accounts roll up where, what counts as a material variance for your business, how your entities relate, what last month looked like, the language your leadership expects.
Document that context once and the variance commentary stops being a clever prompt you re-engineer every month. It becomes a repeatable skill that runs against a stable project. That's the difference between a party trick and something you can actually depend on.
The work isn't clever wording. It's writing down how your business actually runs — once, clearly — so the AI has somewhere to stand.
The shift this year
This isn't just our framing. Gartner put it bluntly in 2025: context engineering is in, prompt engineering is out. The field has moved on from hunting for the perfect phrasing to building the context the model operates within. Prompts are getting commoditized. Context is the durable advantage, because your context is the one thing a competitor can't copy.
If you want a place to begin, pick one workflow and document it end to end — inputs, rules, definitions, the lot. Then point AI at it and watch the quality change. You'll have built your first project, and the skills will come easily after that.
Key Takeaways
- The biggest AI unlock is context, not a smarter model — same model, very different results.
- A project is the context layer; a skill is a repeatable action that runs inside it.
- Generic output usually means the AI has no project to stand in, not that it isn't capable.
- Document one workflow end to end and the skills built on top of it become reliable and reusable.
- Context engineering is the durable advantage — it's the one thing competitors can't copy.