Here is a problem every business deploying AI eventually hits: what happens when the rules change?
You train an AI model on your customer data, your workflow, your pricing structure. Then your pricing changes. Regulations shift. Customer behaviour evolves. Suddenly, the model that worked last quarter is giving bad answers. You either retrain it (expensive, slow) or you live with the decay.
A new survey paper from researchers A Run and Ziluo Ding (arXiv:2607.11906) maps out an alternative: what if your AI could adapt on the fly, inside its own context window, without needing to be retrained at all?
In-Context Learning: The Shortcut
This is called in-context reinforcement learning (ICRL). Instead of updating the model’s weights (the traditional retraining approach), the model learns from the examples and feedback it sees during a single session. Think of it like handing a new employee a playbook instead of sending them back to training. They read the playbook, they follow the rules, and when the playbook updates, they just read the new version.
The twist this survey focuses on is “non-stationarity” — the environment changing in ways that make old experience misleading. If you have been handling support tickets one way for six months and the company introduces a new returns policy, the old context becomes a liability, not an asset.
How AI Can Tell What Still Applies
The survey organises existing research around three questions: what changes (the reward structure, the rules, the interface), how it changes (gradually or suddenly), and how obvious the change is to the AI. The most interesting work involves retrieval-augmented agents that can fetch relevant past experiences while ignoring the stale ones — essentially, an AI with a working memory that knows when to delete old files.
What This Means for Business AI Deployments
Right now, most business AI is static. You deploy a model, it works for a while, and then it drifts. The ICRL approach promises AI that can roll with changes: an automated customer support agent that adopts a new returns policy just by reading the updated document, a financial compliance assistant that incorporates new regulations without a full retraining cycle.
We are not there yet. This is active research, not production software. But the direction is clear: the next generation of business AI will not just be smarter at the moment of deployment, it will stay smart as the world changes around it.
Practical Advice for Now
While we wait for this research to reach production tools, the practical step is to build change awareness into your AI workflows. Give your AI agents access to your current policy documents. Structure your knowledge base so the AI can tell which version of a rule applies. And when you make a change, update the reference material your AI reads — the model itself might still be static, but the context it reads at inference time can be fresh.
Read the full survey: arXiv:2607.11906