A chatbot on a website is not an AI strategy. It is a surface. The useful work is behind it: which documents are in scope, which APIs an assistant may call, which identities it inherits, and how a wrong answer is caught.
Start with the job
Classify incoming documents. Summarize a case file for a person who already has access. Draft against a knowledge base that is maintained. Route an exception. Those are jobs. “Add AI” is not a job.
Then design the connections
- Identity: the model should not see more than the user is allowed to see.
- Retrieval: sources should be named, current, and owned.
- Actions: if an agent can write or send, it needs logs and limits.
- Evaluation: someone has to sample outputs in the real workflow.
This is why MHM treats AI as a layer on infrastructure, software, and security. If those layers are weak, the model will confidently operate on a mess.
When the answer is no AI
If the process is already a clear rule, deterministic automation is cheaper and safer. If the data is not trustworthy, fix the data. Declining a model is part of being a serious technology partner.