
Agentic workflows and RAG in production
AI that answers real users and supports real staff, with a person in the loop where it matters. I have taken a help-centre assistant from a rules-based interpreter to retrieval-augmented workflows.
What it covers
- Reply suggestions and automatic answersAutomatic answers in productionAt StudentTenant.com, reply suggestions and automatic answers cut staff interactions by 53% over seven years, starting from a rules-based interpreter and moving to AI as it matured.Learn more → that take routine questions off your team.
- Lead qualification, so staff time goes to leads that can convert.
- Workflow automation with retrieval-augmented generationRAG: retrieval-augmented generationBefore answering, the AI looks up the relevant passages in your own documents and answers from them. It reduces made-up answers and lets the source be shown. over your own documents.
- LLM tools and MCPMCP: Model Context ProtocolAn open standard that lets an AI assistant use tools and data sources in a controlled way. The assistant gets only the access the signed-in user already has. interfaces where the assistant acts with the permissions of the signed-in user, never more.
- Durable AI workflows: long chains of model calls that survive timeouts, overloaded providers and rate limits, and still finish.
Track record
53%fewer staff interactionsAIML reply suggestions, over 7 years
> 50%of user questionsanswered automatically, instantlyMore than half of user questions get an automatic, instant answer.
3generations in productionrules-based, then conversational AI, then retrieval-augmented workflows
2014first chatbotUmnyy, built in 2014 with tools and curated sources
Figures from StudentTenant.com; the first chatbot was Umnyy, built in 2014.
How it runs
Pick one narrow job
An assistant that does one thing well beats one that does everything badly.
Ground it in your data
Retrieval over your documents, with the source shown.
Measure on real questions
A test set from your own cases, run before every change.
Ship with a human fallback
Unsure or sensitive requests go to a person.
What you receive
- A working assistant or workflow in your environment.
- An evaluation set and a way to rerun it.
- A short note on what it does not do.
A good fit when
Your team answers the same questions all day, or you want AI features in your product without losing control of what they can access.
Planning agents or RAG for real users?
I will tell you honestly whether I am the right person.
Get in touch