Hi everyone 
I’m working on agent architectures for email workflows and I’d love to learn from people actually dealing with this in production (or trying to).
If you’re using agents to manage email — spam filtering, summarizing threads, drafting replies, keeping sales conversations alive, registering new customers into a CRM — what’s failing for you?
Common issues I keep hearing about:
- Agents losing context across long threads
- Hallucinated replies that damage customer trust
- No persistent memory between sessions
- Costs that scale badly with volume
We’ve been building PPAI, an architecture that handles these tasks with persistent state, and I’d genuinely like to test it against real problems — not demos. If you have a use case that’s breaking, tell me about it. Happy to run a pilot with someone who’s actively working on this.
What are you using today, and where does it fall short?
To be honest, I’m not really sure I follow your point. What exactly is the goal here? You mentioned problems you’ve heard about, but not ones you’ve actually run into yourself. Also, is hallucination really an issue in sales conversations in your project ?
"If you’re using agents to manage email — spam filtering, summarizing threads, drafting replies, keeping sales conversations alive, registering new customers into a CRM — what’s failing for you?"
we build a new arquitecture that is more oriented to social task, like manage an email, and we are testing. Yes hallucination is an issue, deoends of the model can invent produtcs, orders, names.. is very common.
Thanks for the great explanation! To add to your point, we actually ran an experiment where we recorded discount negotiations between customers and sales experts and fed that data to our AI. On average, the AI’s suggested price dropped significantly after just 3 or 4 rounds of negotiation. Hope that helps!
I’d be interested to know if your data shows similar results.
you are using an agent? is difficult to say here, because. is hard to understand that can exist a different model than llm. i am runnning a radio qith arounc 20 complex tasks and has been sustained for more than 770.000 without interruption or human interaction— we are trying to see if can setup a model for real business task. you are usiing an agent? you have your own pipeline? what model are you using? we can try use our IA model? i can give you instructions and an endpoint, just to learn how reacts, we believe can solve that task better than LLM agents.
We segmented customers into tiered cohorts based on proprietary lead scoring. We then implemented a RAG architecture to ground LLM outputs in class-specific personas and dynamic price-elasticity data.
PS:English is my second language, so please bear with me if there’s any confusion. Thanks for understanding!
Thanks for the details — that’s a solid setup, and the fact that you still saw price drift after 3-4 rounds despite RAG grounding is exactly the pattern we’ve been studying.
Our view (and the reason we built a different architecture instead of a fix): RAG re-injection has a failure mode that compounds per turn. The model doesn’t hold a view of the task — it reads everything you re-inject and produces text, not a decision. It behaves like a text producer, not a negotiator. So grounding helps per-turn accuracy, but it can’t give the model something it structurally lacks: continuity of intent across turns. We’re assigning tasks to LLMs as if they understood them; they don’t — they complete them.
In our architecture (PPAI) that role split is explicit: an LLM agent can be the “accountant” (discrete, stateless tasks — it’s genuinely good at those), while the negotiator is a persistent entity that owns the relationship and the constraints over time, using the LLM only as its language interface.
To be upfront: we can’t offer you a fix for your current stack, because our conclusion was that the problem is architectural, not promptable. But I’d genuinely like to test this against your scenario. We could run a controlled experiment with fictional customers — same negotiation setup, N rounds, and see whether a persistent entity holds the pricing constraint where a RAG agent drifts. If it drifts too, that’s useful data for both of us. Interested? note: what is your primmary language?
Thanks for the explanation! My primary language is Chinese.
Tomorrow, I’m going to post some of my own ideas from our past projects. I’ll be leaving out the parts that involve my former partners.
I’m looking forward to seeing your great work!