Rebuilding a context-blind email AI into a stateful sales engine

1. 01

   Generating an email that actually reads as one email

   The challenge

   The composer treated AI generation as a set of isolated tasks. A rep generated a hook, then a story, then an outcome, each on its own, then stitched them together by hand. Because the steps never shared context, the story did not build on the hook and the outcome did not pay off the story. The result read like three fragments, not one persuasive message.

   What we shipped

   We rebuilt the composer so the AI always knows what came before. The hook is generated first, and every later step is written with the earlier content in context. The email arrives coherent instead of needing a human to make it make sense.

2. 02

   Letting a rep fix one part without breaking the rest

   The challenge

   There was no way to refine a single section. Asking to rewrite one part meant breaking the whole structure, so reps either accepted output they were not happy with or rewrote it by hand, which defeats the point of the product.

   What we shipped

   Any section can now be regenerated on its own with plain-language instructions: shift the tone, shorten or lengthen it, or rework the story to match a specific hook, all without touching the rest of the email.

3. 03

   Bringing that control to the rest of the app

   The challenge

   The flexibility stopped at the composer. Everywhere else in the product a rep was back to copying text out, pasting it somewhere else to improve it, and pasting the result back.

   What we shipped

   We built a context-aware revision chatbot. When a rep highlights any part of their content, a hook, a value proposition, a line in the email, that selection is attached to the chatbot's context automatically. They can ask it to sharpen, rewrite or critique exactly that piece, get a suggestion that understands where it sits in the sales cycle, and paste it straight back.

4. 04

   Making the founder's sales method run the same way every time

   The challenge

   Output quality depended on whether a given prompt happened to land. The founder's proprietary sales methodology was the product's real differentiator, and it was being applied inconsistently across the generation steps.

   What we shipped

   We audited and rewrote the Langbase pipes underneath the features. Inputs were standardized, system prompts tightened to follow the methodology, and models tuned for consistency, so the output is on-method every time rather than when the prompt gets lucky.

5. 05

   Staying reliable when the model is slow or the data is thin

   The challenge

   Real usage surfaced the ugly cases. A slow upstream response would time out part-way through a generation. A missing or partial persona would fail an otherwise valid run. And a credit could be deducted for work that never actually completed.

   What we shipped

   We made generation resilient. Persona handling degrades gracefully instead of failing the request, credit deduction retries rather than quietly charging for nothing, and timeouts were tuned to the real latency of the AI calls, so a slow response stays a slow response instead of surfacing as an error.

Before and after

Technology stack