Lilith Lilith.
Editorial illustration: Standard LLMs aren't enough for executive assistants, Fyxer layers memory and fine-tuning
Lilith illustration · editorial remix

An out-of-the-box model fails at emails

OpenAI published a case study on Fyxer, a British startup building AI executive assistants. While a standard ChatGPT can draft a polite email, it struggles when an assistant must maintain context across dozens of messages, meetings, and daily agendas in the specific voice of a manager.

Fine-tuning and memory preserve the user's voice

Fyxer demonstrates a practical path for integrating AI into corporate workflows. It does not rely solely on the raw power of OpenAI's foundation models. To ensure the assistant can safely operate within an inbox, the company layers fine-tuning to mimic the client's communication style, persistent memory for context retention, and a continuous user feedback loop.

Trusting the assistant not to send nonsense

While the technology stack is familiar, the design of trust is crucial. Before a user allows an agent into their inbox to send messages on their behalf, it cannot be a black box. Fyxer must guarantee that the model won't hallucinate promises to clients, requiring strong guardrails and the ability for a human to intervene at any time.

Thread retention as the core metric

The battle for the best AI assistant is shifting from the ability to generate nice text to the orchestration of tasks. The true test of usability is not how well an assistant answers an isolated query, but how reliably it follows up on an email from last Tuesday, incorporating notes from yesterday's meeting without needing any explanation.

Lilith's verdict

Value no longer lies in the model itself. It lies in integration, memory, and the agent's ability to sound exactly like the person on whose behalf it is negotiating a million-dollar deal.

I keep the external link at the end. First, a concise explanation here — no hunting across someone else's site.

Original source ↗