What we are exploring
Communication overload is a familiar problem: long threads, scattered channels and a constant trickle of messages that may or may not need attention. This lab investigates whether AI can reduce that load by triaging, summarising and drafting - without quietly taking decisions away from the person responsible. We prototype small flows that classify inbound messages, condense conversations and suggest replies, and we study where those assists genuinely help versus where they add a layer of risk.
A recurring theme in our research is faithfulness. A summary that drops a crucial caveat, or a suggested reply that subtly changes tone, can do real harm in a professional setting. So we focus as much on what these tools should refuse to do - guessing at intent, smoothing over disagreement, hiding messages - as on what they can usefully produce. Everything here is experimental and may change as we learn more.
How we measure summary faithfulness
Summarisation is easy to demo and hard to trust, so we treat faithfulness as the metric that matters most. In our experiments a summary is only useful if a reader can rely on it without being misled, which means it must preserve key facts, commitments and caveats, and link back to the source so nothing is taken out of context. We test summaries against the original threads and look specifically for omissions, invented details and shifts in emphasis.
These are internal, exploratory evaluations rather than formal benchmarks, and we are honest about their limits. They tell us whether a prototype tends to distort meaning, not that it never will. Where a pattern repeatedly produces misleading condensations we discard it, and where it holds up we note the conditions under which it does - short threads, factual content, clear authorship - rather than overclaiming general reliability.
- Key facts, commitments and caveats preserved, not just the gist.
- Source links so any summarised point can be checked in context.
- Checks for invented details, omissions and shifts in emphasis.
- Confidence signals surfaced so low-certainty summaries are obvious.
- No hidden messages - surfacing priority never means concealing the rest.
Where the human stays in control
The firm boundary in this lab is sending. Drafting and summarising can be AI-assisted, but the decision to send anything on a person’s behalf stays with that person. We prototype interfaces that make the assist obvious and the override effortless - suggested replies are clearly marked as drafts, triage decisions show their confidence and reasoning, and nothing is dispatched automatically. The goal is to save time on the mechanical parts while keeping judgement human.
This matters for accountability as well as quality. When a message carries professional or contractual weight, the responsible person needs to own its content, not rubber-stamp a machine’s guess. Our research therefore treats AI here strictly as decision-support, and any future tooling built on these patterns would carry the same constraint forward rather than relaxing it for convenience.
