Recall · hard
Review · easy
A sprint retro has 4 open-ended fields and takes about 10 minutes to fill thoughtfully. An NPS survey with comment boxes takes 5. A project intake form with scope questions takes 15. Multiply by team size and frequency and your organization burns hours per week on form-filling. HubSpot's research shows that surveys over 25 minutes lose three times as many respondents as those under five.
But the time isn't the real cost. HubSpot's research shows the dropout cliff hits at 7-8 minutes — and the responses submitted after that cliff are measurably worse. The respondent is still typing, but they've stopped thinking.
Two cognitive tasks, one textarea
Every open-ended field hides two distinct operations.
Retrieval: “What happened this sprint?” You open Slack, scroll through Linear, skim your PRs, try to reconstruct two weeks of context. This is the slow part. George Miller's classic 1956 paper established that working memory holds about 7 items — and you're asking someone to juggle dozens of events, conversations, and decisions into a coherent narrative.
Judgment: “Is this accurate? What would I emphasize?” This is fast. You read a summary, spot what's missing, fix what's wrong. Recognition, not generation.
These operations have radically different cognitive costs. Retrieval is effortful, error-prone, and degrades with fatigue. Judgment is lightweight, accurate, and stays sharp even at the end of a long day. Yet every form tool conflates them into a single blank textarea.
The mismatch
AI assistants are exceptional at retrieval. Given access to tickets, docs, messages, and code, they synthesize two weeks of work into a structured answer in seconds — with dates, names, and specifics. They're mediocre at judgment. They don't know which detail matters politically, what the subtext of a Slack thread was, or which win the team lead should hear about.
Humans are the opposite. Retrieval under time pressure produces “fine” and “nothing to change.” But give a human a draft of their own work to review, and their judgment is precise and fast. They catch the wrong date, add the context the AI missed, soften the phrasing on a sensitive point.
The 10-second fill works because it matches each party to their strength. The AI retrieves and drafts. The human reviews and approves. The form stops being a memory test and becomes an editorial review.
What this looks like in practice
A respondent gets a form URL in Slack. They hand it to Claude or ChatGPT. The agent reads the form's JSON spec — fields, types, probes — and drafts answers from the respondent's context: merged PRs, closed tickets, recent threads. The respondent reads the draft, changes one sentence, submits.
No browser extension. No plugin. The form itself has a native JSON interface that any AI can read and submit to. That's the architectural choice that makes this work — the agent doesn't need to scrape HTML or click buttons. It reads a structured spec and posts structured data.
The result isn't just faster fills. It's that the responses contain specifics: “Shipped the auth migration Tuesday, zero rollbacks” instead of “went fine.” Not because the respondent tried harder — because the form stopped asking them to do something humans are bad at, and started asking them to do something humans are good at.