Design for a model that might be wrong · States

The drafted reply

Place a drafted reply on the agent's screen, decide whether it ever auto-sends, and name every state it needs, with the error rate as a daily count.

designmediumExercise · 4 min · 39 of 59

On this page

The design lead wants a spinner and a disclaimer; the FDE wants the easy tickets to send themselves. Both proposals skip the state chart, and this exercise is where it gets drawn, with the numbers that decide it.

The situation

The model drafts replies for one tenant's 1,200 support agents, a tenth of the 12,000 seats. Each agent handles 40 tickets a day. Assume its queue sits wholly inside the intents the assistant covers, so every ticket gets a draft (coverage 1.0, against the North Star's 0.60 average): 48,000 drafts a day, about a quarter of the 200,000 requests a day across all tenants (the bill in tokens as a cost model). On the 200-example golden set, 92% of drafts are acceptable; of the 8% that are not, half a point is wrong on facts (a policy, a price, a date: one draft, the most the launch gate's 0.5% allows) and 7.5 points on tone. Latency: first token at about 400 ms, p50 2.1 s to the last token, p95 6 s. Agents currently send a reply every 2 minutes, and the team's guardrail is that reply time must not rise. The tenant's contract has a clause on factual errors in customer-facing text. The design lead has proposed a spinner and a disclaimer; the FDE has asked whether drafts can auto-send for the simplest 30% of tickets.

The question

Where does the draft appear, does it ever auto-send, and which states does the screen need? Decide, with the numbers.

A strong answer · Rubric · Where answers go wrong

Write your answer first, then open

A strong answer

At 8% wrong, auto-send means about 3,840 bad replies a day, 240 of them factually wrong. Never at this rate, and not for the simplest 30% either: that slice is 14,400 drafts a day, the golden set does not score it on its own, and even the overall half point, read off one failing draft, would send about 72 wrong facts a day into a contract clause. The threshold to revisit is a factual-error rate under 0.5% on a golden set stratified by ticket type, with at least 600 clean examples on the slice, and a measured undo rate.

The draft streams into a suggestion area above the composer, inserted with one action. First token at about 400 ms keeps the agent inside Nielsen's 1 s limit, so flow is not broken even though the full draft lands at the 2.1 s p50; nothing here is instant, which would be 0.1 s.

States, with what the agent can do: loading (skeleton under 1 s; keep typing); streaming (Stop); done (Insert); wrong (edit inline, regenerate, report, and the report feeds the eval set); empty and refused (a one-line reason; write the reply); error (one silent retry, then a reason); partial (marked; continue or discard); slow (past 4 s, Cancel; the agent types over it); over budget (the reason and reset date; write the reply).

Metrics: share of drafts sent unchanged, and edit distance before send. Guardrail: reply time must not rise above 2 minutes.

Drop the disclaimer: it moves liability, not the 240 wrong facts a day, and the contract clause is about the error, not the notice. What would change the auto-send answer: shadow-mode data on the slice with no changed number, date or price in at least 600 checked drafts.

Rubric

Turns the error rate into a daily count: 3,840 wrong drafts, 240 on facts.
Makes the auto-send call and names the threshold that changes it, including that the slice must be measured on its own.
Lists every state the screen needs, including refused, slow and over budget, each with what the agent can do in it.
Uses streaming against perceived latency, with the first-token time placed against the 1 s limit.
Names one metric and one guardrail, with the guardrail's number.
Drops the disclaimer and says why: it changes the notice, not the error rate.

Where answers go wrong

"Add a spinner and a disclaimer that AI can make mistakes." A disclaimer moves liability, not error; there is no state for wrong, and no number anywhere. The reviewer's note: the answer designed the done state and labelled it.
"Auto-send the simplest 30%; the model is 92% right." The 92% is measured across the whole set, not on the slice, and 200 examples cannot show any slice under 0.5%. Even at the overall half-point factual rate, 14,400 auto-sent drafts a day put about 72 wrong facts and 1,080 off-tone replies in front of customers, with no agent to catch them.
"Show the finished draft when it is ready, so the agent is not distracted by text appearing." At the p95 that is 6 s of blank panel, past the 1 s limit on at least 2,400 drafts a day, and agents start typing their own reply before it lands. Without a slow state the late draft has nowhere to go but on top of their text.