Sierra's Console Answers 'Why Did Resolution Drop?' Ask It Which Resolution.

Sierra rebuilt its agent console around a chat this week. This is a rewrite of the question the console opens with, from someone who priced the word it depends on.

The short version

Sachi Shah's November 2 post says Sierra redesigned Agent Studio around Ghostwriter so a builder can go from asking why resolution rates dropped to examining the conversations, updating a journey, and testing the fix in one conversation, with the evidence, the changes, and the test results kept together in a session until someone merges them. The first example question has already accepted a definition of resolved, and the definition is where the number moves. The Gorgias benchmark, as Jason Lemkin reported it, counts zero human touch and puts the top five vendors at 64% to 75% with a median of 48%, while Gorgias's own customer reporting counts a resolution after 72 hours without a human. My own per-resolved pricing on a 48-hour definition held for 98% of cases and the other 2% became more than 200 contested cases a week. The rewrite: the first reply to a metric question in an agent console is the definition line, three rates and three deltas before any cause, and a proposed journey change is simulated under all three.

What Sierra shipped

Shah's post is clear about the shape, which is why it is worth arguing with. Ghostwriter is Sierra's agent for building agents. Until now, using it meant moving between the parts of Agent Studio that edit journeys, run simulations, and review performance. Now Ghostwriter stays with you across the studio and sees what you are looking at, a journey, a simulation, or a dashboard, so you can ask about the thing on screen or have it fixed on the spot.

Each session works like a chat thread and keeps the conversation together with the changes it makes. Changes stay in the session until someone reviews and merges them. You can share a session, fork one, or ask Ghostwriter to review a teammate's changes. Related agents sit in one project with the data and channels they share, so a lender's document-collection agent and its application agent get updated together in one session. Before anything ships: line-by-line comparisons, Preview Chat, Simulations, and release governance with required approvals and controlled rollouts.

I wrote about Sierra's six operating choices in The Sierra Playbook: What They Actually Do Differently, and this is consistent with them: price the outcome, own the last mile, treat annotated conversations as regression tests. The session model is good. The line-by-line diff is the thing I asked for in Sierra's Voice Agents Fix a Third of the Details They Catch. Read Back the Diff., applied to the builder instead of the caller. None of that is the subject here.

The subject is the first sentence of the post's own example. "Go from asking why resolution rates dropped."

The word the question depends on

"Why did resolution drop" has already picked a definition of resolved. I know because I priced it, and One Support Agent, Three Resolution Rates is the long version.

The Gorgias benchmark, as Lemkin reported it on October 6, runs 13 vendors on 212 live ecommerce stores and counts a conversation as resolved only with zero human touch, no handover, and no push out of the channel. The top five vendors land at 64% to 75%. The median is 48%. Sierra, in that benchmark, resolves 48% and ties for the top quality score.

Then the paragraph Lemkin wrote that I'd tape to the wall. Gorgias's own customer-facing reporting counts a resolution once the AI resolves it and 72 hours pass without a human agent helping. A customer who gave up and never came back is a resolution under the second rule and a failure under the first.

Mine was a third rule. We priced per resolved ticket. Resolved meant no escalation within 48 hours, in writing. It held for 98% of cases, pilot included. The other 2% were customers who came back after 60 to 72 hours with the same problem, which we had already counted and billed as resolved. At sunset volume that was more than 200 contested cases a week and 600 disputes a month against a projection of 50. At month 20 we rewrote the definition to no follow-up in seven days. It cost revenue and cut disputes by 60%.

Same agent. Same month. Three rates. The number everyone argued about was the definition.

What the chat changes

On a dashboard, the definition sat somewhere you could find it. The column header. The footnote under the chart. The sentence in the contract. You could put your finger on it before the renewal call.

In a chat, the definition sits wherever the agent picked it up. Ghostwriter sees the dashboard you are looking at, which means it sees the number the dashboard chose to show, under the rule the dashboard chose. When you ask why that number dropped, the answer comes back in a sentence, and "resolution dropped because the returns journey started escalating" can be true under the 72-hour count and false under zero touch. Nothing in the sentence says which.

That is not a Sierra problem. It is what happens to any metric the moment the console answers in prose. The design handbook's Disclosure Without Overwhelm has the frame: showing everything makes the user responsible for checking work they never asked to supervise, and the test of a disclosure is whether it changes what a reasonable reader does next. The definition line passes that test. A warehouse of conversations does not.

The rewrite

"Ask the console why resolution dropped" becomes "the console answers with which resolution, under all three definitions, before it names a cause."

One rule, for any agent console that answers in sentences. The first reply to a metric question is the definition line. The rate under zero human touch. The rate under the invoice definition, exactly as the contract words it, quiet window and all. The rate under no related follow-up in seven days, on any channel. The delta under each since the period you asked about. Then, and only then, the cause.

And the second half, for the part of Sierra's post I like most. A session that proposes a journey change shows the simulated before and after under all three. A fix that moves the 72-hour number and leaves zero touch flat has fixed the timeout, not the agent. A fix that moves zero touch and leaves the seven-day number flat has taught the agent to close conversations faster, and you will meet those customers again in column three.

This is the same discipline I argued for in 86% Engaged, 20% Resolved: Kill the Engagement Rate Headline: lead with resolution under a stated definition, and print the hand-off count beside it. The chat-shaped console makes it more urgent, not less, because the headline now arrives as a sentence that sounds like a cause.

Under Enterprise AI Agents the claim is that what matters is how many agents still complete production work after ninety days and what each successful outcome costs. Nine of ten vendors in the Gorgias benchmark lost automation within four weeks. The console that catches that is the one that answers "which resolution" before "why."

This week: take 30 conversations your agent closed last month, weighted to your real ticket mix, and score each three ways. Then ask your vendor's console why resolution dropped and see which of the three columns its answer matches. If it matches one and does not say which, you have found the definition it picked for you.

Related answer: What should an agent console answer first when resolution rate drops?

Sources: Sachi Shah, "Redesigning Agent Studio around Ghostwriter," Sierra blog, November 2, 2026, read in full. Jason Lemkin, "How Much of Your Customer Support Can AI Really Resolve? The Best Get About 70%. The Median Is 48%. Here's the Real Data From 13 Vendors," SaaStr, October 6, 2026, reporting the public benchmark run by Gorgias; the benchmark figures are as he reports them. Falk Gottlob, "One Support Agent, Three Resolution Rates" (October 7, 2026), "86% Engaged, 20% Resolved: Kill the Engagement Rate Headline" (October 15, 2026), "Read Back the Diff" (October 29, 2026), "The Sierra Playbook: What They Actually Do Differently," and "Disclosure Without Overwhelm" in the Design Operating Model.