choir-2.4 · Tool use + escalation routing now in GA

Resolve 84% of tickets. Escalate the rest with full context.

Choir reads your knowledge base, your past tickets, and your billing system — then resolves customer issues end‑to‑end. Refunds, account changes, password flows. Median time to resolution: 28 seconds.

Try in playground → Read the paper
SOC2 Type 2·HIPAA·EU residency
LIVE · ticket #84,217
routing: billing.refunds
MR
Maya R. · customer · 11:42
Hi — my subscription was charged twice this month and I only have one account. Can you fix this?
choir · resolving
ch
choir · 11:42:21 · confidence 0.97
I see two charges on Nov 12 for $49.00 — both for the same Pro plan on account mrubio@…. The second is a duplicate triggered by a retry after a 3DS timeout. I'm refunding the duplicate to your card ending 4242 — settles in 3–5 business days. Per the duplicate-charge policy, no further review needed.
refund issued · $49.00 stripe.refunds.create policy:pol_812
Resolved 28.3s · 1,180 tok
CSAT survey sent
84.1%
auto‑resolution rate
28s
median time to resolution
$0.04
cost per resolution
4.6 / 5
CSAT, post‑resolution
Trusted at 1,408 companies — from Series A to Fortune 500
Linnea Forecast ▲ Vantage Brightwave Halcyon Northwind atrium Quartermast Linnea Forecast ▲ Vantage Brightwave Halcyon Northwind atrium Quartermast
§ BENCHMARKS

Better than the incumbents. By a wide enough margin to matter.

Evaluated on 12,408 anonymized production tickets across 23 SaaS companies. Methodology and per‑customer breakdowns in the paper.
Auto‑resolution rate
higher is better
  • Choir 2.4
    84.1%
  • Zendesk AI
    40.6%
  • Intercom Fin
    52.3%
  • GPT‑4o (raw)
    32.9%
Median resolution time
lower is better
  • Choir 2.4
    28s
  • Zendesk AI
    3m 41s
  • Tier‑1 human
    14m 02s
  • Email‑only
    9h 18m
CSAT (post‑resolution)
higher is better
  • Choir 2.4
    4.62
  • Human tier‑1
    4.21
  • Zendesk AI
    3.18
  • Intercom Fin
    3.71
Cost per resolved ticket
lower is better
  • Choir 2.4
    $0.04
  • Zendesk AI
    $0.31
  • Tier‑1 BPO
    $0.94
  • Tier‑1 in‑house
    $1.40
Eval set v3.2 · 12,408 tickets · 23 companies · published 02 Mar 2026
§ USE CASES

Five plays that handle the bulk of your inbox.

Each play is a deterministic policy + tool chain. You can edit, version, and test them in the Studio.

play.refunds
+38% vs Q3

Refund handling

Detects duplicate charges, applies your refund policy, calls Stripe directly, sends the confirmation email.

charge.duplicate detected
policy: pol_812 matched
stripe.refunds.create $49.00 ok
play.access.reset
2,841 / wk

Password & access

Verifies identity via signed email link, resets the password, removes the lockout state in Auth0.

identity verified · auth0
lock state cleared
reset link issued ok
play.kb.qa
grounded

Product Q&A

Pulls from your docs, release notes, and last 90 days of resolved tickets — always citing the source.

"You can enable SSO in Settings → Security → SAML. Note: requires the Team plan." — docs.v7.2 #144
play.account.change
conf ≥ 0.95

Account changes

Plan switches, seat additions, email changes, billing address — gated by your approval rules.

plan: Pro → Team · seats +4
proration: $148.21 · invoice scheduled
applied · admin notified
play.escalate
15.9% of total

Escalation, with full context

When confidence drops, Choir summarizes everything it knows and hands off to the right human — never cold.

routed to @billing-on-call · attached: timeline, 4 prior tickets, payment trail.
play.outage.triage
live

Outage triage

During incidents, Choir auto‑detects spikes by region or feature, mutes noise, and posts a clear holding response.

incident us‑east · auth · −38% vs Tue
tickets +412 in 6 min · grouped
§ MODEL CARD

choir‑2.4 · production

A fine‑tuned reasoning model with first‑class support for tool calls against your billing, identity, and helpdesk systems.

context window200,000 tokens
reasoningmulti‑step · planner + executor
tool usestripe · auth0 · zendesk · intercom · front
training corpusenterprise support · 4.2B tickets
latency P50 / P95820ms / 1.18s
pricing$0.04 per resolution
residencyUS · EU · APAC
complianceSOC2 Type 2 · HIPAA · GDPR
§ SDK

A single function call.

Stream the resolution, await the result, or both. Tool calls are executed inside Choir against your linked accounts — return values stream back as structured events.

choir‑sdk · 1.4.2
from choir import Choir

client = Choir(api_key="sk_live_…")

resolution = client.resolve(
    ticket_id="tkt_84217",
    channel="email",
    policy="billing.refunds",
    tools=["stripe", "auth0"],
    stream=True,
)

for event in resolution:
    # tool_call / token / done
    print(event.type, event.data)
§ PRICING

Pay only for what gets resolved.

No seat minimums. No SDR call required to see the live‑pricing dashboard. Volume committed pricing kicks in past 100k resolutions/mo.
Sandbox
Free
100 resolutions / month
  • Playground + Studio access
  • All policies, all tools
  • 1 environment
  • No SSO, no SLA
Start free
Usage
most teams
$0.04 / resolution
no seat fees · billed monthly
  • Everything in Sandbox
  • Unlimited seats + 3 environments
  • Zendesk · Intercom · Front integrations
  • 99.9% SLA + email support
  • Daily eval reports
Start with $200 credit
Enterprise
Talk to us
volume + BYOC + custom training
  • Bring your own cloud (AWS, GCP, Azure)
  • Fine‑tuning on your corpus
  • EU / APAC data residency
  • 99.99% SLA · 24×7 support
  • Dedicated solutions engineer
Contact sales
Customer story · Brightwave (Series C, ~28,400 tickets/mo)

“We replaced a 14‑person tier‑1 team in 9 weeks. The 84% number turned out to be conservative — we hit 86.4% in week three and the team finally got to work on retention.”

IC
Ines Calderón
Head of Engineering, Brightwave
−72%
tier‑1 ticket cost, month‑over‑month
+18%
net retention (cohort Q1)
28h
saved per agent per week
9 wk
go‑live from kickoff
§ COMPARISON

Where Choir lands vs the alternatives.

capability
Choir
Zendesk AI
Intercom Fin
Open‑source agent
End‑to‑end resolution (incl. refunds)partialpartialDIY
Tool calls against billing & identityDIY
Confidence‑gated escalation with contextbasicbasic
PII redaction before model invocationadd‑onadd‑onDIY
Bring‑your‑own‑cloud deployment
Fine‑tuning on your corpuslimited
SOC2 Type 2 + HIPAADIY
§ FAQ

For the buyer who actually reads the docs.

How accurate is the auto‑resolution, and what's the false‑confidence rate? +
On the public eval set, Choir's correctness rate (resolutions that the customer accepted and didn't reopen within 14 days) is 96.8% when confidence ≥ 0.95. The false‑confidence rate (Choir confident, customer disputes) is 1.2%. We publish a daily report per customer with these numbers, broken down by policy.
Does it integrate with Zendesk, Intercom, and Front? +
Yes, natively. Choir sits in your existing helpdesk as an AI agent and writes back the same fields your humans do — labels, internal notes, status. Median time from connection to first auto‑resolution is 41 minutes.
What happens when it can't resolve — how is escalation context handled? +
When confidence drops below the threshold you set per policy (default 0.85), Choir produces a structured handoff: a one‑paragraph summary, the timeline of what it tried, the relevant docs/policies it considered, and any tool calls it almost made. The escalated agent sees this as a pinned internal note. No customer message is sent without your sign‑off in escalation mode.
What about PII in tickets — is it redacted before model invocation? +
Yes. PII redaction runs in your VPC (or our redaction edge in your region) before any payload reaches the model. Names, emails, card numbers, addresses, and any custom regex patterns you specify are replaced with stable tokens that re‑hydrate at delivery time. We never store raw PII in logs.
Can we fine‑tune on our specific product corpus? +
On Enterprise. We continuously fine‑tune a private adapter on your resolved tickets, docs, and release notes — refreshed weekly. Most customers see a 3–6 point lift in auto‑resolution rate within the first month, particularly on product‑specific Q&A.
live playground

Drop in a ticket. Watch it resolve.

Try Choir on a real ticket — yours or ours — in the browser. No credit card. The playground uses the same model that ships to production.