Sample engagement · Customer support

Tier-1 deflection for a 4-person CSM team.

Anonymized profile of the kind of company we build for. Same shape an actual engagement takes — discovery, build, shadow, cutover — with the architecture, integrations, and target outcomes we’d commit to.

The client (anonymized · “Frame”)

ARR

$4M

B2B SaaS, dev-productivity tools

Headcount

32

NYC, remote-friendly

CSM team

4

covering 180 paying accounts

Stack: Zendesk for support, Notion for the knowledge base, Slack for routing, Postgres for product data, Hubspot for accounts. Sound like ~30% of the companies we talk to.

The bottleneck

Where 56 hours / month actually go.

14 hours per CSM, every month, on work the team would describe as “answering the same five questions.” That’s the spend we’d quote against.

62%

Tier-1 “how do I…”

Repeating questions whose answers already live in the docs and in past tickets. The agent reads the same questions over and over.

23%

Triage & routing

Decide whether it’s billing, engineering, or onboarding. Find the owner. Pass it on with context. Median 15 minutes per ticket.

15%

Duplicate root-cause

Multiple tickets with the same underlying issue, each one written-up from scratch. The triage doesn’t see the pattern.

Cost of inaction

56 hrs/mo × $80/hr loaded labor = $4,480/mo of toil.

Median first-response time sits at 6h 12m against a 4h SLA. Customer-impact on NRR isn’t measured — that’s the first thing discovery instruments.

What we’d ship

One classifier. Three branches. No black box.

Every decision the agent makes is logged with input, output, model version, and prompt version. Audit trail is the floor, not a feature.

Incoming ticket  (Zendesk webhook)
        │
        ▼
┌─────────────────────────────────┐
│  Classifier  (Claude Haiku)     │
│  fast · cheap · ~120ms p50      │
└──────────────┬──────────────────┘
               │
   ┌───────────┼────────────┐
   ▼           ▼            ▼
[ KB-answer ] [ Routing ] [ Escalation ]
        │         │            │
        ▼         ▼            ▼
  Drafts reply  Posts to    Hands to
  with cited    right Slack human with
  KB anchors    channel +    full context
  → CSM 1-click owner tag    window
  approve
        │
        └──────────────┐
                       ▼
              Postgres run-log
              (latency · accuracy
              · override rate)
  • Drafts, not auto-sendsOn cutover the agent writes the response. The CSM sends it. We turn on auto-execution only after metrics earn it.
  • Cited KB anchorsEvery drafted reply links to the Notion sections it pulled from. CSMs verify in seconds; customers see provenance if you choose to expose it.
  • No prompt-chain monolithsEach step is a discrete prompt with its own log. Debugging is reading three log entries, not unwinding a 4,000-token god-prompt.

Integration footprint

Five integrations. Forty-five minutes of setup on your side, total.

ToolWhat we needWho sets it up
ZendeskAPI key + ticket-update permissionsClient (~15 min)
Notion KBInternal integration token, scoped to KB workspaceClient (~10 min)
SlackBot token, scoped channelsClient (~10 min)
HubspotAPI key, read-only contacts/companiesClient (~10 min)
PostgresConnection string in dedicated schema (theirs or ours)Discussed at discovery

Engagement timeline

Four weeks, mapped.

If we hit end of week three and metrics don’t justify cutover, we run a kill review — not a rescue.

  1. Week 01

    Discovery session · $1,000

    • Day 1 — 3-hour live ride-along with one CSM, watch 5–7 tickets end-to-end
    • Day 2 — pull 200 anonymized tickets, score deflection-potential per category
    • Day 3 — write 1-page spec, build quote (≈$8k for this scope), ROI calc

    Spec + quote in your Google Docs.

  2. Week 02

    Build

    • Days 1–3: integrations (Zendesk, Notion, Slack), classifier prompt v0
    • Days 4–7: draft generator anchored on your real past replies as style references
    • Days 8–10: CSM approve UI (dead-simple Next.js), Postgres logging schema

    Agent runs on staging Zendesk, no outbound traffic yet.

  3. Week 03

    Shadow

    • Agent classifies & drafts for every incoming ticket, posts as internal note
    • CSMs work as usual; we measure agreement rate, edit-distance, miscategorizations
    • 2–3 prompt-update cycles based on real misses, calibrated against your data

    Baseline metrics. Prompt v2.

  4. Week 04

    Live

    • Cutover: agent drafts, CSM 1-click approves; auto-send stays off
    • 48-hour heightened-monitoring window after flip
    • Hand-off doc: how to debug, where logs live, on-call process

    Retainer kicks in only if 4-week metrics hit baseline.

Target outcomes (per discovery framework)

What we’d commit to measuring — and to honouring.

MetricBaselineAfter 4 wksAfter 12 wks
CSM hours / month on Tier-15632 (–43%)14 (–75%)
Median first-response6h 12m<90 min<30 min
Tier-1 deflection rate0%50–60%70%+
Misroute rate (wrong team)18%<8%<4%

These are targets per our discovery framework, not gold-standard results from delivered work. Different baselines, different attainable numbers — that’s what discovery measures.

Pricing — applied to this scope

First twelve weeks ≈ $9,800.

Compare to the bottleneck: 56 hrs × $80 × 12 weeks = $53,760 of toil over the same window. Pay-back inside the first 4–8 weeks if targets hit.

  1. Discovery call (2 hrs)Free
  2. Discovery session (8 hrs)$1,000
  3. Implementation (3 wk build + 1 wk shadow)≈$8,000
  4. Support — first 8 weeks ($100 / wk)$800

What we won’t do

The trust signals you can’t fake.

  • We won’t auto-send replies on cutover.

    That’s your decision after the metrics earn it — not our default.

  • We won’t build if you have <50 repeat tickets / month.

    ROI is too shallow. We’ll say so in discovery and you keep the spec.

  • We won’t use chained-prompt monoliths.

    Each step is its own prompt + log. Debugging is readable, not archaeological.

  • We won’t lock you into our stack.

    Code, prompts, logs are yours. Handover is a one-week doc, not a renegotiation.

Your version of this engagement

Bring one process you’d like to stop doing. The first call is free.