Methodology v1.0-2026-04 · Last updated 2026-04-27

How we project customer-support AI savings.

Every benchmark, every channel modifier, the hidden-cost ranges, the deflection curve, the build-vs-buy formula, and the CI-enforced distribution-bias gate are documented below. Same answers, same projection, every time.

Run the calculator Jump to benchmarks

Volume + cost model

The baseline annual labour cost is the dollar volume the calculator works against. It captures only the agent time directly applied to ticket resolution, not management or floor overhead - those live in the loaded hourly cost.

annual_labour_cost = monthly_tickets × 12 × (AHT_min / 60) × agent_cost_per_hour
weighted_modifier  = Σ (channel_share_i × channel_modifier_i)
deflection         = Σ (complexity_share_i × deflection_benchmark_i) ± 0.15
gross_savings      = labour × deflection × 0.85_residual × weighted_modifier
hidden_costs       = token_pct × gross + integration_one_time + training_one_time + qa_pct × labour
net_savings        = gross_savings − hidden_costs

The 0.85 residual multiplier is a uniform haircut for residual escalation, AI runtime overhead, and supervision drag. It is intentionally conservative - we would rather under-promise than over-promise.

Deflection curve assumptions

Realistic AI rollouts follow an S-curve: ~20% of full deflection in month 1, ~85-90% by month 6, asymptotic near-target by month 12. Anything that promises full deflection from day one is over-promise; the calculator models the steady-state target after month 6.

Per-complexity deflection rates are pulled from the per-industry benchmark table (next section). Simple = FAQ / lookups / status. Medium = bookable steps such as cancellations, address changes, or refunds. Complex = multi-system or judgement calls humans still own.

Build-vs-buy formula (TCO model)

The recommendation is a deterministic score. The engine evaluates seven dimensions - volume, complexity, languages, regulated industry, IP sensitivity, integration depth, and operating hours - and rolls them into a SaaS score and a custom score.

if scoreSaas >= scoreCustom + 1     -> saas
if scoreCustom >= scoreSaas + 1     -> custom
otherwise (within 1 point)          -> hybrid

3-year TCO comparison (display-only):
saas_3y    ≈ savings × 0.35 × 3 + 60_000   (license + setup)
custom_3y  ≈ savings × 0.18 × 3 + 220_000  (build + ops + ongoing eng)
hybrid_3y  ≈ savings × 0.28 × 3 + 140_000  (mid-blend)

Source: app/lib/tools/cs-savings/build-vs-buy.js. Every threshold is editable. The recommendation never invokes an LLM; same inputs, same recommendation, every time.

Hidden costs taxonomy

Four buckets. Together they typically subtract 15-25% from gross savings - vendor pitches almost never include them, so the calculator subtracts them before showing net.

  • Token / API consumption10–15% of gross savings

    Annual, ongoing

    Frontier-model token pricing has dropped 70%+ in 18 months but volume scales with usage.

  • Helpdesk integration$25,000–$35,000

    One-time

    Connectors to Zendesk / Salesforce / Freshdesk + auth. SaaS and custom both pay this.

  • Training data prep$12,000–$18,000

    One-time

    Knowledge-base curation, intent labelling, evaluation set construction.

  • QA + supervision6–10% of agent cost

    Annual, ongoing

    Someone reviews escalations, flags drift, retrains the model. Required for production.

30-industry table

Per-industry deflection benchmarks.

Aggregated from Salesforce State of Service, Gartner CIO Survey, McKinsey State of AI, Crescendo CX Benchmarks, and Zendesk CX Trends. The full citable Dataset is available at the deflection-rate benchmarks page (CC BY 4.0).

IndustrySimpleMediumComplexSources
E-commerce85%55%10%Salesforce, Zendesk, Crescendo
SaaS80%50%10%Salesforce, Gartner:
Retail75%45%8%Salesforce, Zendesk
Banking70%35%5%Gartner:, McKinsey:
Insurance65%30%5%Gartner:, McKinsey:
Healthcare55%25%3%Gartner:, McKinsey:
Legal40%15%2%Gartner:, McKinsey:
Telecom78%45%8%Salesforce, McKinsey:
Travel & Hospitality72%40%8%Zendesk, McKinsey:
Airlines70%35%5%Zendesk, Gartner:
Restaurants85%50%5%Zendesk, Crescendo
Real Estate60%30%5%Gartner:, Crescendo
Automotive65%32%5%McKinsey:, Gartner:
Media & Entertainment78%45%8%Zendesk, Salesforce
Logistics75%40%8%McKinsey:, Gartner:
Manufacturing55%25%5%Gartner:, McKinsey:
Utilities70%35%5%Gartner:, McKinsey:
Energy65%30%5%Gartner:, McKinsey:
Education65%30%5%Zendesk, Gartner:
Government & Public Sector50%20%3%Gartner:, McKinsey:
Nonprofit60%28%5%Zendesk, Gartner:
Professional Services55%25%5%Gartner:, McKinsey:
Financial Services68%32%5%Gartner:, McKinsey:
FinTech78%45%8%Salesforce, Gartner:
Cybersecurity55%25%5%Gartner:, McKinsey:
Gaming82%50%8%Zendesk, Crescendo
Marketplaces80%48%8%Salesforce, Zendesk
B2B Services60%28%5%Gartner:, McKinsey:
Consumer Electronics78%45%8%Zendesk, Salesforce
Agriculture55%25%5%Gartner:, McKinsey:

Channel modifiers

Each channel has a different deflection ceiling.

The base industry rate is multiplied by the channel modifier weighted by your channel mix. Chat deflects best; voice is the hardest because of ASR error rates and intent ambiguity.

  • Chat1.00

    Highest deflection - conversational AI excels in text chat. The default for tier-1 deployments.

  • Email0.85

    Asynchronous, slower - AI summarisation helps but full deflection rarer. Often handled by hybrid agent + AI.

  • Voice0.65

    ASR error rates and intent ambiguity cap voice deflection. Improving rapidly but not yet at chat parity.

  • Social0.90

    Mostly tier-1 question routing. High deflection but lower volume in most industries.

Hidden cost reference

The four buckets we subtract from gross.

  • Token / API consumption

    Annual, ongoing

    10–15%

    of gross savings

  • Helpdesk integration

    One-time

    $25,000–$35,000

    per deployment

  • Training data prep

    One-time

    $12,000–$18,000

    per deployment

  • QA + supervision

    Annual, ongoing

    6–10%

    of agent cost

Honesty floor

buzzi.ai builds custom. The calculator still recommends SaaS.

Our commercial interest is in custom builds. To prove the recommendation engine isn't biased toward our own pocket, we run a CI test on every change: 50 synthetic inputs, SaaS must be the recommendation in at least 40% of them. If the gate fails, the build fails.

  • ≥ 40%

    SaaS recommendation floor

    CI-enforced across 50 synthetic inputs.

  • 74%

    Current SaaS share

    Distribution as of April 2026.

  • distribution-bias.test.js

    Audit test

    __tests__/tools/cs-savings/

Versioning

Methodology v1.0-2026-04 - April 2026.

Refresh cadence is quarterly. Token cost percentages, helpdesk integration ranges, and per-industry deflection benchmarks all get re-validated against the public reports and re-published with a numbered version + change log. Each release re-runs the SaaS≥40% distribution gate.

Last updated: 2026-04-27. Next scheduled refresh: July 2026.

Citations

Where the numbers come from.

  • · Salesforce State of Service 2025-2026
  • · Gartner: AI in customer service 2025
  • · McKinsey: Next frontier of customer engagement 2026
  • · Crescendo CX Benchmarks 2025-2026
  • · Zendesk CX Trends 2026
  • · BLS, ONS, ABS, Eurostat (loaded agent cost defaults by country)

Ready?

Project your savings in 60 seconds.

Ten questions. Free preview. The methodology you just read - applied to your numbers.

Run the calculator