The wrong AI customer support platform locks you in for 2-3 years and undermines your CX strategy. The right one drives 50-80% deflection on tier-1 tickets and reshapes your support economics. Here's the 9-step framework we use for serious enterprise evaluations.
The framework
Step 1: Define your deflection ceiling
What's the realistic max deflection rate for your ticket mix?
- Mostly informational + FAQ tier-1: 70-85% ceiling
- Mixed informational + light transactional (status checks, account info): 55-75% ceiling
- Heavy transactional (refunds, exchanges, account changes): 40-60% ceiling
- Complex multi-step + regulated: 25-45% ceiling
This number anchors the rest of the evaluation. A platform that promises 80% on heavy-transactional work is overpromising.
Step 2: Lock the integration must-haves
List the integrations you need day-one:
- Ticketing: Zendesk, Intercom, Salesforce Service, Freshdesk
- CRM: Salesforce, HubSpot
- Knowledge base: Confluence, Notion, Helpscout, internal wiki
- Commerce: Shopify, BigCommerce, Magento, custom OMS
- Identity / auth: Okta, Auth0, custom
- Payment / refund: Stripe, PayPal, Adyen, custom
If a platform doesn't have a mature integration for your top 3, it's a non-starter — regardless of how impressive the AI is.
Step 3: Pricing model fit
Three main pricing models in 2026:
- Outcome-based (per resolution): Sierra, Decagon, Forethought. Better aligned incentives; harder to forecast.
- Bundled in ticketing platform: Intercom Fin, Zendesk AI, Service Agentforce. Easier procurement; usually lower deflection.
- Per-conversation + platform fee: Ada, some others. Hybrid model.
Pick the model that aligns with your CX strategy:
- "I want vendor incentives aligned to deflection" → outcome-based
- "I want predictable cost + simple procurement" → bundled
- "I'm somewhere in between" → hybrid
Step 4: Build the shortlist (3-4 vendors)
Standard 2026 shortlist for serious mid-market + enterprise evaluations:
- Sierra — modern AI-native leader
- Decagon — modern challenger, often better mid-market pricing
- Your stack-aligned option: Intercom Fin, Service Agentforce, or Zendesk AI
- One wildcard: Ada, Forethought, or vertical specialist if your industry has one (e.g., Hippocratic AI for healthcare)
Step 5: Run the RFP
Standard RFP sections:
- Company + product overview
- Reference deployments (3-5 in your industry + scale)
- Pricing model + 12/24/36 month TCO projections
- Integration depth for your must-have stack
- Security + compliance (SOC 2, ISO 27001, GDPR, vertical-specific)
- Implementation timeline + resources
- SLA + support model
- Data ownership + portability terms
Don't accept vague answers. "We integrate with everything" isn't a yes — make them name your specific stack components.
Step 6: Pilot 2-3 vendors in parallel
The most-important step. Don't pick from RFP responses alone.
- Duration: 4-12 weeks per pilot. Longer is better — deflection rates rise over the first 90 days as the platform learns your patterns.
- Scope: Same workflow tier across all vendors, same volume, same time period.
- Measurement: Deflection rate, customer satisfaction (CSAT) on AI-handled vs human-handled, time-to-resolution, escalation reasons, false-positive rate on auto-resolves.
Step 7: Calculate true TCO at projected volume
Vendor pricing pages lie. Build the real TCO:
- License + per-resolution fees (or bundled pricing)
- Implementation cost (one-time or amortized)
- Internal staffing to operate the platform (typically 1-3 FTE equivalents)
- Ongoing optimization labor (prompt tuning, knowledge curation, workflow refinement)
- Change management for your human support team
True TCO is typically 1.3-1.8× the vendor list price. Forecast accordingly.
Step 8: Negotiate
The list prices are negotiable:
- 15-30% discounts at volume commitments
- Implementation cost concessions
- Performance-based pricing terms (lower per-resolution rate tied to deflection milestones)
- Multi-year commitments for additional discount
Don't sign first-quote. Standard enterprise procurement leverage applies.
Step 9: Plan the rollout
The platform won't deflect 60% out of the box:
- Weeks 1-4: Implementation + integration wiring. Limited live traffic.
- Weeks 5-12: Soft launch — 10-30% of tier-1 traffic, heavy human review of outputs.
- Weeks 13-26: Scale up — 50-80% of tier-1 traffic, continuous optimization.
- Months 7-12: Maturity — full deflection rate, optimization stabilizes.
Resource the change management. The platform works; the human support team adaptation needs investment.
Common patterns by company stage
Series-B SaaS, 5-15 person CX team:
- Sierra or Decagon (greenfield, no legacy migration)
- Budget: $200-500K/year all-in
- Timeline: 6-9 months end-to-end
Mid-market e-commerce, 30-100 person CX team:
- Sierra, Decagon, or Ada (depending on migration story)
- Budget: $500K-2M/year all-in
- Timeline: 9-15 months end-to-end
F500 enterprise, 200+ person CX team:
- Sierra or Decagon for new use cases, Ada for legacy chatbot migrations, plus stack-aligned option (Service Agentforce for Salesforce shops)
- Budget: $2-10M/year all-in
- Timeline: 12-26 months end-to-end
Decision flowcharts
I'm on Intercom + can't migrate: Intercom Fin is the natural pick. Pilot it; if deflection lands above 50%, you're done.
I'm on Salesforce Service Cloud: Service Agentforce + Sierra side-by-side pilot. Pick by deflection × procurement comfort.
I'm greenfield + want max deflection: Sierra vs Decagon head-to-head. Pilot both for 6-12 weeks.
I'm migrating from legacy chatbot + procurement comfort matters: Ada vs Sierra. Ada wins if migration risk is the top concern; Sierra wins if deflection ceiling is.
What we'd skip
- Picking from demos. Demos are sales-team-designed; real workflow performance is different.
- Skipping the pilot. The wrong platform locks you in for 2-3 years; the pilot is cheap insurance.
- Buying the cheapest option to "save money." The wrong platform burns more in opportunity cost than the right one costs in fees.
- Underinvesting in change management. Tools don't auto-deploy themselves successfully.
Bottom line
AI customer support platform selection in 2026 is a 12-26 week structured process. Define your ceiling, lock your integrations, build a 3-4 vendor shortlist, run a real pilot, calculate true TCO, negotiate, plan the rollout. Don't shortcut — the platform you pick will shape your CX economics for the next 3 years. The right answer for most modern deployments is Sierra or Decagon. The right answer for many in-stack-bound deployments is the platform-native AI. Both can be defensible; the framework above tells you which fits your situation.
Sierra vs Decagon vs Intercom Fin → · Best AI for customer support → · How to evaluate AI agent →