Sierra Agent SDK vs Decagon 2026: CX Bots That Bill Per Fix

Two of the biggest names in AI customer support just stopped selling seats — and every business owner with a support team should recalculate their budget. The Sierra vs Decagon fight this month isn’t about who has the smarter model. It’s about who will bill you only when a ticket actually gets resolved. Sierra, Bret Taylor’s CX agent company, and Decagon, its most aggressive rival, both shipped new agent-builder tooling alongside outcome-based pricing tiers where the line item is a fixed price per resolution, not a monthly license per human. With enterprise genAI spend under a microscope after the Pentagon and EU platform news cycle, procurement teams are re-bidding CX contracts on resolution rates instead of license counts. Vendors who can’t quote a per-fix number are getting cut from the shortlist.

What’s new in the Sierra vs Decagon race

The headline change is commercial, not technical. Both companies now lead with outcome-based AI pricing: you’re quoted a rate per resolved conversation — typically low single-digit dollars for simple deflections, higher for multi-step transactional fixes like a refund, an address change, or a subscription downgrade. If the agent hands off to a human, most contracts say you don’t pay for that conversation. That is a fundamentally different risk profile than the Zendesk-style per-agent-per-month seat license, where you pay the same whether the software solves anything or not.

Alongside the pricing, both shipped developer surfaces that matter more than the pricing page. The Sierra AI agent SDK pushes the “agent as code” model: you define your agent’s behavior in a declarative config, wire tools to your actual backend APIs, and run a simulation suite against recorded transcripts before anything reaches a customer. Decagon’s AI support platform leans into admin-editable operating procedures — support leads write the escalation logic in structured natural language, and the platform compiles that into enforceable guardrails plus a QA layer that grades every conversation. Both directions converge on the same insight: an agent you can’t test, version, and audit is an agent you can’t put on outcome-based billing, because neither side would trust the invoice.

The second change is measurement. Once you’re paying per fix, “resolution” becomes a contract term that has to be defined, instrumented, and disputed. Both vendors now expose AI agent resolution rate telemetry — resolved vs. deflected vs. escalated, plus customer satisfaction attached to each — because the number is now the invoice. That shifts leverage toward buyers, and it’s why this news cycle matters to a ten-person company and not just the Fortune 500.

Why it matters

• Support cost becomes variable, not fixed. Seat licenses are a bet you make in January. Per-resolution pricing scales with actual volume, which is enormously better for seasonal businesses, product launches, and anyone whose ticket volume spikes unpredictably.
• The vendor now shares your downside. If the agent can’t handle your weird edge cases, the vendor eats the cost of that failure instead of collecting a license fee for shelfware. That aligns incentives in a way SaaS support tooling never has.
• Your knowledge base becomes the actual product. Resolution rate is mostly a function of whether the answer exists in a machine-readable form. Companies with clean docs, clear refund policies, and real API access to order data will hit 60–70% autonomous resolution. Companies with tribal knowledge in a Slack channel will hit 20% and blame the model.
• Headcount math changes overnight. If an agent resolves at $2 per ticket and your loaded cost per human-handled ticket is $6–$12, the arithmetic is brutal and obvious. The realistic near-term outcome for most small teams isn’t firing people — it’s not hiring the next two reps while volume doubles.
• “Replace Zendesk with AI agents” is now a real line item, not a pitch. But it’s usually wrong as stated: most buyers keep the ticketing system of record and put the agent in front of it. The seat count shrinks; the platform stays.
• Contract risk moves to definitions. Who decides a ticket was “resolved”? What happens when the customer reopens it 48 hours later? Under outcome-based AI pricing, that clause is worth more than the headline rate.

How to use it today: piloting AI customer support agents in 2026

1. Pull your last 90 days of tickets and cluster them. You cannot evaluate a per-resolution quote without knowing your own mix. Export from your current helpdesk and count how many tickets fall into your top 10 intents.

   curl -s -G "https://yourcompany.zendesk.com/api/v2/search.json" \
     --data-urlencode "query=type:ticket created>2026-06-01" \
     -u "$ZD_EMAIL/token:$ZD_API_TOKEN" \
     | jq -r '.results[] | [.id, .subject, (.tags | join("|"))] | @tsv' \
     > tickets_90d.tsv
   
   cut -f3 tickets_90d.tsv | tr '|' '\n' | sort | uniq -c | sort -rn | head -20

   If your top 10 intents cover 70%+ of volume, you are an excellent candidate. If your volume is a long tail of bespoke problems, autonomous resolution will underperform and per-fix pricing will look expensive relative to the deflection you get.

2. Compute your true cost per human-handled ticket. This is your negotiating floor. Loaded salary (salary × 1.3 for benefits and overhead) divided by tickets actually closed per rep per year.

   Loaded cost per rep/yr:      $52,000 × 1.3 = $67,600
   Tickets closed per rep/yr:   18 per day × 235 days = 4,230
   Cost per human ticket:       $67,600 / 4,230 = $15.98
   
   Vendor quote:                $2.20 per resolved ticket
   Assumed resolution rate:     55%
   Blended cost per ticket:     (0.55 × $2.20) + (0.45 × $15.98) = $8.40

   Run this before you take a demo. Walking in with a blended-cost model is the difference between negotiating and being sold to.

3. Define your agent as code, not as a chat prompt. Both platforms want structured behavior definitions. Here’s the shape of a Sierra AI agent SDK–style config — the pattern generalizes across vendors.

   agent:
     name: order-support
     persona: |
       You are a support agent for Northgate Supply. Direct, warm, never
       apologetic more than once. Never invent policy. Never promise a
       refund timeline you cannot verify from the tools available.
   
     tools:
       - name: lookup_order
         endpoint: https://api.northgate.com/v1/orders/{order_id}
         auth: bearer:${NORTHGATE_API_KEY}
       - name: issue_refund
         endpoint: https://api.northgate.com/v1/refunds
         method: POST
         requires_confirmation: true
         max_amount_usd: 150
   
     procedures:
       - id: refund_request
         when: customer requests a refund
         steps:
           - verify order exists via lookup_order
           - if order age > 30 days, escalate to human
           - if amount > 150, escalate to human
           - otherwise call issue_refund and confirm in writing
   
     escalation:
       always_escalate_on:
         - legal threat
         - chargeback mention
         - three consecutive failed resolution attempts

4. Build an evaluation set before you launch. Take 200 real resolved tickets, strip PII, and use them as a regression suite. This is the single highest-leverage thing you can do, and almost nobody does it.

   {
     "case_id": "refund-042",
     "transcript": [
       {"role": "customer", "text": "Ordered a pump on the 3rd, still nothing. I want my money back."}
     ],
     "context": {"order_id": "NG-88210", "order_age_days": 12, "amount_usd": 89.00},
     "expected_outcome": "refund_issued",
     "must_not": ["promise delivery date", "escalate to human"],
     "must_include": ["refund confirmation", "reference to order NG-88210"]
   }

5. Instrument resolution yourself. Do not accept the vendor’s dashboard as the sole source of truth on the number you’re being billed for. Log every conversation to your own store and compute a 7-day reopen rate.

   SELECT
     date_trunc('week', closed_at)         AS wk,
     count(*)                              AS agent_resolved,
     count(*) FILTER (WHERE reopened_at IS NOT NULL
                      AND reopened_at < closed_at + interval '7 days')
                                           AS reopened_7d,
     round(100.0 * count(*) FILTER (WHERE reopened_at IS NOT NULL
           AND reopened_at < closed_at + interval '7 days') / count(*), 1)
                                           AS reopen_pct
   FROM conversations
   WHERE resolved_by = 'agent'
   GROUP BY 1 ORDER BY 1 DESC;

   A reopen rate above 15% means you’re paying twice for the same fix. That’s your leverage in the renewal conversation.

6. Negotiate the definition, then the rate. Ask for these four clauses in writing: (a) a resolution that reopens within 7 days is not billable; (b) escalated conversations are never billable; (c) you get raw conversation-level export, not just aggregates; (d) a pilot period with a volume floor of zero. Vendors competing head-to-head right now will give you most of this.

7. Start with one intent, not your whole queue. Pick the highest-volume, lowest-risk intent — order status is the classic — and run the agent on it exclusively for 30 days. Measure resolution rate and CSAT against your human baseline before expanding scope.

How it compares

Dimension	Sierra	Decagon	Zendesk / Intercom (incumbent)
Primary pricing model	Per resolved conversation	Per resolved conversation	Per seat, with AI resolution add-ons
Developer surface	Agent SDK, declarative config, simulation suite	Natural-language operating procedures, admin console	App framework and webhooks around a ticketing core
Who configures it	Engineering-leaning; technical teams get the most out of it	Support ops leads can configure without engineers	Support admins
Voice support	Yes, voice is a first-class channel	Yes, voice and chat	Varies; often a separate product or partner
System of record	Sits in front of yours; not a ticketing replacement	Sits in front of yours; not a ticketing replacement	Is the system of record
Best fit	Mid-market to enterprise with real backend APIs	Teams wanting fast configuration without dev cycles	Teams that need the full helpdesk stack first
Main risk	Implementation depth; needs clean API access	Guardrails written in prose can drift without testing	Paying for seats that AI is making redundant

The honest framing: this is not a three-way race. Sierra and Decagon are competing to be the reasoning layer; the incumbents are competing to keep the system of record. Most businesses will end up with both, and the interesting question is which line item grows.

What’s next

Expect the per-resolution rate to fall and the definition of “resolution” to get contested. The first wave of outcome-based contracts is priced generously to the vendor because nobody has good baseline data yet. As buyers instrument their own reopen rates and start disputing invoices, expect tiered rates by intent complexity — pennies for order status, dollars for a multi-step return with a shipping label — and expect audit clauses to become standard. If you’re signing in the next quarter, get the export rights now while competitive pressure is high.

Watch for the platform squeeze. Zendesk, Intercom, and Salesforce all have their own resolution-priced offerings, and they already own your ticket history. The pitch will be “you don’t need a second vendor.” The counter-pitch from Sierra and Decagon will be that a company whose revenue model is seats cannot credibly optimize for eliminating seats. Both arguments are partly true. The deciding factor for most buyers will be whether the incumbent’s agent can actually hit competitive resolution rates on their specific queue — which you can only learn by running both against the same evaluation set.

The deeper shift to watch is agents moving from deflection to transaction. Answering “where’s my order” is table stakes. The revenue event is an agent that can process the return, offer the retention discount, and upsell the replacement part — at which point the CX agent stops being a cost center and starts being measured on margin. When vendors begin pricing on recovered revenue rather than resolved tickets, the category has genuinely matured, and that’s the version of this that should interest business owners most.

Frequently Asked Questions

Is outcome-based AI pricing actually cheaper than seat licenses?

Usually yes, but only if your resolution rate is decent. At a 55% resolution rate and a $2–3 per-fix rate, most businesses land 30–50% below their fully loaded human cost per ticket. Below about 35% resolution, you’re paying for an AI layer plus the same human headcount, and the math stops working. Run the blended-cost calculation above with your own numbers before signing anything.

What counts as a “resolution,” and can vendors game it?

It varies by contract, which is exactly the problem. The common definition is a conversation that closes without human involvement and without the customer reopening within a defined window. Insist that window be at least 7 days and that reopens are non-billable. The gameable failure mode is an agent that gives a plausible-sounding non-answer and closes the ticket — your reopen rate and CSAT are the detectors.

Should I replace Zendesk with AI agents entirely?

Almost certainly not in one move. Keep your ticketing system as the record of truth and put the agent in front of it as the first responder. What actually shrinks is your seat count on the incumbent, since fewer humans touch fewer tickets. Ripping out the helpdesk is a separate, much larger project — and doing both at once means you can’t tell which change caused which outcome.

Sierra vs Decagon — which should a small business pick?

If you have engineering capacity and clean backend APIs, the Sierra AI agent SDK gives you more control and better testability. If your support lead is the most technical person who’ll touch this, Decagon’s procedure-based configuration will get you live faster. For companies under roughly 50 employees, honestly evaluate whether your incumbent’s built-in AI resolution tier gets you 80% of the value at a fraction of the implementation effort — run all three against the same 200-ticket evaluation set and let the numbers decide.

What’s a realistic AI agent resolution rate to expect?

For a business with well-documented policies and API access to order and account data, 50–70% on the top intents is achievable within a quarter. For a business whose answers live in people’s heads, expect 20–30% until the knowledge base is fixed. Vendors quoting 80%+ in a demo are usually measuring deflection — conversations that ended — not resolution. Ask specifically which one their number represents.

How long does implementation actually take?

Two to six weeks to a limited production pilot on one or two intents, assuming your APIs exist and someone can make decisions about refund and escalation policy. The bottleneck is almost never the model — it’s getting a human to write down what the agent is allowed to do. Budget more time for policy definition than for engineering, and start assembling your evaluation set on day one.