What 'Custom Pricing' Really Means in AI Customer Support (And Why It Matters)

If you have spent any time evaluating AI customer support platforms, you have noticed something: most of them do not publish pricing. Instead, you see "Contact Sales," "Custom Pricing," or "Let's Talk." After your third discovery call that could have been a pricing page, you start to wonder what everyone is hiding.

They are not hiding one thing. They are hiding several things, and understanding what "custom pricing" actually means in this market will save you months of procurement pain and potentially six figures in overspend.

This post breaks down the real cost ranges across major AI support vendors, explains the business model mechanics behind opaque pricing, and gives you seven questions that will accelerate any vendor evaluation.

Why AI Support Vendors Hide Pricing

There are four legitimate reasons and two less legitimate ones.

Legitimate reasons:

Value varies by deployment. An AI agent resolving 5,000 password resets per month delivers different ROI than one handling 500 complex billing disputes. Vendors argue that flat pricing would either overcharge simple use cases or undercharge complex ones.
Infrastructure costs vary. Large language model inference is not free. A vendor processing 100-word tickets in English has different costs than one handling 2,000-word technical tickets in six languages. Some variability in pricing is rational.
Enterprise procurement expects negotiation. If you publish $200K and a Fortune 500 company was willing to pay $400K, you left money on the table. If a 50-person startup sees $200K, they never call. Custom pricing lets vendors play both ends.
Bundling and scoping. Most deals include some combination of implementation services, integrations, custom model tuning, and ongoing support. Quoting a single number without knowing the scope would be misleading.

Less legitimate reasons:

Price discrimination. Some vendors price based on what they think you can afford, not what the product costs to deliver. A well-funded Series D startup might pay 3x what a bootstrapped company pays for the same product.
Hiding unfavorable comparisons. If Vendor A charges $200K and Vendor B charges $60K for similar capability, Vendor A would rather you not know that until you are deep in the sales process and switching costs feel high.

What Vendors Actually Charge: The Range Table

Based on publicly available information, customer reports, G2 and Gartner Peer Insights reviews, and disclosed contract ranges, here is what the major AI support vendors typically charge as of early 2026.

Vendor	Typical Annual Range	Pricing Model	Minimum Commitment	Setup Timeline	Key Requirement
Decagon	$95,000 - $590,000	Per-conversation (~$0.99)	~$50,000/year	~6 weeks	Dedicated Agent Engineers required
Sierra AI	$150,000 - $350,000+	Outcome-based	Not publicly disclosed	Weeks to months	Enterprise-grade deployment team
Forethought (acquired by Zendesk, March 2026)	$40,000 - $160,000	Tiered by ticket volume	20,000+ tickets/year	30 - 90 days	Minimum ticket volume threshold
Twig	Usage-based at $5/ticket	Per resolved ticket	No minimum (free tier: 100 tickets/mo)	~30 minutes	None

A few observations worth making about this table.

The range within a single vendor is enormous. Decagon's spread from $95K to $590K is a 6x difference. That range exists because pricing depends on conversation volume, complexity, and how many Agent Engineers (Decagon's term for implementation specialists) are assigned. If you are quoted $200K by Decagon and your peer company got $95K, neither of you is being lied to — you just have different scopes.

Outcome-based pricing sounds appealing but is hard to audit. Sierra AI's outcome-based model means you pay when the AI successfully resolves a ticket. The challenge is agreeing on what "successfully resolved" means. If the AI sends a response and the customer does not reply within 72 hours, is that resolved? Different definitions produce very different bills.

Acquisition changes the math. Forethought's acquisition by Zendesk in March 2026 means its pricing and packaging will likely shift. If you are a Zendesk customer, bundling could work in your favor. If you are not, the product's roadmap independence is uncertain.

Usage-based pricing removes volume risk. When you pay per resolved ticket, your AI cost scales linearly with value delivered. If deflection drops in a slow month, your bill drops too. With fixed contracts, you pay the same regardless.

The Real Cost Is Rarely the License Fee

The sticker price on a vendor contract is typically 50-70% of what you will actually spend in Year 1. Here is where the rest goes:

Hidden Cost	Typical Range	Who Bears It
Implementation / professional services	$10,000 - $75,000	Usually included in contract or charged separately
Internal engineering time (integrations)	80 - 400 hours	Your team
Knowledge base preparation	40 - 160 hours	Your team
Agent training on new workflows	20 - 60 hours	Your team
QA and tuning (first 90 days)	40 - 120 hours	Split between vendor and your team
Opportunity cost during ramp-up	Varies	Your team

When a vendor says "6-week implementation," translate that into your team's time: a project manager, a support operations lead, possibly an engineer, all spending 10-20 hours per week for six weeks. At a blended internal rate of $75/hour, that is $13,500 to $27,000 in soft costs before the AI answers a single ticket.

Vendors that offer managed implementation — where they do the setup and tuning work — reduce this burden significantly. The tradeoff is less internal control during configuration. For most teams under 10,000 tickets per month, the managed approach is faster and cheaper.

How "Per Conversation" Pricing Gets Complicated

Decagon's per-conversation pricing (approximately $0.99 per conversation) seems straightforward until you dig in. Consider these scenarios:

Customer asks two questions in one chat session. Is that one conversation or two?
AI partially resolves but customer calls back. Do you pay for both the AI conversation and the human handling?
AI asks a clarifying question and customer abandons. Is that a billable conversation?
Bot-to-bot interactions (automated ticket routing that triggers the AI). Do those count?

Every per-conversation vendor defines "conversation" differently. Decagon's model charges when the AI engages, regardless of resolution. At $0.99 per conversation and 8,000 conversations per month, you are at roughly $95K annually — which aligns with their disclosed minimums. But if your definition of "conversation" is broader than theirs, the math shifts.

How Outcome-Based Pricing Gets Complicated

Sierra AI's outcome-based approach has its own ambiguities:

Resolution attribution. If the AI provides information but a human agent closes the ticket, who gets credit?
Partial resolution. The AI handles steps 1-3 of a 5-step process and the agent finishes. Is that an AI outcome?
CSAT thresholds. Some outcome-based contracts only count resolutions above a CSAT threshold. If the customer rates 3/5, the resolution might not count — but you still benefited from the deflection.
Renegotiation triggers. As your deflection rate improves, the per-outcome cost might increase because the vendor argues they are delivering more value. Read renewal terms carefully.

7 Questions to Ask During Procurement

These questions will save you time and surface the information vendors do not volunteer. Bring them to every evaluation call.

1. "What is your all-in cost for my specific volume, including implementation?"

Do not accept a range. Give them your monthly ticket volume, your helpdesk platform, your top 5 ticket categories, and ask for a specific annual number that includes everything needed to get live. If they cannot give you a number within 20% accuracy on the first call, their pricing process is going to be slow.

2. "What happens to my bill if deflection underperforms by 50%?"

This separates usage-based vendors from fixed-fee vendors. With usage-based pricing, underperformance means a lower bill. With fixed pricing, you pay the same. With outcome-based pricing, you pay less — but check whether there is a minimum commitment that kicks in.

3. "What does my team need to provide during implementation, measured in hours?"

Vendors will say implementation is "turnkey" or "white glove." Push for specifics: how many hours of your support ops lead's time? Your engineer's time? Your content team's time? Multiply by your internal rates to get the true all-in cost.

4. "How do you measure and report resolution quality, not just resolution volume?"

Any vendor can close tickets. The question is whether closed tickets stay closed. Ask for their quality framework — do they measure customer reopens, CSAT on AI-resolved tickets, escalation rates? Quality scoring that evaluates multiple dimensions (accuracy, completeness, tone, policy compliance) is more reliable than a single pass/fail metric.

5. "What is the contract term, and what are the renewal economics?"

Most enterprise AI vendors push for annual contracts. Some push for multi-year. Ask explicitly: What is the renewal price increase cap? Is there an auto-renewal clause? Can you downgrade mid-term if volume drops? A 2-year contract at $200K/year with a 15% annual escalator means you are committing $430K, not $400K.

6. "Can I run a paid pilot before signing an annual contract?"

A 30-day pilot at actual volume is worth more than three months of demos. If the vendor will not offer a pilot — or only offers a pilot on a sanitized subset of tickets — ask why. Vendors confident in their product welcome pilots. Those that do not are worried about what real-world performance will reveal.

7. "What do I need to do if I want to leave after Year 1?"

Data portability, conversation history export, knowledge base ownership, and transition support. If the vendor built custom models on your data, who owns them? If your helpdesk integration depends on their middleware, what breaks when you disconnect? The exit terms tell you how confident the vendor is that you will want to stay.

A Framework for Comparing Apples to Oranges

Since every vendor structures pricing differently, normalize them into a single metric: effective cost per resolved ticket.

Here is the formula:

Effective Cost Per Resolved Ticket = Total Annual Cost / Total AI-Resolved Tickets Per Year

Run this for each vendor using their quoted price and your projected deflection:

Scenario: 2,000 tickets/month, 35% deflection	Annual AI-Resolved Tickets	Total Annual Cost	Effective Cost/Ticket
Decagon (low end)	8,400	$95,000	$11.31
Decagon (mid range)	8,400	$250,000	$29.76
Sierra AI (low end)	8,400	$150,000	$17.86
Forethought (low end)	8,400	$40,000	$4.76
Twig (per ticket)	8,400	$42,000	$5.00

Now compare the effective cost per AI-resolved ticket against your cost per human ticket. If the AI costs more per ticket than a human, you are paying a premium for speed and availability, not savings. That might be worth it for 24/7 coverage — but you should make that tradeoff consciously.

When "Custom Pricing" Actually Works in Your Favor

Custom pricing is not inherently bad. It works in your favor when:

You have leverage. High volume, a recognized brand (for case studies), or a competitive evaluation gives you negotiating power.
Your use case is narrow. If you only need AI on one channel or one ticket type, a custom scope should mean a lower price than a full-platform license.
You negotiate success metrics into the contract. Tie a portion of the fee to actual deflection performance. Some vendors will agree to performance-based pricing if you push.

It works against you when:

You are the smaller fish. Vendors optimize for their largest customers. If you are below their typical deal size, you may get less attention during implementation.
You do not know the market rate. Without public pricing or peer data, you cannot tell if you are getting a fair deal. This post should help with that.
You are time-pressured. Custom pricing processes take 2-6 weeks. If you need to deploy next month, you have less room to negotiate.

The Transparency Trend

The market is slowly moving toward pricing transparency. Vendors that publish pricing tend to have shorter sales cycles, higher trial-to-paid conversion, and lower customer acquisition costs. As a buyer, you should reward transparency — it signals confidence in the product's value delivery.

If you want to see what transparent AI support pricing looks like — no sales call required — our pricing page shows exactly what each plan costs. For head-to-head comparisons with specific vendors, see Twig vs. Decagon and Twig vs. Sierra AI.

The vendors who publish pricing are betting that informed buyers make better customers. The vendors who hide pricing are betting that uninformed buyers pay more. Decide which kind of relationship you want with your AI provider.