Two years ago, buying AI meant paying per seat or per hour. In 2026, the reference price for an AI support agent is a number per outcome: $0.99 per resolution for Intercom Fin (Fin), $0.50 per resolved conversation for HubSpot’s Customer Agent since April 2026 (Resolve247), $1.50–2.00 per automated resolution at Zendesk (Macha). Salesforce charges $2 per conversation or $0.10 per action (Salesforce). Even consulting is moving: around a quarter of McKinsey’s fees were tied to outcomes by late 2025 (report).
For a buyer, this is good news and a trap at the same time. Good, because the vendor shares the risk. A trap, because “per outcome”, “per conversation”, “per action” and “per credit” sound alike and produce very different bills. This guide explains the six pricing models you will meet, what the major platforms charge, how to compare offers on cost per successful outcome, and what belongs in the contract.
The six AI pricing models at a glance
| Model | How you pay | Example (2026) | Who carries the risk | Works best for |
|---|---|---|---|---|
| Fixed build price | One price for an agreed scope | MVP packages at $3,999 / $7,499 / $14,999+ (HouseofMVPs) | Vendor (scope and overruns) | Clear, stable scope |
| Build plus retainer | Setup fee plus a monthly fee | About $5k build plus about $1k per month, the most common hybrid in 2026 (TheCrunch) | Shared | Systems that need ongoing care |
| Subscription per agent or location | Flat monthly fee | AI receptionist Rosie at $49 / $149 / $299 for 250 / 1,000 / 2,000 minutes (Rosie) | Vendor, within the plan’s volume | Standard products |
| Usage-based | Per minute, conversation, document or credit | Voice AI at $0.09–0.31 per minute all-in (Techsy) | Buyer | Variable volumes |
| Outcome-based | Per resolution, booking or lead | Fin $0.99; HubSpot $0.50 per resolution, $1 per recommended lead | Vendor | Standardised, measurable processes |
| Hybrid with success fee | Smaller fixed fee plus a bonus | Distyl ties part of its fees to client results (Channel Dive) | Shared | Larger or strategic projects |
Foundation Capital describes the same shift as a spectrum: paying for access, then for volume, then for a completed process, then for an outcome. Its warning is worth repeating: outcome pricing only works for standardised processes; elsewhere it leads to “margin surprises” for one side or the other (Foundation Capital).
What the platforms charge in 2026
These are the numbers every AI offer is now compared against:
| Platform | Price | Notes |
|---|---|---|
| Intercom Fin | $0.99 per outcome; lead qualification $9.99 | Minimum 50 outcomes per month on third-party helpdesks; 14-day trial (Fin) |
| Zendesk AI | $1.50 per automated resolution with commitment, $2.00 without | Copilot at $50 per agent per month (Macha) |
| HubSpot Breeze | $0.50 per resolved conversation; $1 per recommended lead | Credits at $10 per 1,000; since 14 April 2026 (Resolve247) |
| Salesforce Agentforce | $2 per conversation or $0.10 per action; voice $0.15 | Add-ons $125 per user per month; Agentforce Max from $550 (Salesforce) |
| Microsoft Copilot Studio | $200 per 25,000 credits; $0.01 per credit pay-as-you-go | (CloudZero) |
| Botpress | Plus $150 per month (250 conversations), Team $750 (1,500) | Overage $0.50–0.65 per conversation (Botpress) |
| n8n | €20 / €50 / €667 per month, billed annually | Self-hosted community edition is free (n8n) |
| Voice infrastructure | Retell $0.12–0.18, Vapi $0.13–0.31, Synthflow $0.11–0.15, Bland $0.09–0.14 per minute | Vapi HIPAA add-on $2,000 per month (Techsy) |
The maths that matters: cost per successful outcome
Headline unit prices are not comparable when the units differ. Take a support team with 5,000 conversations per month where the AI resolves 60%, that is 3,000 conversations:
| Pricing | Calculation | Monthly cost | Cost per resolution |
|---|---|---|---|
| HubSpot, $0.50 per resolved conversation | 3,000 × $0.50 | $1,500 | $0.50 |
| Intercom Fin, $0.99 per outcome | 3,000 × $0.99 | $2,970 | $0.99 |
| Zendesk, $1.50 per automated resolution | 3,000 × $1.50 | $4,500 | $1.50 |
| Salesforce, $2 per conversation | 5,000 × $2 | $10,000 | $3.33 |
The lower your resolution rate, the more per-conversation pricing hurts: at 40%, $2 per conversation becomes $5 per resolution. Per-action pricing behaves similarly, because a complex case can trigger many actions.
Two more adjustments before you decide. First, add the fixed layer: helpdesk or CRM seats, minimums and add-ons. The cheapest unit price inside a platform you would otherwise not use is rarely the cheapest total. Second, compare with your own baseline cost per ticket. An AI resolution only creates value if it is cheaper than the human handling it replaces — and if quality holds. If you are still building that baseline, start with our guide to calculating AI ROI.
Example: one phone workload, four price tags
Voice AI shows how much the pricing model shapes the bill. Take a dental practice or a service company with 600 inbound calls a month and an average call length of 2.5 minutes, so about 1,500 minutes:
| Pricing model | Example rate | Monthly cost for this workload | What you also need |
|---|---|---|---|
| Raw infrastructure, per minute | $0.09–0.31 per minute all-in (Techsy) | about $135–465 | Your own build, integrations and maintenance |
| Product subscription, minute bundles | Rosie: $299 for 2,000 minutes (Rosie) | $299 | Fits only if the product’s workflow fits yours |
| Per call | Smith.ai AI plan: about $1.67–2.00 per call (Smith.ai) | roughly $1,000–1,200 at that rate | Little set-up effort |
| Agency bundle | For example $500 per month for 1,000 minutes, at a cost of about $97 (Call Supplai) | depends on the overage rate | Integration and care included |
The infrastructure price looks cheapest, but it excludes the work that makes the agent useful: calendar and CRM integration, testing, monitoring and changes. The per-call product is the most expensive per unit and the simplest to start. The agency bundle sits in between and shows typical service margins: the provider earns on integration and care, not on minutes. None of these is “right” in general; the right one depends on how standard your call flows are and how much integration you need. That is also why outcome components, such as a fee per confirmed booking, are increasingly added on top of a base fee.
How agencies and service providers price AI work
Platform fees cover the software. Getting an agent to work with your systems, data and processes is a separate budget:
| Service | Typical 2026 range | Source |
|---|---|---|
| Process audit | $1.5–2k fixed, 1–2 weeks | TheCrunch |
| Single automation | $3–15k | TheCrunch |
| Connected processes in one department | $15–30k | TheCrunch |
| Custom AI agent | $10–50k+ | Lets-Viz |
| Retainers | $2–4k / $4–8k / $8–15k per month | TheCrunch |
| Senior hourly rates | North America $180–280; Eastern Europe $100–160 | TheCrunch |
| Outcome-based support outsourcing | Crescendo: $1.25–2.99 per resolution plus about $2,900 per month | Macha |
| Productised creative subscription | Designjoy $4,995 per month; Superside from $15k per month | Designjoy, Superside |
New structures are appearing at the edges. Tenex bills for delivered story points instead of hours (Built In). Fixed-price studios such as HouseofMVPs lock the price when the scope is signed, split payment 50/50, absorb overruns and let the client pay model API costs directly (HouseofMVPs).
Why vendors are cautious about “unlimited” plans: AI companies run at gross margins of roughly 50–60%, against 80–90% for classic SaaS (SaaS Mag). Every unlimited promise is a bet on your usage.
Choosing the right model for your process
Match the pricing model to the nature of the work:
- Standardised, high-volume, measurable from shared data (order status, appointment booking, password resets): outcome-based or usage-based with a cap. You pay only for what works.
- Custom integration with an evolving scope (an agent that acts in your ERP, a new internal tool): fixed build price plus retainer. Outcome fees here invite disputes about what changed and why.
- Unpredictable volume (seasonal businesses, campaigns): usage-based with monthly caps, or tiered subscriptions.
- Strategic projects with shared upside: a fixed fee plus a capped success bonus tied to one metric both sides can see.
- Marketing services: percentage-of-ad-spend fees are considered fair only with a cap and a target CPA or ROAS; about 80% of DTC agencies surveyed use a fixed retainer or retainer plus percentage (Foxwell).
Budget predictability is the hidden criterion. Gartner forecasts that spending on AI platforms and models will grow 63% in 2026, which makes budgets harder to predict and pushes buyers towards cost controls (Gartner). Whatever model you choose, a monthly ceiling is non-negotiable.
What to put in the contract
- The definition of the outcome. “Resolved” might mean no human handoff and no reopening within seven days; “booked” an appointment in the calendar that is not cancelled within 24 hours; “qualified” a lead that meets agreed criteria. HFS Research expects the definition of “resolved” to become the next battleground in SLAs (HFS).
- The measurement source and audit rights. Which system counts, who can see the raw data, how disputed cases are sampled.
- Caps and floors. A monthly maximum for you, and any minimums for the vendor (such as Fin’s 50 outcomes per month on third-party helpdesks).
- Model and minute costs. Included with a 20–30% buffer, re-billed with a 15–25% markup, or paid by you directly to the provider; some providers add a 15–20% premium for model drift and token price spikes (Digital Applied).
- Price review triggers. Model prices fall; volume tiers should adjust automatically.
- Quality guardrails. Customer satisfaction, correct escalation and complaint rates, so that outcomes cannot be gamed by closing tickets too early.
- Ownership and exit. Who owns code, prompts, evaluation sets and data, and how you leave. Our guide to avoiding AI vendor lock-in has a checklist.
Typical mistakes
- Comparing headline prices across different units. $0.99 per resolution and $2 per conversation are not “about the same”.
- Ignoring seats, minimums and platform dependence. The per-outcome price only applies inside the vendor’s ecosystem.
- Paying for outcomes the vendor cannot control. Revenue guarantees are a red flag; the FTC banned Air AI from marketing business opportunities after misleading claims about earnings and growth (FTC).
- No usage cap. A misconfigured loop or a traffic spike can multiply the bill.
- Percentage-of-spend fees without a ceiling. They reward spending, not results.
- Outcome metrics without quality guardrails. Speed and volume improve while customers quietly leave.
- Long fixed commitments in a falling market. Model and minute prices keep dropping; a three-year contract without a price review clause locks in today’s costs.
- Forgetting your own costs. Whatever the pricing model, someone on your side reviews quality, updates the knowledge base and handles escalations. Put those hours into the comparison, too.
Conclusion
Outcome-based AI pricing is a real improvement for buyers: you pay for results, and the vendor has a reason to make the system work. But it only works when the outcome is standardised, measurable from shared data and defined in the contract. Compare every offer on cost per successful outcome, add the fixed platform layer, cap the variable costs and keep ownership of your data and configuration.
We price our own work the same way: a fixed price for the build, a transparent retainer for care, and outcome-based components in Lotus Libre Voice and support agents in Lotus Libre Agents where the result is measurable. If you are weighing offers — a platform agent, an agency quote or a support platform like those in our AI support agent guide — book a free 15-minute audit and we will compare them with you on real cost per outcome.