What are the main business models in conversational AI?
In our conversational AI market deck, you will find everything you need to understand the market
SUMMARY
Conversational AI companies mainly make money through subscriptions, per-seat software, usage-based APIs and voice, outcome fees, and hybrid enterprise contracts; advertising, commerce and implementation services are secondary revenue streams.
The market is splitting into two businesses. One sells people better software; the other sells companies work performed by machines, and the billing model changes with that shift.
The familiar SaaS seat is still surprisingly resilient. Mainstream workplace assistants cluster around roughly $20 to $25 per employee per month, but that pricing unit becomes less useful once the AI works independently of a specific employee.
Falling inference prices cut both ways. They put pressure on companies that simply resell tokens, while improving the economics of products that keep charging a subscription or a fixed price for a useful outcome.
Customer service is moving fastest toward outcome pricing because the work is measurable. A resolved ticket, qualified lead or saved cancellation can be priced against labor savings or revenue rather than against the number of messages exchanged.
Voice remains more stubbornly usage-based. Minutes still track infrastructure cost closely because speech, model inference, telephony and listening time all keep consuming resources while a call is active.
Hybrid enterprise pricing is becoming the practical middle ground: a recurring platform or seat fee keeps revenue predictable, while credits, actions, minutes or outcomes let the vendor participate when AI usage expands.
AI also creates a problem for incumbent SaaS vendors. If automation reduces the number of human agents a customer needs, vendors that rely only on seats can lose revenue even while their product becomes more valuable.
Advertising and shopping add a new consumer-internet layer to conversational AI, but they are still secondary to subscriptions and enterprise revenue. Their importance rises if assistants become a major starting point for product discovery and purchasing.
The strongest pricing power sits above the model layer. Companies that control workflows, permissions, business data and measurable outcomes can keep charging for value even when the underlying intelligence gets cheaper and easier to switch.
What does conversational AI actually sell now?
Conversational AI companies currently make money in five core ways: subscriptions, employee seats, usage, completed outcomes and hybrid enterprise contracts, with advertising and commerce starting to appear around the edges.
The confusing part is that the same chat or voice interface can hide completely different economics. ChatGPT and Claude mostly sell access to an assistant. OpenAI and Anthropic APIs sell inference. Retell sells voice capacity by the minute. Intercom Fin and Zendesk increasingly charge when an AI agent gets a job done. Salesforce mixes its existing software relationship with credits consumed when agents take actions.
Those models also correspond to different stages of AI autonomy. When a person spends all day using an assistant, a monthly subscription or employee license still makes sense. Once the AI starts handling thousands of customer conversations on its own, the number of human users becomes much less relevant. Usage and outcomes then become easier units to sell.
So the conversational AI market has already split into two broad businesses. One sells people better software. The other increasingly sells companies machine-performed work.
Why do people still pay monthly for ChatGPT and Claude?
Consumer conversational AI subscriptions are working today because the providers have found enough differences in usage intensity to support a surprisingly wide range of prices.
OpenAI currently charges $8 per month in the US for ChatGPT Go, $20 for Plus and $200 for Pro. Anthropic charges $20 for Claude Pro and starts Claude Max at $100 per month. A consumer can therefore spend 10 or 25 times more than the basic paid tier without moving into a traditional enterprise contract.
That is more revealing than the familiar $20 headline price. Conversational AI demand clearly does not stop there. Heavy users will pay much more for additional model access, higher limits, priority capacity and advanced tools.
The subscription also gives providers a useful economic buffer. Most subscribers will never consume exactly the same amount of inference, so light users help absorb the cost of heavier users. Very intense users can then be pushed toward $100 or $200 plans instead of making the standard tier uneconomic.
We should expect more segmentation here rather than one universal AI subscription price. Providers are learning how much access different groups want and putting a price on each level.
| Consumer plan | Current US price | What mainly changes |
|---|---|---|
| ChatGPT Go | $8/month | More everyday usage |
| ChatGPT Plus | $20/month | More advanced models and tools |
| ChatGPT Pro | $200/month | Much higher premium capacity |
| Claude Pro | $20/month | Higher Claude usage |
| Claude Max | From $100/month | Far higher usage limits |
If you want more recent data on this point, please see our latest conversational AI market report.
This market map, featured in our conversational AI market deck, highlights top companies and startups in the conversational AI market
Why do companies still pay per employee for conversational AI?
Companies still buy conversational AI by the employee because workplace copilots currently behave much more like productivity software than autonomous labor.
ChatGPT Business costs $20 per user per month with annual billing. Claude Team costs $25 per member per month annually. Microsoft 365 Copilot Business is priced at $21 per user per month for eligible small and midsize businesses, while Microsoft's newer Business Standard bundle with Copilot is $23.50 per user per month.
We therefore keep finding the same rough price zone: around $20 to $25 per worker per month for mainstream workplace AI. That is surprisingly conventional pricing for technology with such ambitious productivity claims.
There is a practical reason. A procurement team already understands licenses. If 500 employees need an AI assistant, buying 500 seats is simple. The vendor also gets predictable recurring revenue instead of a bill that jumps every time employees have an unusually busy week.
The cracks are already appearing, though. OpenAI Business now supports flexible credits beyond included usage, and Enterprise customers can use credit-based or token-based arrangements. As assistants perform more work without a person continuously prompting them, the seat becomes a weaker measure of consumption.
For now, we would still call per-seat SaaS one of the main conversational AI business models. Its strongest territory is the human copilot, where one employee remains clearly attached to one license.
Can token-based AI APIs stay attractive when model prices keep falling?
Token-based conversational AI APIs remain a strong model-provider business, but building an application whose main value is reselling those tokens looks increasingly fragile.
The latest OpenAI pricing moves show how quickly the underlying unit can deflate. GPT-5.6 Luna now costs $0.20 per million input tokens and $1.20 per million output tokens after an 80% price cut. GPT-5.6 Sol is currently offered at promotional pricing of $4 per million input tokens and $20 per million output tokens. Anthropic's Claude Opus 4.8 list price is $5 per million input tokens and $25 per million output tokens.
Take a small interaction with 2,000 input tokens and 500 output tokens. The raw model bill is about $0.001 on Luna, $0.018 on the current Sol pricing and $0.0225 on Opus 4.8. The difference between the cheapest and most expensive of those examples is more than 22 times, even though all three can sit behind an interface that simply looks like "AI chat" to the user.
The longer trend is even more aggressive. Stanford's AI Index calculated that the inference cost required to reach roughly GPT-3.5-level performance fell by more than 280 times between late 2022 and late 2024. Recent model releases have kept pushing cost per unit of useful intelligence down.
For OpenAI or Anthropic, token pricing still works because enormous volumes accumulate across thousands of developers. An application company faces a different problem. Customers can see competing models getting cheaper, so a simple markup on inference becomes hard to defend.
Cheaper models are great news for companies charging $20 per month, $0.99 per completed task or another value-based price. Their input cost can fall while the customer's willingness to pay changes much more slowly. That gap is where some of the best conversational AI economics are starting to appear.
If you want more recent data on this point, please see our latest conversational AI market report.
As this chart shows, and as featured in our conversational AI market deck, search interest in conversational AI has increased sharply
Why are customer-service AI companies charging per resolution now?
Customer-service conversational AI is moving fastest toward outcome pricing because a resolved problem has an obvious economic value and can directly replace human workload.
Intercom currently charges $0.99 when Fin produces a standard resolution, successful procedure handoff or disqualification. Its newer sales product charges $9.99 when Fin successfully qualifies and routes a prospect. The tenfold difference inside the same product is revealing: Intercom is already pricing different AI conversations according to the value of what happened at the end.
Zendesk has gone through an even clearer transition. Its bot pricing previously relied partly on monthly active users and Answer Bot resolutions. Today, AI-agent usage revolves around automated resolutions and a resolution allowance. Zendesk also introduced resolution tiers that distinguish assisted escalations, contained resolutions and verified resolutions. The invoice is gradually being tied to the amount and quality of work performed by the AI.
Sierra has built its commercial model around the same idea from the start. Its customers agree on outcomes such as successful resolutions, saved cancellations, upsells or other completed processes. Sierra said in its year-two review that it had passed $150 million in ARR, and its current website says 40% of the Fortune 50 now partner with the company. We cannot attribute that growth entirely to pricing, but outcome-based AI has clearly moved beyond a small pricing experiment.
There is still a measurement problem. A customer who stops replying may be satisfied, may have given up, or may have called another channel. Zendesk's recent introduction of contained and LLM-verified resolutions shows how much work goes into defining what "solved" actually means. Sierra also allows blended consumption pricing where a clean outcome is difficult to measure.
The model works best when the result is easy to observe: resolve a ticket, book an appointment, qualify a prospect, collect a payment or save a cancellation. In those situations, paying for successful work feels much more natural than paying for messages.
| Company | Current AI billing logic | Example |
|---|---|---|
| Intercom Fin | Pay per outcome | $0.99 standard outcome |
| Intercom Fin for Sales | Pay per qualification | $9.99 qualification |
| Zendesk AI Agents | Resolution allowance and tiers | Charge linked to resolution type |
| Sierra | Negotiated outcomes | Resolution, retention, upsell and other agreed results |
Why does voice AI still charge by the minute?
Voice conversational AI still charges by the minute because call duration tracks the vendor's real costs unusually well.
Retell currently advertises AI voice agents at roughly $0.07 to $0.31 per minute. Its detailed calculator shows what sits underneath that number: voice infrastructure, text-to-speech, the chosen language model, telephony and optional features such as denoising or quality assurance all add variable cost while the call remains active.
ElevenLabs packages the same economics inside subscriptions. ElevenAgents Business costs $990 per month and includes 12,375 call minutes. That works out to exactly $0.08 per included minute.
A ten-minute call therefore costs around $0.80 at that ElevenAgents ratio. On a $0.31-per-minute configuration, the same call represents $3.10. Once a company handles hundreds of thousands of calls, those differences get expensive fast.
This makes unlimited voice much harder to offer than unlimited-looking text chat. Silence even costs money on Retell because the speech system remains active and listening.
Voice vendors may eventually attach more pricing to booked appointments, qualified leads or completed service requests. For now, minutes remain one of the cleanest ways to make sure revenue grows alongside infrastructure cost.
This chart, included in our conversational AI market deck, shows annual VC investment in conversational AI startups
Are hybrid pricing models becoming the default in enterprise conversational AI?
Hybrid pricing is currently becoming the safest model for enterprise conversational AI because it gives vendors recurring revenue and still lets the bill grow when AI usage takes off.
Intercom combines traditional helpdesk seats with Fin outcomes. ElevenLabs sells a monthly platform subscription with included voice minutes and then meters additional consumption. OpenAI Business charges for seats while allowing companies to buy flexible credits beyond included usage.
Salesforce has pushed the idea further with Flex Credits. Its current rate card assigns 20 credits to a standard Agentforce action and 30 to a standard voice action. At Salesforce's published price of $500 per 100,000 Flex Credits, those examples correspond to roughly $0.10 and $0.15 of credit consumption.
Enterprise contracts also contain work that rarely appears in a self-service pricing calculator. Sierra's current product pages describe an expert agent-development team that has supported hundreds of deployments. Retell's enterprise offer includes dedicated implementation support. Large deployments involve integrations, permissions, testing, evaluation, security work and ongoing optimization.
A hybrid contract can absorb all of that much better than one simple meter. The base fee pays for the platform, administration, integrations and vendor relationship. Usage or outcomes then increase revenue as the AI performs more work.
A large part of enterprise conversational AI is heading here: predictable recurring revenue underneath, variable AI revenue on top.
If you want more recent data on this point, please see our latest conversational AI market report.
If AI replaces support agents, don't software vendors lose seat revenue?
AI can absolutely shrink a software vendor's human-seat base, so successful incumbents need a second revenue meter that grows as machines do more work.
Imagine a support platform with 1,000 licensed human agents. If conversational AI lets the customer operate with 600 people, pure seat pricing exposes 40% of the vendor's original licenses. The customer got more efficient, while the software company potentially got smaller.
Salesforce gives us an early look at the escape route. At the end of its fiscal 2026, Agentforce had reached $800 million in annual recurring revenue, up 169% year over year. Salesforce had closed more than 29,000 Agentforce deals, and production accounts had grown nearly 50% in a single quarter.
More interestingly, Salesforce said Agentforce and Data 360 had already produced 2.4 billion "agentic work units," its measure for tasks completed by agents. More than 60% of fourth-quarter bookings across Agentforce and Data 360 came from expansions inside existing customers.
Those figures show how an incumbent can monetize AI without depending entirely on another employee license. Salesforce already owns the account, the CRM data and much of the workflow. It can add a new consumption stream each time software performs work inside that installed base.
For traditional SaaS companies, this transition is becoming urgent. The seat can survive, but relying on the seat alone becomes increasingly dangerous as AI gets better.
This chart, included in our conversational AI market deck, breaks down Cognigy's playbook in conversational AI
Will ChatGPT-style assistants make serious money from ads and shopping?
Advertising is already becoming a real conversational AI revenue model, while shopping commissions remain much earlier and should still be treated as an emerging business.
OpenAI began testing ads on ChatGPT Free and Go accounts in the US and has since expanded the program to the UK, Mexico, Brazil, Japan and South Korea. Plus, Pro, Business, Enterprise and Edu remain ad-free. That creates a familiar consumer internet ladder: advertising helps finance free or cheap access, while subscribers pay to remove ads and receive more capacity.
Shopping has developed in parallel. Instant Checkout allows eligible purchases to happen directly inside ChatGPT, and OpenAI's original merchant structure included a small fee on completed purchases. Product discovery itself remains separate from advertising, and merchants can also send product feeds so their catalogs appear more accurately.
The distribution is already meaningful. OpenAI says Shopify's catalog makes millions of Shopify merchants available for product discovery in ChatGPT, while major retailers including Target, Sephora, Nordstrom, Lowe's, Best Buy, Home Depot and Wayfair have integrated through its commerce infrastructure.
We would still rank ads and transaction fees well below subscriptions and enterprise AI revenue today. They become much more interesting if conversational assistants turn into a major starting point for product discovery. At that point, the assistant controls valuable purchase intent before the customer reaches Google, Amazon or a merchant's own website.
Where are the best margins in conversational AI?
Subscriptions and outcome pricing have the best structural margin upside in conversational AI, while pure token resale and labor-heavy implementation have much less room to separate revenue from cost.
We cannot rank actual company margins cleanly because most vendors do not disclose the gross margin of each conversational AI product. We can, however, see how each pricing model reacts when inference becomes cheaper.
Suppose an AI company receives $0.99 for a successful resolution and the underlying model cost required to solve it falls by half. The customer still values the solved problem, so the vendor has a chance to keep part of that saving. A $20 subscription has the same property as long as average usage remains under control.
Token resale behaves differently. When model providers cut token prices by 50% or 80%, customers quickly expect lower prices from intermediaries as well. Voice has some of the same pressure because customers can compare per-minute rates directly.
Services are even harder to scale. More implementation projects usually require more engineers, solution architects or consultants. Software revenue can grow much faster than the team delivering it; services revenue usually cannot.
The best margin position comes from buying a commodity-like input such as inference while selling something customers value in a much less commodity-like way: access, a completed workflow or a measurable business result.
If you want more recent data on this point, please see our latest conversational AI market report.
This chart, included in our conversational AI market deck, shows annual funding in conversational AI startups
What stops conversational AI products from becoming commodities?
Workflow integration currently protects conversational AI companies far more than the chat interface itself.
Building a respectable chat interface has become easy. Recreating everything behind an enterprise deployment is much harder. The AI may need permission to read a CRM record, check an order, change a subscription, issue a refund, update a ticket, write back to a database and hand the conversation to the correct employee when something goes wrong.
Current products increasingly reflect that reality. ChatGPT Business connects with systems including Microsoft 365, Google Drive, Slack, GitHub, Linear and Figma. Intercom can run Fin over its own helpdesk and over other support environments. Salesforce puts Agentforce directly next to CRM records, Data 360 and existing business processes. Sierra deploys agents across chat, SMS, WhatsApp, email and voice while connecting them to customer systems.
Once a company has built hundreds of permissions, workflows, evaluations and exception rules around one platform, replacing the model is usually easier than replacing the whole operating layer.
That changes where value accumulates. Model quality still matters, especially for difficult tasks. But the strongest conversational AI businesses increasingly own the place where the AI is allowed to act, the data needed to act correctly and the measurement system that proves whether the action worked.
Who has the strongest pricing power in conversational AI today?
Conversational AI companies have the strongest pricing power when they can charge for business value, while generic inference and infrastructure face the most visible price pressure.
Intercom provides a simple example. A normal Fin outcome costs $0.99, while a successful sales qualification costs $9.99. The underlying conversation may use similar AI infrastructure, yet the second event commands roughly ten times the price because a qualified buyer is worth more than a routine support resolution.
Zendesk is making the same idea more granular by separating resolution tiers according to how much automation and value the AI delivered. Sierra negotiates different business outcomes directly with customers. These vendors have some room to move prices independently of tokens because the customer is comparing the bill with labor savings or new revenue.
At the other end, API customers can compare OpenAI, Anthropic, Google and open-model providers on cost per million tokens. Voice developers can compare per-minute rates between platforms. Switching still involves work, but the price itself is highly visible.
Enterprise platforms sit somewhere in between. Salesforce, Microsoft and OpenAI can bundle AI with data, identity, administration and existing software relationships. That gives them more pricing power than a standalone model wrapper, even when the underlying model is available elsewhere.
As conversational AI matures, pricing power should increasingly follow control of the workflow rather than ownership of the conversation box.
If you want more recent data on this point, please see our latest conversational AI market report.
This chart, included in our conversational AI market deck, compares the main business model options for conversational AI enterprise platforms
What are the main business models in conversational AI?
The main conversational AI business models today are subscriptions, per-seat software, token APIs, usage-based voice or messaging, outcome pricing and hybrid enterprise contracts; ads, commerce and implementation services add secondary revenue streams.
Consumer assistants are still strongest as subscriptions. ChatGPT and Claude have shown that monthly AI access can support prices ranging from low-cost plans to $100-plus power-user products.
Human copilots still fit the SaaS model. Around $20 to $25 per employee per month remains a common price zone for mainstream business assistants.
Model providers make money from tokens because inference itself is what they sell. Falling token prices make that model increasingly difficult for undifferentiated application companies to copy with a simple markup.
Voice AI still fits usage pricing because every minute consumes speech, model and communications infrastructure. Customer-service agents are moving much faster toward resolutions and business outcomes because successful automation can be measured against human labor or revenue.
For large enterprise deployments, the strongest model currently looks hybrid. A recurring platform or seat fee gives the vendor predictable revenue, while credits, minutes, actions or outcomes let revenue increase when AI performs more work.
Conversational AI is gradually changing what software companies get paid for. Access still dominates consumer assistants and copilots, but autonomous products are pushing the market toward completed work. The companies in the strongest position will keep the recurring software relationship while adding a second meter tied to how much useful work the AI actually performs.
| Business model | Where it works best today | Main strength | Main weakness |
|---|---|---|---|
| Consumer subscription | General assistants | Predictable recurring revenue | Heavy-user cost |
| Per-seat SaaS | Workplace copilots | Simple enterprise purchasing | Weakens as agents replace users |
| Token/API usage | Model providers | Scales directly with developer demand | Strong price deflation |
| Per-minute/message | Voice and communications infrastructure | Revenue follows usage | Easy price comparison |
| Outcome pricing | Customer service and transactional agents | Tied closely to customer value | Outcomes can be difficult to define |
| Hybrid platform + usage | Enterprise AI | Combines recurring and variable revenue | More complicated contracts |
| Ads and commerce | Mass-market assistants | Monetizes free users and purchase intent | Still early |
| Implementation services | Complex enterprise deployments | Helps close and expand large accounts | More dependent on human labor |
OUR METHODOLOGY
We approached the question by separating conversational AI into the units customers actually pay for: access, employee seats, inference usage, voice or messaging usage, completed outcomes, enterprise platform access, implementation work, advertising and commerce.
For each model, we looked for recent, observable commercial evidence: current prices, billing mechanics, product packaging, changes in the monetized unit, disclosed adoption and examples of AI moving from assisting a person to performing work more autonomously.
We compared companies operating at different layers of the market rather than treating one pricing page as representative of the whole category. Consumer assistants, model providers, workplace copilots, customer-service agents, voice platforms and enterprise software all have different cost structures, so their billing models need to be read separately.
We gave the greatest weight to direct commercial evidence such as published prices, rate cards, product documentation, company disclosures and reported customer adoption. Stanford HAI's AI Index is the main independent source used for the longer-term comparison of inference-cost declines.
We assessed each example for what it actually demonstrates. A low token price is evidence about inference economics, not proof that every application should charge by the token; an outcome fee is evidence that customers will pay for completed work, but it does not tell us the vendor's exact gross margin unless that margin is disclosed.
The broader conclusions come from patterns that recur across several companies: seats remain effective for human copilots, usage pricing follows infrastructure cost, outcome pricing becomes more attractive when AI performs measurable work, and hybrid contracts let enterprise vendors keep recurring revenue while adding a second meter for machine activity.
Key sources include OpenAI on ChatGPT consumer tiers, Anthropic pricing, ChatGPT Business pricing and credits, Stanford HAI's AI Index, Intercom on Fin outcomes, Zendesk on automated-resolution tiers, Sierra on outcome-based pricing, Retell AI pricing, ElevenLabs agent pricing, Salesforce's Agentforce Flex Credit rate card, Salesforce's FY2026 results, OpenAI on ChatGPT advertising, and OpenAI on Instant Checkout and commerce.
This chart, featured in our conversational AI market deck, illustrates revenue distribution by customer segment in the conversational AI market
Related blog posts
