When Does a Custom AI Chatbot Cost Less Than a Per-Resolution AI Agent?
A worked comparison at 1,000, 10,000 and 50,000 conversations a month, using list prices from Zendesk, Intercom, HubSpot and Salesforce. Per-resolution agents win at low volume, and here is where the break-even sits.
A custom AI chatbot costs less than a per-resolution AI agent only above about 3,400 to 17,800 conversations a month against a $0.99 price, or about 8,100 to 75,000 against a $0.50 price, by our own arithmetic at a 50% resolution rate. Below about 1,600 conversations a month, every per-resolution list price we checked beats every custom option in our model. If that is your volume, buy the vendor's agent.
A per-resolution AI agent is a help desk or CRM vendor's built-in assistant, which you switch on and pay for each conversation it settles without a person. A custom AI chatbot is a language model you choose, connected to your systems through their APIs (the connection points software offers to other software). You pay for it through a build fee and monthly running costs, whether it settles anything or not. So one cost moves with results and the other is mostly fixed, and the break-even is the volume where the two lines cross.
Key facts at a glance
- Against a $0.99 price, custom wins from about 3,400 conversations a month for the cheapest documents bot, and from about 17,800 for the costliest agent that takes actions. Against $0.50, the range is about 8,100 to 75,000 (our arithmetic, 50% resolved).
- At 1,000 conversations a month, per-resolution fees come to $250 to $1,000, and the cheapest custom option to $1,552 (our arithmetic).
- At 50,000 conversations a month, per-resolution fees come to $12,500 to $50,000, and custom to $4,400 to $14,500 (our arithmetic).
- List prices per resolution, checked 5 November 2026: HubSpot $0.50, Gorgias $0.90 to $1.00, Intercom $0.99, Zendesk $1.50 to US$2.00.
- Below about 10,000 conversations a month, model usage is not the main custom cost. Build and maintenance are.
How we ran the numbers
We are Webify.AI, and we build custom chatbots, so we have an interest in this answer. That is why every input is listed below and the arithmetic is labeled as ours. No vendor or study published these totals. Vendor figures are public list prices. Negotiated prices could be lower, which would push every break-even higher.
Two assumptions sit under everything. First, half of all conversations end as billable resolutions, on both sides; we test 30% and 70% later. Second, the chatbot serves web chat, which carries no per-message channel fee; we treat WhatsApp separately.
What the vendors charge
The units differ, so the last column converts each one to a fee per conversation at 50% resolved. Per-resolution vendors bill only resolved conversations. Freshworks and Salesforce bill for sessions, actions or conversations whether or not they are resolved.
Swipe the table sideways to see every column.
| Vendor agent | Billing unit | List price | Fee per conversation (our arithmetic) |
|---|---|---|---|
| HubSpot Customer Agent | Resolved conversation | $0.50 | $0.25 |
| Gorgias AI Agent | Resolved conversation | $0.90 on annual billing, $1.00 monthly | $0.45 to $0.50 |
| Intercom Fin | Outcome | $0.99 | $0.495 |
| Zendesk AI agents | Automated resolution | $1.50; US$2.00 pay-as-you-go | $0.75; $1.00 |
| Freshworks Freddy AI Agent | Session | $49 per 100 sessions | $0.49, every session |
| Salesforce Agentforce, Flex Credits | Action | $500 per 100,000 credits, 20 credits an action, so $0.10 an action | $0.40 at four actions, every conversation |
| Salesforce Agentforce, conversations | Conversation | $2 | $2.00, every conversation |
Zendesk's blog quotes $1.50 per automated resolution, while its pricing page lists pay-as-you-go at US$2.00 and showed the committed price as £1.30 when we checked, so we run both dollar figures. Four actions per conversation is our assumption.
What a custom chatbot costs each month
Build and maintenance below are our planning estimates from our AI agent cost guide, at our hourly rates of $25, $35, $45 and $60. They are estimates, not quotes. Our chatbot cost guide shows how they are built up.
Build and maintenance
Swipe the table sideways to see every column.
| What the chatbot does | Build | Build over 24 months | Maintenance a month | Fixed cost a month |
|---|---|---|---|---|
| Answers from documents | $8,400 to $16,800 | $350 to $700 | $1,050 to $2,100 | $1,400 to $2,800 |
| Reads business systems | $22,800 to $45,600 | $950 to $1,900 | $1,050 to $2,100 | $2,000 to $4,000 |
| Takes actions | $42,000 to $84,000 | $1,750 to $3,500 | $1,050 to $2,100 | $2,800 to $5,600 |
We spread the build evenly over 24 months with no interest. The guide gives 20 to 40 hours a month of maintenance for the first months after launch, and we hold that flat for all three types. That may overstate a simple documents bot and understate any chatbot at 50,000 conversations a month.
Model usage
Model providers charge by the token, a small chunk of text the model reads (input) or writes (output). Our mid-priced example is Claude Sonnet 5.5 or gpt-6.1-sol, at $2 input and $10 output per million tokens. Our small example is Claude Haiku 5.5 or gpt-6-luna, at $0.10 and $0.50. The token counts below are our assumptions, not measurements, and we count no savings from caching.
Swipe the table sideways to see every column.
| What the chatbot does | Model calls per conversation | Input tokens per call | Output tokens per call | Cost per conversation, mid-priced | Cost per conversation, small |
|---|---|---|---|---|---|
| Answers from documents | 4 | 4,000 | 500 | $0.052 | $0.0026 |
| Reads business systems | 6 | 5,000 | 500 | $0.09 | $0.0045 |
| Takes actions | 10 | 6,000 | 500 | $0.17 | $0.0085 |
The first row works out like this: 4 calls × 4,000 tokens = 16,000 input tokens, or $0.032 at $2 per million, plus 4 × 500 = 2,000 output tokens, or $0.02 at $10 per million, for $0.052. You pay it on every conversation, including those a person ends up handling. The totals use the mid-priced model, because a small model has to pass your tests first.
Hosting
We use DigitalOcean's server and managed database list prices as an example, with our own sizing, rounded up for backups and logs. Up to 5,000 conversations a month, two 4 GiB servers at $24 plus a 2 GiB database at $30.45 is $78.45, so we use $100. Up to 25,000, two 8 GiB servers at $48 plus a 4 GiB database at $60.90 is $156.90, so we use $200. Above that, four 8 GiB servers plus an 8 GiB database at $122.10 is $314.10, so we use $400.
Channel fees
Web chat on your own site has no per-message fee, so it counts as $0 in our totals. WhatsApp is different. Meta charges per delivered message, with 1,000 free service messages a month per business number, and has billed service messages again since 1 October 2026. Twilio's table of Meta's fees lists them at $0.0034 in the US and $0.022 in the UK. At four replies per conversation on one number (our arithmetic):
Swipe the table sideways to see every column.
| Conversations a month | Billable replies | WhatsApp fee at the US rate | WhatsApp fee at the UK rate |
|---|---|---|---|
| 1,000 | 3,000 | $10 | $66 |
| 10,000 | 39,000 | $133 | $858 |
| 50,000 | 199,000 | $677 | $4,378 |
Meta bills the delivered message whoever wrote it, and Intercom's pricing page also lists WhatsApp as "Charged per message sent". So if the vendor's agent would answer on WhatsApp too, the fee lands on both sides and the break-even does not move. Routing through Twilio adds its own $0.005 per message.
A help desk route can add a fee. Zendesk's pricing page shows Suite Professional including 1,000 monthly active users on Sunshine Conversations, its API for outside bots, then $50 per extra 2,500. If every conversation is a different customer, that adds $0, $200 and $1,000 at our three volumes, which we add to the custom side in the Zendesk break-evens.
The totals at 1,000, 10,000 and 50,000 conversations
Monthly, our arithmetic, on web chat, with the mid-priced model and 50% resolved. Custom ranges run from our low to our high estimate.
Swipe the table sideways to see every column.
| Option | 1,000 a month | 10,000 a month | 50,000 a month |
|---|---|---|---|
| HubSpot at $0.50 | $250 | $2,500 | $12,500 |
| Salesforce Flex Credits, four actions | $400 | $4,000 | $20,000 |
| Gorgias at $0.90 | $450 | $4,500 | $22,500 |
| Freshworks at $0.49 a session | $490 | $4,900 | $24,500 |
| Intercom at $0.99 | $495 | $4,950 | $24,750 |
| Zendesk at $1.50 | $750 | $7,500 | $37,500 |
| Zendesk at $2.00 | $1,000 | $10,000 | $50,000 |
| Salesforce at $2 a conversation | $2,000 | $20,000 | $100,000 |
| Custom, answers from documents | $1,552 to $2,952 | $2,120 to $3,520 | $4,400 to $5,800 |
| Custom, reads business systems | $2,190 to $4,190 | $3,100 to $5,100 | $6,900 to $8,900 |
| Custom, takes actions | $3,070 to $5,870 | $4,700 to $7,500 | $11,700 to $14,500 |
At 1,000 conversations a month, the vendor wins. The highest per-resolution bill, $1,000, is still below the cheapest custom option, $1,552. Salesforce's $2 per-conversation price is the exception: a documents bot beats it from about 800 conversations a month, although Salesforce's Flex Credit pricing comes to $400 on our assumption.
At 10,000, it depends on the bot and the price. A documents bot at $2,120 to $3,520 beats everything except HubSpot at $2,500, where the two overlap. An agent that takes actions, at $4,700 to $7,500, beats $0.99 only at the bottom of its range and loses to HubSpot outright. Over 24 months, Intercom's list price comes to $118,800, against $50,880 to $84,480 for the documents bot and $112,800 to $180,000 for the action agent.
At 50,000, custom wins on this arithmetic against every price but one. HubSpot at $12,500 sits inside the action agent's $11,700 to $14,500 range. If maintenance doubled at that volume, the agent would cost $12,750 to $16,600 and HubSpot would win outright.
Where the break-even sits
The formula is: break-even conversations = (fixed monthly cost + hosting) ÷ (vendor fee per conversation − model cost per conversation). For Intercom against the cheapest documents bot, that is ($1,400 + $100) ÷ ($0.495 − $0.052), or about 3,400 conversations a month, which is about 1,700 resolutions.
Swipe the table sideways to see every column.
| Vendor price | Answers from documents | Reads business systems | Takes actions |
|---|---|---|---|
| HubSpot, $0.50 | 8,100 to 15,200 (4,000 to 7,600) | 13,800 to 27,500 (6,900 to 13,800) | 40,000 to 75,000 (20,000 to 37,500) |
| Gorgias, $0.90 | 3,800 to 7,500 (1,900 to 3,800) | 6,100 to 11,700 (3,100 to 5,800) | 10,700 to 20,700 (5,400 to 10,400) |
| Intercom, $0.99 | 3,400 to 6,800 (1,700 to 3,400) | 5,400 to 10,400 (2,700 to 5,200) | 9,200 to 17,800 (4,600 to 8,900) |
| Zendesk, $1.50 | 2,200 to 4,300 (1,100 to 2,100) | 3,300 to 6,600 (1,600 to 3,300) | 5,300 to 10,300 (2,700 to 5,200) |
| Zendesk, $2.00 | 1,600 to 3,100 (800 to 1,600) | 2,400 to 4,600 (1,200 to 2,300) | 3,600 to 7,200 (1,800 to 3,600) |
The table shows conversations a month, with resolutions in brackets, at 50% resolved. It is our arithmetic, rounded to the nearest hundred, and the Zendesk rows include the Sunshine Conversations user fee.
How far the answer moves
Swipe the table sideways to see every column.
| Change, against Intercom's $0.99 | Answers from documents | Reads business systems | Takes actions |
|---|---|---|---|
| Base case | 3,400 to 6,800 | 5,400 to 10,400 | 9,200 to 17,800 |
| 30% resolved | 6,500 to 12,200 | 10,600 to 20,300 | 25,200 to 47,200 |
| 70% resolved | 2,300 to 4,500 | 3,500 to 7,000 | 5,700 to 11,100 |
| Small model | 3,000 to 6,100 | 4,300 to 8,600 | 6,200 to 11,900 |
| Build spread over 12 months | 4,200 to 8,400 | 7,800 to 15,100 | 14,600 to 29,200 |
| Build spread over 36 months | 3,100 to 6,200 | 4,400 to 8,800 | 7,400 to 14,300 |
- The resolution rate moves the answer in conversations, but barely in resolutions: the cheapest documents bot needs about 2,000 resolutions at 30%, 1,700 at 50% and 1,600 at 70%.
- A low price can close the door. At HubSpot's $0.50 and 30% resolved, the fee is $0.15 per conversation, less than the $0.17 an action agent spends on mid-priced tokens, so it never breaks even.
- Who builds it matters. KORE1, a US staffing firm, lists AI and machine learning engineers at $140 to $240 an hour. Our build and maintenance hours at those rates push the action agent's break-even against $0.99 to about 23,600 to 80,000 conversations.
- A documents bot may need less maintenance than we assumed; our chatbot cost guide plans less for that type. At half our figure, its break-even against $0.99 falls to about 2,200 to 4,200 conversations.
Is a few thousand resolutions a month the right rule of thumb?
Only for simple bots at about $1 or more per resolution. At $0.99, the break-even is about 1,700 to 3,400 resolutions a month for a documents bot, 2,700 to 5,200 for one that reads your systems and 4,600 to 8,900 for one that takes actions. At $0.50, it is 4,000 to 37,500. So the rule holds for the first type, stretches for the second, and is too low for the third and for HubSpot's price.
Comparing only the token bill makes custom look far better than it is. At 10,000 conversations, a documents bot's mid-priced tokens cost $520 against Intercom's $4,950, but its full cost is $2,120 to $3,520, mostly build and maintenance.
Running the model on your own server does not change this much. DigitalOcean lists a 48 GB GPU server at $1.57 an hour, about $1,146 a month at 730 hours, or $2,292 for two (our arithmetic). At 50,000 documents conversations, hosted tokens cost $130 on the small model and $2,600 on the mid-priced one, so the server is no obvious saving. The stronger reason to run your own model is a data rule, which our private AI page covers.
What the arithmetic leaves out
Seats and platform fees
Seats usually cancel out, because your team needs a help desk either way. Plan upgrades do not. HubSpot's agent needs a Professional or Enterprise subscription, and Service Hub Professional starts at $90 a seat plus a required $1,500 onboarding fee, against $7 for Starter: $830 more a month for 10 seats at the lowest listed prices (our arithmetic).
The custom route can force an upgrade too. Zendesk includes a low-volume version of Sunshine Conversations from Suite Professional upward, at US$115 a seat against US$55 for Suite Team on annual billing: $600 more a month for 10 seats (our arithmetic). Fin needs no seats of its own but has a minimum, such as 50 outcomes a month, on another help desk. At Gorgias, each automated interaction also counts as a helpdesk ticket.
Resolution rate
Intercom says Fin averages 76% across more than 12,000 customers, and HubSpot says its agent averages 70%. We have not tested either claim, so we used 50%. The gap between your custom chatbot's rate and the vendor's can matter more than the fees: at 10,000 conversations a month, 10 points is 1,000 conversations your team handles. Test both on the same sample of real conversations.
What counts as a resolution
Definitions differ, so a billed resolution is not always a solved problem: HubSpot counts no hand-off to a person for 72 hours, Intercom counts no further help requested and also bills workflow handoffs at $0.99, and Zendesk bills only after an AI check of the conversation.
Who carries the risk
- Performance: if a vendor's agent resolves little, you pay little. A custom chatbot costs its build and fixed monthly costs even if it resolves nothing.
- Usage: the vendor absorbs model costs, including long conversations and retries. On a custom chatbot, those tokens are your bill.
- Growth: here the risk flips. A per-resolution bill rises with every resolution, while a custom bill is mostly flat.
- Price and terms: a vendor can change its price or definition, and Salesforce's pricing page says it is "subject to change". On the custom side, providers retire models, as Anthropic's price list shows, so you pay for migration and retesting.
- Scope: our build figures are estimates, and your scope will move them.
Time and reasons other than cost
A first version scoped to one task with one or two integrations typically takes six to twelve weeks, and a vendor's agent can usually start sooner. Some reasons to build have nothing to do with fees: one chatbot across several systems, data kept in your country or on your own servers, or actions in systems the vendor's agent cannot reach. Our AI chatbot integration page covers those, and our guide to which platforms integrate custom AI chatbots sets out each platform's route. Those reasons can justify a build below the break-even, as long as you budget it as a capability, not a saving.
What we would do at each volume
- Under about 1,600 conversations a month: buy the vendor's agent. No custom option in our model beats a per-resolution price on cost.
- About 1,600 to 10,000: start with the vendor's agent unless you have a reason other than cost, measure your billed resolutions for a few months, then rerun the formula. A documents bot can win here against $0.99 or more. An agent that takes actions usually needs a higher vendor price to win.
- Above 10,000: compare a vendor quote and a custom quote side by side. Custom usually wins against list prices of $0.99 or more, while a $0.50 price stays competitive, especially against agents that take actions. Ask the vendor for a better price first: Intercom says enterprise discounts may be available, and Zendesk says committed usage costs less per resolution than pay-as-you-go.
We build custom chatbots and connect them to Zendesk, Salesforce, HubSpot, WhatsApp, Genesys Cloud and other systems, with the chatbot and connectors built as one project. If a platform's own agent is the better buy, we say so. The first call is free: bring your monthly conversations, billed resolutions and plan, and we can run this formula together. Our AI chatbot integration page explains how a project runs.
Frequently asked questions
How many conversations a month does a custom AI chatbot need to beat a per-resolution agent?
By our arithmetic, at 50% resolved, about 3,400 to 17,800 against $0.99 a resolution, 2,200 to 10,300 against $1.50, and 8,100 to 75,000 against $0.50. The low end is a documents bot at our low estimate; the high end is an agent that takes actions at our high estimate.
Do per-resolution agents win for a small business?
On cost, usually yes. Below about 1,600 conversations a month, every per-resolution list price we checked beats every custom option in our model, and up to a few thousand they still beat most builds. At those volumes, build a custom chatbot only for something the vendor's agent cannot do, not to save money.
Does a cheaper model change the answer?
Mostly for agents that take actions. Against $0.99, a small model moves the action agent's break-even from about 9,200 to 17,800 conversations to about 6,200 to 11,900, if it passes your tests on real conversations. A documents bot's break-even moves by only about 10%.
Do WhatsApp fees change the break-even?
Not if the vendor's agent also answers on WhatsApp, because the fee is owed on the delivered message either way. Meta's Business Agent is a different case, billed at $2.00 per million tokens, which Meta puts at about 4 to 5 cents a message. Since 15 January 2026, WhatsApp's terms also bar general-purpose AI assistants from the API when AI is the primary function, though businesses using AI to serve their own customers are still allowed.
Are list prices the real prices at high volume?
Not always. Intercom says enterprise discounts may be available through sales. Any discount raises the break-even: 20% off a $0.99 price moves the range from about 3,400 to 17,800 conversations to about 4,400 to 26,500 (our arithmetic). Get the quote before you compare.
Can you start on a vendor's agent and move later?
Yes, and at low volume we think that is the sensible order. You learn your real volume, your resolution rate and the questions customers ask, which are the inputs this comparison needs.
Sources
- Intercom: Pricing, checked 5 November 2026.
- Intercom: Fin pricing, checked 5 November 2026.
- Intercom: Fin, checked 5 November 2026.
- Zendesk: Pricing plans, checked 5 November 2026.
- Zendesk: Understanding outcome-based pricing, checked 5 November 2026.
- Zendesk: About automated resolutions for AI agents, checked 5 November 2026.
- Zendesk: About Sunshine Conversations in Zendesk Suite, checked 5 November 2026.
- Zendesk: Sunshine Conversations subscription and usage, checked 5 November 2026.
- HubSpot: AI Customer Agent, checked 5 November 2026.
- HubSpot: Service Hub pricing, checked 5 November 2026.
- Salesforce: Agentforce pricing, checked 5 November 2026.
- Gorgias: Official information about Gorgias, checked 5 November 2026.
- Gorgias: Pricing, checked 5 November 2026.
- Freshworks: Purchase and manage Freddy AI add-ons, checked 5 November 2026.
- Anthropic: Claude pricing, checked 5 November 2026.
- OpenAI: API pricing, checked 5 November 2026.
- Meta: WhatsApp Business Platform pricing, checked 5 November 2026.
- Twilio: WhatsApp pricing, checked 5 November 2026.
- Twilio: WhatsApp price table, CSV, checked 5 November 2026.
- TechCrunch: WhatsApp changes its terms to bar general-purpose chatbots, checked 5 November 2026.
- DigitalOcean: Droplet pricing, checked 5 November 2026.
- DigitalOcean: Managed database pricing, checked 5 November 2026.
- DigitalOcean: GPU Droplet pricing, checked 5 November 2026.
- KORE1: Tech contractor hourly rates 2026, checked 5 November 2026.
Key takeaways
- Below about 1,600 conversations a month, every per-resolution list price we checked beats every custom option in our model, so buy the vendor's agent.
- Against a $0.99 price at 50% resolved, custom breaks even at about 3,400 to 17,800 conversations a month, depending on what the chatbot does (our arithmetic).
- Against a $0.50 price, the break-even is about 8,100 to 75,000 conversations a month, so the rule of thumb of a few thousand resolutions is too broad.
- Build and maintenance, not model tokens, are the main custom cost below about 10,000 conversations a month.
- A vendor's agent leaves most of the performance risk with the vendor, while a custom build gives you a flatter bill and carries that risk yourself.