How AI customer support is priced: seats, conversations or credit
pricing · · 5 min
AI customer support is priced three ways: per seat, per conversation handled, or on usage credit drawn down as models run. Each moves a different risk onto you. OnDuty meters systems, people and monthly credit, and a model call you run through your own assistant spends none. Here is what each meter rewards and where it surprises you.
Three ways AI customer support gets priced
Most AI customer support is sold on one of three meters: a seat for each person who uses it, a fee for each conversation the AI handles, or credit that is drawn down as models run. The meter decides who carries the risk, so it is worth reading before you read the price.
Per seat
A seat price is predictable and easy to approve. It also charges you for headcount rather than work, so a team of three with heavy volume looks cheap and a team of thirty with light volume looks expensive. It fits a business where the number of people using the tool is stable and the AI is a helper to them.
Per conversation or per resolution
This meter follows volume, which sounds fair until volume is the success you wanted. The bill grows with every customer the AI talks to. If the price is per resolution, ask who defines resolved, because the party holding the definition also holds the invoice.
Usage credit
Credit is closest to the real cost, since every model call has one. The trade is a bill that moves. It needs three answers in writing: whether unused credit expires, what happens at zero, and whether anything you run yourself is metered.
How OnDuty prices it
OnDuty meters three things: systems, people and monthly credit. Every plan includes the whole Brain, the layer that holds one record per customer across the systems a business runs, so a lower plan means fewer systems and people, not a thinner Brain. Current prices are read live from Stripe on the pricing page and are deliberately not repeated here.
A person is anyone with app access or an MCP key. The customers your agents talk to never count, so a busy week of customer conversations never adds a seat.
Credit pays for the model calls OnDuty runs for you: agents answering customers, and the Ask chat. Your own ChatGPT or Claude reading the Brain over the Model Context Protocol (search, fetch and the context call) spends no credit on any plan. That is the one meter we chose not to run, because the assistant a team already pays for should not be charged twice.
Included credit resets on the first of the month and does not roll over. Purchased credit lasts 12 months. Auto top-up is on by default, which is worth knowing before the first month ends.
WhatsApp is a capability, one switch rather than a quantity. It is an add-on on Pro and Team and is included from Operations upward. Enterprise has no published price on purpose: residency, guardrails written with you and dedicated support are negotiated, and the pilot described on the enterprise page is six weeks, one workflow, one team, with success criteria written before week one.
What happens when credit runs out
A support agent that goes quiet mid-conversation is the failure every credit model has to answer for. On OnDuty, when credit runs out an agent hands the customer chat to a person with a holding message. The conversation is not dropped and the customer is not left in silence. Your own assistant over MCP keeps working, since it spends no credit.
That choice has a cost: a person has to be there to take the chat. If nobody can staff the handoff, set a top-up before the credit runs out, not after.
Questions to ask any vendor before you compare prices
The same six questions expose the real price of any of the three meters.
- What exactly is billable: a seat, a conversation, a message, a resolution or a model call?
- Who defines a resolution, and can you see the cases behind each one?
- Does included usage roll over, and does purchased usage expire?
- What does the AI do at the limit: stop, keep going and bill, or hand off to a person?
- Is the assistant you already pay for metered when it reads your data?
- Are there setup fees or a minimum term, and are they waived on annual billing?
Why we publish no savings figure
You will not find a cost per conversation or a percentage saved on this site. OnDuty publishes a number as an outcome only when it comes from an instrumented deployment it can name, and none exists yet. What the pricing page does show is an estimate of what a month of credit buys, labeled as an estimate and drawn from typical usage or from a workspace's own usage, never presented as a result. Use it to size a month, then run a week and read your own usage.
Which meter fits which business
Choose by the shape of your load. If volume is steady and the team is small, a seat price is easy to budget. If volume spikes with campaigns or seasons, a per-conversation price punishes your best weeks, and a credit model lets you size a month and top up when you need to. If several AI tools will read the same customer data, prefer the vendor that does not charge you for your own assistant reading it.
And if you only need one bot answering one list of questions on one channel, you do not need a Brain. A plain chatbot on a per-seat price will serve you better than we will.
Questions
How much does AI customer support cost?
It depends on the meter, which is why the meter matters more than the headline price. OnDuty has a Free plan that needs no card, and paid plans priced by systems, people and monthly credit. The current prices are on the pricing page, read live from Stripe.
Is there a free plan?
Yes, with its own limits and no card needed. The Brain's chat is on every plan, and an unknown plan resolves to Free, never to unlimited.
Does credit expire?
Included credit resets on the first of the month and does not roll over. Purchased credit lasts 12 months. Auto top-up is on by default.
What happens when credit runs out?
An agent hands the customer chat to a person with a holding message. Your own assistant reading the Brain over MCP keeps working, because it spends no credit.
Why does Enterprise have no published price?
It is negotiated, because residency, guardrails written with you and dedicated support differ for every company. The pilot is six weeks, one workflow and one team, with success criteria written before week one.
When would you tell me not to buy OnDuty?
If you need one chatbot on one channel, a seat-priced chatbot is the better buy. If you need an AI that answers from inside your own network, that is not built: conversations are still processed in OnDuty's cloud. And if you need a SOC 2 report before you can sign, there is none yet.
Updated