A support lead reviews the monthly bill and sees a low subscription price. Then they open the conversation logs. Many chats were handed to a human, several ended without an answer, and the plan covered fewer customer questions than expected. A chatbot support cost model shows the gap between the price on the pricing page and the support coverage you actually received.
This guide is for support and CX leads forecasting monthly AI support spend. The useful output is cost per supported conversation, with AI usage, human handoffs, and unresolved conversations kept separate.
What a chatbot support cost model should measure
Start with this calculation:
Cost per supported conversation = total support cost / supported conversations
The denominator needs a clear definition. A supported conversation may be one that the agent resolved, one that a human completed after handoff, or both. Pick the definition before comparing plans.
Then split the monthly cost into three parts:
- AI usage cost: message credits, model usage, or another billable unit used while the agent handles the conversation.
- Human handoff cost: the people, seats, or support capacity needed when the agent passes the thread to your team.
- Unresolved conversation cost: chats that consume usage but do not produce an answer or a completed support outcome.
A low headline price is useful only when you know how much customer support coverage it buys. If a plan handles 500 conversations but your team must finish 200 of them manually, the subscription price is only one part of the cost.
The billing units behind chatbot support costs
Chatbot pricing usually measures one of six things. The same support queue can produce very different bills depending on the unit.
| Billing unit | What it measures | What changes the bill | Question to ask before signing |
|---|---|---|---|
| Seats | Human users, AI agents, or both | Team size and the number of agents you deploy | Does a seat mean a human teammate, an AI agent, or either one? |
| Messages | Customer messages, AI replies, or both | Average turns in each conversation | Does the platform count the customer message, the AI reply, or the pair as one unit? |
| Conversations | A chat thread or session | Conversation volume and session rules | When does a conversation start and end, and does a reopened chat count again? |
| Resolutions | Chats marked completed or solved | Resolution rate and the vendor’s definition of resolved | Does a human handoff count as a resolution, and who can mark it resolved? |
| Usage | Model requests, credits, compute, or another metered resource | Model choice, response length, and request volume | What does one unit buy, and what happens when the included allowance runs out? |
| Flat monthly tier | A fixed plan with included limits | The tier you select and whether demand passes its limits | Which limits apply to agents, messages, training, handoff, and actions? |
Seat-based pricing
Seat pricing can look predictable because the monthly amount follows team size. Check whose activity counts. Some plans charge for human agents. Others also limit the number of AI agents or websites you can operate.
A seat plan becomes harder to forecast when a support lead needs seasonal teammates, a second AI agent, or access for an implementation team. Ask for the price of each additional seat and whether a viewer has the same cost as someone who replies to customers.
Message-based pricing
Message pricing needs a precise definition. A customer message is one unit of activity from the visitor. An AI reply is another possible unit. A complete conversation might contain six customer messages and six AI replies, which is twelve billable messages if the platform counts both sides.
That difference makes average conversation length more useful than visitor count. Two sites can have the same number of visitors and very different message bills if one needs a short answer and the other has long troubleshooting threads.
Conversation and resolution pricing
Conversation pricing charges for a thread or session. The rules matter. A session may end after a period of inactivity, when the visitor closes the widget, or when the agent hands the chat to a person.
Resolution pricing is even more dependent on definitions. You need to know whether a resolved chat means the agent answered a question, the visitor clicked a helpful button, or a human marked the issue complete. Also check what happens when a visitor reopens the conversation. A reopened chat may be a new billable event or part of the original case.
Usage-based pricing and monthly tiers
Usage-based pricing meters a resource such as model requests or message credits. Model choice can change the cost of each request. OpenAI’s API pricing is one example of a model-usage pricing page. Its rates are separate from a hosted support platform’s subscription, so do not combine the two without checking how the platform passes model costs to you.
A flat monthly tier includes a set allowance. That allowance may cover message credits, websites, training data, human handoff, team members, or Agent Actions. A tier is easy to budget when your volume stays below the limits. It becomes less useful when peak demand pushes you into an upgrade or an add-on.
How the model works in a real support conversation
Consider a customer asking, “Where is my order?” The agent replies with the order-status process. The customer then provides an order number, asks about a delayed shipment, and receives another reply. If the customer asks for a human, the thread moves to the support team.
For a platform that counts both sides, the four customer messages and four AI replies in that exchange use eight message units. The handoff is a separate operational cost because a person still needs to review the thread and finish the case. If the conversation ends without an answer, those eight units remain usage even though the support outcome is unresolved.
This is why visitor count is a poor forecast input by itself. Count the messages in completed conversations, the messages in handed-off conversations, and the messages in unresolved conversations separately. The same widget traffic can produce different bills when customers need more turns to reach an answer.
What good cost coverage looks like
There is no universal benchmark for a good handoff rate or unresolved rate in the information available for this forecast. Treat the figures below as illustrative assumptions, not performance benchmarks.
Assume 1,000 monthly conversations, four user messages per conversation, and four AI replies per conversation. Assume an 18% handoff rate and a 7% unresolved rate. That produces 8,000 message credits, 180 handoffs, and 70 unresolved conversations.
1,000 × (4 + 4) = 8,000 credits
1,000 × 18% = 180 handoffs
1,000 × 7% = 70 unresolved conversations
A useful coverage forecast reports all three outputs. It does not call the 82% that avoided handoff a resolution rate, because 7% of the conversations in this example still have no completed support outcome. Your own conversation logs should replace these assumptions before you use the model for a buying decision.
Build the forecast in three separate cost buckets
Use three worksheets or three columns in one forecast. Do not let a high deflection rate hide a costly handoff queue.
AI usage cost + human handoff cost + unresolved conversation cost
= total monthly support cost
For a message-credit model, collect these inputs:
- Monthly conversations
- Average customer messages per conversation
- Average AI replies per conversation
- Handoff rate
- Resolution rate
- Unresolved conversation volume
- Price of the plan and any extra credits
AssistLoop counts each user message and each AI response as one message credit. The credit cost for the selected model is shown before you choose the model. That lets you test the effect of model mix instead of treating every message as identical.
Estimate credits like this:
Monthly credits = conversations × (average user messages + average AI replies)
For example, assume 1,000 monthly conversations, an average of four user messages, and an average of four AI replies. The estimate is 8,000 credits:
1,000 × (4 + 4) = 8,000 credits
If 18% of those conversations need a human, that is 180 handoffs. If 7% remain unresolved, that is 70 conversations that used AI capacity without producing a completed answer. The AI usage estimate is still 8,000 credits. Handoff and unresolved volume sit beside it because they create different costs for your team.
A dollar example needs the current plan and model price. Use the plan price, included credits, and any extra-credit purchase price from the AssistLoop pricing page, then divide the resulting monthly cost by supported conversations. Do not divide by all widget sessions if many sessions are empty, unresolved, or handed off without completion.
Last verified: August 2026. Pricing, limits, and usage terms should be checked again before publication or purchase.
Why usable coverage matters more than the plan price
AssistLoop’s Free plan includes 150 message credits per month and no human handoff. That makes it suitable for testing whether your training content covers common questions. It does not cover a support process that needs live escalation.
For a paid tier, compare the limits that affect your queue:
| AssistLoop plan | Annual-billing equivalent | Message credits per month | AI agents | Human handoff | Training documents | Credits roll over? |
|---|---|---|---|---|---|---|
| Free | $0 | 150 | 1 | No | 1 MB | No |
| Basic | $25/mo | 3,000 | 1 | Yes | 50 MB | No |
| Pro | $109/mo | 15,000 | 2 | Yes | 50 MB | No |
| Business | $325/mo | 50,000 | 5 | Yes | 50 MB | No |
Last verified: August 2026. The AssistLoop pricing page shows annual-billing equivalent prices above. Monthly billing is higher. Credits do not roll over. AssistLoop also lists 1,500 extra credits for a one-time $10 purchase, an extra agent for $5 per month, and an extra team member for $12 per month.
The Free plan can be the right choice for a content test. It is the wrong choice when your success measure includes a human taking over the conversation. For that workflow, handoff availability belongs in the forecast from the start.
Unused capacity matters too. If your plan includes 15,000 credits and you use 9,000 in a quiet month, the remaining 6,000 do not move into the next month. A seasonal spike can then push you toward an upgrade or an extra-credit purchase even though the previous month looked underused.
Model peak-month demand separately from average demand. A plan that fits your yearly average may still leave the support team exposed during onboarding campaigns, product launches, or holiday traffic.
The checks that prevent a misleading chatbot cost estimate
Put each billing definition in writing before you compare plans. The words below should have a precise answer in the contract or pricing documentation:
- Message
- Conversation
- Resolution
- Seat
- Model request
- Handoff
Then check how the platform behaves in the cases that affect your queue. Does selecting a different model change usage cost? Are overages available, or does the agent stop when the allowance is used? Do credits expire at the end of the billing cycle? Can unused capacity roll over?
Ask what happens when the agent cannot answer. Can a visitor request human handoff? Does the team receive the full conversation history? Does the handoff include useful context such as the customer’s earlier messages? Does an unresolved chat still count toward usage? These details affect coverage even when they do not appear in the headline price.
Review the ceilings on the workflow itself. Count AI agents, training sources, team members, integrations, and actions. A low-cost tier can become unusable if it cannot hold the content you need, give access to the people who answer escalations, or call the endpoint that completes the support task.
For a plain explanation of metered billing, see Stripe’s guide to usage-based pricing models. Stripe uses usage-based pricing to describe billing tied to consumption. The exact unit still belongs to each provider, so use the guide for the concept and the provider’s current pricing page for the terms.
Source checked: August 2026.
Apply the model to AssistLoop before you buy
Start with message credits. Estimate monthly conversations, multiply by the average number of user messages and AI replies, then compare the result with the plan allowance. Run a second version for your busiest month. This catches the difference between a plan that fits average usage and one that covers the queue when demand rises.
Next, identify the conversations that need people. AssistLoop paid plans include human handoff. The team receives the full thread and context in the shared inbox, so the handoff cost belongs in your support forecast rather than in a vague assumption that the agent will answer everything. Review the handoff volume and the unresolved conversations in your conversation logs.
Answer coverage depends on the training sources you give the agent. AssistLoop can train on website content, uploaded PDF, DOCX, and TXT files, pasted text, and exact Q&A pairs. Train your agent on your own data to organize those sources around the questions your customers actually ask. Do not attach a guessed deflection rate to the setup. Measure answer coverage in your own conversation logs.
Agent Actions are a separate cost-coverage question. If the agent needs to capture a lead, book a meeting, or call an API for order status, check the Agent Actions feature. An answer-only workflow and a workflow that completes an API request can have different support value even when both use the same number of message credits.
Use this decision rule: choose the plan that covers expected usage and the conversations that still need people. Do not choose the plan with the lowest advertised monthly price.
Check the current AssistLoop pricing, then create your AI agent if the limits fit your forecast.
Last verified: August 2026.
FAQ
What is the most useful metric in a chatbot support cost model?
Cost per supported conversation is the most useful starting point. Calculate it after separating AI usage, human handoffs, and unresolved conversations so the average reflects the support outcome.
How do message credits affect chatbot support costs?
A platform may count customer messages, AI replies, or both. Average conversation length therefore matters more than visitor count when estimating monthly usage.
Is a free chatbot plan enough for customer support?
AssistLoop’s Free plan includes 150 message credits per month and no human handoff. It can test whether your content answers common questions, but it does not cover a support process that needs live escalation.
What should you verify before comparing chatbot pricing?
Verify the definition of each billing unit, usage limits, model costs, handoff rules, rollover terms, and overage behavior. Those details determine how much support coverage the advertised price provides.
How often should a chatbot support cost forecast be updated?
Review it monthly while volume and handoff rates are changing, and update it whenever conversation volume, model mix, handoff rate, or plan pricing changes. Recheck pricing before making a purchase decision.
Pricing claims in this article were last verified: August 2026.
Written by