An AI chatbot can cost less than $50 per month as self-service software, hundreds or thousands of dollars per month once usage and human seats increase, or substantially more when an enterprise deploys AI across high-volume service operations and business workflows. Custom development belongs to a separate category and can extend into six-figure projects.

Those figures are not contradictory. They describe different systems and, increasingly, different units of work.

A small-business chatbot may answer questions from a knowledge base. A customer-service AI agent may resolve support issues and hand exceptions to people. An enterprise agentic platform may authenticate users, retrieve account data, invoke tools, update records and execute workflows before it sends its final response.

That is why the most useful way to estimate AI chatbot cost is not to look for one market-wide average. Start with the work the AI must perform, identify what the vendor actually bills for that work, and then add the implementation and operating costs that do not appear in the advertised starting price.

Editorial independence and pricing disclosure: This article is independently researched and written. It is not sponsored, and the publisher has no affiliate relationships with the companies discussed. Company and product names are used solely for identification and comparison; their inclusion does not imply affiliation or endorsement. All trademarks belong to their respective owners.

Pricing and product information reflect publicly available sources reviewed on September 7, 2026, and may change. Prices are in USD unless stated otherwise. Billing terms, included usage and availability vary by product, region and contract. Monthly prices under annual billing are monthly equivalents. Cost examples use stated assumptions and are not vendor quotes, market averages or guarantees of costs, savings or performance. Confirm current terms directly with each vendor before purchasing.

AI Chatbot Cost at a Glance: SMB vs. Enterprise

AI chatbot pricing ranges from low-cost SaaS subscriptions to substantial enterprise operating budgets because vendors sell different capabilities and meter different units of work. SMB buyers often begin with subscription and usage costs. Enterprise buyers may also need to model human seats, agent actions, workflows, integrations, governance and ongoing operations.

Deployment profile	Possible pricing structure	Main cost drivers
Basic SMB chatbot	Monthly subscription or usage allowance	Plan limits, AI responses, message or conversation allowances
Growing support chatbot	Platform + human seats + metered AI usage	Agent seats, sessions, outcomes, resolutions or usage credits
Integrated enterprise support AI	Seats + AI consumption + implementation	Volume, integrations, security, data and support requirements
Agentic enterprise platform	Platform/access + usage tied to actions, workflows or credits	Task depth, tools, grounding, workflows and autonomous activity
Custom conversational AI	Development + infrastructure + model/API usage + operations	Engineering scope, architecture, data, integrations, testing and maintenance

These profiles are planning examples, not measured spending averages or restrictions based on company size. Agentic capabilities can also be available in products used by smaller businesses.

Current published pricing shows how different the entry points can be. Tidio lists Starter at $24.17 per month on annual billing, including only 50 one-time Lyro AI conversations; recurring Lyro usage requires a paid AI quota. Meanwhile, Chatbase lists its Hobby plan at $40 per month on monthly billing. Freshdesk starts at $19 per agent per month when billed annually, as a helpdesk/email-AI entry point rather than a web-chat AI package. Intercom starts at $29 per full seat per month on annual billing, with Fin AI usage charged separately.

But comparing those numbers alone tells you very little.

Chatbase uses message credits. Freshdesk meters additional Freddy AI Agent use in sessions. Intercom charges for Fin outcomes. Zendesk uses automated resolution allowances and tiers. Salesforce offers pricing structures based on conversations, resolutions, users or Flex Credits tied to agent actions. Microsoft Copilot Studio uses Copilot Credits, with different activities consuming different amounts.

There is no universal unit called “one chatbot interaction.”

Why AI Chatbot Prices Range From Almost Free to Six Figures

Part of the confusion around AI chatbot pricing comes from treating fundamentally different purchases as versions of the same product.

Self-service chatbot SaaS

At the simplest end, a company subscribes to hosted software that handles tasks such as FAQs, basic lead capture or knowledge-base answers.

The vendor operates most of the underlying platform. Setup may consist largely of supplying approved content, configuring instructions and adding the chatbot to a website or another channel.

This is where many of the lowest advertised prices come from.

The trade-off is that the advertised subscription is usually defined by limits: message credits, AI conversations, responses, users, data capacity or other allowances. The plan price therefore answers only the first budgeting question.

Customer-service platforms with AI

The next category is an existing support or customer-service platform with AI incorporated into the workflow.

The company may pay for human support seats while separately paying for or drawing down an included allowance for AI work. That is the basic structure visible in products such as Intercom, Freshdesk and Zendesk.

Here, the AI is not merely a website widget. It sits inside a broader operation containing support agents, tickets, knowledge, routing, escalation and reporting.

That matters because the cost is now partly driven by the human support system surrounding the AI.

Enterprise conversational and agentic platforms

At the enterprise level, conversational AI can extend beyond answering questions.

An AI system may retrieve information from company data, authenticate a customer, decide which process applies, invoke an external tool, update a business record or trigger a workflow. Some platforms also support autonomous agents whose activity is not initiated by a live customer conversation.

The conversation becomes an interface to business operations rather than the whole workload.

That changes the economics. The relevant question is no longer just:

How many conversations will the chatbot have?

It becomes:

How much work does the AI perform to complete each business task?

Custom-built conversational AI

Custom development is another purchasing category entirely.

Instead of primarily paying for access to a finished SaaS application, the organization takes responsibility for more of the engineering stack: application logic, integrations, model access, retrieval infrastructure, hosting, security, testing, monitoring and maintenance.

Current Clutch chatbot-development marketplace guidance gives indicative minimums of $1,000–$10,000 and enterprise examples of $100,000–$250,000+ in its FAQ. These are marketplace indications, not guaranteed quotes or representative completed-project costs; the lower bound does not establish the cost of a production-ready AI system.

Those figures illustrate the breadth of custom-development scope. They should not be read as an average market price: Clutch is a service-provider marketplace, and the projects represented differ substantially in architecture, integrations, data requirements and compliance needs.

The Four Variables That Determine What You Actually Pay

A practical cost framework is:

Real AI chatbot cost = platform/access + human seats + metered AI work + implementation/integration + required add-ons + ongoing operations

Four variables determine how those components behave.

1. How you acquire the system

First establish what you are actually buying.

A self-service chatbot subscription, a customer-service platform, a configurable enterprise agent platform and a custom software project are different acquisition models.

Enterprise buyers in particular now have a substantial middle path between “buy a chatbot” and “build everything yourself”: configure agents, workflows, tools and data connections on an enterprise platform while the vendor provides much of the underlying runtime.

That option can reduce the amount of infrastructure an organization must own without making implementation trivial.

2. What the vendor bills

Current chatbot and AI-agent platforms meter a range of units, including:

• human seats;
• AI-generated responses;
• message credits;
• conversations;
• sessions;
• outcomes;
• resolutions;
• actions;
• workflows; and
• broader consumption credits.

Even similar-sounding terms can have different definitions.

An Intercom outcome is not a Freshdesk session. A Salesforce resolution should not automatically be equated with an Intercom outcome. A Microsoft Copilot Credit and a Salesforce Flex Credit are unrelated billing currencies.

The billing definition has to come before the dollar comparison.

3. How much work the AI performs

Conversation volume is a useful starting point, but model choice, response length and billing definitions also matter. Agentic systems add variation from tool use, workflows and autonomous behavior. Count only the units the selected vendor actually meters.

One customer request might require one generated answer. Another may require the agent to retrieve account data, perform several reasoning or routing steps, call two tools, update a record and launch a workflow.

Both may appear as one request in a customer-service report.

They do not represent the same amount of AI work, but a reasoning step or tool call is not automatically a separately billable action.

For enterprise budgeting, therefore, workload should include task depth: how much vendor-metered work is required to complete an average business task.

4. How complex the deployment is

Finally, consider everything surrounding the AI. Cost can be affected by requirements such as:

• CRM and helpdesk integrations;
• order, account or ERP systems;
• identity management;
• permissions;
• retrieval and grounding;
• multilingual or voice channels;
• audit logs;
• data handling and residency;
• security reviews;
• SLAs;
• analytics;
• evaluation; and
• custom workflows.

These requirements often explain the difference between a modest-looking software price and a significantly larger first-year enterprise budget.

What Does an AI Chatbot Cost for an SMB?

The sources reviewed here do not establish a representative market-wide average SMB spend. Current vendor pricing combined with a transparent workload calculation is more defensible.

Basic website chatbot

A small business handling FAQs, lead capture or a modest knowledge base can enter the market in the tens of dollars per month.

For example:

Product example	Current published starting point	What matters beyond the price
Tidio Starter	$24.17/month, annual billing	50 one-time Lyro AI conversations; recurring AI use requires a paid quota
Chatbase Hobby	$40/month, monthly billing	Includes 700 message credits, not 700 conversations
Freshdesk Growth	$19/agent/month, annual billing	Helpdesk/email-AI entry point; separate session usage. Web-chat AI requires Freshdesk Omni
Intercom Essential	$29/full seat/month, annual billing	Fin outcomes are separately metered

The units matter immediately.

Chatbase explains that AI responses consume different numbers of message credits depending on the model. Its Hobby plan’s 700 credits therefore should not be rewritten as “700 conversations.”

Likewise, Tidio maintains separate usage concepts. Its documentation for monthly billable conversations says the human conversation quota covers person-to-person communication, while Lyro AI usage is separate.

A small company can therefore start cheaply, but the right question is still:

What will our actual workload consume?

Growing customer-support deployment

Costs change when a company adds human users and more AI-handled volume. Consider a transparent Intercom example.

Assume the company uses Essential on annual billing, needs three full seats and generates 300 qualifying Fin resolutions or Procedure handoffs for chat/email on Intercom per month.

Intercom currently lists Essential at $29 per full seat per month annually. Its Fin outcome documentation lists resolutions and Procedure handoffs for chat/email on Intercom at $0.99 each. A billed Procedure handoff is not necessarily an autonomously resolved issue.

The arithmetic is:

3 seats × $29 = $87/month

300 modeled outcomes × $0.99 = $297/month

Illustrative monthly-equivalent recurring subtotal = $384/month

This is a calculated example, not evidence that an SMB normally spends $384 per month.

It also excludes implementation labor, optional products, other channels, taxes and any different outcome mix.

More importantly, it demonstrates why the advertised $29 seat price is not the same as the operating cost of the AI deployment.

What Does an Enterprise AI Chatbot Cost?

There is no single vendor price that represents all enterprise deployments because enterprise is not a pricing model.

A large deployment can combine human licenses, AI consumption, agent actions, workflows, data access, integrations, security controls, implementation work and ongoing operational ownership.

That makes cost-stack thinking more useful than plan-price thinking.

Enterprise cost layer	What to budget for
Platform and seats	Base product, support-agent users and required plan tier
AI consumption	Outcomes, resolutions, sessions, conversations, credits or other metered use
Agentic work	Actions, workflows, tools, grounding or other activity triggered by tasks
Integration and data	Connections, data preparation, identity, knowledge and system access
Governance and security	Permissions, auditability, privacy, retention, testing and approval controls
Operations	Monitoring, evaluation, knowledge maintenance, troubleshooting and optimization

Platform and human seats

Traditional support-platform licensing can still be material.

Freshdesk publishes Enterprise at $89 per agent per month annually. Intercom lists Expert at $132 per full seat per month on annual billing. Zendesk currently lists Suite Professional at $115 per agent per month when paid yearly.

At dozens or hundreds of human users, the seat layer alone can become a meaningful recurring expense.

But seat count increasingly fails to explain the full economics of enterprise AI.

AI consumption

AI usage may be separately metered even where the product plan includes access to AI capabilities.

For example, Freshworks publishes $49 for each additional pack of 100 Freddy AI Agent sessions. Its add-on documentation clarifies that the initial 500-session allocation for new qualifying customers is a one-time complimentary allocation, not 500 new free sessions every month.

Zendesk illustrates another approach. Its 2026 automated-resolution documentation groups AI-agent resolutions into tiers funded by a resolution allowance. Zendesk introduced resolution tiers in May 2026 and is transitioning affected accounts to resolution allowances. Verify the account’s model and contract before applying historical per-resolution prices.

Because the current dollar economics can depend on the customer’s plan, allowance and contract, it would be misleading to invent a universal “Zendesk price per resolution.”

When an Enterprise Chatbot Becomes an Agentic Conversational Platform

The biggest difference in enterprise AI economics is not simply that more people are chatting with the system.

The AI may be doing more work.

A conventional generative chatbot might look like:

question → retrieve information → answer

An agentic conversational system can look more like:

request → understand goal → retrieve data → select tools → execute actions → verify result → continue workflow → respond

And some agents can run workflows in response to events without an active customer conversation.

That shift is visible directly in current enterprise pricing models.

Salesforce: actions can become the unit of work

Salesforce’s Agentforce pricing supports multiple commercial models rather than one universal “Agentforce price.”

For Flex Credits, Salesforce Help currently lists $500 per 100,000 credits and 20 Flex Credits for a standard Agentforce action. At that published rate, one standard action corresponds to $0.10.

The key point is not that “Agentforce costs ten cents.” It is that a business task can require multiple actions.

A service request could involve verifying the customer, retrieving an order, checking eligibility, updating a record and starting a downstream workflow.

The number of incoming conversations may remain unchanged while the billable work per conversation increases under action-metered pricing. Task depth affects usage cost when the selected billing model meters the additional work. Under a qualifying resolution or outcome model, several actions can fall within one billable result.

Microsoft: one interaction can consume multiple capabilities

Microsoft offers 25,000-Copilot-Credit packs at $200 per pack per month, alongside pay-as-you-go and pre-purchase options.

That does not mean a pack includes 25,000 conversations.

Microsoft’s Copilot Studio billing documentation lists standard metered consumption rates to different activities, including:

• classic answer: 1 Copilot Credit;
• generative answer: 2;
• agent action: 5;
• tenant graph grounding: 10; and
• additional rates for agent flows and AI tools.

Eligible employee-facing use under an authenticated Microsoft 365 Copilot license may be included, subject to licensing conditions and fair-use limits. External and autonomous usage must be assessed separately.

Microsoft gives the example that a grounded generative response can consume both the grounding and generative-answer components within one interaction.

That makes task depth a practical budgeting variable.

Conversation volume and task depth are separate

For a conventional chatbot, a buyer might begin with:

monthly conversations × relevant billing rate

For an agentic system, a more useful intermediate measure is:

Agentic workload = business tasks × average vendor-metered work per task

The exact meaning of “work” depends on the platform. It may translate into actions, credits, requests, workflows or other consumption units.

Two organizations can therefore handle 20,000 AI-assisted customer requests per month and still have very different costs.

One system may answer questions.

The other may authenticate users, query proprietary data, call tools and change records.

SMB vs. Enterprise AI Costs Side by Side

The following SMB and enterprise profiles are illustrative planning examples, not measured spending averages or restrictions based on company size. A high-volume SMB can have enterprise-like consumption, while a very large company can deploy a narrowly scoped chatbot.

Cost dimension	Illustrative bounded deployment	Illustrative complex enterprise deployment
Primary job	Answer or route relatively bounded requests	Answer, resolve and potentially execute business processes
Billing profile	Subscription, messages, credits or conversational usage	May combine seats, outcomes, actions, workflows and consumption pools
Human users	Small team	Seat costs can be material across large operational teams
Task depth	Bounded in this example	One request may require several systems or actions
Integrations	Standard connectors may be enough	Business-critical and proprietary systems may be involved
Knowledge/data	Smaller or simpler source set	Multiple governed repositories and permission boundaries
Security	Standard SaaS controls may suffice	SSO, granular permissions, auditability and data controls may be requirements
Deployment	Configuration-led	Architects, developers, system owners and business stakeholders may be involved
Ongoing operations	May be handled alongside another role	Evaluation, monitoring and governance may require dedicated ownership
Cost predictability	Easier with bounded use	Depends on both volume and work performed per task

The useful dividing line is therefore not employee count.

It is operational complexity and workload.

How the Major AI Chatbot Pricing Models Differ

No pricing model is universally cheapest.

Pricing model	What is billed	Main advantage	Main budget risk
Fixed/tier subscription	Access to a plan	Predictable starting price	Allowance may be too small for production use
Per-seat	Human user	Simple to model	Cost rises with team size
Per-response/message credit	AI-generated activity	Flexible at modest volume	Model and response patterns change consumption
Per-conversation	Vendor-defined interaction	Connects cost to usage	Conversation rules vary
Per-session	Time-bounded interaction	Can bundle multiple exchanges	Session does not equal a ticket or customer
Per-outcome	Vendor-defined result	Links billing to a defined result	Outcome definitions and types differ
Per-resolution	Vendor-defined successful resolution	Aligns with automation success	Qualification and tier rules differ
Per-action/workflow credit	Work performed by an agent	Exposes operational work	One request can trigger many units
Consumption pool	Shared credits/capacity	Flexible across capabilities	Difficult to forecast without workload modeling
Hybrid	Two or more models	Supports complex platforms	Several independent cost drivers can compound

The right question is not:

Which vendor has the lowest price per unit?

It is:

How many of that vendor’s units will our workflow consume?

Why Billing Definitions Matter More Than the Advertised Rate

This is where otherwise sensible vendor comparisons become misleading.

Intercom outcomes

Intercom’s current Fin documentation says only one Fin outcome is billed per conversation even when Fin takes multiple actions. For chat/email on Intercom, resolutions and Procedure handoffs cost $0.99; sales disqualifications also cost $0.99, while qualifications cost $9.99. A billed handoff does not necessarily mean the issue was resolved autonomously.

Therefore:

$0.99 per outcome ≠ $0.99 per conversation in every use case.

Freshdesk sessions

Freshworks defines an AI Agent session using time windows.

In Freshdesk Omni, a Chat AI Agent session covers interactions within 24 hours of conversation start. The Email AI Agent uses a 72-hour window from the first customer email. Standalone Freshdesk prices should not be treated as prices for the Omni chat product.

Therefore:

one session ≠ one response, one ticket or necessarily one customer.

Chatbase message credits

Chatbase says each AI-generated response consumes message credits based on the model used. Current documentation shows model-dependent consumption ranging from lower-cost one-credit responses to substantially higher credit usage for some models.

Therefore:

700 credits ≠ 700 conversations.

Tidio human and AI conversations

Tidio’s human conversation quota and Lyro AI usage are separate.

Therefore:

a Tidio billable conversation and a Lyro AI conversation should not be collapsed into one allowance.

Zendesk automated resolutions

Zendesk’s current resolution system measures AI-agent usage through automated resolutions and groups resolutions into tiers.

Zendesk introduced resolution tiers in May 2026 and is transitioning affected accounts to resolution allowances. Verify the account’s model and contract before applying historical per-resolution prices.

Salesforce conversations, resolutions and actions

Salesforce publishes multiple commercial structures, including conversations, Help Agent resolutions and action-oriented Flex Credits. Under qualifying resolution pricing, multiple actions can be included in one billable result.

Even where two structures happen to share the same headline dollar amount, they are not the same billing event.

Microsoft Copilot Credits

Microsoft’s billing documentation describes credits that can represent different combinations of answers, actions, grounding, flows and tools.

Therefore:

25,000 credits ≠ 25,000 chatbot conversations.

The procurement rule is simple:

Never compare two chatbot prices until you know precisely what event causes each vendor to bill you.

Hidden Costs That Change the Chatbot Budget

The vendor invoice is only part of total cost of ownership.

AI consumption	Vendor bill	Recurring	Varies with volume and whichever additional work the model meters
Human seats	Vendor bill	Recurring	Can become material for larger teams
Implementation	Vendor, partner or internal labor	Mostly upfront	Configuration and deployment require real work
Integrations	Vendor, partner and/or internal	Both	Connections must be built and maintained
Knowledge/data preparation	Often internal or services	Both	AI quality depends on reliable source material
Security/governance	Platform + internal effort	Both	More access and autonomy require stronger controls
Evaluation/observability	Platform and/or internal	Recurring	AI behavior needs ongoing testing and monitoring
Tool/API consumption	Vendor or third party	Recurring	Agentic workflows may use other paid systems
Human escalation	Internal cost	Recurring	Automation does not eliminate exception handling
Change management	Internal labor	Upfront + ongoing	Processes and staff responsibilities change

Microsoft’s AI-management guidance supports planning for lifecycle monitoring, data and model maintenance, governance and cost oversight. Count internal labor once if it is already included in implementation or managed-service costs.

This distinction prevents a common budgeting error:

vendor invoice ≠ total economic cost.

Knowledge-base cleanup is a good example. A vendor may not charge separately for the employee hours spent cleaning source material, resolving contradictory policies and setting permissions, but the organization still incurs that cost.

Agentic deployments add another concern: cost can be created by the agent’s behavior itself.

AWS’s Agentic AI Lens recommends per-cycle, per-task and per-day budget controls because autonomous agents can enter reasoning loops, repeatedly invoke tools or accumulate expensive context.

AWS separately recommends tracking costs by agent, workflow, reasoning phase and tool invocation, including metrics such as cost per task completion.

Conversation volume is a useful starting point, but model choice, response length and billing definitions also affect cost. Agentic systems add variation from tool use, workflows and autonomous behavior. Count only vendor-metered units, and put controls around activity that can create additional consumption.

Worked AI Chatbot Cost Scenarios

The most defensible calculation method is:

business workload → vendor-specific billing translation → calculated cost

The following scenarios use published prices and explicit assumptions. They are illustrations, not market averages or vendor quotes.

Scenario 1: Small support team using Intercom

Assumptions

• 3 Essential full seats
• annual billing
• 300 qualifying Fin resolutions or Procedure handoffs per month for chat/email on Intercom

Published price

• $29/full seat/month
• $0.99 per modeled resolution or Procedure handoff

Calculation

3 × $29 = $87/month in seats

300 × $0.99 = $297/month in Fin usage

Calculated monthly-equivalent recurring subtotal: $384/month

Sources: Intercom plans and billing cadence; Fin outcome definitions and rates.

Excluded: implementation, other products, channels, taxes and different outcome types.

The point of the scenario is not the $384 total. It is that AI consumption can outweigh the apparent seat price even in a small deployment.

Scenario 2: Growing support team using Freshdesk

Assumptions

• Freshdesk Pro
• annual billing
• 10 agents
• 2,000 Freddy Email AI Agent sessions per month
• initial one-time 500-session allocation already exhausted

Published price

• $55/agent/month annually
• $49 per additional 100 AI Agent sessions

Calculation

10 × $55 = $550/month in agent licenses

20 session packs × $49 = $980/month in AI usage

Calculated monthly-equivalent software/usage subtotal: $1,530/month

Sources: Freshdesk pricing; Freddy AI add-ons and session definitions.

This subtotal excludes implementation, taxes, other add-ons and Freshdesk Omni. It assumes all 20 packs are purchased and consumed in the applicable billing cycle. Auto-recharge pack sizing is inconsistent in the vendor documentation; verify it in the account before enabling auto-recharge.

A key qualification remains: 2,000 Freshdesk sessions do not necessarily equal 2,000 customers, tickets or individual AI responses.

Scenario 3: Enterprise agentic workload using action-based pricing

Now model a different type of deployment.

Assumption

• 20,000 AI-handled business tasks per month

Illustrative gross standard-action consumption at published list rates, assuming no included credit balance or discounts is applied. Sources: Salesforce Agentforce pricing and standard-action Flex Credit rates:

• 20 Flex Credits per standard action
• $500 per 100,000 Flex Credits
• calculated standard-action price: $0.10

Now vary task depth.

Assumption	Actions per task	Monthly actions	Calculated usage cost
Simple workflow	2	40,000	$4,000/month
Moderate workflow	4	80,000	$8,000/month
Deeper workflow	7	140,000	$14,000/month

At four standard actions per task:

20,000 tasks × 4 actions = 80,000 actions

80,000 × 20 Flex Credits = 1.6 million Flex Credits

1.6 million ÷ 100,000 × $500 = $8,000/month

Calculated annualized consumption: $96,000

The request volume did not change between the three rows.

Only the work performed per request changed.

That is the economic difference between simple chatbot-volume modeling and agentic workload modeling.

The example excludes implementation, Data 360 and other Salesforce services, voice, human licenses where applicable, support and third-party systems. It models gross consumption, not a minimum invoice for 20,000 tasks; included or promotional credits, discounts and negotiated terms may change the amount billed.

First-Year Enterprise TCO Is More Than 12 Monthly Invoices

Recurring software cost and first-year total cost of ownership are different numbers.

A useful framework is:

First-year TCO = one-time implementation + 12 months of recurring platform/consumption + ongoing operating cost

A November 2025 Forrester Total Economic Impact study of Agentforce for Customer Service provides a useful example of how those categories can combine.

The study was commissioned by Salesforce. Forrester interviewed six decision-makers and constructed a composite organization, and its disclosure explicitly says readers should use their own estimates rather than assume the modeled results will apply to another company.

For the composite organization, the risk-adjusted model included:

• $40,496 in initial implementation and training-related costs;
• $165,000 in Year 1 Agentforce consumption; and
• $37,663 in Year 1 ongoing management within the combined implementation, training and management category.

The model therefore shows:

Initial + Year 1 costs = $243,159

This is the undiscounted sum of the study’s initial and Year 1 modeled cost categories, not a complete quote for a new deployment or a present-value figure. Some interviewees received discounted implementation support; additional Data Cloud costs may apply.

That figure is not an enterprise AI-agent benchmark. The useful finding is what is inside it.

The study’s interviewees reported implementation timelines ranging from two weeks to six months, with implementation teams involving roles such as administrators, architects, developers, application managers and customer-service managers. Ongoing management among the interviewed organizations ranged from 0.05 to 0.75 FTE, covering activities such as support, troubleshooting and performance monitoring.

This supports a more realistic procurement question:

Who will connect, test, monitor and improve the AI after we buy the software?

Packaged Chatbot vs. Agent Platform vs. Custom Build

For enterprise buyers, “buy versus build” is no longer quite enough. Microsoft’s cost-and-benefit evaluation framework also considers extending existing solutions; the options below are a planning framework whose fit depends on the workflow and requirements.

Path	What the buyer receives	Best suited to	Main trade-off
Packaged chatbot/SaaS	Prebuilt application with bounded configuration	Standard FAQs, support and lead workflows	Less architectural flexibility
Configurable agent platform	Vendor runtime plus tools for agents, actions, workflows and integrations	Deeper enterprise automation without owning the whole AI stack	Greater implementation and governance responsibility
Custom build	Organization controls substantially more of the application and architecture	Proprietary workflows, unusual constraints or strategic system control	Potentially substantial engineering and operating responsibility

When packaged SaaS makes sense

Packaged software is easiest to justify when:

• the workflow is common;
• the necessary integrations already exist;
• fast implementation matters;
• standard security controls are sufficient; and
• the business does not need to own the underlying architecture.

When an agent platform makes sense

A configurable platform becomes attractive when the business needs to:

• connect enterprise data;
• create custom actions;
• orchestrate workflows;
• expose agents through several channels;
• apply corporate identity and governance; or
• expand into more agentic use cases over time.

This path can reduce the need to build foundational infrastructure while still requiring real implementation work.

When custom development makes sense

Custom development becomes more plausible when:

• workflows are highly proprietary;
• existing platforms cannot support essential integrations;
• deployment architecture is constrained;
• unusual control over models or data is required; or
• the business logic itself creates strategic differentiation.

There is no reliable universal conversation or action threshold at which custom development suddenly becomes cheaper.

And model inference should not be mistaken for custom-system TCO. OpenAI’s current API pricing, for example, illustrates that foundation-model use can be metered separately by input, cached input, output and other product-specific units. Those charges are only one layer beside engineering, hosting, retrieval, integrations, monitoring, evaluation and maintenance.

How to Compare AI Chatbot Vendors Without Being Misled by Starting Prices

Ask every shortlisted vendor the same questions:

1. What exactly creates a billable event? Is it a message, session, conversation, outcome, resolution, action or credit?
2. What usage is actually included? Distinguish access to the AI feature from included AI consumption.
3. Can one customer request create multiple billable actions or units?
4. What happens when an action fails or retries?
5. Are AI charges separate from human-agent seats?
6. Are retrieval, grounding, workflows, tools or external-model use billed separately?
7. Does autonomous or background agent activity consume the same allowance?
8. Are development, testing or evaluation activities billable?
9. Which integrations, security controls and governance features require a higher tier or separate implementation?
10. What would our invoice look like for one representative business task at current, 2× and expected 12-month volume?

Is an AI Chatbot Worth the Cost?

ROI should be calculated from the economics of the target workflow rather than a generic industry savings percentage.

A useful starting formula is:

Net value over the assessment period = avoided costs + monetized capacity/service benefits + incremental contribution profit − total incremental deployment and operating costs

Count each benefit once. To express ROI as a percentage, divide net value by total incremental costs and multiply by 100. Keep capacity gains distinct from realized payroll savings, and use contribution profit rather than revenue alone for incremental sales benefits.

Potential value can include:

• fewer human-handled interactions;
• shorter handling time;
• more service capacity;
• reduced backlog;
• deferred hiring;
• expanded service hours; or
• incremental contribution profit where the use case supports it.

But labor efficiency is not automatically cash savings.

In a hypothetical example, if an AI system saves 1,000 employee hours and the company keeps the same staffing level, the benefit may appear as increased capacity, faster response or avoided future hiring rather than an immediate payroll reduction.

Likewise, failed automation can create:

• escalations;
• rework;
• quality-review effort;
• customer recovery work; or
• financial and operational risk.

That is why the business case should model the specific workflow using variables such as:

• monthly task volume;
• successful AI completion rate;
• escalation rate;
• human handling time;
• cost of human handling;
• incremental business value;
• implementation cost;
• recurring TCO; and
• error/rework cost where material.

The research for this article did not identify a sufficiently strong independent benchmark showing that every chatbot deployment reduces costs by one universal percentage.

That number should therefore be calculated from the organization’s own operating data.

Frequently Asked Questions

How much does an AI chatbot cost for a small business?

A basic small-business chatbot can begin in the tens of dollars per month using current SaaS products. For example, Tidio Starter is $24.17 per month on annual billing with only 50 one-time Lyro AI conversations; recurring Lyro usage requires a paid AI quota. Chatbase Hobby is $40 per month on monthly billing with 700 message credits, whose consumption depends on the model. Real cost depends on the included allowance and how the vendor meters AI activity, so the starting subscription should not be treated as the complete monthly budget.

How much does an enterprise AI chatbot cost?

There is no single vendor price that represents all enterprise deployments. Enterprise deployments can combine platform licenses, human seats, high-volume AI consumption, actions or workflows, data access, implementation, security and ongoing operations. Some products publish consumption pricing while other enterprise components require sales-assisted or negotiated pricing.

Can one chatbot conversation generate multiple usage charges?

Yes, depending on the billing model. Some vendors meter a conversation or one outcome, while agentic platforms may measure several actions, capabilities or credits used to complete the work behind a single request. The vendor’s exact definition determines whether one interaction creates one or several billable units. For example, Intercom bills only one Fin outcome per conversation, whereas Microsoft documents additive metered capabilities, subject to licensing inclusions.

What is per-resolution chatbot pricing?

Per-resolution pricing bills based on a successful resolution as the vendor defines it. The qualification rules are platform-specific. A Zendesk automated resolution, Salesforce Help Agent resolution and Intercom outcome should not automatically be treated as interchangeable units.

Is flat-rate or usage-based chatbot pricing cheaper?

Neither is universally cheaper. Fixed pricing can improve predictability when the workload fits comfortably inside the allowance. Usage pricing can be economical for low or variable demand but may rise quickly at scale. The correct comparison requires translating the same business workload into each vendor’s billing system.

Is it cheaper to build or buy an AI chatbot?

Buying packaged software is generally the simpler route when the required workflow already exists in the product. A configurable agent platform provides a middle option for organizations that need custom tools and workflows. Full custom development may make sense where proprietary requirements justify greater engineering and operating responsibility. There is no universal volume at which building becomes cheaper.

Conclusion

The difference between SMB and enterprise AI chatbot cost is not simply that one buyer is small and the other is large.

The cost model changes with the work.

A bounded SMB chatbot may primarily consume messages, credits or conversations. A growing customer-service operation may combine human seats with sessions, outcomes or resolutions. An enterprise agentic system may make the conversation only the starting point, with additional economics determined by tools, actions, workflows, autonomous tasks, integrations and governance.

The most reliable budgeting sequence is therefore:

purchase model → billing unit → workload → task depth → deployment complexity → total cost of ownership

SMB buyers should focus on whether the product supports the required workflow at predictable usage levels.

Enterprise buyers should model the entire operating environment around the AI—not only how many people talk to it, but what the system does after the conversation begins.

References

1. Intercom. “Pricing.” https://www.intercom.com/pricing
2. Intercom Help Center. “Fin and Intercom Plans Explained.” https://www.intercom.com/help/en/articles/9061614-fin-and-intercom-plans-explained
3. Intercom Help Center. “Fin AI Agent Outcomes.” July 30, 2026. https://www.intercom.com/help/en/articles/8205718-fin-ai-agent-outcomes
4. Freshworks. “Freshdesk Pricing.” https://www.freshworks.com/freshdesk/pricing/
5. Freshworks Support. “Manage Freddy AI Add-ons.” Updated August 4, 2026. https://crmsupport.freshworks.com/support/solutions/articles/50000011515-manage-freddy-ai-add-ons
6. Chatbase. “Pricing.” https://www.chatbase.co/pricing
7. Chatbase. “Frequently Asked Questions.” https://www.chatbase.co/docs/faq/faq
8. Tidio. “Pricing.” https://www.tidio.com/pricing/
9. Tidio Help Center. “The Monthly Conversations Limit.” https://help.tidio.com/hc/en-us/articles/11874534946460-The-monthly-conversations-limit
10. Zendesk. “Pricing.” https://www.zendesk.com/pricing/
11. Zendesk Help. “About Automated Resolution Tiers.” Accessed September 7, 2026. https://support.zendesk.com/hc/en-us/articles/9570369117338-About-automated-resolution-tiers
12. Zendesk Help. “Upgrading From Automated Resolutions to Resolution Allowances.” Accessed September 7, 2026. https://support.zendesk.com/hc/en-us/articles/10479528943130-Upgrading-from-automated-resolutions-to-resolution-allowances
13. Salesforce. “Agentforce Pricing.” https://www.salesforce.com/agentforce/pricing/
14. Salesforce Help. “Agentforce Pricing.” May 19, 2025. https://help.salesforce.com/s/articleView?id=004811240&language=en_US&type=1
15. Microsoft. “Copilot Studio Pricing.” https://www.microsoft.com/en-us/microsoft-365-copilot/pricing/copilot-studio
16. Microsoft Learn. “Billing Rates and Management — Microsoft Copilot Studio.” https://learn.microsoft.com/en-us/microsoft-copilot-studio/requirements-messages-management
17. Clutch. “Top Chatbot Companies / Chatbot Development Cost Guidance.” Accessed September 7, 2026. https://clutch.co/developers/artificial-intelligence/chatbots
18. OpenAI. “API Pricing.” https://developers.openai.com/api/docs/pricing
19. Forrester Consulting. “The Total Economic Impact™ of Agentforce for Customer Service.” Commissioned by Salesforce. November 2025. https://tei.forrester.com/go/Salesforce/Agentforce/?lang=en-us
20. Amazon Web Services. “AGENTCOST07-BP01: Implement Automated Cost Controls With Intelligent Cutoffs.” https://docs.aws.amazon.com/wellarchitected/latest/agentic-ai-lens/agentcost07-bp01.html
21. Amazon Web Services. “AGENTCOST05-BP01: Establish Agent-Level Reasoning Cost Tracking and Attribution.” https://docs.aws.amazon.com/wellarchitected/latest/agentic-ai-lens/agentcost05-bp01.html
22. Salesforce. “Agentforce Help Agent Announcement.” https://www.salesforce.com/news/stories/agentforce-help-agent-announcement/?bc=OTH&ver=1786463721
23. Microsoft Learn. “Manage AI.” https://learn.microsoft.com/en-us/azure/cloud-adoption-framework/ai/manage
24. Microsoft Learn. “Evaluate Costs and Benefits of AI Solutions.” https://learn.microsoft.com/en-us/training/modules/evaluate-costs-benefits-ai-powered-business-solution/