AI Chatbot for Business: What It Resolves and What It Costs

A team buys an AI chatbot to cut ticket volume. Six months later the widget answers policy questions competently and every request that needs something to actually happen still lands in a human queue. Volume is unchanged. The pilot is quietly not renewed.

That outcome is common, and it is almost never a model problem. An AI chatbot for business succeeds or fails on what sits around the model: what knowledge it can reach, whether it can act in your systems, and how it behaves when it is unsure.

Gartner predicts agentic AI will autonomously resolve 80% of common customer service issues without human intervention by 2029, with a 30% reduction in operational costs. Getting there depends on choices made before deployment, not after.

An AI chatbot for business is worth buying when it resolves requests rather than routing them, and what it costs depends on the billing unit more than the sticker price. Both are below. Start free at enjo.ai to test one on your own tickets.

TL;DR

• An AI chatbot for business is worth buying when it resolves requests end to end, not when it answers well and routes anyway.
• Cost is decided by the billing unit, not the sticker: per seat, per resolution, per conversation, or per reply produce very different invoices at the same volume.
• At 2,000 AI-handled conversations a month, published rates put the same workload anywhere from $295 to roughly $2,000.
• The three questions that decide it: what knowledge can it reach, what can it do in your systems, and what does an escalation carry to the human.
• Start on one high-volume request type, measure autonomous resolution on it, expand when the number holds.

What an AI Chatbot for Business Actually Is

An AI chatbot for business understands natural language, retrieves accurate information from connected knowledge sources, and performs structured actions such as creating tickets, triggering workflows or updating records. Unlike rule-based bots, it combines LLM reasoning, real-time retrieval and deterministic execution to resolve requests across websites, Slack and Microsoft Teams. The distinction is set out in full in our chatbot vs conversational AI comparison.

From Scripted Trees to Reasoning Systems

Early chatbots followed strict decision trees: if the user says X, respond with Y. They broke easily, needed manual scripting, and could not interpret natural language. They were useful only for predictable questions.

Modern systems use large language models that parse intent from open-ended questions, reason through multi-step instructions, hold context across a conversation, and adapt to unusual phrasing, typos and internal jargon.

Language models alone do not make a reliable chatbot. What enterprises need is structured retrieval, deterministic workflows and guardrails around the model. The result is closer to a service interface: a programmable layer between people and systems.

How It Learns Your Organization

An AI chatbot does not learn your organization the way a model trains on a dataset. It becomes effective by connecting to three things: retrieval sources such as Confluence, SharePoint, Notion and product docs; workflow systems such as Jira, ServiceNow, HRIS and identity tools; and the surfaces where people actually ask, meaning the website, Slack and Microsoft Teams.

The capability emerges from three layers working together. The retrieval layer pulls the correct policy or answer in real time, using embeddings, metadata and permission checks. Conversation memory tracks context, identity and the evolving problem state across a session. Workflow execution runs the actions: create tickets, route approvals, reset passwords, check device health, update records.

That combination moves a system from answering "what is our VPN policy" to completing "create a Jira incident and attach the troubleshooting logs." See Enjo's full integrations list for what that connects to in practice.

Which Type of AI Chatbot Your Business Needs

"AI chatbot" now covers three genuinely different things.

A website chatbot is customer-facing. It answers product and policy questions, captures leads, deflects routine support requests and escalates the rest with context attached. Sessions are short and transactional.

A Slack or Teams agent is internal. It lives where employees already work, creates tickets inline, retrieves internal answers with permission checks, and runs approval or access flows. Our guide to AI agents for Slack and Teams covers that internal surface in depth, including identity mapping and approval flows.

A workflow engine is back-end. It executes deterministic multi-step tasks, connects actions in sequence, and guarantees auditability. This is where enterprise-grade accuracy comes from.

Most organizations end up running all three. The question worth asking a vendor is whether those three share one retrieval layer or three separate ones, because that decides whether a customer and an employee get the same answer to the same question.

Where AI Agents Fit

AI agents sit one layer above chatbots. A chatbot interprets a question and returns an answer. An agent carries a request from intent to outcome: it works out the sequence of steps, executes them across connected systems, and verifies each completed before reporting back.

In IT support that is the difference between replying "try resetting your VPN" and reviewing the diagnostic logs, inferring the likely cause, triggering the workflow, updating the Jira or ServiceNow ticket with structured context, and notifying the right team if a person is still needed. The user sees one conversation. Underneath, the agent orchestrates work across several tools.

That is the difference between answering and resolving. Our guide to AI support agents covers the agent layer in depth.

What an AI Chatbot for Business Actually Delivers

The case for automation is usually made in adjectives, which is why it collapses at the first budget review. The version that survives is made in numbers a finance team can check a quarter after launch, and there are four of those worth committing to.

The first is the share of requests that never reach a person. Aurora reached 63% autonomous resolution across frontline IT requests, and Aptean handles 300K+ cases a year with 37% fully resolved by AI. Everything else on this list follows from that number, and no other metric substitutes for it.

The second is what coverage costs. Delivery Hero runs employee support across 70+ countries with 30% deflection and responses 80% faster than before, which turns a staffing problem into a configuration one. Nobody is rostered against a time zone that generates eleven tickets a week.

The third is where the team spends its hours. Employees ask variations of the same 30 to 40 topics, and once the AI holds those, the volume equivalent of 120 agents at Aptean is handled by AI rather than hired for. The people who stay work the cases that need judgment, which is also the work they wanted.

The fourth is quieter and shows up in the ticket queue rather than a dashboard. A chatbot that captures intent, attaches the logs and selects the right issue type removes the triage step entirely, so nobody reads a message purely to decide who should read it. Because every retrieval query and workflow execution is logged, the same system that makes support faster is also the one that makes it auditable.

How Much Does an AI Chatbot for Business Cost Per Month?

Published rates put a business AI chatbot between roughly $50 and $2,000 a month for a ten-person support team handling around 2,000 AI conversations. The spread is not a quality difference. It is the billing unit, and there are four of them in the market.

Per seat. You pay for every human agent whether or not the AI resolved anything. Zendesk Suite Team is $55 per agent per month on annual billing, Suite Professional $115, with Suite Enterprise plus Copilot quote-only. The Copilot add-on is $50 per agent per month on top, and the AI agents themselves bill separately on automated resolutions at a rate Zendesk does not publish on its pricing page. Freshdesk runs $19, $55 and $89 per agent per month on annual billing.

Per resolution or outcome. You pay each time the AI finishes something. Fin charges $0.99 per outcome with a 50-outcome monthly minimum, so about $49.50 is the floor before anything else, plus $29 per helpdesk seat per month if you run it on Intercom rather than another helpdesk. The model is only forecastable once you know your resolution rate, and almost nobody does before they deploy.

Per conversation or per action. Salesforce Agentforce sells Flex Credits at $500 per 100,000, with a standard action consuming 20 credits, so $0.10 an action, and a voice action 30 credits. A conversation is however many actions it takes.

Per session. Freshworks meters Freddy AI Agent in sessions: 500 included once per account, then $49 per 100 sessions, and unused sessions expire at the end of the billing cycle. Freddy AI Copilot is a separate $29 per agent per month on Pro and Enterprise.

Per reply. You pay for responses the AI actually sends. This is how Enjo bills, with unlimited human agent seats and unlimited AI agents on published plans, so the invoice scales to support volume rather than headcount.

What that costs at 2,000 AI conversations a month

Ten human agents, 2,000 conversations reaching the AI, 60% resolved without a person. Published rates only:

• Per outcome at $0.99: 1,200 resolutions, $1,188 a month, no seat cost on a non-Intercom helpdesk.
• Per seat plus sessions: ten Freshdesk Pro seats at $55 is $550, plus 1,500 chargeable sessions at $49 per 100 is $735. About $1,285 before Copilot.
• Per action: at ten actions a conversation, 20,000 actions at $0.10 is $2,000.
• Per seat plus unpublished resolution meter: ten Zendesk Suite Team seats is $550, and the AI resolution charge is not published, so this line cannot be budgeted from the pricing page.
• Per reply: Enjo Standard is $295 a month for 3,000 AI Replies, which covers the 2,000 above with room.

The arithmetic that matters is not this month. It is what happens at twice the volume, because three of the five models above rise in step with success.

Building it in-house

A prototype is the retrieval pipeline. The system is edge cases, hallucination control, escalation logic, audit trails, SOC 2 evidence, RBAC, multi-source knowledge sync and permanent model maintenance as foundation models change underneath you. Published research puts the expensive components at access control, SSO, compliance logging and multi-source ingestion rather than retrieval, with ongoing personnel cost for a single system at $75,000 to $150,000 a year (RagAboutIt, 2025).

What Enjo costs

The Free plan is $0 a month with 200 AI Replies, unlimited human agent seats, unlimited AI agents, unlimited channels, Inbox, Help Center and Insights, with no credit card. At the limit, requests auto-escalate to human agents in Inbox, so nothing is dropped and the escalation model proves itself before you pay for anything. Paid plans start at $95 a month for 1,000 AI Replies, and the full detail is published at enjo.ai/pricing.

Whichever model a vendor uses, one question separates the sales deck from the invoice: what does a resolved request cost at twice our current volume? The answer is usually available in thirty seconds if the vendor is comfortable giving it.

The Top 5 AI Chatbots for Business, Compared

Five tools cover the realistic shortlist for a business buyer, and they split on where they run and what they bill for: Enjo (AI replies, permanent free tier), Fin (per outcome, any helpdesk), Zendesk AI agents (seats plus an unpublished resolution meter), Freshdesk Freddy AI Agent (seats plus expiring session packs), and Salesforce Agentforce (Flex Credits per action, inside the Salesforce platform).

Enjo resolves requests across Slack, Microsoft Teams, website chat and the Help Center portal, reads Confluence, SharePoint, Google Drive, Notion, Guru and past tickets into one index, and takes action in Okta, Jira, ServiceNow, Salesforce and custom APIs. Enjo Platform is built agent-first: the AI Support Agent is configured before any queue exists, and Inbox is the helpdesk for escalations and human collaboration rather than the primary work surface.

Fin is a single-product AI agent for customer service messaging, sold standalone and as a layer on Salesforce, HubSpot, Freshworks and others. It bills per outcome with a 50-outcome monthly minimum, and Copilot for human agents is a separate $35 per user per month.

Zendesk AI agents run inside Zendesk on Zendesk knowledge and Zendesk objects, included with every Suite and Support plan. The per-resolution rate is not published on the pricing page, so the AI line cannot be modelled before a sales conversation.

Freshdesk Freddy AI Agent handles conversations on email and chat inside Freshworks. Sessions are sold in packs of 100 at $49 after the first 500, and unused sessions expire each billing cycle rather than rolling over.

Salesforce Agentforce is Salesforce's native agent layer, scoped to the Salesforce platform and billed on Flex Credits, where a standard action is 20 credits and a voice action is 30, so cost tracks workflow complexity rather than conversation count.

Where It Delivers, by Function

Policy and answer retrieval. Employees ask variations of the same 30 to 40 topics: VPN access, leave policy, procurement rules, device allowances, security protocols. A precise chatbot identifies the exact policy segment, applies permission filters, returns a context-aware answer with a link to the source, and flags low-confidence responses for review.

Ticket creation and updates. Ticketing is an expensive bottleneck because people dislike portals, which means ticket data arrives incomplete. A chatbot turns conversation into structured work: it captures intent, attaches screenshots or logs, selects the right issue type, enriches metadata, and updates as troubleshooting progresses until the request is resolved.

HR queries, onboarding and approvals. Parental leave, tax documents, day-one onboarding tasks. For approval-heavy flows the chatbot collects the data, requests approval, notifies stakeholders and updates the HRIS.

Troubleshooting and device support. Understand the issue, suggest verified steps, collect logs or device metadata, trigger backend checks, and escalate with pre-filled technical context. This removes the early diagnostic work that tier-one agents traditionally do.

Lead capture and customer questions. Interpret buying intent, answer product and integration questions, separate support from sales, capture lead information and route qualified users onward.

What This Looks Like in Production

Three deployments show the same approach against three different constraints.

At Aurora, the constraint was that engineers would not leave Slack to file a ticket. Employee IT support runs where they already work, 63% of frontline requests resolve without a human, and the rest escalate with the full conversation, account context and suggested next steps attached.

At Aptean the constraint was knowledge scatter rather than volume. 5M+ documents across Salesforce case feeds, knowledge bases and SharePoint were unified into a single index before any automation was switched on, and the result is 300K+ cases a year with 37% fully resolved by AI. That is the volume equivalent of 120 agents handled by AI, which is a headcount the company never had to add as it grew.

At Delivery Hero the constraint was languages and time zones. Support runs across 70+ countries and 95,000+ Slack members, with 30% deflection and responses 80% faster, because the knowledge layer answers in the requester’s language without a separate team behind each one.

What the three have in common matters more than what separates them. Each started with one high-volume request type, measured autonomous resolution on it, and expanded only when the number held. None of them began with a broad rollout, and the ones that do are the pilots that quietly do not renew.

How an AI Chatbot Actually Works

A modern AI chatbot works because five layers operate together, and a deployment fails at whichever one is weakest. Most teams are further from production than the adoption numbers suggest. Forrester's 2026 assessment found roughly three-quarters of enterprise leaders reporting agentic AI adoption while only a small minority ran it in meaningful production beyond chatbot-like use.

Language Understanding

Models interpret messages, extract intent and generate responses. They are good at identifying what someone wants, parsing messy phrasing, summarizing, and guiding troubleshooting steps. Without grounding, they guess. The reliable pattern is hybrid: the model understands the question, retrieval fetches the source of truth, the model rewrites it in context, and deterministic logic executes the task and verifies the outcome.

The Retrieval Layer

This is the backbone. It connects to Confluence, SharePoint, Notion, PDFs, API-based sources, product documentation and policy pages, and it needs three capabilities. Structured indexing, so documents are chunked, embedded and tagged with metadata covering permissions, product area, last-updated date and classification. Permission-aware filtering, so people see only what they are allowed to see, filtered at query time rather than from a static snapshot. And real-time freshness, because policies change and documentation evolves, which means continuous indexing rather than periodic syncs. That last one is the clearest dividing line between consumer chatbots and enterprise-ready systems: one answers from current organizational truth, the other from a frozen copy of it. Our AI knowledge base guide covers how to structure that layer.

Deterministic Execution

Most legacy chatbots failed because they could not act, not because they misunderstood. An execution engine calls APIs reliably, creates and updates tickets in Jira or ServiceNow, triggers approvals, provisions access, posts updates to Slack or Microsoft Teams, and verifies each step with a success or failure check. Determinism matters because actions must be predictable, auditable, reversible, logged and governed. A model can recommend the next step; it should not improvise the execution. Without this layer, enterprises get silent failures and compliance gaps. Agentic workflows covers how these sequences are built and constrained.

Conversation State and Identity

The system has to know who is asking, what they asked before, and which systems they may reach. That means identity mapping from a Slack or Teams account through the internal directory to permissions, so finance policies reach finance staff only. It means state management across multi-turn conversations, holding intent, the selected ticket, unresolved sub-steps and workflow progress. It means context window management, keeping the active conversation small but relevant. And it means channel-aware behavior, because a website session is short, a Slack session is threaded, and a Teams session may be a group channel.

Security Controls

Enterprises need specific controls, not a general assurance: single sign-on through Okta, Azure AD or Google Workspace for identity consistency across surfaces; role-based access control fine-grained enough to separate general employees, IT admins, HR leaders and workflow creators; audit logs recording retrieval queries, workflow executions, ticket actions, escalations and admin changes; encryption at rest, in transit and across private-link deployments; and governance workflows so configuration changes follow approval and logging.

Enjo runs on SOC 2 Type II, ISO 27001 and GDPR compliance, with Okta SSO, granular RBAC and full audit trails. Guardrails enforce what an agent may say and do.

Channels and Why Behavior Should Differ

AI chatbots operate as a distributed layer across surfaces, each with its own interaction model. Deployments fail when one generic widget is dropped into places where the interaction pattern does not fit.

On a website, users want concise answers, minimal friction and a clear next step. A good website chatbot summarizes, avoids nested clarifying questions, and routes to the right action without pushing. Sessions are often under 30 seconds. Our ai website chatbot guide covers setup and integrations.

In Slack or Microsoft Teams, employees expect threading and persistence. A good internal agent holds state, references earlier messages, runs commands inside the thread and adapts to the workspace's pace. The real value is behavioral: people stop context-switching to a portal. Our Jira and Slack integration guide covers the IT case.

Inside workflow execution, correctness matters more than personality. The engine verifies each step, logs actions, handles retries and prefers reliability over creativity.

The unification point is the retrieval layer. A customer asking a pricing question on the website and an employee asking about discount eligibility in Slack should not get contradictory answers. One knowledge layer across surfaces is what prevents that.

How It Integrates Into Support Workflows

An AI chatbot is operationally valuable only when it integrates with the systems that already anchor support. It should behave like a workflow orchestrator rather than a parallel channel.

Multi-Step Flows

Agentic behavior is often described as autonomy. In practice it is more disciplined: the chatbot decomposes a request into steps, executes them in order, verifies outcomes and keeps a complete audit trail.

A well-governed flow has clear preconditions, deterministic transitions, built-in error recovery and user-facing transparency about which step is in progress. That replaces the legacy pattern where one access request generates a week of email and ticket back-and-forth.

What to Automate and What to Hand Off

Over-automation is among the most common deployment failures. The last two rows are where it starts. The cost of getting the boundary wrong is not one bad answer. Gartner's September 2026 survey of 3,566 customers found only 27% would try a chatbot again after a negative experience, so a confident wrong answer on an ambiguous request costs more than a handoff.

A Four-Week Rollout

The order matters more than the speed.

Week one, connect the channel and the knowledge. One surface, one knowledge source, answers only. Nothing automated yet. The goal is to see what the retrieval layer can and cannot answer before any action is wired.

Week two, add triage rules. Classify and route every request on arrival, so no human reads a message purely to decide who should read it. Priority signals get caught here too.

Week three, roll out the second surface. Website and Slack or Teams on the same retrieval layer, so the answers match. This is where answer drift appears if the layers are separate.

Week four, automate the first workflow. One deterministic multi-step flow with explicit fallbacks and a complete audit trail. Then measure resolution before adding the next request type.

Automating before triage works means automating requests you have not yet classified, which is how pilots produce impressive dashboards and unchanged queues.

Governance and Accuracy Tuning

Accuracy does not hold on its own. Policies change, workflows shift and integrations break.

Each row needs a named owner or it is nobody's job.

Insights reports which request types resolve, which escalate, and which gaps recur.

Common Challenges and How to Avoid Them

Deployments rarely fail because the model was wrong. They fail in five recognizable ways, and each one has a fix that costs less than the failure does.

The most common is a chatbot answering from a frozen copy of your documentation. The policy changed in March, the index was built in January, and the answer is confidently wrong in a way nobody notices until a requester escalates. Continuous indexing rather than periodic syncs fixes it, and citations on every answer make a stale source visible the moment it is used.

The second is automating before triage works. Automating a request type you have not yet classified produces a dashboard that looks impressive while the human queue stays exactly as long as it was. Classification and routing come first, automation second, which is why the rollout above spends a week on each.

The third is over-automation at the boundary. Gartner’s September 2026 survey of 3,566 customers found only 27% would try a chatbot again after a negative experience, so a confident wrong answer on a sensitive HR case is not one bad answer, it is a requester who stops using the system. Decide explicitly which request types the AI may attempt, and hand off everything past that line with full context attached.

The fourth is a permission leak, and it is the one with real consequences. If identity is not mapped from the Slack or Teams account through to the internal directory and role, a finance policy reaches someone who should never have seen it. Filtering has to happen at query time rather than from a static snapshot, because the snapshot is always a little behind the org chart.

The fifth is slower and harder to see. Policies change, integrations break and content goes stale, and accuracy bends downward from roughly month three unless every review area has a named owner. A governance table nobody is accountable for is a document, not a control.

Start With One Request Type

Point an AI chatbot at one high-volume, well-defined request type, measure autonomous resolution on it, and expand when the number holds.

Enjo has a permanent free tier with 200 AI Replies a month and unlimited seats, which is enough to run week one and week two of the rollout above before any procurement conversation. Full pricing is published, with no credit card and no sales call to get started.

Frequently Asked Questions

What is an AI chatbot for business?

An AI chatbot for business understands natural language, retrieves answers from your connected knowledge sources, and performs structured actions such as creating tickets or triggering approvals. The distinction from a consumer chatbot is that it operates inside your systems with permission checks and an audit trail.

How does an AI chatbot work?

In four layers. A language model interprets the request. A retrieval layer pulls relevant context from connected sources such as Confluence and SharePoint. An orchestration layer decides whether to answer, ask a follow-up or trigger a workflow. An action layer executes tasks in tools like Jira or ServiceNow. Guardrails and audit logs run across all four.

What is the difference between an AI chatbot and a traditional chatbot?

Traditional chatbots follow scripted decision trees and break on anything off-script. AI chatbots use language models and retrieval to interpret novel questions and execute multi-step actions. In practice, traditional bots deflect simple questions while AI chatbots resolve requests.

How much does an AI chatbot for business cost?

Pricing models vary more than headline prices, so compare the billing unit rather than the sticker. Per-seat, per-resolution, per-conversation and per-reply models produce very different invoices at the same volume. Enjo publishes its pricing, with a permanent free tier at 200 AI Replies a month. Ask any vendor for the cost per resolved request at twice your current volume.

Can an AI chatbot replace human agents?

No. It resolves repetitive, rule-driven requests such as password resets, policy lookups and access provisioning, which is where Gartner expects 80% of common cases to land by 2029. Ambiguous, sensitive or judgment-heavy issues still need people. The working pattern is hybrid: AI takes the predictable volume and escalates the rest with full context attached. If you are shortlisting vendors, see our guide to the best AI chatbots for customer service.

How much does an AI chatbot cost per month?

Between roughly $50 and $2,000 a month for a ten-person team handling around 2,000 AI conversations, depending entirely on the billing unit. Per-outcome pricing at $0.99 puts 1,200 resolutions at $1,188. Seat-based pricing plus metered AI sessions lands near $1,285 on Freshdesk Pro. Flex Credits at $0.10 an action reach $2,000 at ten actions a conversation. Enjo Standard is $295 a month for 3,000 AI Replies, and the Free plan is $0 for 200.

What are the top 5 AI chatbots for business?

Enjo, Fin, Zendesk AI agents, Freshdesk Freddy AI Agent and Salesforce Agentforce cover the realistic shortlist. They differ less on answer quality than on where they run, which knowledge sources they read, whether they can act in other systems, and what they bill for. Enjo bills per AI reply with unlimited human agent seats and a permanent free tier at 200 replies a month.