AI agents and chatbots have the widest cost spread of any software category in 2026. The same business need, such as “answer customer questions and update our CRM”, might be quoted at $1,500 by a freelancer and $80,000 by an agency. Both can be honest numbers. They describe different products, different risk levels and different amounts of testing.
This guide explains what is in each price, what you pay each month after launch, and how to compare quotes. For broader context, see how much it costs to build an app.
AI agent development cost at a glance
| Project type | US agency | Freelancer or small team | AI-native fixed price (Tantito) |
|---|---|---|---|
| FAQ or website chatbot (rule-based or light AI) | $5,000-$20,000 | $500-$5,000 | not offered separately |
| Chatbot grounded in your documents (RAG) | $30,000-$80,000 | $1,200-$12,000 | $2,400 |
| Single-workflow agent | $15,000-$40,000 | $1,200-$6,000 | $2,400 |
| Multi-tool agent with actions in your systems | $30,000-$150,000 | $5,000-$25,000 | $6,500 |
| Enterprise agent platform (compliance, many integrations) | $72,000-$300,000 | n/a | scoped by quote |
Sources: Groovyweb’s 2026 guides cite $30,000-$80,000 for LLM chatbots with $400-$6,000 per month to run, and $30,000-$150,000 for agents, with enterprise builds at $72,000-$300,000. Fiverr’s guides put average chatbot gigs at roughly $216-$371, and single-workflow agent gigs run $1,200-$2,500 plus $400-$800 per month according to gig-economy writeups. The single-workflow agency and the small-team rows are our estimates interpolated from those ranges. Tantito prices are from our pricing page. These are vendor-published numbers; read them as anchors.
Chatbot vs agent: what you are actually buying
The words are used loosely, and cost follows the difference.
- A chatbot answers questions. Good ones are grounded in your own documents so they do not invent policy. The cost is mostly in ingesting and structuring your content, testing answer quality, and wiring the chat widget or channel.
- An agent also acts. It can look up an order, update a record, book an appointment, send an email or hand off to a human. Every action is a tool the agent can call, and each tool needs permissions, input checks, error handling and tests. That is why an agent costs two to three times what a chatbot does.
Ask any vendor which one they are quoting. A “chatbot” quote that quietly excludes actions, handoff and testing is a different product from an “agent” quote that includes them.
Cost by builder type
| Builder | Typical hourly rate | What you get at the low end | What you get at the high end |
|---|---|---|---|
| US agency | $100-$200 | Prototype with limited testing | Production agent with evals, monitoring and security review |
| Western Europe | $70-$100 | Similar to US at 20-40% lower cost | |
| Eastern Europe | $40-$60 | Competent builds, good value for defined scope | |
| Latin America | $35-$55 | Same, often with time-zone overlap for US clients | |
| India | $20-$50 | Wide quality range | |
| Freelancer | $30-$130 | Notebook-style demo or no-code flow | Solid single-workflow agent |
| AI-native fixed-price studio | fixed | Single-workflow agent from $2,400 | Multi-tool agent at $6,500 |
Hourly ranges follow the sources cited in our app cost guide (Bestarion, Arc, Uvik, Capital Numbers, Lemon.io). The quality notes are our judgment, not measurements.
What drives AI agent cost
1. Number of tools and systems it can act on
Each connected system (CRM, helpdesk, calendar, payment, email, database) needs an integration, scoped permissions and failure handling. We estimate $1,000-$5,000 per system at agency rates. Read-only lookups are cheaper than actions that change data.
2. Accuracy and risk tolerance
A bot suggesting a recipe can be wrong occasionally. A bot quoting insurance terms or issuing refunds cannot. High-stakes agents need larger test sets, approval steps for risky actions and human review of edge cases. This is where much of the agency price goes, and it is not something AI coding removes.
3. Your data
Clean, current documents are cheap to ground an agent on. Contradicting PDFs, outdated wiki pages and scattered spreadsheets cost time to untangle. Budget for content cleanup before you budget for the model.
4. Channels
A website widget is the baseline. WhatsApp, Slack, SMS and email each add setup, provider fees and conversation-handling quirks. WhatsApp Business also has its own per-conversation pricing from Meta.
5. Guardrails and handoff
Refusing off-topic requests, resisting prompt injection, escalating to a human and logging conversations are the difference between a demo and something you can put in front of customers.
6. Evaluation and testing
Serious builds include a test suite of real questions with expected answers, run on every change. It is invisible to users and a frequent line item that cheap quotes omit.
7. Ongoing operations
Models change, documents change and users find new ways to break things. Plan for monitoring, prompt and knowledge updates, and spend tracking.
Monthly running costs
Build price is half the picture. After launch you pay for:
| Cost | Typical range | Notes |
|---|---|---|
| Model usage (tokens) | A few cents per conversation (our estimate) | Varies by model, prompt length, tools used |
| 1,000 conversations per month | about $10-$80 in model fees (our estimate) | |
| 10,000 conversations per month | about $100-$800 in model fees (our estimate) | Caching and smaller models reduce this |
| Hosting, database, vector search | $0-$100 per month at small scale (our estimate) | |
| Channel fees (WhatsApp, SMS) | Per-message or per-conversation | Check the provider’s current rates |
| Monitoring and maintenance | $299-$990 per month at Tantito; $400-$6,000 at agencies |
Model prices change frequently, so check the provider’s pricing page (Anthropic, OpenAI) before you commit. The per-conversation numbers above are our rough estimates for support-style conversations of a few hundred words, and they swing widely with design choices.
A practical safeguard: a cost cap per conversation and a monthly spending alert on your model account. Without it, a loop bug or an abusive user can run up a bill quickly.
How AI coding agents changed the cost of building agents
Building an AI agent used to mean a lot of custom plumbing: orchestration code, retrieval pipelines, connectors, chat interfaces and evaluation harnesses. Open frameworks and AI coding agents shortened all of it.
What got cheaper. The scaffolding around the model, such as chat UI, API clients, document ingestion, logging, dashboards and tests, is the kind of well-specified work that coding agents do quickly under review. A 2023 controlled experiment found developers using GitHub Copilot completed a task 55.8% faster. Fixed-price studios now publish agent builds from $1,500 to $7,500, where agencies quote $30,000 and up.
What stayed expensive. A 2025 randomized trial by METR found experienced developers on mature codebases were 19% slower with AI tools, even though they felt faster. It was a small study of early-2025 tools, but it reinforces a point that applies strongly here: the hard part of an agent is not generating code. It is:
- Defining what the agent should and must not do.
- Getting accurate, current data into it.
- Testing behavior on messy, adversarial real-world inputs.
- Managing risk when it can change records or message customers.
- Operating it after launch as models and policies change.
Those tasks need judgment, and cutting them is how cheap agents become expensive incidents.
Worked examples
These are illustrative scenarios built from the ranges above. They are not past Tantito clients, and the quotes are our estimates, not real bids.
Example 1: Website support chatbot for a 20-person software company
Scope: Answers questions from the help center and pricing page, collects the visitor’s email when it cannot answer, hands off to the support inbox. Website channel only.
| Builder | Estimated build | Estimated monthly run |
|---|---|---|
| US agency | $30,000-$50,000 | $400-$1,500 |
| Freelancer | $1,200-$6,000 | $100-$400 |
| AI-native fixed price | $2,400 | $50-$150 model and hosting on your accounts |
Example 2: Order-status and returns agent for an online store
Scope: Looks up orders in the store backend, explains return policy, starts a return request, escalates unusual cases to a human. Website chat and email.
| Builder | Estimated build | Estimated monthly run |
|---|---|---|
| US agency | $40,000-$80,000 | $1,000-$4,000 |
| Eastern Europe | $18,000-$35,000 | $500-$1,500 |
| Freelancer | $5,000-$15,000 | $200-$800 |
| AI-native fixed price | $6,500 | $100-$400 model, hosting and care plan options |
Example 3: Appointment-booking agent on WhatsApp for a clinic
Scope: Answers questions about services and hours, checks the calendar, books and reschedules, sends reminders. Handoff to reception for anything medical. Health-data compliance adds scope not included here.
| Builder | Estimated build | Estimated monthly run |
|---|---|---|
| US agency | $35,000-$90,000 | $800-$3,000 |
| Latin America | $12,000-$30,000 | $300-$900 |
| Freelancer | $2,500-$12,000 | $200-$600 |
| AI-native fixed price | $6,500 (multi-tool) | plus WhatsApp conversation fees |
Example 4: Internal knowledge assistant for a 50-person firm
Scope: Answers employee questions from policy documents, wiki and shared drives, respects document-level access, cites sources. Slack channel.
| Builder | Estimated build | Estimated monthly run |
|---|---|---|
| US agency | $40,000-$100,000 | $1,000-$5,000 |
| India | $12,000-$30,000 | $300-$1,000 |
| AI-native fixed price | $2,400-$6,500 depending on number of sources and access rules | $100-$300 |
Hidden costs and red flags in AI agent quotes
- No evaluation suite. If there is no test set, nobody knows how good the agent is.
- No cost cap or usage dashboard. You should see spend per conversation and per month.
- Model account owned by the vendor. It should be yours, with your keys, so you can switch vendors.
- “Trained on your data” without explanation. Most business agents use retrieval over your documents, not model training. Ask which one is meant.
- No human handoff. Every production agent needs a way out.
- Fixed build price, open-ended “AI fees”. Markups on model usage can dwarf the build cost over a year.
- No guardrails against prompt injection. An agent with tools can be manipulated by hostile input.
- Vague accuracy claims. “95% accurate” means nothing without the test set and definition behind it.
- Lock-in to a proprietary platform. Check whether you can export prompts, data and code.
Questions to ask any AI agent vendor
- Is this a chatbot or an agent, and which actions can it take?
- How do you test answer quality, and can I see the test set?
- Whose accounts hold the model keys, the data and the code?
- What does it cost per 1,000 conversations on my expected volume?
- How do you prevent it from taking harmful or out-of-scope actions?
- When and how does it hand off to a human?
- What happens when my documents or policies change?
- What do monitoring and maintenance cost per month?
- What is covered by your warranty after launch?
Where Tantito fits
Tantito is an AI-native studio: AI agents do the building and one senior engineer verifies each delivery. Our AI agent and chatbot packages are priced as follows.
- Single workflow, $2,400: one agent job, such as answering from your documents or handling one process.
- Multi-tool agent, $6,500: actions in your systems through tool calls.
Delivery takes 1 to 3 weeks depending on the tier.
Both include grounding on your documents and data, a website, WhatsApp, Slack or email channel, guardrails, a test suite, human handoff, a usage and cost dashboard, and deployment on your own accounts. The price is fixed, a working preview arrives within 5 business days or the deposit is returned, the code is yours, and bugs are fixed free for 30 days. See AI chatbot development for the chatbot-focused page and business automation for workflow automation, which starts at $1,500. Care is $299 per month.
We are not the right fit for agents that need regulated-industry certification, on-premise models or a large multi-team rollout. For those, a specialist agency or in-house team is the safer choice.
Use the software cost calculator, check pricing, or request a quote. If you are comparing builders more broadly, read fixed price vs hourly software development.
Sources
Compiled November 2026. Most published agent pricing comes from vendors with a commercial interest, so read the ranges as market anchors.
- Groovyweb: AI chatbot development cost 2026
- Groovyweb: AI agent development cost guide 2026
- Fiverr: chatbot developer cost guide
- Fiverr: AI automation expert cost guide
- Dev.to: what AI automation services actually cost in 2026
- HouseofMVPs pricing and AutomateNexus, published AI-native fixed prices
- Clutch: software development pricing, updated September 2026
- Anthropic API pricing and OpenAI API pricing, for current model rates
- METR: early-2025 AI and experienced open-source developers
- Peng et al.: the impact of AI on developer productivity (GitHub Copilot)