In 2026, custom AI chatbots cost $5k–$50k, AI agents $20k–$120k, voice AI agents $20k–$120k, RAG systems $30k–$300k, and AI workflow automation $10k–$80k. Agencies typically charge $50–$150 per hour, and running costs land at $500–$5,000 a month. Most mid-market projects we scope come in between $30k and $100k.
We scope AI projects every week, which means we get asked for these numbers every week — and watch people arrive with quotes that differ by 4x for what sounds like the same system. This guide is the pricing conversation we have on those calls, written down: what each type of system costs, why identical-sounding scopes price so differently, and where the money actually goes.
A companion piece, what drives AI development cost, digs into the underlying cost mechanics. This one is organised the way buyers actually ask: by the thing you want to build.
Cost by solution type
| Solution | Typical range | Timeline | What moves the price |
|---|---|---|---|
| Custom AI chatbot | $5k–$50k | 3–8 weeks | Number of data sources, permission-aware retrieval, write access to CRM or helpdesk |
| AI agent | $20k–$120k | 6–16 weeks | How many systems it touches, approval workflows, memory, eval depth |
| Voice AI agent | $20k–$120k | 6–16 weeks | Latency target, languages, telephony integration, outbound compliance |
| RAG system | $30k–$300k | 8–24 weeks | Corpus size and messiness, retrieval quality bar, per-user permissions |
| Workflow automation | $10k–$80k | 4–12 weeks | Document variety, integration count, human review requirements |
Three notes on reading that table honestly. First, the bottom of each range assumes clean data, one or two integrations, and a team that already knows what it wants. Second, compliance-heavy industries — healthcare, finance, anything touching PII at scale — should expect to sit in the upper half regardless of technical complexity, because review, audit and data-handling requirements are real work. Third, these are build costs; running costs are a separate line we cover below, because the industry has a habit of forgetting to mention them.
Where each range comes from
A chatbot at $15k–$30k is the most common configuration we ship: grounded in your documents via retrieval, honest about what it does not know, integrated with one or two systems. The version at $50k adds permission-aware retrieval across multiple sources — the details are in our AI chatbot development work.
An AI agent costs more than a chatbot for one structural reason: it acts rather than answers. Acting means tool access, approval checkpoints, state handling, and evals that prove it does the right thing — the engineering around the model, which is most of the budget.
For RAG development, the corpus is the cost. A thousand well-structured pages and a million scanned PDFs with handwritten annotations are different projects wearing the same acronym. Retrieval quality targets compound this: each additional point of answer accuracy costs more than the last.
AI workflow automation is the range with the friendliest floor: a single document-processing pipeline that removes a genuinely painful manual task can land near $10k–$20k and pay for itself in months. The top end is multi-step automations across several systems with review queues and audit trails.
AI chatbot pricing in detail
Chatbots are the most-quoted item on this page and the one where quotes diverge most, so it is worth breaking the range apart. AI chatbot pricing splits into three tiers that look similar in a demo and cost very differently to build:
| Tier | Build cost | Monthly | What it does |
|---|---|---|---|
| Off-the-shelf, configured | $0–$3k setup | $50–$500 | Crawls your public site and docs, answers from them, no system access |
| Custom, grounded | $15k–$30k | $200–$1,500 | Retrieval over your own content, one or two integrations, escalation to a human |
| Custom, permission-aware | $30k–$50k | $500–$3,000 | Multiple sources, per-user permissions, write access to CRM or helpdesk |
The jump from row two to row three is almost entirely permissions. A chatbot that answers from public documentation is a retrieval problem. A chatbot where each user may only see what they are entitled to see is an authorisation problem wearing a retrieval costume, and it has to be designed in rather than added later.
Running cost scales with conversation volume, at roughly $0.02–$0.15 per conversation on 2026 model pricing depending on how much context each answer needs. For a fuller treatment of deflection rates, escalation design and the maintenance nobody quotes, see AI chatbot for your website. If you are pricing an agent rather than a chatbot, what an AI agent actually costs in 2026 covers the four pricing models and a worked twelve-month total.
What actually drives the price
- Integration count. The strongest single predictor. Every system the AI must read from or write to adds auth, schemas, error handling and testing. Two integrations to five is not 2.5x the work, but it is close.
- Data quality. If your documents are clean and your databases consistent, you are at the bottom of the range. If the first month is data archaeology, you are not.
- Compliance. HIPAA, GDPR, financial regulation: each adds design constraints, review cycles and documentation that are genuinely necessary and genuinely billable.
- Latency targets. A chatbot that responds in three seconds is standard work. A voice agent that responds in 700ms is architecture. Speed is a feature you pay for.
- Evaluation depth. Evals — fixed test sets run on every change — are the difference between knowing your system works and hoping it does. Skipping them lowers the quote and raises the total cost.
The hidden costs
The quote is not the cost. Budget for:
- LLM API usage. Usage-based and therefore success-based: the more your system is used, the more it costs. A busy customer-facing chatbot can run hundreds to thousands of dollars a month in inference alone.
- Monitoring and observability. Someone has to notice when answer quality drifts. Tooling plus a few hours of attention a month.
- Model upgrades. Providers retire models on their schedule, not yours. Each migration needs eval runs and occasionally prompt rework.
- Retraining and content refresh. RAG indexes go stale; vision models meet new products; automations meet new document formats.
As a planning figure, $500–$5,000 a month covers most custom systems, scaling with volume. Anyone quoting a build with zero ongoing cost is deferring the conversation, not the cost.
Build custom or buy a product?
Genuinely not everything should be custom. If an off-the-shelf product does 90% of what you need and your workflow can bend to it, buy it — a $50–$500 monthly subscription beats a $50k build. Custom wins when the value lives in your specific data, your specific systems, or a workflow that products do not serve: the calculation we walked through for CRMs in build-or-buy terms applies to AI systems unchanged.
How to spend less without getting less
- Start with scoping, not a build. A one-to-two-week paid scoping engagement that examines your data and systems produces a real number and often removes half the imagined scope. Ours is credited toward the build if you proceed.
- Ship the narrow version first. One workflow, one integration, real users. The second phase is then priced on evidence instead of optimism.
- Fix the data before the AI. An engineer cleaning data at AI-project rates is the most expensive data cleaner you will ever hire.
- Reuse the connectors. Integration work built as MCP servers is reusable across every future AI project, which makes project two dramatically cheaper than project one.
How we price it
Our engagements start with a free consultation, then a paid scoping phase that produces a fixed build quote — and the scoping fee is credited toward the build if you go ahead. It aligns incentives: we do not have to pad quotes against unknowns, and you do not commit five figures on a guess. If the scoping concludes the project is not worth building, that is the deliverable, and it is a cheap one. The full process is on our AI development services page, or talk to us about your numbers directly.
Common questions
How much does it cost to build an AI agent?
A production AI agent typically costs $20,000–$120,000 in 2026. Simple single-workflow agents with one or two integrations sit at the low end; agents that plan multi-step work across several business systems, with memory, evals and human approval checkpoints, sit at the top. Compliance-heavy industries such as healthcare and finance can push past that range.
How much does a custom chatbot cost?
Custom AI chatbots run $5,000–$50,000. A chatbot grounded in your own documents with basic integrations lands around $15,000–$30,000; adding permission-aware retrieval, multiple data sources, or CRM and helpdesk write access moves it toward the top of the range.
How much does an AI chatbot cost?
An AI chatbot costs $50–$500 a month off the shelf, or $15,000–$30,000 to build custom and grounded in your own content. Permission-aware chatbots that respect who is asking, pull from several sources and write back to your CRM or helpdesk run $30,000–$50,000, plus $200–$3,000 a month to operate.
What does RAG development cost?
RAG systems cost $30,000–$300,000 depending on corpus size and messiness, retrieval quality targets, permission handling, and evaluation depth. A single-source knowledge assistant sits near the bottom; enterprise deployments over millions of documents with per-user permissions and continuous evals sit at the top.
What are the ongoing costs of an AI system?
Budget $500–$5,000 per month for a typical custom system: LLM API usage, hosting, monitoring and observability, plus periodic eval runs and model upgrades. High-volume voice agents and vision systems can run higher; usage-based LLM costs are the line item that most often surprises teams.