An AI agent costs from a few hundred dollars for a configured off-the-shelf chatbot to $6,000–$40,000 for a custom single-purpose agent, $20,000–$120,000 for a multi-tool agent wired into your systems, and $150,000 or more a year for an enterprise programme. Model fees for a support chat run from under a cent to about 12 cents per conversation. Integrations, data, evals, voice and team location move the number most.
Market figures below come from providers’ own pricing pages. Still choosing a use case? Start with our AI/ML deployment roadmap.
What each type of AI agent costs to build
Most quotes fall into six tiers. Price follows effort, and effort follows how many systems the agent touches.
| Type | What you get | Timeline and effort | Typical build price | Example |
|---|---|---|---|---|
| Off-the-shelf agent you configure | A vendor platform trained on your content, channels connected, handoff rules set; no custom code | 1–2 weeks | $0–$3,000 setup, plus subscription or per-resolution fees | Website and WhatsApp agent for product and order questions |
| Workflow automation with an AI step | A trigger, one or two model calls (classify, extract, draft) and writes into your tools | 1–3 weeks; 0.5–1.5 engineer-weeks | $500–$5,000 per workflow | Invoice emails turned into accounting entries |
| Custom single-purpose agent | One job: retrieval over your documents, one to three integrations, guardrails, an eval set, an admin view | 4–8 weeks; 4–10 engineer-weeks | $6,000–$40,000 | Support agent that answers from your policies and checks orders |
| Multi-tool agent with integrations and MCP | Several systems via APIs or MCP servers, actions with approval steps, memory, handoff, audit logs, observability | 8–16 weeks; 10–30 engineer-weeks | $20,000–$120,000 | Sales agent that qualifies leads, books meetings and updates the CRM |
| AI product or SaaS MVP | Accounts, billing, web or mobile app, admin panel, AI features, analytics | 10–16 weeks; 10–40 engineer-weeks | $10,000–$150,000 | Subscription legal research and drafting app |
| Enterprise programme | Several agents on a shared platform, single sign-on, data governance, security review, private cloud | 6–12 months for the first wave | $150,000 and up a year | Agents across support, finance and operations |
These ranges assume blended rates of roughly $25–$100 an hour, about $1,000–$4,000 per engineer-week. Onshore agencies in the US, UK and the Gulf bill above that band, so their quotes sit near or beyond the top of each range.
The model rarely pushes a project up a tier. Integrations do, along with whether the agent takes actions (refunds, bookings, record updates) or only answers, the languages it handles, the state of your data and compliance work.
The Model Context Protocol (MCP), an open standard for connecting AI applications to external systems, cuts integration work when your tools already have MCP servers. Where they don’t, building and securing them is billable work.
What an AI agent costs to run each month
Running costs come in five lines: model tokens, hosting, a vector database, monitoring and evals, and maintenance, plus per-minute charges for voice. They are separate from the build price.
Model APIs charge per million tokens, and output costs more than input:
| Model | Input per 1M tokens | Output per 1M tokens | Typical role in an agent |
|---|---|---|---|
| Claude Haiku 5.5 (Anthropic, prompts up to 100,000 tokens) | $0.10 | $0.50 | Routing, classification, simple replies |
| Claude Sonnet 5.5 (Anthropic) | $2 | $10 | Most support and sales agents |
| Claude Opus 5.5 (Anthropic) | $4 | $20 | Hard reasoning and complex tool use |
| gpt-6-luna (OpenAI) | $0.10 | $0.50 | High-volume, simple turns |
| gpt-6.1-sol (OpenAI) | $2 | $10 | General agent work |
| Gemini 3.5 Flash-Lite (Google) | $0.30 | $2.50 | High-volume agentic tasks and translation |
Sources: Anthropic pricing, OpenAI pricing and Gemini API pricing. Some Gemini prices change on set dates, so check before you budget.
Worked example: a support chat agent
Assume each model call sends a 3,000-token system prompt and tool definitions, 2,000 tokens of retrieved help-centre passages and about 1,000 tokens of conversation so far, and gets back a 250-token reply. Six customer messages and two order lookups make eight calls: 48,000 input tokens and 2,000 output tokens.
- Claude Sonnet 5.5: 48,000 × $2 per million is $0.096, plus 2,000 × $10 per million is $0.02, so about $0.12 a conversation.
- With prompt caching: at steady traffic the fixed 3,000-token prefix is read from cache on all eight calls at $0.10 per million, bringing it to about $0.07.
- Claude Haiku 5.5: about $0.006 a conversation.
At 5,000 conversations a month that is about $580 on Sonnet, $350 with caching, $30 on Haiku, or roughly $95 if a router sends 80% to Haiku and the rest to cached Sonnet. Measure with your own transcripts: Anthropic notes its newer tokeniser produces about 30% more tokens for the same text, and Anthropic’s engineers measured agents using about four times the tokens of chat, and multi-agent systems about 15 times.
Hosting, data, monitoring and maintenance
- Hosting and database: managed Postgres such as Supabase Pro starts from $25 a month. Staging, backups and logging cost more.
- Vector database: the pgvector extension in your existing Postgres adds no licence fee. Pinecone has a $20 flat Builder plan and a $50 monthly minimum on Standard.
- Monitoring and evals: Langfuse is open source and free to self-host; its cloud plans are $29 (Core) and $199 (Pro) a month. Re-running evals after each change also costs tokens.
- Maintenance: often the largest line after launch. Knowledge goes stale, real conversations expose gaps and providers retire models. Plan for a few engineer days a month at first.
How to cut running costs without losing quality
- Use the smallest model that passes your evals. Claude Haiku 5.5 input costs $0.10 per million tokens against $4 for Opus 5.5, a 40-fold gap.
- Route by difficulty. A small model handles routine turns; a larger one takes the hard ones.
- Cache fixed prompts. Anthropic bills cache hits at 0.1x the base input price (0.05x on Sonnet 5.5 and Opus 5.5). OpenAI lists cached input for gpt-6.1-sol at $0.10 per million against $2.
- Batch anything that can wait. Anthropic’s and Google’s batch APIs cost 50% of standard prices, and OpenAI’s batch rates are half its standard ones.
- Use retrieval (RAG) before fine-tuning. Retrieval keeps answers current without retraining. Fine-tune only when evals show a gap retrieval cannot close.
- Trim context. Summarise long histories, cap retrieved passages and tool outputs, and set output limits.
- Stay single-agent until evals say otherwise. Multi-agent designs multiply token use.
Voice agents versus chat agents: what changes
A voice agent has the same reasoning core with harder edges: it must take turns naturally, cope with interruptions, reply in under about a second, connect to phone numbers and call transfer, and understand accents on noisy lines. Expect several extra weeks of build and testing over the same agent in chat.
Running costs are billed per minute. Google lists Gemini 3.8 Live at $0.005 per minute of audio in and $0.018 per minute out. ElevenLabs charges $0.08 a minute plus model usage for its realtime speech models. A US line on Twilio adds $0.0085 a minute inbound, $0.014 outbound and $1.15 a month per local number.
If the agent speaks for half of a four-minute call, that is roughly $0.09 on a raw speech model and phone line, or about $0.40 on a managed voice platform, before long-call context growth. The same conversation in chat costs a few cents. Voice pays off where customers already call, such as bookings and after-hours enquiries. Tell callers they are speaking to an AI, and check recording and outbound-call consent rules in each market.
Build, buy or hire: the ways to get an agent
| Option | Cost | Speed | Main risk | When it fits |
|---|---|---|---|---|
| Off-the-shelf agent | Subscription or per-resolution fees that grow with volume | Days | Limited customisation and control over actions; lock-in | Standard support questions on a helpdesk you already use |
| Freelancer | Lowest hourly cost | Fast to start | One person; uneven testing and security; no cover when they move on | Small, well-defined automations and prototypes |
| Onshore agency (US, UK, Gulf) | Highest rates | Weeks | Budget overrun on long projects | Regulated work needing on-site workshops or local contracts |
| Offshore agency | A fraction of onshore rates | Weeks | Time-zone overlap, communication, IP terms; vet shipped work | Defined scope, cost-sensitive budgets, ongoing development |
| In-house team | Full salaries plus benefits, all year | Months to hire | Scarce AI skills; key-person dependence | AI is your core product and the work never stops |
Buying is often the right start. Intercom’s Fin charges $0.99 per outcome with no seats required, and you get a finished product on day one. At 5,000 resolved conversations a month, though, that is about $4,950 a month, against $95–$580 of model fees in the worked example. A custom agent wins once volume is steady, or when it must take actions in your own systems that a packaged tool won’t.
Why rates differ so much by country
The same scope can cost two to four times as much depending on where the team sits, and most of the gap is labour cost. The US median wage for software developers is $135,980 a year, and wages make up 70.0% of private-sector employer compensation costs. That makes a median developer about $194,000 a year to employ, roughly $93 an hour, before office, management, tools and margin.
In Pakistan, Payscale’s survey puts the average base salary for a senior software engineer at about PKR 1.54 million a year (about $5,500). The sample is small and widely spread, so treat it as a direction, not a precise figure.
An agency rate also carries project management, QA, tools, overhead and margin, so offshore rates sit well above local salaries yet well below US ones. A low rate is no bargain if the work takes twice the hours.
How to scope a fixed-price pilot
A fixed price works when the scope is narrow and the finish line is written down; hourly billing suits open-ended exploration.
- Pick one job with a number attached, such as order-status questions that take staff a known number of hours a week.
- Collect 50–200 real examples and write down what a correct answer or action looks like. This becomes the eval set.
- Fix the boundaries: channels, languages, systems it may read, actions it may take and what it must hand to a person.
- Agree acceptance criteria, such as the share of eval cases answered or handed off correctly, response time and escalation rules.
- Set a running-cost target per conversation and choose models to meet it.
- Time-box it to four to eight weeks, with a demo on your real data by week two or three.
- List what is out of scope and how change requests are priced.
- Decide in advance what happens next: go live, iterate once with a clear target, or stop.
How Pixelpk prices and builds AI agents
Pixelpk Technologies is an AI and software development company founded in 2018 and headquartered in Lahore, with offices in Malaysia, the UK, Australia, Saudi Arabia and the US. We build and run our own AI products: Aivastark, one support and sales agent for website chat, Instagram, WhatsApp, Messenger, Telegram and email, and Pakistan LawBot, with cited legal research, drafting of 13 document types and contract analysis in English and Urdu. We also built Khurshid Answers, which gives cited answers over a scholar’s writings in English, Urdu, Arabic and Roman Urdu.
Our prices:
- WhatsApp and website AI chat agent (Aivastark setup): from $390
- AI automation: from $490 per workflow (1–2 weeks), or a pack of three for $1,290
- AI readiness audit and roadmap: from $1,500 (2–3 weeks)
- AI agent or assistant: from $6,000, about PKR 1.7 million (4–8 weeks)
- AI product or AI SaaS MVP: from $10,000, about PKR 2.8 million (10–14 weeks)
- AI-powered dedicated developer: from $25 an hour, starting in 1–2 weeks
- AI receptionist (phone voice agent) for small businesses: $149 or $299 a month plus $490 or $990 setup
The price covers senior engineers who use AI coding agents and AI tools as part of the work, so you don’t buy or learn the tools. You own the code and IP, and everything is tested before launch. Engagement models are Extended Team, Managed Dedicated Team and Custom Solution Development; our AI solutions page covers the services. For a quote, contact us or book a free strategy call.
Frequently asked questions
How much does it cost to build an AI agent?
A custom single-purpose agent, such as a support or sales agent that answers from your documents and uses one to three integrations, typically costs $6,000–$40,000 and takes four to eight weeks. Configuring an off-the-shelf agent can cost a few hundred dollars. Multi-tool agents that act across several systems cost $20,000–$120,000.
How much does an AI chatbot cost to run each month?
Model fees for a support chat agent run from under a cent to about 12 cents per conversation, depending on the model, conversation length and caching. At 5,000 conversations a month that is roughly $30–$600. Add hosting from about $25 a month, optional vector database and monitoring plans, and maintenance time.
How much does an AI voice agent cost?
A voice agent costs more to build than the same agent in chat, because of turn-taking, interruptions, telephony and testing on real calls. Running it is billed per minute: a four-minute call costs roughly $0.09 on a raw speech model with a phone line, or about $0.40 on a managed voice platform.
How much does an AI MVP cost?
An AI product or SaaS MVP usually costs $10,000–$150,000 and takes 10–16 weeks. The low end is a focused web app with accounts, one core AI feature and basic billing from an offshore team. The high end adds mobile apps, several AI features, complex integrations or an onshore team. Model and hosting fees are extra.
Is it cheaper to build or buy an AI agent?
Buying is cheaper to start and often right for standard support questions on a helpdesk you already use. Per-resolution pricing grows with volume: at $0.99 per outcome, 5,000 resolutions cost about $4,950 a month. A custom agent costs more up front but far less per conversation, so it wins at steady volume.
Why do quotes for the same AI agent vary so much?
Quotes differ because of scope and labour cost. Scope covers integrations, actions, languages, data preparation, evals and compliance, and vague briefs get padded quotes. Labour cost depends on location: a median US developer costs an employer roughly $93 an hour before overhead and margin, so onshore agencies often charge two to four times offshore rates.