AI agent development cost: what drives the price
AI agent development cost comes in two parts: a one-off build and a monthly running bill. The build is driven by how many systems the agent connects to, how much it may do without a person approving, the channels and languages it works in, and how much testing the risk level demands. The running bill covers model usage, hosting, messaging or telephony fees, monitoring and the human time spent reviewing the agent's work.
This guide breaks down each cost driver, shows how token pricing turns into a cost per conversation, compares off-the-shelf tools with custom builds, and lists what to bring to a scoping conversation so the quote reflects your situation. Where we refer to our own prices or timelines, you will see a bracketed placeholder rather than a number: those depend on scope, and we quote them after a proper look.
The two bills: building the agent and running it
Most budget surprises come from treating an agent like a website: pay once, then forget it. An agent behaves more like a new team member. It needs to be set up, trained on your processes, given access to the right systems and then supervised. The build cost is usually the larger number at the start, but the running cost is the one that decides whether the agent pays for itself over a year.
| Cost component | One-off or recurring | What drives it | Red Balloon range |
|---|---|---|---|
| Discovery and design | One-off | Number of use cases, process complexity, risk review | Quoted after scoping |
| Build and integrations | One-off | Systems connected, read or write access, legacy software | Quoted after scoping |
| Evaluation and pilot | One-off | Size of test set, languages, risk level, shadow-mode length | Quoted after scoping |
| Model usage | Recurring | Conversation volume, length, model tier, tool calls per task | Billed at provider cost |
| Messaging or telephony | Recurring | Channel, message categories, call minutes | Billed at provider cost |
| Monitoring and support | Recurring | Review cadence, number of agents, response time agreed | Quoted monthly after scoping |
A clear scope keeps both bills honest. The more precisely you can say what the agent will and will not do, the less time goes into building edge cases nobody needed. That is why the first phase of any agent project we run is a written specification, as described on our page about AI agent development.
What drives the build cost
Integrations and access
Every system the agent touches adds work: a CRM, a calendar, a booking or practice-management system, an ERP, a payment gateway. Read-only lookups are simpler than writes, because writes need validation, permission checks and rollback rules. Modern software with documented APIs is quicker to connect than desktop accounting tools or in-house systems that need a custom bridge or exported files. Ask for a list of integrations with an effort estimate for each.
How much the agent decides on its own
An agent that drafts replies for a person to send is cheaper to build safely than one that acts on its own. Each extra decision the agent makes without approval needs rules, tests and a fallback. Many projects start in draft-and-approve mode and move individual intents to full automation once the logs show they are handled reliably, which spreads the cost over time.
Channels and languages
A web chat agent in English is the simplest case. WhatsApp adds template approval and message categories. Voice adds speech-to-text, text-to-speech, telephony and latency tuning, and usually costs more to build and run than text. Hindi, Hinglish and regional languages need testing on real customer messages or recordings, because accuracy varies by model and accent. Each language and channel needs its own test cases.
Evaluation, guardrails and compliance
Testing is where cheap quotes cut corners. A proper build includes a set of real questions with expected answers, cases the agent must refuse, and a rerun of that set after every change. Agents in healthcare, lending or insurance need stricter refusal rules and more human review. India's Digital Personal Data Protection Act applies to the personal data in agent conversations, so notice, purpose limits and processor terms with model providers belong in the scope.
How running costs work: tokens, caching and model choice
Language models are billed per token, roughly a short word or part of a word, with separate prices for input (what you send the model) and output (what it writes). As of November 2026, Anthropic's pricing documentation lists Claude Sonnet 5.5 at $2 per million input tokens and $10 per million output tokens, and Claude Haiku 5.5 at $0.10 and $0.50 for prompts up to 100,000 tokens. Prices change often, so check the provider's current page before budgeting.
Agents use more tokens than chatbots. Each time the agent calls a tool, the conversation so far, the instructions and the tool result are sent to the model again. A task with five tool calls can send the same context five times. That is why the cost per resolved task matters more than the price per token, and why an estimate should come from a sample of your real conversations rather than a guess.
To see how the arithmetic works, take a purely illustrative task that sends 20,000 input tokens in total and receives 1,000 output tokens. On Sonnet 5.5 list prices, that is $0.04 of input plus $0.01 of output, so about $0.05 before caching or discounts. The same task on Haiku 5.5 costs a twentieth of that. Your real token counts will differ, so treat this as a method, not a benchmark.
Two provider features cut the bill. Anthropic's documentation prices cache reads at a small fraction of the base input price (on Claude Sonnet 5.5, $0.10 per million tokens against $2 for fresh input), which helps when the same long instructions are sent on every call. Its Batch API gives a 50 percent discount on input and output for work that does not need an instant reply, such as overnight document processing. Routing simple requests to a smaller model and keeping the larger model for hard cases is the third lever.
Running costs people forget to budget
- Messaging fees. Meta has charged per delivered message on the WhatsApp Business Platform since 1 July 2025, with prices by category and country. From 1 October 2026, service messages beyond a free tier of 1,000 per business phone number each month are charged too.
- Telephony and speech. Voice agents pay for call minutes, speech-to-text and text-to-speech on top of model usage, often from three different providers.
- Hosting and storage. Servers, a vector database for document search, and storage for logs and transcripts.
- Monitoring tools. Tracing and evaluation platforms that record every conversation and tool call, usually priced per seat or per volume.
- Human review. Someone on your side samples conversations every week. Budget their hours; they are a real cost and the most important control.
- Model changes. Providers retire model versions on their own timelines, and each switch means rerunning your test set and sometimes adjusting prompts.
- Content upkeep. When prices, policies or products change, the documents the agent reads must change too, or it will quote old information.
Off-the-shelf, low-code or custom: how the cost profile differs
| Option | Upfront cost | Ongoing cost | Good fit | Watch out for |
|---|---|---|---|---|
| Off-the-shelf agent product | Low | Per-seat or per-conversation subscription | Standard support or booking flows on supported tools | Limited integrations, data held by the vendor, price rises at scale |
| Low-code platform build | Moderate | Platform fees plus model usage | Workflows with some agent steps, teams that want to edit flows | Complex logic gets hard to test and version |
| Custom build | Highest | Model usage, hosting, support | Several systems, strict guardrails, regulated data | Needs an owner and a maintenance plan |
The cheapest option at the start is not always the cheapest over two years. A per-conversation subscription that looks small at pilot volume can overtake the cost of a custom build once traffic grows, while a custom build carries maintenance you have to plan for. Low-code tools sit in between; our comparison of n8n, Zapier and Make covers how their pricing models behave as volume rises.
Whichever route you choose, insist on owning the parts that hold value: the prompts, the test set, the conversation logs and the integration code or flow definitions. If those sit in your accounts, you can change supplier or model without starting again. If they sit in the vendor's, the switching cost becomes part of the price you pay later.
Building for India or abroad: what changes the price
The same agent can cost different amounts depending on where its users and data are. Data residency is one factor. Anthropic's documentation, for example, applies a 1.1 times price multiplier on newer models when inference is restricted to the United States, and other providers price regional processing differently. If personal data must stay in a specific country, the model choice narrows and the cost can rise.
Regulation is the second factor. Agents serving Indian users work under the DPDP Act and sector rules from regulators such as RBI or IRDAI. Agents serving patients in the United States may need a business associate agreement with every vendor that handles protected health information under HIPAA, and agents serving Europeans fall under GDPR and the EU AI Act's transparency duties. Each adds design, documentation and review time to the build.
Language and channel habits are the third. Indian deployments lean on WhatsApp and voice in several languages, which adds testing and messaging fees. Deployments for buyers abroad more often run on web chat, email or helpdesk tools in one language. Teams abroad that want an Indian partner for the build should also ask about time-zone overlap for support, which affects how the monthly support line is priced.
How to keep AI agent costs under control
- Start with one narrow, high-volume task and prove it before adding more.
- Estimate token use from a sample of your real conversations, not from a vendor's price page.
- Route simple requests to a small model and keep the larger model for hard cases.
- Cache long, repeated instructions and batch work that can wait.
- Set daily spend alerts so a loop or traffic spike cannot run up the bill unnoticed.
- Review cost per resolved task every month against the cost of doing the same work by hand.
- Keep your test set current, so a model switch to a cheaper option can be checked in hours rather than weeks.
Cost control also depends on the workflow around the agent. If leads arrive with missing fields or documents are out of date, the agent spends tokens asking questions it should not need to ask. Fixing the data and the CRM processes underneath often reduces the running cost more than switching models does.
How to scope and budget a first agent
A useful quote needs four things from you: monthly volumes for the task, a sample of real conversations or tickets, a list of the systems involved with notes on API access, and a list of actions the agent must never take. With those, a partner can estimate the build, the token cost per task and the review effort, and show you where the agent breaks even.
For reference, our agent projects run through discovery, design, build, pilot and handover, over a timeline set during scoping, and are priced as an Automation Build with a fixed scope and an hours-saved target, followed by an optional monthly support plan quoted to the review cadence you need. For a broader view of what agents can do before you commit, see our AI agent examples for business.
If you have a task in mind, start with a scoping call: we go through your volumes, systems and risk list and tell you whether an agent, a simpler workflow or nothing at all is the right answer. You can read how we build, test and hand over custom AI agents, or begin with the Growth Audit, a two-week paid diagnostic whose fee is credited to the next engagement.