Cost guide
What an AI agent actually costs to build
There is no list price, and anyone who quotes one before asking which systems the agent is allowed to write to has guessed at the part that matters most.
Get a scoped estimateWhy two quotes for the same agent look nothing alike
Most quotes differ because the two teams heard different products. One priced an assistant that answers questions from your help centre. The other priced something that issues the refund, writes the outcome back to your CRM, and can be audited when a customer disputes what it said.
The question that moves the number more than any other is whether the agent is allowed to act. Reading is cheap and forgiving. Writing brings permissions, idempotency, an approval path and a trace of every step — and that is where most of a budget goes.
Cost drivers
What actually moves the number
None of these are line items you can negotiate away with a discount. They are decisions about what the agent is allowed to do, and each one either adds weeks to the build or removes them.
Write access to your systems
Reading order history is a few days of work. Issuing the refund through your billing API, with idempotency keys so a retried run cannot refund twice, is not.
The number of integrations
Each system brings its own auth, rate limits and sandbox. A Salesforce write with per-user OAuth is a different week's work from a Slack webhook.
The state of your documents
Retrieval over a maintained Confluence space starts quickly. Four years of scanned PDF contracts need an OCR and clean-up pass before the agent reads anything useful.
Approval checkpoints
Deciding that a refund above your limit waits for a person is one line of policy and several weeks of build: queue, notification, audit record, timeout behaviour.
The evaluation set
Real past tickets, scored before launch, so a model upgrade can be measured instead of guessed at. Assembling it takes longer than teams expect.
Where the model runs
A hosted API ships fastest. If patient or payment data cannot leave your network, you are also buying GPU capacity and the people to keep it running.
Tracing and disputes
Every step, tool call and prompt logged, so when a customer says the agent promised a refund you can read back the actual run.
Channels
Chat and email share most of their plumbing. Voice adds telephony, a latency budget and interruption handling, and behaves like a separate project.
Conversation volume
Model spend scales with traffic, and an agent that re-reads a whole policy document every turn runs far hotter than one retrieving the right paragraph.
Scope tiers
Four honest scopes, smallest first
Most teams should start further up this list than they expect. What each tier excludes matters more than what it includes, so the exclusions are written out.
- TIER ONE — PILOT
- One job, proven on your data. Ships in 2–3 weeks.
- ₹1L – ₹1.5L$1.1k – $1.7k
- A working agent against real past cases, run in a sandbox, with a written verdict on whether it holds up. Suits teams who need evidence before a budget request. Does NOT include any write to a live system, guardrail config or a UI.
- TIER TWO — INTERNAL AGENT
- One agent, staff-facing only. Ships in 5–8 weeks.
- ₹4L – ₹6L$4.5k – $6.8k
- Retrieval over your own documents with citations, read-only tool access, behind your SSO, plus a small admin view for reading transcripts. Suits ops and support teams. Does NOT touch customers, and takes no action that changes a record.
- TIER THREE — CUSTOMER-FACING AGENT
- Acts in your systems, with escalation. Ships in 3–4 months.
- ₹14L – ₹18L$16k – $20k
- Write access to one or two systems of record, an approval path for anything above your limit, full tracing, an evaluation set and a staged rollout. Suits real support or sales volume. Does NOT include voice or self-hosted models.
- TIER FOUR — AGENT PLATFORM
- Several agents, regulated environment. Ships in 6–12 months.
- From ₹40LFrom $45k
- Multiple agents sharing tools and policy, self-hosted or private-cloud models, role-based access, retained audit trails and integration into an enterprise stack. Suits healthcare, fintech and anyone whose data cannot leave. Does NOT start here without a passed pilot.
Indicative ranges for scoping conversations, not quotes. They reflect what work of this shape has cost us to deliver — your figure comes out of a paid discovery, against a written scope, and is fixed before any code is written. Gas, licences, cloud and third-party audit fees sit outside these numbers.
How it is scoped
From a vague ask to a defensible number
Discovery exists so the quote survives contact with your systems. The timelines below are typical for a first agent; the second one is faster, because the guardrails and the tracing already exist.
Discovery
1 weekWe pick the single job the agent should own, list every system it must touch, and write down the actions it must never take alone.
Pilot
2–3 weeksThe agent runs against real past cases. This is where scope shrinks, and occasionally where a project stops, which is a cheap thing to learn in week three.
Build and harden
4–10 weeksIntegrations, the approval path, tracing and the evaluation set. Nothing speaks to a customer until it has been scored against real cases and red-teamed.
Run and tune
OngoingSomeone reads transcripts weekly, tightens prompts and permissions, and cuts model spend once you can see where the agent is actually spending its time.
Worth budgeting for now, or worth waiting
Worth reading before the budget conversation rather than after it.
Worth budgeting for now if…
- A person does this task end to end, most days
- You can name the system of record the agent must update
- You have tickets, transcripts or docs to ground the answers in
- Someone senior will own the approvals when the agent is unsure
Wait, if…
- The workflow is fixed and rule-based — AI Automation costs far less
- Your documentation is stale, and fixing it first is cheaper
- Nobody will have time to read transcripts after launch
- One person handles the entire volume in an hour a day
Frequently Asked
Questions
Common questions about scoping, cutting and running an AI agent build.
Write access, mostly. An agent that answers questions from your help centre is a contained piece of work. An agent that issues refunds, updates the CRM and cancels the order needs permissions, idempotency, an approval path and a trace of every step, and each of those is real engineering. After that the drivers are the number of integrations, whether your documents need cleaning before retrieval works at all, and whether compliance forces the model onto infrastructure you control rather than a hosted API.
Cut breadth before depth. One job the agent owns completely is worth more than several it half-handles, and it is also the only version you can measure honestly. Voice is usually the first thing to drop, then the second and third integrations, then the custom admin interface, which a transcript export replaces for a while. What should not be cut is the evaluation set and the escalation path. Those are the parts that prevent a bad launch, and retrofitting them costs more than building them.
Data preparation, red-teaming and their own time. Retrieval quality is decided by the documents, so somebody has to find the current version of the returns policy, delete the outdated copies and tag what is left. Red-teaming gets skipped almost every time: someone has to attempt prompt injection and unsafe outputs before a customer stumbles into them. And the internal effort is real, because your ops lead will spend weeks defining exactly what the agent may decide without asking.
Three buckets, and only one of them tends to surprise people. Model usage scales with conversation volume and with how much context each turn carries, so a badly built agent can run several times hotter than a careful one for identical work. Infrastructure stays modest unless you self-host models, where GPU capacity dominates everything else. Maintenance is the bucket teams underestimate: prompts drift, providers change model behaviour without notice, and an agent nobody tunes decays within a couple of quarters.
Phase it, in almost every case. The pilot tells you whether the task is genuinely automatable, which is not obvious until an agent has tried it on your real cases. An internal agent then lets your own staff find the failure modes before a customer does, and the guardrail work carries over unchanged. The exception is a regulated environment where a security review will happen once regardless, in which case splitting the build mostly means going through that review twice.
Have a project in mind?
Fixed price after a paid discovery — no hourly billing. A real engineer reads every enquiry, and we reply within 24 hours.