Cost guide

How much does an AI agent cost in 2026?

An agent pilot on one real task runs $5K–$25K at the direct rate, quoted fixed once the scoping sprint (a fixed scope at a fixed price, with an end date) has written down the tools, the permission boundaries and the approval gates. The sprint and the monthly production run are priced on their own below, and every figure becomes a written proposal within 48 hours.

Request this option See the full rate card

Direct rate, in ₹ or $ · Itemised proposal in 48 hrs

The range

What it costs,
at the direct rate.

Priced as a scoping sprint that decides whether the task is agent-shaped at all, a pilot on one real task, and a monthly run once the agent has earned unattended operation.

One-offScoping and evaluation sprint$1.5K–$5Kone-time
ProjectAgent pilot$5K–$25Kfixed quote, one-time
RetainerProduction run and monitoring$500–$3Kper month

Indicative: every figure is refined into an itemised proposal within 48 hours.INR figures shown excl. GST.

Every option includes a Prism workspace for requests, approvals, tasks and the weekly review.

What moves the price

Why the range
is a range.

How many tools it can call

Every system the agent can reach is a permission boundary: a credential scoped to that step, a rule for what it may read or write, and a test that it holds. An agent that looks up a CRM record and drafts a reply is one boundary. One that reads inventory, calls a supplier and raises a purchase order is several, each wired and tested. We count tools, not steps, when we size the pilot.

Biggest lever

Whether the task is agent-shaped at all

An agent is worth building when the task has many steps, several tools and a path that changes with what it finds. Much of what is pitched as an agent is one grounded answer or one fixed sequence, cheaper to build as a chatbot or a workflow. The sprint tests the shape first and writes down the cheaper build where one exists, the largest saving on this page.

Often forgotten

What it is allowed to change

Reading is cheap to make safe; writing is not. A read-only agent that researches, reconciles and drafts for a person to send needs scoped credentials and a log. An agent that sends, spends or writes to a live system needs an approval gate, a threshold you set, a reversal path and a record of who approved what. Each write action is priced as the guardrail around it, not as the call itself.

Risk lever

The scenarios it must survive

Before it runs unattended, the agent runs supervised against scenarios drawn from the real job, including the traps where the obvious next step is the wrong one. A tidy task with few ways to fail needs a short suite. A task where a wrong call costs money or a customer needs a long one, scored against your baseline. The production run re-runs it whenever a tool changes.

Hidden lever

How often the systems underneath change

A live agent stays safe only while its tools behave as they did at evaluation. The monthly run is sized on run volume and how often the connected systems change: a stable ERP is one retainer, a stack of SaaS tools with breaking changes is another. Model usage sits outside both figures: the provider bills it, and the proposal shows it as a pass-through.

Retainer lever

Where the data has to stay

The model is chosen in the sprint on fit, cost and data residency. Regulated data or a residency rule can mean a model hosted in a particular region, or a deployment inside your own cloud, with infrastructure to set up and keep patched. That work is a visible line, settled by the sprint's model recommendation before the pilot is priced.

Scope lever

The options

Which option
fits your case.

Each option below is one way to buy AI agents. Read what it is and who it fits, then request the one that matches your case; the proposal itemises it.

One-off

Scoping and evaluation sprint

$1.5K–$5Kone-time

What it is

Discovery that decides whether you need an agent, then writes the design: tools, permission scope per call, approval thresholds, evaluation scenarios and a model recommendation. It ends in an itemised pilot scope, or a written case for a cheaper build.

Who it fits

Anyone with a multi-step task they suspect could run itself, and anyone quoted an agent elsewhere who wants a second reading. It does not fit a task you already know is one question or one fixed path.

Included

The task tested for agent shape: many steps, several tools, a varying path
Tools, permission scope per call and approval thresholds agreed in writing
Evaluation scenarios from the real job, traps included
A model recommendation on fit, cost and data residency
An itemised pilot scope, or a written case for a cheaper build

Not included, quoted separately

The pilot build
Model and API usage during scenario testing

Turnaround: 1 to 2 weeks
Moves the number: How many systems the agent would need to touch: one CRM lookup and six back-office systems are different sprints.

Request this option →

Project

Agent pilot

$5K–$25Kfixed quote, one-time

What it is

A working agent on one real task from your operation, supervised, every tool call scoped to least access, approval gates before anything irreversible, and runs scored against your own baseline. Handed over with its audit trail, runbook and full ownership of the code.

Who it fits

Teams with a task the sprint has confirmed is agent-shaped and the systems to wire it into. It does not fit a buyer who wants unattended operation from the start, which is earned, or a blanket key to the stack, which we do not wire.

Included

One real task running within 14 days of NDA, every step logged
Tool integrations wired to live systems, each call scoped to least access
Approval gates before anything irreversible, external or above a set threshold
Supervised runs scored against your baseline; unattended operation only where earned
Audit trail, runbook, admin training and full ownership of the code

Not included, quoted separately

Model and API usage
Hosting
Further tasks or tools, priced after the pilot has earned them

Turnaround: 3 to 6 weeks; working agent within 14 days of NDA
Moves the number: Tool count: every system the agent can call is a permission boundary to design, wire and test.

Request this option →

Retainer

Production run and monitoring

$500–$3Kper month

What it is

The retainer that keeps a live agent inside its guardrails as the tools and data underneath it change: logs watched, the evaluation suite re-run on every change, thresholds tuned on what the logs show, and a monthly report against the pilot baseline.

Who it fits

Any agent that writes to live systems or runs at volume. A read-only agent on a stable stack can be run in-house from the runbook.

Included

Run and log monitoring, with anomalies raised to your team
Evaluation suite re-run whenever a tool, API or data source changes
Permission scope and approval thresholds adjusted on what the logs show
Model upgrades tested before they touch the live agent
A monthly report against the baseline agreed at the pilot

Not included, quoted separately

Model and API usage
Hosting
New tools or tasks (priced as extensions)

Turnaround: Monthly, 3-month minimum
Moves the number: Run volume and how often the systems underneath change: a stable ERP and three SaaS tools with quarterly breaking changes are different retainers.

Request this option →

Cheaper or dearer

What moves a quote
down, or up.

The pilot is quoted once the sprint has written down the tools, gates and scenarios, so every item below is a proposal line, not a surprise after it.

Brings the quote down

One task with a clear finish line and one system to read from
Tools with documented APIs and service accounts you can scope narrowly
A read-heavy task: research, reconciliation, or drafting for a person to send
A baseline you already measure for the evaluation to score against
A team that will run the agent in-house from the runbook

Pushes the quote up

Several back-office systems, each a permission boundary to wire and test
Write actions with money attached: purchase orders, refunds, live-system updates
Legacy systems without APIs, reached through exports or a browser
Regulated data, residency rules or a deployment inside your own cloud
A run volume where a rare wrong call stops being rare

What you get

What the fee
actually buys.

The fee buys an agent on one task, the boundaries it runs inside and the evidence that it stays there. No accuracy figure, hours-saved number or completion rate is promised, because those are measured against your baseline after the pilot, not sold before it.

A written verdict on the shape

The sprint says in writing whether the task needs an agent, a chatbot or a workflow, and scopes the cheaper build where one exists.

A permission map you approve

Each tool the agent may call, the access that call carries and the actions that wait for a person, agreed before anything is built.

Approval gates where you set them

Sends, spends and writes to live systems hold for sign-off above thresholds you choose. The agent proposes and explains; a person decides.

Evaluation against your own baseline

Supervised runs on real scenarios, traps included, scored against how the work is done today. Unattended operation is granted only where the scores earn it.

An audit trail for every run

What it was asked, what it called, what it decided and who approved it, so a wrong call is visible rather than silent.

The runbook, the training and the code

How to adjust the scope, add a tool, read the logs or switch it off, with the code, prompts and pipelines owned by you.

AgenciesAgency partners buy at the partner rates on agencies.pixelcrayons.com; volume terms are agreed under NDA.

Request the partner card →

Questions

Frequently
asked.

Because the real price depends on things nobody knows until the task is mapped: how many tools it must call, what it may change and where the approval gates sit. A quote given before that is a guess dressed as a number. The sprint writes them down and ends in a fixed pilot price, or a written case that a chatbot or a workflow would do the job for less.

No. Usage is billed by the provider on your own account, in scenario testing and in production, so you can see it, cap it and change models without asking us. The pilot fee buys the engineering: tool integrations, permission scoping, approval gates, evaluation and handover. The proposal shows usage as its own pass-through line.

Tool count first, then write access. An agent reading from one system and drafting for a person is the bottom of the range. One that reaches several back-office systems, some without APIs, and can spend or write to live records is the top. Each system is a boundary to wire; each write action needs a gate, a threshold and a reversal path. A deployment inside your own cloud adds infrastructure on top.

No. The handover gives your team the runbook, the training and full ownership; a read-only agent on a stable stack can run in-house. The run earns its fee where the agent writes to live systems or the tools underneath it change often, because every change should trigger a re-run of the evaluation suite. The proposal says which you are.

As an extension, priced after the pilot has earned it against your baseline. A new tool is a new permission boundary to wire and test; a new task is a new set of scenarios and gates. Each is its own line, so you can see what the addition costs and sequence it once the first task's results are in.

Because an agent chooses its own path, so the guardrails cannot sit on one step. A chatbot answers one question; a workflow runs one mapped path with a checkpoint where judgement sits. An agent decides at each step what to call next, so the checkpoint has to cover a class of action wherever it falls: anything irreversible, external or above a threshold. Designing and testing that boundary across every tool is the difference.

Other cost guides
and the service itself.

How much does an AI chatbot cost in 2026?Cost guide: the options, what moves the price and what the fee buys.How much does a website cost in 2026?Cost guide: the options, what moves the price and what the fee buys.About AI agentsThe service page: what the work is, how it runs and the proof behind it.AI agents on the rate cardThe same ranges beside every other service, in one card.

Last updated 4 September 2026

What it costs,at the direct rate.

Why the rangeis a range.

How many tools it can call

Whether the task is agent-shaped at all

What it is allowed to change

The scenarios it must survive

How often the systems underneath change

Where the data has to stay

Which optionfits your case.

Scoping and evaluation sprint

Agent pilot

Production run and monitoring

What moves a quotedown, or up.

What the feeactually buys.

A written verdict on the shape

A permission map you approve

Approval gates where you set them

Evaluation against your own baseline

An audit trail for every run

The runbook, the training and the code

Frequentlyasked.

Other cost guidesand the service itself.

What it costs,
at the direct rate.

Why the range
is a range.

Which option
fits your case.

What moves a quote
down, or up.

What the fee
actually buys.

Frequently
asked.

Other cost guides
and the service itself.