How Much Does It Actually Cost to Build a Custom AI Agent in 2026?
A custom AI agent costs $15K to $50K to build for one workflow in 2026, $50K to $100K across several systems, and $4 to $400 per 1,000 tasks to run.
Feb 12, 2026
Short answer: it costs $15,000 to $50,000 to build a custom AI agent that does one job well in a growing business, with a senior team, in two to three months. An agent that reads and writes across several systems, with approvals and audit records, costs $50,000 to $100,000. Keeping it accurate after launch costs $5,000 to $10,000 per month if you have no engineers of your own. Running it is the small number: model fees for 1,000 agent tasks come to between $4 and $400 a month, using API prices read from OpenAI, Anthropic and Google on 14 October 2026.
If you want a range for your own scope first, the AI development cost calculator takes about a minute.
AI agent cost by complexity tier
Most quotes differ because they price different agents. Sort the agent you need into one of three tiers before you compare numbers.
| Tier | What the agent does | Build cost | Typical time | Running cost |
|---|---|---|---|---|
| Pilot on a workflow tool | Answers or drafts inside one tool; a person checks every output; no writes to business systems | A few thousand dollars of freelance time at $30 to $150 an hour (Upwork, read 14 Oct 2026) | 1 to 4 weeks | Tool plan plus model fees |
| First production agent (KUMO Starter Build) | One workflow end to end: retrieval with citations, one or two systems read or written under written rules, handoff to a person, a test set, deployed in your cloud | $15,000 to $50,000 | 2 to 3 months | Model fees, hosting, logging |
| Agent across several systems | Writes to several systems, approvals by role, audit records, rollback when a step fails | $50,000 to $100,000 | 3 to 5 months | As above, plus more monitoring |
| Ongoing engineering | Failure review, test reruns after model changes, new tools, knowledge refresh | $5,000 to $10,000 per month | From launch | Included |
The pilot tier is a real option. If the work is drafting replies or summarising documents and a person approves everything, a workflow tool and a freelancer can prove the idea before you spend a production budget. The step up to the second tier happens when the agent must act on its own inside systems that hold money, customer records or commitments.
What 1,000 agent tasks cost to run
Model fees are the line people overestimate. Take one task as five model calls, each reading about 6,000 tokens of instructions, context and tool results and writing about 400 tokens. That is 30 million input tokens and 2 million output tokens per 1,000 tasks.
| Model | Price per million tokens, input / output | Cost per 1,000 tasks | Source, read 14 Oct 2026 |
|---|---|---|---|
| OpenAI GPT-6 Luna | $0.10 / $0.50 | $4 | developers.openai.com pricing, standard |
| Anthropic Claude Haiku 5.5 | $0.10 / $0.50 (prompts up to 100,000 tokens) | $4 | platform.claude.com pricing |
| Google Gemini 3.8 Flash | $0.75 / $3.75 through 31 Dec 2026, then $1.50 / $7.50 | $30, then $60 | ai.google.dev pricing, paid tier |
| Anthropic Claude Sonnet 5.5 or OpenAI GPT-6.1 Sol | $2 / $10 | $80 | platform.claude.com; developers.openai.com |
| Anthropic Claude Opus 5.5 | $4 / $20 | $160 | platform.claude.com pricing |
| OpenAI GPT-6 Astra or Anthropic Claude Fable 5.1 | $10 / $50 | $400 | developers.openai.com; platform.claude.com |
At $0.08 a task on a mid-tier model, the bottom of the Starter band, $15,000, would pay for 187,500 tasks. The build is where the money goes: connecting the agent safely to your systems and proving it behaves.
For comparison on people, the US median software developer wage was $135,980 a year in May 2025 (Bureau of Labor Statistics, Occupational Outlook Handbook), about $11,330 a month before benefits and recruiting, for one engineer with no testing or product cover.
What moves the build cost inside a tier
Operations, not apps. Reading an invoice and issuing a credit are two different builds even inside the same billing system. Count each read and each write the agent performs.
How much it may do alone. An agent that drafts for a person is cheaper than one that changes records. Anything near money, customer commitments or access rights needs allowed-action lists, limits and a way to undo.
Approvals. "Human in the loop" is a feature: a queue, an owner, what the reviewer sees, and what happens if nobody approves in time.
Source quality. Conflicting policies, stale pages and answers buried in attachments add work. Cleaning the sources first is cheaper than prompting around them.
Testing. A production agent needs a test set of real, awkward and unsafe requests with pass criteria, rerun after every change. A demo skips this, which is why demos are cheap.
Volume. High volume pays for caching, batching and routing easy requests to a cheaper model. Each is build work that lowers the monthly bill.
Costs people miss after launch
Plan from month one for model fees, hosting, logging and any workflow tool plan, then for engineering time: reviewing failed answers, refreshing knowledge, rerunning tests when a model version changes, and fixing connectors when another system changes its API. Without engineers of your own that is the $5,000 to $10,000 per month band, falling once the agent is stable. Software maintenance cost in year one breaks the categories down.
Two more: ownership and portability. Ask who owns the code, the prompts, the test set and the cloud accounts. If the answer is the vendor, budget for rebuilding when you leave.
Build, buy or pilot
Buy a platform when the job is common to every company (meeting notes, document search, writing help) and seat pricing beats a build at your headcount. Copilot Studio vs ChatGPT Business vs a custom agent prices that choice for a 50-person company.
Pilot first when nobody has measured how often the task happens or how good a draft has to be. A few weeks on a workflow tool answers both.
Build when the agent must act inside your own systems under your rules, when the data cannot leave your cloud, or when the workflow is how you compete. If you are weighing a fuller cost breakdown with three sample quotes for one brief, AI agent development cost walks through it.
What the Starter band buys at KUMO
A KUMO Starter Build ($15,000 to $50,000) for an AI agent is a launched first production version: one workflow end to end, retrieval with citations, one or two systems read or written under stated rules, handoff into the queue your team already uses, a test set with agreed pass criteria, monitoring, deployment in your own cloud account, and a handover another engineer can continue from. The code is yours from day one.
How KUMO builds agents into operations is on AI workflow automation. KUMO's founders built Volopay's first production version as its founding engineers, and KUMO is rated 4.9 on Clutch. To talk through your scope, book a free 30-minute call with an engineer.
Frequently asked questions
How much does it cost to build an AI agent? $15,000 to $50,000 for a first production agent that does one job inside your business, built by a senior team in two to three months. Agents that act across several systems cost $50,000 to $100,000.
What is the AI agent development cost per month after launch? Model fees are usually $4 to $400 per 1,000 tasks at 14 October 2026 API prices, plus hosting and logging. Engineering upkeep without your own team is $5,000 to $10,000 a month.
How much does AI agent development cost with a freelancer? Upwork lists AI agent developers at $30 to $150 an hour (read 14 October 2026). That suits a pilot; ask who reviews failures and who holds the keys before you rely on it.
How much does it cost to add natural-language AI agents to a website platform? If the agent only answers from your content, it sits in the Starter band or below. If it acts on customer accounts, bookings or payments, expect the top of the Starter band or the $50,000 to $100,000 band.
Is a custom AI agent worth it compared with off-the-shelf tools? It is when the agent must work inside your systems under your rules. For general writing and search, a seat tool is usually cheaper. Run the numbers for your headcount before deciding.
