The Real Cost of Building AI Agents In-House: A 12-Month Budget

The real cost of building AI agents in-house is mostly people. In my worked 12-month budget for a first agent, a lean team costs about $316,000, a typical two-engineer team about $619,000, and a heavier team about $1.09 million. Model spend is under 3% of the typical case. Expect about seven months from approval to production traffic.

What is the real cost of building AI agents in-house?

An in-house agent has three cost layers. The first is the team: salaries plus benefits, for every month from approval, including the months before anything ships. The second is the hidden work around the model: data, integrations, evals (repeatable tests with known answers), security review, on-call, and model migrations. The third is the run cost after launch, which never reaches zero.

In the base budget below, people are 87% of year-one cost and model tokens are 2.8%. Model spend would have to rise about 13 times before it matched the cost of the one engineer you hire.

Token costs scale with volume. Team costs scale with calendar time. A first agent has low volume and a long calendar, so the calendar wins. Google engineers made the same point about machine learning a decade ago in Hidden Technical Debt in Machine Learning Systems: “it is common to incur massive ongoing maintenance costs in real-world ML systems.” Agents add tool calls, permissions, and vendor model churn on top.

Still choosing between buying and building? Start with the scorecard in build vs buy AI agents.

The team you need, and what it costs loaded

A first production agent needs two engineers, part-time testing, part-time security, and real time from the person who owns the workflow. One engineer is a single point of failure. More than three rarely ships the first agent faster, because hiring, data access, and security review set the pace. That is my judgment.

How I loaded the salaries. BLS Employer Costs for Employee Compensation, released September 9, 2026, puts total compensation for full-time private-industry workers in management, professional, and related occupations at $82.23 per hour worked in June 2026. Wages were $55.74 of that, or 67.8%. So for BLS wage figures, loaded cost is about wage ÷ 0.678, a 1.475 multiplier. The other 32.2% is paid leave, supplemental pay (bonuses and overtime, $3.88), insurance, retirement, and legally required benefits. Stack Overflow’s figure is total pay including bonuses, so for it I divide by wages plus supplemental pay instead: $82.23 ÷ $59.62 = 1.379. ECEC excludes recruiting, laptops, software seats, and office space, so I budget those separately.

The Stack Overflow figure is self-reported by 42 US respondents, fielded June 23 to August 5, 2026. It also counts perks, which may overlap other benefit lines. I cannot size that overlap, so treat the AI engineer line as an upper bound. I use it because the BLS has no “AI engineer” category.

The hidden work: a cost category table

Every row is work the agent needs before it is safe in production. The ranges are my estimates for a first agent touching two or three systems, not survey data.

For two systems read and one written, the engineering rows add up to roughly 15 to 29 engineer-weeks. The base budget gives about 50 before launch. The gap covers ramp-up, waiting on access, and rework; I assume a new team turns a third to a half of its time into shipped work.

Data. In February 2025, Gartner predicted that through 2026, organizations will abandon 60% of AI projects unsupported by AI-ready data. In its survey of data management leaders, 63% did not have, or were unsure they had, the right data practices for AI. On-call. Google’s SRE book caps on-call at 25% of an engineer’s time. On that rule, a single-site 24/7 rotation with a primary and a secondary needs at least eight engineers. Two engineers can cover business hours. They cannot sustain a 24/7 rotation within those limits.

How long until the first agent is in production?

Plan on about seven months from approval to production traffic.

Month 3 is optimistic for the hire. SHRM’s 2026 recruiting benchmarks, drawn from over 4,600 organizations, put median time-to-fill at 39 calendar days for nonexecutive positions. I found no reliable public figure for AI engineers specifically.

Outside data agrees. In Gartner’s Q4 2023 survey of 644 respondents, it took 8 months on average to go from AI prototype to production. Only 48% of AI projects made it at all. That survey predates most agent work. MIT NANDA’s The GenAI Divide (July 2025) found that top-performing mid-market companies reported an average of 90 days from pilot to full implementation, while enterprises took nine months or longer.

Decision rule: if the business needs the agent working inside 90 days, treat an unhired team as substantial schedule risk. SHRM’s 39-day median alone is 43% of that window. Add notice and onboarding, which I assume take two to four weeks each.

A worked 12-month budget for a first in-house agent

Everything here is hypothetical except the BLS figures, survey figures, and model prices. The back-office agent reads from two systems, writes to one, and reaches 20,000 tasks a month in production.

Model cost per task. Claude Sonnet 5.5 costs $2 per million input tokens and $10 per million output tokens on Anthropic’s pricing page as of October 15, 2026. A cache hit costs $0.10 per million. Each task makes 5 calls of 10,000 input tokens (half of them cached) and 800 output tokens. One call is $0.0100 + $0.0005 + $0.0080 = $0.0185. Five calls are $0.0925. Add 20% for retries: $0.111 per task. I ignored cache-write charges.

The checks. The five people lines sum to $536,985, which is 86.7% of the total. Model lines sum to $17,430, which is 2.8%. Dividing the AI engineer line by the model total gives 13.2, the “13 times” figure above.

What does $619,415 buy? One agent live for five months, a test set, tracing, a security review on file, and a team that knows the workflow. It does not buy a second agent. If the software engineer was pulled off product work, the real price also includes whatever that engineer would have shipped instead.

Sensitivity: team size and model spend

The base case hides two big assumptions: team size and token appetite. The table moves each one alone.

Where do 3.75 and 15 come from? Anthropic reports that in its data, agents typically use about 4 times the tokens of chat, and multi-agent systems about 15 times. Both ratios compare to chat, so multi-agent versus single agent implies about 15 ÷ 4 = 3.75 times. The 15x row is a stress assumption, not a measured ratio.

Each engineer added or removed for a full year moves the total by $200,560 to $275,847, or 32% to 45% of base. A multi-agent redesign moves it by $47,932. Team size is the lever. Model choice matters only at high volume or with multi-agent designs.

The lean case hides a failure mode: one engineer means no review on changes and no cover for vacation. Detect it early. If changes ship without a second reader, you are running the lean case whatever the budget says.

What it costs to keep the agent running

Year two, same agent, 20,000 tasks a month. The team shrinks but stays.

Why half an engineer? Assume one or two forced migrations and about 20 prompt or tool changes a year. Then the recurring column in the category table adds up to roughly 13 to 28 engineer-weeks a year. That is my estimate. The total is $17,623 a month, or $0.88 per task, which is eight times the $0.111 token cost. Tokens are 13.2% of it.

Model migrations are the forced part of that work. Anthropic gives at least 60 days’ notice before retiring a publicly released model, and requests to retired models fail. Claude Sonnet 4.5 was deprecated on September 30, 2026, with a tentative retirement date of November 30, 2026. OpenAI’s deprecations page shows the same churn: on June 11, 2026, it announced that the original GPT-5 and o3 snapshots shut down on December 11, 2026. Tooling retires too. OpenAI’s Evals platform and Agent Builder are scheduled to shut down on November 30, 2026. Change control for prompts, models, and tools covers the release process, and agent evals in CI covers the gate.

Decision rule: below about 137,000 tasks a month, non-token costs dominate the run cost. At that volume, all tokens, including $111 a month of eval runs, come to about $15,292 a month, equal to everything else. Below it, cut cost by sharing the team across agents, not by switching models. The monthly run-cost model breaks down the per-task lines.

Where in-house builds win, and where they lose

The first agent is the expensive one. Its test runner, tracing, security patterns, and tool connectors carry over to agent two. That reuse is the whole economic case for in-house.

In-house wins when:

• You can name agents two and three. If the roadmap is one agent, the year buys five live months.
• Data cannot leave your environment and no vendor can contractually run where it must.
• The workflow logic is your edge, such as pricing or underwriting rules built from your own history.
• You already run a platform team with on-call. That removes the eight-engineer problem, because the agent joins an existing rotation.

In-house loses when:

• You need value inside 90 days. Hiring and security review alone use most of that window.
• The agent must act around the clock and you have no 24/7 rotation. Design it to stop and hand work to a human queue, or do not build it yet.
• Nobody owns migrations. Anthropic’s 60-day minimum notice, with no test set, means a scramble every time a model retires.
• The team would come off revenue work. Price that trade explicitly.

Short on people? AI agency vs in-house team covers who builds first. Who owns the AI agent covers the roles to name before launch.

Bottom Line

Building AI agents in-house is a staffing decision, not a model decision. People are about 87% of year one in my base budget.

Budget for the hidden work: data, evals, security review, on-call, and migrations. Plan about seven months to production.

Next step: fill in the 12-month table with your own salaries and volume. Then compare it with three fixed-scope quotes, using AI agent development cost to check what each quote includes.

FAQ

Is model API spend the biggest cost of an in-house agent?

Not at first-agent volumes. When it grows, timing is often the cheapest lever. Anthropic’s Batch API processes requests asynchronously at a 50% discount on input and output tokens, which suits overnight backlogs.

What happens when the model our agent uses is retired?

Calls to it fail. Unless safety or compliance concerns require a faster timeline, OpenAI commits to at least 6 months’ notice for generally available models, at least 3 months for specialized variants such as Codex models, and as little as 2 weeks for preview models. Keep preview models out of business-critical paths.

Can we hire contractors instead of full-time engineers?

Yes. In Stack Overflow’s 2026 survey, 14.9% of respondents described themselves as independent contractors, freelancers, or self-employed. Budget their rates, sourcing time, and onboarding as separate lines. Do not apply the employee benefits multiplier on top, because a contractor’s rate may already include overhead. Write handover and test-set ownership into the contract.

Does the second in-house agent cost as much as the first?

No. My estimate is about half the build effort, because the test runner, tracing, security patterns, and connectors already exist. Run cost also grows less than one for one.

Three Ways to Work With TMA

Need an agent built? We deploy production AI agents in your infrastructure. Working pilot. Real data. Measurable ROI. → Schedule Demo

Want to co-build a product? We’re not a dev agency. We’re co-builders. Shared cost. Shared upside. → Partner with Us

Want to join the Guild? Ship pilots, earn bounties, share profit. Community + equity + path to exit. → Become an AI Architect