Build In-House or Hire an Agent Agency: The Real Numbers

Most build-versus-buy conversations for agentic AI start with the wrong question. The debate gets framed as platform economics — license fees versus engineer salaries — when the actual decision is about which organization owns the failure modes. An agent that books travel, moves money, or closes tickets can be wrong in ways a dashboard cannot. Whoever owns the runbook when it misbehaves at 3 a.m. is the one who really bought it.

Both paths are legitimate. Either can pay for itself, and either can quietly drain a budget for eighteen months while producing demo videos. What separates the two is not vision or model choice; it is fully loaded cost, time to a first production agent, and the ongoing tax of keeping that agent trustworthy. This piece prices both honestly and names the conditions under which each one wins.

So what does each path actually cost, and when should you refuse to hire an agency at all?

What an In-House Agent Team Really Costs in Year One

The salary number is the smallest part. Robert Half's 2026 guide puts the AI and ML engineer midpoint at $170,750, with employer taxes and benefits typically adding 25%–40% and agency recruiting fees another 15%–25% of first-year base. That is one engineer, and one engineer does not ship a governed agent platform.

A credible in-house pod for a regulated enterprise looks like this: a tech lead, two agent engineers, a platform or DevOps engineer, a security/compliance partner at roughly half time, and a product owner who actually understands the workflow being automated. Independent market analyses put a fully loaded six-person U.S. AI team at $1.2M to $2.5M annually, and that is before GPU credits, model API spend, observability tooling, and the audit budget.

The multiplier bites harder at the senior end. Industry hiring benchmarks put a senior AI engineer's year-one loaded cost at roughly $300K to $460K once recruiter fees, tooling, and inference spend are counted. Miss four of the seven line items and the plan lands 35–50% under.

Two more costs almost never appear in the original slide:

• Ramp. A senior AI hire produces useful platform code somewhere around month three to five. A whole pod finding its rhythm takes longer.
• Retention risk. PwC's 2025 AI Jobs Barometer found workers with AI skills command a 56% wage premium. Your compensation bands will be tested every quarter.

What a Senior-Led Agency Actually Prices

Agencies come in three flavors that get lumped together and shouldn't be. Body shops rent engineers by the hour. Systems integrators sell decks and staff. A senior-led execution firm — the category Automatic.co sits in — sells a delivered agent running inside your systems, on a fixed scope, with the architecture handed to your team at the end.

Pricing for that third category typically lands in one of two shapes. A discovery-plus-first-agent engagement for a mid-complexity workflow usually runs in the low-to-mid six figures over eight to sixteen weeks. A platform build with three to five agents, shared tooling, an approval framework, and a runbook typically lands between $400K and $900K over roughly two quarters. Those are honest ranges, not floors. Regulated deployments — HIPAA, SOC 2, GLBA, air-gapped — add roughly 25–40% for validation, evidence collection, and control mapping.

You are paying for four things that are hard to hire individually: an AI agent architecture the compliance team will actually sign off on, a tool registry and approval framework that survives the second and third use case, forward-deployed engineers who have already made the mistakes on someone else's system, and a clean handover so your internal team runs it in year two.

Where the math breaks against the agency: any workflow where an average model called from a scripted pipeline would do. If you don't need agent behavior, don't buy it. A retrieval-augmented FAQ bot is not worth $600K to anyone.

Time to a First Production Agent

Speed is not a vanity metric here; it is a compounding one. Every month the platform is not in production is a month of paying the pod without recovering value, and it is a month where organizational patience erodes. The MIT NANDA "GenAI Divide" report captures the gap starkly: mid-market organizations move from pilot to full implementation in about 90 days on average, while large enterprises typically take nine months or longer.

The nine-month figure is where in-house builds usually land, and often overshoot. A candid sequence looks like this: two months to hire and onboard, two months for architecture and tool integration, two to three months to a hardened first agent, another two to the second. A senior-led agency compresses the first two stages to roughly zero and shortens the middle by working from patterns already field-tested elsewhere.

Deloitte's 2026 State of AI research is consistent with this: initial AI pilots deliver learning quickly, while scaling to strong returns typically takes six to twelve months or longer. Whichever path you pick, budget the calendar honestly.

The Governance Overhead Nobody Prices

The line item that quietly kills in-house programs is not the build; it is the tax of keeping agents in production. Gartner has flagged the risk directly, predicting that over 40% of agentic AI projects will be canceled by the end of 2027, citing escalating costs, unclear business value, and inadequate risk controls.

The governance stack that keeps an agent trustworthy is not glamorous, and it does not get built by accident:

• Decision lineage. Every tool call, every model output, every human override, indexed and replayable. The value of an audit log is exactly zero until the day it saves you.
• Approval gates. Well-placed approval gates around irreversible actions — payments, external comms, production changes.
• Versioning and rollback. Prompt, tool, and policy versions pinned so a bad Tuesday rollout can be reverted without archaeology.
• Drift and calibration monitoring. Because model providers ship silent updates and your prompt was tuned to yesterday's behavior.
• Human-in-the-loop that is not theater. Genuine override authority, not a rubber stamp. The human-in-the-loop myth is what happens when this is skipped.

None of this is optional in a regulated environment, and all of it needs an owner. If the plan does not name that owner and their budget, the plan is incomplete.

When You Should Not Hire an Agency

An agency engagement is the wrong answer more often than agency websites will tell you. Do not hire one if any of the following is true.

The workflow is not ready. Bad processes make terrible agents. If the human version of the process lives in tribal knowledge, disputed spreadsheets, and Slack DMs, an AI readiness assessment or a workflow mapping exercise will pay back faster than a build. RAND's analysis of more than 2,400 enterprise AI initiatives found 80.3% fail to deliver their intended business value — and the ones that fail usually fail on the process side, not the model side.

The ROI is thin. If the workflow moves fewer than a couple thousand transactions a month, or the human step it replaces takes minutes rather than hours, the math often will not clear a six-figure build. Say so out loud before the SOW is signed.

You already have the pod. If you employ two or three engineers who have shipped agent systems, plus a platform lead, plus an executive who will hold the line on scope, an agency mainly buys you calendar. That may still be worth it. Often it is not.

Leadership wants a demo, not a system. This is the failure mode behind the MIT NANDA finding that 95% of generative AI pilots deliver no measurable P&L impact despite $30–40B in enterprise investment. No vendor can save a program whose sponsor only wants a video for the board.

The Hybrid Pattern That Usually Wins

For most regulated enterprises the honest answer is not build or buy; it is buy the first two agents and the platform pattern, then run the third one yourself. A senior-led firm delivers the architecture, the tool registry, the approval framework, and a runbook. Your team pairs on delivery, then owns operations and the next backlog of workflows. The agency's incentive is a clean handover; yours is a team that gets stronger with each release.

Two internal disciplines make this pattern hold. First, an automation roadmap with adult supervision — a prioritized queue where each candidate workflow has a named owner, a measurable baseline, and a kill criterion. Second, a clear line between agent code your team maintains and vendor code that stays vendor-neutral, so no single model or framework becomes a lock-in.

Making the Call

The decision is not really about cost per engineer. It is about which organization is better positioned to absorb the cost of being wrong in production. An in-house team wins when you have a steady, multi-year backlog of proprietary workflows and executive patience measured in years, not quarters. A senior-led agency wins when the calendar matters more than the org chart, when the compliance surface is unfamiliar, and when the first agent needs to be right rather than instructive.

The failure mode both paths share is treating the launch as the finish line. An agent that only suggests is a search box. An agent that acts without lineage, approvals, and a rollback plan is a liability with a friendly interface. Whichever path you choose, price the governance, name the owner, and set a kill criterion before the first pull request. The organizations that clear the GenAI divide are the ones that did the boring parts on purpose.