How to Size and Budget a Team for an AI Agent Engagement

Quick answer

A minimum viable team for even a narrow, single-workflow AI agent is three roles: a technical owner who scopes and prioritizes, an engineer who owns the integration into your real systems, and a subject-matter reviewer who spends part-time hours validating outputs against reality. Whether those roles sit in-house or inside an agency, the roles don't disappear. And budget correlates far more with integration complexity and the ongoing evaluation effort than with headcount or model cost.

The roles exist whether you build in-house or hire an agency

The mistake teams make when comparing in-house vs. agency cost is comparing "our headcount" to "their invoice" as if the invoice includes something extra. It doesn't. It's the same roles, just on the vendor's payroll instead of yours. Four roles show up in essentially every real AI agent engagement:

• Technical owner / PM: scopes the workflow, prioritizes edge cases, makes the ship/no-ship call.
• Integration engineer: connects the agent to your actual systems (CRM, ticketing, TMS, whatever holds the source-of-truth data). This is usually the largest single line item.
• Subject-matter reviewer: the person who can say "that answer is actually wrong" for your domain. Almost always underestimated in time commitment.
• Eval/monitoring owner: someone accountable for the agent's accuracy over time, not just at launch. Often the same person as the technical owner on a small team, but the responsibility has to be assigned explicitly.

Sizing in-house by project maturity

For a pilot (one workflow, internal or low-risk external use): one part-time technical owner, one engineer at 50-100% allocation for the integration work, and a few hours a week from a subject-matter expert for review. This is usually achievable without a dedicated hire if you already have the engineering capacity.

For a scaled deployment (customer-facing, multiple workflows, or regulated data): the team needs to become closer to full-time across all four roles, plus dedicated eval/QA time that scales with the number of workflows in production. Not just at launch, but ongoing, since model and prompt changes need to be re-validated.

Sizing an agency engagement

You're paying for their version of the same four roles, packaged into a rate. What you should be evaluating isn't "is their day rate competitive" in isolation. It's whether their proposal explicitly staffs the subject-matter-review and post-launch eval roles, or whether those get quietly left as "your team's responsibility" after handoff. That gap is where agency engagements go over budget: the SOW covers build, not the ongoing evaluation loop that a production agent actually needs.

Where the budget actually goes

Line item	Typical share of budget	Notes
Model/inference cost	Smallest share	Rarely the constraint people expect it to be
Integration engineering	Largest share	Connecting to real systems, not writing prompts
Evaluation & monitoring (ongoing)	Recurring, often underbudgeted	Continues after launch, not a one-time cost
Subject-matter review time	Recurring, easy to undercount	Often informally "borrowed" from someone's day job

Common sizing mistakes

The most common mistake is treating "we hired an AI agency" as equivalent to "we no longer need internal time". Someone on your side still has to own the subject-matter review and the go/no-go decisions, and that time needs to be budgeted, not assumed to be free. The second most common mistake is sizing the team for the build phase and forgetting that eval and monitoring are ongoing costs, not launch costs. See our ROI framework for how that recurring cost factors into the payback math. For the underlying cost breakdown that team sizing rolls up into, see our AI agent cost guide; if you're still deciding whether to staff in-house or hire out at all, that's the build-vs-buy call covered in our agency vetting checklist.

FAQ

How many engineers do you need for a first AI agent project? Often one, if they can own the integration work end to end. The bottleneck is rarely a need for more engineers, it's needing dedicated subject-matter review time that a solo engineer can't provide themselves.

Is it cheaper to hire in-house or use an agency for an MVP? For a single pilot workflow, an agency is usually faster and often cheaper than a new hire, since you're not carrying salary overhead for a role you may only need at full intensity for a few months. For an ongoing, multi-workflow program, in-house tends to be more cost-effective over 12+ months.

What percentage of budget should go to ongoing monitoring and eval? There's no universal percentage, but it should be a named, recurring line item from the start. Not something added retroactively after the first production incident, which is when most teams discover they under-budgeted it.

Do you need a dedicated ML engineer, or is a generalist engineer enough? For most agent projects built on top of existing foundation models, a strong generalist engineer who's comfortable with APIs, prompting, and evals is enough. A dedicated ML/research engineer becomes necessary only if you're doing custom fine-tuning or training, not for typical agent + retrieval + tools builds.

Accelate scopes team size and budget against your actual workflow's integration and eval needs, not a generic day-rate estimate, so you know what you're staffing before the project starts.