Picking the wrong staffing model kills more AI initiatives than picking the wrong model. We’ve watched companies hire a full-time AI lead before shipping their first workflow, and watched others retain a consultancy on a three-year contract because no one inside the building knew what to ask for. Both failures are about capacity, not technology.

There are three usable shapes: in-house build, agency or consultancy hire, and fractional embedding. Each is right in a narrow set of conditions and wrong elsewhere. The decision matrix below is what we walk through with teams trying to figure out which shape fits.

The three shapes

In-house build

You hire a full-time AI engineer (or an AI-capable engineer) and put the work on their plate. They own the build, the production system, and the eval discipline.

When this fits. You have or expect to have more than three workflows in flight within 12 months. You can pay $200K+ all-in for the right person. You have a pipeline of AI work that’s clear enough to attract a senior engineer — they need to know what they’re building before they accept the role.

When it doesn’t. You haven’t shipped one workflow yet. Hiring before the first ship is the most common mistake we see. The new engineer arrives, has no existing system to learn from, has to invent the architecture, the eval discipline, and the workflow selection process — alone. They burn out or leave inside 18 months because the role was scoped before anyone knew what the role actually involved.

Agency or consultancy hire

You retain a firm that brings a team — usually a senior lead plus mid-level engineers — to deliver a defined scope. The engagement is typically project-based with a defined endpoint, though many agencies push for ongoing retainers.

When this fits. You have one specific workflow that needs to ship and you don’t plan to do more for at least six months. The agency’s value is compressing the time-to-first-ship by bringing pre-built infrastructure (eval harnesses, monitoring patterns, model-swap discipline) you don’t need to invent yourself.

When it doesn’t. You have ongoing AI work and the agency’s economics depend on you continuing to need them. Most agencies pyramid — the senior on the sale is not the senior on the build, and the build engineers rotate off the project as soon as a higher-priority client appears. For ongoing work, the agency is structurally misaligned with making your team self-sufficient. They make less money when you don’t need them.

Fractional embedding

You retain a senior operator on a fractional basis — typically 4–6 hours a week — to work alongside your team. They’re not running the work themselves; they’re providing senior judgment, eval discipline, and model-swap decisions while your team executes.

When this fits. You have at least one in-house engineer who can do the implementation work but lacks the AI-specific operating experience. The fractional senior provides the experience without the cost of a full-time hire, and your engineer learns the discipline by doing the work rather than receiving a deliverable.

When it doesn’t. You have no in-house engineer to receive the embedding. The fractional model assumes there’s someone on your side who can execute between sessions; without that person, the senior is doing all the work themselves on a four-hour-a-week budget, which means nothing ships.

The five-factor decision matrix

Score each of the five questions below. The pattern of your answers points to the right shape.

Factor 1 — How many workflows are in flight?

  • One — agency fits well if you can name what you want.
  • Two to four — embed or agency, depending on factor 3.
  • Five or more, or growing — in-house starts to make sense.

Factor 2 — How long is your time horizon?

  • Three to nine months, defined endpoint — agency.
  • Ongoing, no end date — embed (best) or in-house (if scale warrants).
  • Indefinite but uncertain — embed. The fractional commitment is reversible in a way the others aren’t.

Factor 3 — Do you have an existing senior AI/eng person?

  • Yes, senior, AI-capable — embed adds the AI-specific discipline they may not have without another full-time hire.
  • Yes, senior, not AI-experienced — embed teaches them the discipline; the work itself goes through them.
  • No senior, just mid-level — agency is the safer bet because the build needs senior judgment that mid-levels alone can’t reliably provide.
  • No engineering team at all — agency, with the explicit understanding that operating the system long-term requires hiring an engineer.

Factor 4 — Is the budget per-project or recurring?

  • Per-project — agency model fits the cash structure.
  • Recurring monthly budget — embed (most aligned) or in-house (over a longer horizon).
  • Capped, one-time — agency, but be careful about the size of the cap. Real builds are 4–8 weeks of senior time. Capping below that produces a deck.

Factor 5 — Will the work be intermittent or continuous?

  • Intermittent (one push, then quiet, then another push) — embed handles this elegantly because the fractional commitment scales up and down.
  • Continuous (always something in flight) — in-house if the volume justifies it; otherwise embed.

Reading the matrix

Add up your answers. Most teams will see a clear pattern.

If three or more answers point to agency: scope a defined project with one firm and protect against scope creep with a clear shipping definition (running on day 60, eval harness in place, runbook handed off).

If three or more point to embed: retain a fractional senior and pair them with the engineer who’ll inherit the work. The right person is someone who’s shipped this pattern before, not someone who can talk about it.

If three or more point to in-house: hire — but only after you’ve shipped at least one workflow with one of the other shapes. The hire is much easier when you can describe what they’ll do on day one with concrete reference to a running system.

Two anti-patterns

Hiring full-time before shipping one

We get pulled into AI hiring conversations regularly. The company has decided to hire a Head of AI or Principal AI Engineer before they have any AI in production. The recruiter wants to know what to look for in a candidate.

The honest answer: don’t hire yet. The role description will be vague because no one inside the company knows what the role actually involves. The candidate will join and spend three months figuring out what to build. That figuring-out work is exactly what an audit or a fractional embed delivers in two to four weeks. Do that first. Then hire with a real job description.

Retaining an agency for ongoing work

Agencies are economically incentivized to keep you on retainer. Their pricing assumes you don’t internalize the discipline they’re providing. After the initial project ships, they will propose extending into “optimization,” “new workflows,” “ongoing eval review,” and similar work that could be done in-house if you had the people.

For one-shot builds, the agency is great. For ongoing work, the math doesn’t favor you. Either hire (if the volume is real) or embed (if you want senior judgment without full-time cost). Don’t pay agency rates indefinitely for work that doesn’t require an agency’s mode of delivery.

Agencies make less money when you don’t need them. Plan for the day you don’t — or pick a shape where the incentives align with your independence.

What to do with this

Write down your answers to the five factors before writing a JD or signing a contract. The right shape is usually obvious once the answers are on paper. If they’re mixed, the most likely fit is fractional embedding — it’s the most flexible shape and the easiest to exit if it isn’t working.

If you’ve already picked a shape and it isn’t working, the most common fix is moving from agency to embed. The transition is easier than it sounds: you keep the systems that were built, and the embed senior helps your team operate them rather than rebuilding from scratch.

The previous chapter on eval harnesses and AI workflow governance covers what your in-house team needs to operate, regardless of which shape built the system originally. The earlier chapter on the readiness audit is also worth re-reading through the staffing lens — several of the diagnostics directly affect which shape will work for your team.