The wedge in one sentence: agencies ship the agent. Fractional ships the agent AND the org redesign, eval bar, vendor calls, hiring, and team handoff. Per-hour, agencies are cheaper. Per-outcome, the math flips once you account for the strategy layer.
Same dimensions, both sides. Third column names what to look for when scoring vendors.
Deliverable: Production AI shipped to your repo, eval bar live, team trained, vendor calls owned, hiring run, adoption defended.
Engagement: 6-month minimum, 1 to 3 days per week.
Code lives where: Your repo, your cloud, your accounts. Model-agnostic.
Handoff: Runbook, eval suite, named patterns, observability. Team runs on-call after.
Sits where: Inside the exec staff. Reports into CEO, CTO, or COO.
Deliverable: The agent, the workflow, or the automation, built to spec.
Engagement: Project fee, fixed scope.
Code lives where: Varies. Some agencies retain ownership or run the agent on their stack.
Handoff: Demo, sometimes documentation. Eval, observability, and production reliability often skipped.
Sits where: External vendor. Coordinates with one PM on your side.
Does the code ship to your repo? Or does the agency keep it on their infrastructure?
Is there an eval suite at handoff? No eval = no production reliability after the demo.
Who owns vendor calls and budget defense? If the answer is "you," and you don't have an AI lead, you bought half the work.
What happens at month 6? If the agency stays embedded forever, you're paying for the strategy layer they don't provide.
If your room looks like this, hire an agency, not a Fractional.
Someone owns strategy, eval bar, vendor calls, adoption. You need extra hands to ship a defined automation faster than in-house can.
One document pipeline. One Zapier-style integration. One agent for one workflow. Project fee, fixed deliverable, 4 to 8 weeks. Fractional is overkill.
Braintrust or equivalent already running. You'll wire the agency's output into your reliability bar yourself.
A small one-shot experiment to see if AI moves a metric. Fractional is wrong shape for a single-shot test. Agency or Scoped Build fits.
If two or more show up, the build-only model leaves the strategy layer unowned.
Agency can build. No one will own budget defense, vendor consolidation, or adoption cadence. That work doesn't disappear, it goes unowned.
Agent works. Adoption stalled. That's an org-design problem, not a build problem. Fractional fixes both layers.
Each has its own agency, stack, eval. No allocation logic across them. Fractional pulls the portfolio together with Three-Horizon and a kill list.
Agencies leave with the knowledge. Fractional trains the team to run without me. Eval suite, playbook, named patterns, runbook. Not a permanent dependency.
Quoted in writing within 48 hours. Honest framing on which model fits, even when it isn't Fractional.