Generative AI agencies: hire for creation systems that ship

Give the agency pipeline design, prompt/tooling patterns, eval harnesses, and the first production slice for a named workflow. Keep brand voice ownership, legal/IP policy, publish approval, product priorities, and long-term custody of keys, asset stores, and repos on your side. Hybrid works well: agency ships a bounded generator; your team wires it into marketing ops or the product surface and staffs ongoing review. Write the split into the SOW so a polished reel never leaves you without a maintainable system.

Treat a golden-path reel as marketing, not evidence. Require evals on your brand kit and hard cases (off-brand tone, false claims, IP edge cases), documented human review rates, and cost-per-successful-output tracking. Ask what breaks when inputs are messy or guidelines change, and who owns the eval suite after handoff. Prefer a short paid pilot against your constraints over a free demo that only works on their sample set.

Choose generative AI when the job is creating or adapting content and media. Choose LLM integration when the job is grounding models in your systems of record for answers, copilots, or tools. Many strong shops cover both — ask which workstream owns your first 90 days and how they measure success (creative quality/throughput vs faithfulness/latency).

Pilots are typically fixed-scope for one workflow and often land in the low–mid five figures depending on formats, volume, and integration depth. Production hardening (access control, monitoring, brand-kit refresh, cost ceilings, handoff) should be priced as a separate phase. Ask what is in-scope for the pilot versus explicitly out — and never let the pilot “become” the product without written ownership and ops.

Almost always yes at the start. Even strong generators need humans for brand risk, factual claims, and edge cases until your eval suite and refusal rules are proven on real traffic. Plan review capacity into the pilot; agencies that promise fully automatic publish without a quality gate are selling wishful thinking. Reduce review load over time with measured pass rates — do not pretend day one needs zero humans.