Choose an AI automation agency or freelancer

Scope the workflow before the supplier

Describe one trigger, the current process, the intended output, responsible owner and systems involved. List the data classes, approvals, exception paths and decisions that must remain human. A supplier cannot estimate or govern “add AI to the business” as reliably as a bounded operational flow.

Ask the discovery phase to produce an interface map, verified constraints, non-goals, failure modes, measurement plan and staged release path. This deliverable should remain useful even if you choose another implementation partner. A proposal built mostly around tools or a demo leaves the most expensive operating questions unanswered.

Compare named capacity, not company size

For a freelancer, identify backup arrangements, availability, review discipline and the point where specialist help would be added. Direct senior delivery can reduce handoffs, but one person still has finite capacity. For an agency, name who performs discovery, engineering, security review and maintenance rather than treating the logo as the team.

Match capacity to the release. A single integration may not need several parallel roles. A programme spanning data, identity, user interfaces and change management may. Ask how work is reviewed, how decisions are recorded and how ownership transfers if a team member changes.

Make access and data governance explicit

List every account, credential, dataset and production action the supplier may touch. Use organisation-owned accounts, least privilege, separate environments and revocable access. Personal accounts and shared passwords make handover and incident response fragile, whatever delivery model you choose.

For model-assisted steps, define permitted inputs, expected output, deterministic validation and human review. Record provider, model, configuration and retention assumptions that matter. A responsible partner should be willing to block automation when authority, evidence or approval is missing.

Demand production evidence and recovery

A prototype demonstrates a path, not reliable operation. Acceptance should cover successful runs, expected failures, retries, duplicate prevention, observability, alerts and manual recovery. Require a runbook and rollback path before the workflow performs meaningful external actions.

Agree how value will be measured from a baseline. Possible measures include success rate, processing time, correction volume and exception rate, with named sources and periods. Do not accept invented ROI or hours saved. The first release may be valuable because it makes failures visible before it reduces labour.

Choose the maintenance model before launch

Decide who monitors incidents, rotates credentials, reviews provider changes, tests upgrades and answers user questions. Compare a retained freelancer, agency support plan and trained internal owner on response expectations and knowledge transfer. A build without an operating owner is not complete.

Use a short paid discovery or limited production slice when uncertainty is high. Judge the partner on clarity, evidence, restraint and handover quality. Keep the option to expand only after the first workflow survives real use. This makes the agency-versus-freelancer decision reversible and tied to observed delivery.

Decision criteria

Use the same questions for every option before choosing.