Skip to content
SwitchyardAI automation agency · Australia Book 15 min

Judging an agency

How to judge an AI automation agency before you sign

Eight questions, the answers that should reassure you, and the ones that should not. Then the contract clauses worth reading twice.

The eight questions

Ask these in the first or second meeting. A good agency answers most of them without hesitation, because they are the same questions it asks itself before taking on a project.

Questions to ask an automation agency
QuestionA good answerWorth worrying about
Who owns what you build?The workflows, prompts, configuration and code built for us are ours, and the platform accounts are in our name or will be transferred on a set date.Vague references to “our platform” without saying what happens to your automations if you leave. Licence fees that grow with every new automation.
Where does our data go?A simple diagram per automation: which systems are read, which AI service is called, whether anything is stored, and in which country it is processed.“It’s all secure” without detail. Personal information flowing through a consumer AI account.
Which AI models do you use, and on what terms?The provider and the account type, and whether that provider’s business terms allow your inputs to be used for training.The same personal ChatGPT or similar login used for several clients.
Where is the human in this?A named approver for every step that sends, pays, deletes or changes a record, with an easy way to approve or reject.Fully autonomous sending to customers from day one.
What happens when it breaks?Failures are logged, a named person is alerted, and the automation stops rather than guessing.No answer, or “it won’t break”.
Who maintains it and what does that cost?A written maintenance option with a clear scope, and documentation good enough that someone else could take over.Maintenance only available from the original agency, at an unstated rate.
How do we leave?A handover checklist: documentation, credentials, exported configurations, and a clean switch-off procedure.No exit clause in the contract.
How will we know it worked?The current process is measured before the build (time per item, error rate, delay) and the same measures are taken after.Savings figures that are estimated but never measured.

Contract clauses worth reading twice

  • Intellectual property. Look for a clause assigning the deliverables to you on payment. If the agency keeps ownership of reusable components, the clause should still give you a perpetual right to use and modify what was built for you.
  • Accounts and credentials. Who holds the admin login for each platform, and how it is handed over. Credentials shared over email or chat are a bad sign.
  • Data handling. What the agency may access, for what purpose, how long it keeps copies, and how it tells you about a data breach. If you are covered by the Privacy Act, this should match your own obligations.
  • Acceptance. What counts as finished. “The automation runs” is weaker than “the automation processes the agreed test set with the agreed accuracy and every exception reaches a person”.
  • Change of scope. How extra work is priced. Fixed-price stages with a written change process are easier to manage than open hourly billing.
  • Termination and exit. Your right to end the arrangement, what is handed over, and how long the agency supports the transition.
  • Third-party costs. Which subscriptions and usage-based AI charges you will pay directly, and roughly how much, so the total cost is visible before you start.

None of this is legal advice. For a contract of any size, have it read by your lawyer; the list is to help you ask the right questions first.

Who maintains it, really

Automations are not “set and forget”. Software you connect to changes its forms and fields. A supplier changes its invoice layout. An AI model is updated and behaves slightly differently. A staff member who approved everything leaves. Each of those can quietly break an automation, or worse, make it quietly wrong.

There are three realistic models. Your own staff maintain it, which works when the automations are built on a platform they already understand and the documentation is good. The agency maintains it under an ongoing arrangement, which works when the scope and response times are written down. Or a mix: staff handle routine changes and the agency handles anything structural. Decide which before the build, because it changes how the automation should be built.

Red flags in the sales process

  • A demo built on someone else’s data, with no attempt to look at yours.
  • Savings promised before anyone has measured your current process.
  • Pressure to commit to a large build without a smaller first stage.
  • No mention of a human approval step anywhere in the proposal.
  • An unwillingness to name the AI services and where they process data.
  • Every process in your business described as a good candidate for automation.

The last one matters. Some processes are too rare, too variable or too sensitive to automate sensibly, and an agency that says so is more likely to be honest about the ones it does take on.

How Switchyard answers these

Deliverables built for you belong to you, accounts are set up in your organisation’s name wherever the platform allows, and the handover is documented. Each automation has a named human approval step for anything that sends, pays or changes a record, and is logged. AI services and data flows are listed per automation during the audit. Maintenance is described and priced before the pilot starts, as part of the scale stage. Ask us to show any of this on a call.