“Custom AI agent development” is the most oversold phrase in the market, because “custom” can mean anything from a system-prompt edit on a shared platform to a system engineered against your actual queue. Before signing anything, pin down which parts of the proposal are genuinely built for you, and, just as importantly, which parts sensibly are not.
What should be custom
- The process design. The agent’s job description comes from mapping your workflow, not from a template gallery. If nobody watched your operators work, the agent is custom to somebody else’s business.
- The integrations. Real connections into your claims platform, your transport management system, your health records boundary, or your ticket system, with your permissioning. “Integrates with 400 tools” means generic connectors; ask which of your systems has been wired before.
- The retrieval library. The documents the agent looks answers up in: your policies, your records, your history, with citations. This is what separates an agent from an interface to someone else’s model.
- The eval suite. Test cases drawn from your history: real claims, real invoices, real edge cases. A vendor benchmark tells you nothing about your Tuesday queue.
- The ownership. Code, prompts, evals, runbooks: with handover documentation. If “custom” ends in a per-seat licence for your own workflow, it was a product sale wearing a services hat.
What should not be custom
Be equally suspicious of too much invention. Nobody should be hand-rolling their own foundation model (the underlying AI), vector database (the search index behind retrieval), or orchestration framework (the plumbing that sequences agent steps) for a mid-market engagement; proven components exist and your budget should go to the five items above. A team that builds everything from scratch is running a research project on your invoice.
Reading a proposal in five minutes
Count the pages about your workflow versus the pages about their platform. Find the eval section; if there is none, stop. Find the handover section; if it describes a subscription, stop. Then find the number that says what your operators stop doing, and in what week.
Svolta handles this through the Capacity Workflow. A free Consultation checks whether one workflow is worth measuring. Workflow Baseline is $2,995 AUD ex GST per workflow and documents the current state, evidence method, systems, exceptions, and fixed Launch scope. Workflow Launch is $14,995 AUD ex GST per workflow, including all scoping expenses, and puts the approved workflow into production. Workflow Run is $995/month AUD ex GST per workflow. If a configured off-the-shelf tool genuinely covers the case, that decision belongs in Workflow Baseline before build money is committed.