8 Questions to Ask an AI Agent Developer Before You Hire One

What exact job will the agent perform?

Do not accept an agent that handles everything. Name the trigger, the input, the output, and the success event. After-hours intake, a verified appointment, and a CRM task are different deliverables. If those are mixed into one sentence, the scope is not ready.

Do not accept “an AI agent that handles everything” as a specification. Name the trigger, input, output and success event. An after-hours intake request, a verified appointment and a CRM task are different deliverables.

Are you configuring software or building a custom system?

Ask which parts are vendor-hosted, which are custom, and which subscriptions stay on. A low monthly rate is not the cost of implementation. You should know what still works if you stop paying the developer, including the phone number and the prompts.

Ask which components are vendor-hosted, which are custom and what subscriptions are necessary. A low subscription rate is not the cost of implementation. You should know what continues working if you stop paying the developer.

Which integrations are confirmed to work?

Name the phone, calendar, CRM, and messaging tools, then check the permissions before you sign. Works with any CRM is not a test. A preview-only project should say in writing that no integration is live. Do not treat a demo as a connection.

List the exact phone, calendar, CRM and messaging provider; verify permissions and APIs before signing. “Works with any CRM” is not a test result. For a preview-only project, the brief should explicitly state that no integration is live.

What can the agent say, and what is prohibited?

Approve one source of truth and a list of forbidden actions. Medical, legal, and hazardous situations need a qualified person. A receptionist should not guess availability or invent a quote. If the rule is not written, the agent should hand the call over.

Approve a source of truth and specific prohibited actions. Sensitive, medical, legal or hazardous situations need qualified human handling. A receptionist should not guess availability or invent a service quote.

How will human escalation and outages work?

Ask for demos of unavailable staff, a dropped connection, an unknown question, and a frustrated caller. The fallback should record the request without pretending a transfer or a booking worked. Human transfer and fallback are part of the product, not a failure.

Request demos of unavailable staff, interrupted connections, unknown customer questions and frustrated callers. A safe fallback should still record a legitimate request without pretending a transfer or booking succeeded.

Who owns the data, prompts, and accounts?

Put the domain, phone number, recordings, transcripts, prompts, access logs, retention, and what happens when the contract ends in writing before you sign. Recording and promotional texts depend on consent and local rules. An inbound call is not permission to market.

Confirm domain, phone number, caller information, recordings, transcripts, configuration, access logs, retention and termination rights in writing. Recording and promotional messaging requirements depend on jurisdiction, consent and platform policy.

Which acceptance tests must pass before launch?

Test a routine request, an out-of-scope question, a calendar outage, emergency language, a caller who wants a person, and a temptation to invent a fee. Launch only after the owner signs the results. A smooth demo of the happy path is not that test.

1. Routine request produces accurate approved answers and authorized records.
2. Out-of-scope inquiry triggers human handling.
3. Calendar outage never creates a false confirmation.
4. Emergency language triggers approved escalation guidance.
5. A caller can request a human and reach a real fallback.
6. No unauthorized fee, discount, diagnosis or policy promise is fabricated.

What do setup, usage, and support actually cost?

Ask for separate lines for design, configuration, model usage, phone minutes, texts, CRM connectors, and maintenance. Vendor ads are examples, not a build price. Mandile SEO quotes AI agents as a custom scope. Phone and usage stay separate from that quote.

Demand separate line items for design, configuration, API/model tokens, telephone minutes, SMS, CRM connectors and continuing maintenance. Ask what happens when vendor rates, APIs or business policies change. See our AI agent pricing guide for independently published vendor examples, not universal build-cost averages.

How do you compare two AI agent developers?

Use one defined use case, evidence that the integrations work, a tested human escalation, a quoted list of recurring charges, written ownership, and signed launch criteria. If a cell says we will boost it, ask for the page, the system, and the test. Price alone is not the comparison.

What should be in the statement of work?

Include scope, exclusions, dependencies, test data, review meetings, acceptance scenarios, who may deploy, incident response, and ongoing costs before you pay. A preview export is valid when it is labeled non-live. It is not consent to record calls or contact customers.

Include a scope and exclusions list, dependencies, mock/test data plan, review meetings, acceptance scenarios, deployment permissions, incident response, documentation and ongoing costs. A preview/export deliverable is valid—but it must be labeled non-live.

People also ask

Related Mandile SEO resources

• AI agent pricing and hidden costs
• Downloadable implementation brief template
• Voice-agent prelaunch tests