How to hire an AI automation agency without buying a demo

"should i hire an ai engineer or work with an ai development agency?"

That question, typed exactly like that, appears in the Search Console data for our own site. For an owner comparing proposals, a few sales calls can make it harder to answer: each vendor may be describing a different scope.

Consider the owner of a 15-person plumbing company who has watched three agency demos in two weeks. This is a composite, not a client. In each demo an assistant answered a question, booked an appointment and sent a friendly follow-up. All three looked finished. None of them showed what happens when a customer gives half an address, asks for a person, or books a slot another technician already took. The owner can't tell which vendor built a working system and which one built a presentation.

What a demo proves

A demo proves that a path exists. It doesn't show how the system behaves on an ordinary day, when requests arrive incomplete and the office is busy.

The Federal Trade Commission has dealt with the extreme version of this. In February 2025 it finalized an order against DoNotPay, which had marketed an "AI lawyer". The FTC alleged the company never tested whether the product worked at the level of a human lawyer. The order bars that claim without evidence to back it. A buyer can ask the same kind of question in a sales call: what exactly does the AI do here, how was that tested, and which parts are ordinary software?

Hire an engineer or work with an agency?

Compare the work, not the job title.

An internal hire makes sense when the system will change every week and someone needs to own those decisions for years. You are paying for continuity and knowledge of the business. Budget for more than salary: recruiting, management time, and the design, security and infrastructure work one person may not cover.

An outside team makes sense when the job is a defined system with a beginning and an end. Connect the website form to the CRM, qualify the request, book the appointment, stop the follow-up when the customer confirms. You are paying for several skills over a fixed scope. You still need someone inside the business who can say how the work actually runs.

A hybrid works when you already have that person and a small technical team, and you need specialist integration work for one project.

If nobody owns the process today, neither option fixes it. In that case, mapping the process is the first job, and it doesn't need AI.

Write one workflow and send it to every vendor

The quickest way through a stack of proposals is to stop comparing platforms and compare one workflow. Write it in four lines:

• Input: a new inquiry from the website form or a missed call.
• Output: a CRM record with the service, location and preferred time filled in, assigned to a named person.
• Owner: the person who reviews exceptions and can pause the automation.
• Acceptance check: what you will watch happen before you pay the final invoice.

Send the same four lines to every vendor and ask each one to quote against it. The proposals become comparable. If a vendor only wants to talk about a platform subscription, that tells you something too.

Ask to see the exceptions

The polished demo shows a clean request and a calendar with open slots. Ask for a second demonstration using anonymized examples from your own inbox:

• a customer who gives a neighborhood instead of an address
• the same request sent twice through two channels
• a calendar with no open slots that week
• a customer who asks for a person twice
• a CRM update that fails

For each one, ask what the customer sees, what your staff see, and where the request sits until someone acts on it. A useful answer shows a record a person can find and fix. If the vendor has to open application logs to answer, your front desk won't be able to either.

Settle ownership before the build starts

An automation project leaves behind more than code. There are accounts, workflow configuration, prompts, API credentials, message templates and customer records. The agreement should say who owns each one and how you get it back if you change providers.

Don't rely on the words "work for hire" to cover custom software. The US Copyright Office's Circular 30 explains that a commissioned work counts as a work made for hire only if it falls into one of nine categories listed in the Copyright Act and both parties sign a written agreement saying so. Software isn't one of the nine. Ask your attorney whether your contract needs a written assignment of rights instead. This is a question for a lawyer, not a vendor.

How we start

Our AI systems and automation work begins with that single workflow. Before anything is built, we go through recent inquiries with the person who handles them and record where each one arrived, what information was missing and who it waited on. The proposal names the integrations, the business rules and the checks that define done. The demonstration uses your difficult cases rather than ours. Handover covers account access, alerts and a pause control your team can use without contacting us.

For inquiry, booking and follow-up work specifically, our AI revenue systems service explains how the first phase is scoped.

What to expect after the first workflow

You should be able to pick any inquiry from last week and see where it came from, who owned it and what happened next. Your team should be able to correct a record and switch the automation off without raising a support ticket.

Whether sales change is a separate question. It needs a baseline taken before launch, such as response time, unassigned requests and booked jobs, measured the same way afterwards.

If you are deciding between an engineer and an agency, bring one workflow and a few anonymized inquiries to a call. Book a call and we'll say which model we think fits, including when that isn't us.

Sources

• Federal Trade Commission, FTC Finalizes Order with DoNotPay That Prohibits Deceptive 'AI Lawyer' Claims, February 11, 2025.
• US Copyright Office, Circular 30: Works Made for Hire, revised August 2024.