AI Agency Red Flags: 7 Signs to Catch on the Sales Call

The red flags worth catching are habits, not credentials. Every one of them shows up on the sales call. A vendor bills by the hour with no ceiling. It demos on its own tidy documents. It sells a chain of no-code tools as a system. It won't write down what "done" means. It can't say what the system costs to run. It sends its senior people home after you sign. And it says yes to everything on your list.

Each habit can be exposed with one question in about a minute. A vendor worth hiring answers all seven with something you can check. That might be a clause, a log, a number or a name.

Below are the seven signs. Each comes with what it costs you, the question to ask, and what a good answer sounds like. Then comes what to get in writing before you sign.

Agent washing, and McDonald's AI

Gartner counted. Thousands of vendors sell "agentic AI", which is AI that takes steps on its own. Gartner estimates only about 130 of them are real. The rest relabel chatbots, assistants and RPA scripts. RPA is software that clicks through screens by fixed rules. Gartner has a name for that: agent washing.

We have a shorter name for all of it. McDonald's AI.

It looks right on the outside. You open the bag in month three. (The burger in the photo and the burger in the bag have never met.) It is not a company, and we will name no one. It is a demo that runs on the vendor's tidy documents. Then comes a go-live that slips while the hours keep billing.

Why AI vendors are hard to judge

Most buyers can't judge AI work yet, and the vendor knows it. When you can't judge the food, you judge the picture. And the picture is great.

The regulators have seen the same gap. In September 2024 the FTC announced Operation AI Comply. The FTC is the U.S. agency that polices false and unfair business claims. This was a sweep against deceptive AI claims. One company in it advertised "the world's first robot lawyer". The FTC alleged it never tested whether its chatbot's output matched a human lawyer's. It also alleged the company never hired or retained any attorneys. The FTC chair at the time put it plainly. There is "no AI exemption from the laws on the books."

Logos, case studies and the word "agentic" cost a weak vendor exactly what they cost a strong one. So stop judging the picture. Judge the habits. A brochure can't answer a follow-up question.

Seven AI agency red flags, and the 60-second test for each

1. The meter is running

On the call. "We'll scope it as we go." "We bill monthly against the roadmap." No one says how many hours, because the hours are the product.

What it costs you. Hourly billing sets the vendor's margin against your timeline. And nothing in the contract settles it. Say the rate is $165 an hour and the estimate is 360 hours. That is $59,400. If the work runs 90 hours over, that is another $14,850. The vendor is paid for every one of them. (This is a made-up case. Check your last invoice.)

Ask. "If this takes longer than you estimated, who pays for the extra?"

Good sounds like. One fixed price, priced on the outcome, with a ship date in the contract. If the miss is on the vendor's end, the vendor pays.

2. The demo runs on their data

On the call. A slick demo. The documents are tidy. The questions are well formed. The answers are right every time.

What it costs you. You are buying a result that was never produced on your business. The first real file shows you what you bought. By then you have signed.

Ask. "Can you run it on one of my real files, while I watch?"

Good sounds like. "Yes. Send us your messiest one." A demo on sample data is fine, as long as it is called one. A click-through mock-up on dummy data shows the flow. It proves nothing about whether the answers are right.

3. A "system" made of Make and Airtable

On the call. "We use no-code, so it's flexible and you can change it yourself." No-code tools like Make and Airtable are set up by clicking, not by writing code. The diagram of the system is four logos and some arrows.

What it costs you. Most teams are tinkerers. They wire Make and Airtable together and call it a system. They ship it without a thought for token costs or the strain of growth. Or for the audit trail a buyer will one day demand. It holds while volume is small. Then you pay twice: once to build it, and once to rebuild it as purpose-built code.

Ask. "Show me the log for one run from last Tuesday. What happens when volume is ten times higher?"

Good sounds like. A named stack, with the reasons written down. A stack is the set of tools and code a system is built on. You want purpose-built code, on infrastructure made for live use. No-code has a place. It is fine for linking tools at low volume. It is fine for quick prototypes with people who don't write code.

4. Nobody will write down "done"

On the call. "We'll know it when we see it." "We iterate until you're happy." It sounds generous. It is a blank check.

What it costs you. With no end point, no one can be late and no one can be finished. The invoices keep coming. And the fight about whether it is done starts after you have paid.

Ask. "Write me the sentence that says this is finished."

Good sounds like. Win conditions in the scope, written as yes/no, that a stranger could check. Each phase has its own date and is accepted before the next one starts. Done is agreed by both sides. The vendor builds to it, at no extra cost, until it is finished.

5. Nobody mentions the token bill

On the call. You ask what it costs to run. "It's pennies." Or: "It depends on usage."

What it costs you. A token is a small chunk of text, and models bill by the token. Send every query to the priciest model on the list and you have the lazy model bill. Take a live system in our builds that handles 1,000+ AI interactions a day. Its model costs typically run $200 to $1,200 a month, all in. The top of that range is $14,400 a year. A bill ten times larger at the same volume is $144,000 a year. And no one has told you why.

Ask. "What will this cost to run each month at my volume, and whose account gets the invoice?"

Good sounds like. A range, with the levers named. One is tiered routing: each job goes to the cheapest model that can do it. It typically cuts token costs by 40 to 70%. Another is prompt caching. It stores the parts of a prompt that repeat. That can cut input-token cost by 60 to 90%. Then come fallback providers, which are backups if one model goes down. Last come budget alarms, and circuit breakers that stop a runaway loop. And the bills run on your own accounts, so no one marks up your tokens.

6. The senior people vanish after the sale

On the call. The person selling is sharp, fluent and technical. Three days after you sign, you meet "your delivery lead". This person has not heard of that conversation.

What it costs you. You paid for the people in the pitch. You got the people in the budget.

Ask. "Which engineers will build this? Can they join a shared channel with my team this week?"

Good sounds like. Names, and a channel. Your team talks to the engineers who build it, in a shared Slack channel. Every milestone is tracked where you can see it. The people are picked by what the build needs, not by whose calendar is open.

7. They say yes to everything

On the call. Voice agents? Yes. Computer vision? Yes. Your 2009 ERP? Yes. Yes. Yes.

What it costs you. A yes with no plan behind it is a bet, and you placed it. You become the training budget.

Ask. "Take the hardest thing on my list. How would you build it, and who builds it?"

Good sounds like. A plan for that one item, with names. "We custom build to your requirements, full stack, from the ground up. These engineers build it, and this specialist joins for the part that calls for one."

The yes worth paying for

Anyone can say yes on a call. The yes worth paying for has a plan and names behind it.

Ours is the same for every need a buyer brings. We custom build to your requirements, full stack, from the ground up. We bring our team, or the specialists the requirements call for. The scope says who does what before you sign.

Treat the names as the test. Say a yes can't tell you who will build the hard part. That is a yes you will pay for twice.

How to choose an AI development company

The seven questions sort the vendors. The contract decides what you really bought. Before you sign, get these in writing.

Then ask for one reference you can call, from a build the same weight as yours. Ask that client what happened the first time something went wrong. What the vendor did next tells you more than any case study. Who owns the code by default? And which clauses change it? See who owns the code an AI vendor writes.

When not to use this list

The list is built for buying production work. Don't use it on everything.

Say you are paying a lab to explore what is possible. It might be a research sprint, or a prototype to learn from. Then signs 1, 4 and 7 are features, not flags. Those are hourly billing, no fixed definition of done, and a yes to every idea. Cap the hours and keep the question open. Don't expect a system at the end of it.

Maybe you are buying a landing page or a basic CRM setup. Then you don't need an AI vendor, and you don't need this list. Hire a good freelancer.