AI & automation

How to choose an AI agency (and spot the dream-sellers)

Ten questions to ask an AI agency, the warning signs of a dream-seller, and what a fair proposal for a small business should include before you sign.

October 21, 20266 min read

A good AI agency talks about your enquiries, your response times and your numbers before it talks about AI. It proposes a small first project you can measure, keeps your accounts and data in your name, and tells you plainly what the system will not do. A dream-seller does the opposite: guaranteed revenue, a "proprietary platform" it won't name, and no clear answer on usage fees or how you leave.

Below are ten questions that separate the two, the red flags to watch for, and what a fair proposal looks like.

What an AI agency actually does for a small business

For most small and mid-sized businesses, an "AI agency" sets up existing tools. Nobody is building you a custom model, and nobody should be charging you as if they were. The real work falls into four parts:

finding where you lose customers or hours today;
choosing and configuring tools: a CRM, a website chat assistant, an AI receptionist, follow-up sequences;
writing what the AI relies on: your FAQ, rules for qualifying leads, the messages it sends;
testing, measuring and adjusting after launch.

The third part is where cheap projects fail. A chat assistant only knows what it was given. If an agency never asks for your service area, your pricing rules or how you answer common objections, the assistant will fill the gaps by guessing.

Your chatbot speaks for you

In early 2024, British Columbia's Civil Resolution Tribunal held Air Canada responsible for wrong information its website chatbot gave a customer (Moffatt v. Air Canada, 2024 BCCRT 149). The bot told a grieving passenger he could apply for a bereavement fare after travelling, which contradicted the airline's actual policy. The tribunal found negligent misrepresentation and declined to treat the chatbot as something separate from the company. As one McCarthy Tétrault lawyer summed it up, it makes no difference whether the information comes from a static page or a chatbot.

The takeaway for a business owner: whatever the agency's bot says on your site is your statement. Choosing an agency is partly choosing who controls quality.

Ten questions to ask

Which measurable problem do we fix first? A good answer names one task (answering after-hours enquiries, following up on quotes) and one number to watch (response time, share of quotes followed up).
Which tools will you use, and whose name are the accounts in? You want recognizable products and accounts in your business's name, or a written transfer clause at minimum.
Where is my customers' data stored, and is it used to train a model? A serious agency knows this for every tool it installs.
What will the assistant never do? Good answers: it never confirms a final price, never promises a date, and hands off to a person as soon as a question falls outside the approved FAQ.
How does the AI know what to say? Listen for "a knowledge base you approve," not "it learns on its own."
How will you test before launch? Ask how many test conversations, which difficult scenarios, and whether you get to see them.
What is the full first-year cost? Setup, subscriptions, usage charges for messages, call minutes and texts, and maintenance.
Who watches it after launch? You want a named person, a review schedule for conversations, and a turnaround time for fixes.
Can I speak with two current clients? Ideally your size, ideally your industry.
How do I leave if it doesn't work? Notice period, export of contacts and conversations, transfer of accounts.

Red flags

Guaranteed results

"We'll triple your sales" or "replace your receptionist next month." No one can guarantee a business result that depends on your offer, your market and your follow-through. In the United States, the Federal Trade Commission launched "Operation AI Comply" in September 2024, targeting companies that used AI hype to sell income promises and services they couldn't deliver. Then-chair Lina Khan put it bluntly: there is no AI exemption from the laws on the books. Canadian advertising rules don't carve out AI either.

A "proprietary system" with no name

If an agency won't say which platform it uses, it is often because you could buy it directly for less, or because leaving would mean losing your data. Methods can be confidential. The name of the software you depend on should not be.

Automating everything at once

A first project that touches the website, phones, email, social media and invoicing in one go can't be measured and is hard to fix. Sensible projects start with one or two automations and add from there.

Nobody asks about personal information

If no one asks what customer data the system will handle, no one has thought about it. Canada's federal, provincial and territorial privacy commissioners expect organizations using generative AI to limit the personal information they put into prompts, check that outputs are accurate enough for their purpose, and make sure people know when they are dealing with an AI tool. If you serve customers in Quebec, the province's private-sector privacy law adds a specific duty: when a decision about someone is based exclusively on automated processing, you must tell them, and explain the main factors on request.

A long contract before anything is proven

A 12- or 24-month commitment signed before the first workflow has been tested puts all the risk on you.

Silence on usage fees

Many platforms bill call minutes, text messages and AI credits on top of the subscription. A proposal that doesn't estimate them is a surprise waiting on your first invoice.

The flawless demo you can't touch

Prepared demos always go well. Ask to type your own hard questions: a request outside your service area, a pricing question, an angry complaint. Watch what it does when it doesn't know.

What a fair proposal includes

A fair proposal for a small business fits in a few pages and covers:

Diagnosis: what the agency understood about your enquiries, response times and losses, with numbers where possible.
First-project scope: one or two automations, described step by step.
Tools by name: who pays for what, and whose name each account is in.
Content: the FAQ, rules and messages to be written, by whom, and who approves them.
Guardrails: what the AI won't do and when it hands off to a person.
Test plan and acceptance criteria.
Metrics tracked and the date of the first review.
Full first-year cost, including estimated usage fees.
Data: where it is stored, how long it is kept, which subcontractors see it.
Exit terms: notice, export, account transfer.

If any of these is missing, ask for it in writing before you sign.

Agency, freelancer or do it yourself

Do it yourself if the need is simple (an AI assistant for your team, an automatic reply to web enquiries) and someone on staff is patient with software.
Hire a freelancer for a single, well-defined task when you already know what you want.
Hire an agency when the project spans several systems at once (website, CRM, phones, ads) and you need one party to own the whole thing and keep watching it.

The ten questions apply in all three cases. The same instincts help when you hire for marketing in general; our list of marketing agency red flags covers the ad and SEO side.

Where PNE fits

PNE's AI Growth Systems work follows this order: audit your lead flow, set up the CRM, train a chat assistant on your services and rules, write the follow-up sequences, test end to end, then spend 30 days adjusting. You approve every message before it goes live. We don't promise revenue figures; we commit to deliverables and clear reporting.