How to Choose an AI Agency: 12 Questions to Ask Before You Sign

Short answer

To choose an AI agency without getting burned, require three things before you sign: a demo on your own customer questions, a written answer on where your data goes and who owns the code, prompts and knowledge base, and a price that covers both setup and monthly operation. A serious agency answers these without hedging; one that promises numbers before seeing your data is telling you something else. General development contract terms (IP assignment, fixed price per phase, exit clause) are covered in our guide to choosing a software development company; this one covers what is specific to AI.

The short answer: what sets a good AI agency apart

An AI agency rarely sells you a language model. It sells what it builds around one: the knowledge base, the instructions, the integrations with your tools, the testing and the ongoing review. That, not the name of the model, is what determines the result, and what you should judge it on.

It agrees to a demo on your real questions, not only on a prepared script.
It states in writing which model it uses, where your data goes and whether it is used for training.
It leaves you owning the code, the prompts and the knowledge base, in accounts under your name.
It publishes or writes down a price that separates setup, the monthly plan and what makes the bill move.
It explains what the agent does when it does not know, and which actions it cannot take.
It proposes starting with one measurable task rather than a full “AI transformation”.

Why AI needs extra questions

The market moves fast and so do the labels. In a June 25, 2025 press release, Gartner describes “agent washing”, the rebranding of existing assistants, automation tools and chatbots as agents without substantial agentic capabilities, and estimates that only about 130 of the thousands of vendors claiming agentic AI are real. The same release predicts that over 40% of agentic AI projects will be canceled by the end of 2027 because of escalating costs, unclear business value or inadequate risk controls.

Adoption keeps growing all the same: according to Statistics Canada, 19.2% of Canadian businesses reported using AI to produce goods or deliver services in the 12 months before the second-quarter 2026 survey. Many businesses are buying their first AI project right now, often with no checklist for vetting vendors. The 12 questions below are that checklist.

The 12 questions to ask before you sign

Send them in writing to every agency and compare the answers. How precise the answers are matters as much as what they say.

Question	Reassuring answer	Red flag
1. Can you demo it on our real customer questions?	Yes, using a sample of your emails or calls, tricky questions included	Only a canned demo, built for another industry
2. Which language model do you use, and can you switch?	The provider is named, and switching models is planned without a rebuild	“Our proprietary AI”, and nothing more
3. Where does our data go, and through which subprocessors?	A written list of providers and processing locations, before signing	A vague answer or “in the cloud”
4. Are our conversations used to train a model?	No, and it is in the contract or the terms of the providers used	No clear answer
5. Who owns the code, the prompts and the knowledge base?	You do, in a repository and accounts under your name, with a written assignment	Everything stays inside the agency’s platform
6. What does the agent do when it does not know?	It says so and hands off to a person with the history	“It always finds an answer”
7. Which actions can the agent take, with which permissions?	A precise list of actions, limited access, confirmation before important actions	Admin access to your systems
8. How do you test before launch and after every change?	A reference set of questions, re-run on every change	Ad hoc manual checks
9. What does it cost to run, and what makes the bill go up?	A written monthly plan, defined included usage, overages agreed in advance	Usage billing with no cap or estimate
10. Who reviews conversations and corrects the agent, and how often?	A named person, a written frequency, documented fixes	Nobody after launch
11. Does the agent disclose it is AI, and does it handle the languages we need?	Yes, from the first message, including the human hand-off in each language	“Customers won’t notice the difference”
12. What happens if we stop?	You leave with the code, prompts, knowledge base and logs	Penalties, or content withheld

The points that deserve a written answer

Data. Know where your data goes before anyone touches it. If you serve customers in Québec, Law 25 requires a privacy impact assessment for any project to acquire or develop an information system involving personal information, and a written agreement before personal information is communicated outside Québec. Elsewhere in Canada and in the US your obligations differ, but the question is the same: an agency that cannot list its subprocessors leaves you unable to assess your own risk.

Ownership. Under Canada’s Copyright Act, the author of a work is the first owner of its copyright (section 13). An agency is not your employee: without a written assignment, the code and text it produces for you do not automatically belong to you. With AI, add the system prompt, the prepared knowledge base and the test sets to the list: that is often where the value of the work sits.

Actions. OWASP lists “excessive agency” among the top ten risks for LLM applications (LLM06:2025): excessive functionality, excessive permissions or excessive autonomy for high-impact actions. Ask for the list of actions and access rights, and require confirmation before anything that commits the business.

Disclosure. In the European Union, Article 50 of the AI Act has required since August 2, 2026 that people be told they are interacting with AI unless it is obvious. In California, Business and Professions Code section 17941 makes it unlawful to use a bot to mislead someone about its artificial identity in order to incentivize a sale. An agent that introduces itself as an AI assistant avoids both problems.

General information, not legal advice. For an agent that handles health, financial or legal information, have the project reviewed by a lawyer in your jurisdiction.

Red flags

Guaranteed numbers (“double your sales”) before the agency has seen your data.
A refusal to name the language model or the providers used.
A “custom-trained model” pitched to a small business, with no explanation of what that means, what it costs and what it adds over a well-built knowledge base.
No price, no range and no monthly cost estimate before signing.
A demo you cannot reproduce with your own questions.
An agent with no hand-off to a person, or whose conversations you cannot review.
Code, accounts and content hosted only on the agency’s side, with no way to export them.
Testimonials or clients you cannot verify.

Agency, freelancer or subscription tool: when to choose which

An agency is not always the right answer. Choose based on what the tool has to do, not on the label.

Option	The right choice when…	Limit to plan for
Subscription tool (off-the-shelf chatbot)	Your questions are standard, your website or help desk platform already offers it, and the tool does not need to act in your systems	Few custom integrations; the bill follows volume
Freelancer	The task is technical and well bounded, and someone on your side can manage the project and take over later	One person for design, testing and follow-up; plan for continuity
Agency	The agent has to connect to several tools (website, CRM, calendar, email, text) and be reviewed every month	Higher setup cost; insist on ownership and an exit path
In-house team	AI is at the core of your product and you have the volume to justify hiring	Hiring time and fixed cost

If a subscription tool solves your problem, it is often the cheapest option. A good agency will tell you so.

Comparing proposals

Two AI agency proposals can only be compared if they describe the same agent. Bring each one back to the same grid before looking at the total.

The target task and success metric (appointments booked, requests resolved, qualified leads), with a measured baseline.
Knowledge base preparation: who gathers, cleans and updates your documents.
The list of integrations (website, CRM, calendar, email, text) and permitted actions.
Supported languages, human hand-off included.
Testing: how many reference questions, and re-runs on every change.
The monthly plan: what it covers, included usage, what happens with exceptional volume.
Ownership of the code, prompts and knowledge base, and the exit terms.
The hourly rate for out-of-scope work.

Start small: a first engagement in five steps

Pick one task that is expensive today: missed evening calls, quotes with no follow-up, repetitive questions.
Measure the baseline for two weeks (number of requests, response time, appointments booked).
Prepare the knowledge base and the list of permitted actions, then test on your real questions.
Launch on a single channel, review conversations every week and correct.
Compare with the baseline, then decide whether to extend to other tasks.

How ZeniTech answers these questions

ZeniTech is a web, app, automation and AI agency based in Québec City, Canada, founded by Alexandre Blais, working remotely with businesses across Canada and the United States. Our answers to this guide’s questions are the ones already published on our site: prices are public; the code is yours from the first commit, in a repository created in your account, with the domain, hosting and accounts in your name too; custom projects are split into fixed-price phases, and you can stop after any of them.

Our 12 industry AI agents are powered by Orvel AI, our agent layer built on top of several leading language models, so there is no dependency on a single provider. Your data is never used to train anyone’s model, and we will state in writing where your data goes before you sign. The agent answers from your data and rules in English or French, and when it does not know, it says so and hands off to a human with the full conversation. The monthly plan includes reasonable usage; exceptional volume is agreed in advance.

Service	Price (CAD, before taxes)	Approx. USD
AI agent setup (powered by Orvel AI)	CAD 3,000 to 7,500	≈ USD 2,100 to 5,300
AI agent monthly plan	CAD 750 to 1,500/month	≈ USD 530 to 1,050/month
Automation of a simple / complex workflow	CAD 750 / 2,500, monitoring CAD 250/month	≈ USD 530 / 1,760, monitoring ≈ USD 180/month
CRM setup	CAD 3,500, then CAD 45/user/month (3-user minimum)	≈ USD 2,460, then ≈ USD 32/user/month
Custom software / app, first version	From CAD 15,000 / 25,000	≈ from USD 10,500 / 17,600
Work outside the agreed scope	CAD 125/hour	≈ USD 88/hour

Prices as published on zenitech.dev/en/pricing, in Canadian dollars before taxes. USD equivalents at the Bank of Canada daily rate of October 1, 2026 (USD 1 = CAD 1.4243), rounded; our published prices are in CAD. Free 30-minute consultation, no commitment.

Frequently asked questions

How do I choose an AI agency?

Ask for a demo on your real customer questions, a written answer on which model is used, where your data goes and who owns the code, prompts and knowledge base, and a price that separates setup from the monthly plan. Be wary of guaranteed numbers before the agency has seen your data.

What questions should I ask an AI agency before signing?

The most revealing ones: can you demo it on our questions, which model do you use and can you switch, where does our data go, is it used for training, who owns what you build, what does the agent do when it does not know, which actions can it take, how do you test, and what happens if we stop?

How much does an AI agency charge?

It depends on integrations and ongoing review. At ZeniTech, whose prices are published, AI agent setup costs CAD 3,000 to 7,500 (about USD 2,100 to 5,300), then CAD 750 to 1,500 a month (about USD 530 to 1,050), before taxes. Always get the monthly cost in writing, not just the setup fee.

Should I hire an AI agency or a freelancer?

A freelancer suits a well-bounded technical task, if someone on your side can manage the project and take over later. An agency suits an agent that must connect to several tools and be reviewed every month. If a subscription tool already solves your problem, it is often the cheapest of the three.

Who owns an AI agent built by an agency?

Not automatically you. Under Canada’s Copyright Act, the author is the first owner of copyright, and an agency is not your employee, so you need a written assignment. Make sure it covers the code, the prompts, the knowledge base and the test sets, and that the repository and accounts are in your name from the start.

Is my data safe with an AI agency?

It is if you know where it goes. Ask the agency for a written list of its subprocessors and their processing locations, and whether any of them train on your data. If you serve customers in Québec, Law 25 requires a privacy impact assessment for a new system handling personal information and a written agreement before sending it outside Québec.

What is agent washing?

It is the rebranding of an existing assistant, automation tool or chatbot as an “AI agent” without real agentic capabilities. In a June 2025 press release, Gartner estimated that only about 130 of the thousands of vendors claiming agentic AI are real. Ask which exact action the tool performs, and in which system.

Can a Canadian AI agency work with a US business?

Yes, remotely. Ask the same 12 questions, plus three practical ones: the currency of the quote, overlapping working hours, and the governing law of the contract. ZeniTech works remotely from Québec City and publishes its prices in Canadian dollars; the USD figures on this page are approximate conversions.

Free consultation

Get a fixed price for your project in 30 minutes.

We look at what the project has to achieve, tell you what it costs, and tell you when you do not need it.

Book a call →+1 581-748-7017

Sources

Market ranges compiled on October 2, 2026. They are not a ZeniTech quote.

How to choose an AI agency: 12 questions to ask before you sign