To choose an AI agent and automation agency, judge how it works rather than how it demos. A good one starts by understanding your process, connects to the tools you already use, tells you plainly who owns what it builds, and agrees up front how you will measure success. If a vendor jumps straight to a polished chatbot demo without asking how your business runs, that is the first warning sign. The questions below are meant to be asked in the first call or two, before you sign anything.
What should a good diagnosis look like before anything is built?
A good diagnosis is a structured look at your current process, and the agency should ask for it before proposing a solution. Expect questions about where requests come in, who handles them, how long things take, what goes wrong, and which tasks are repetitive enough to automate safely. An agency that quotes a solution in the first conversation is guessing, and you will pay for that guess later in rework. Diagnosis can be short for a simple project, but it should exist and you should see its output in writing.
The output matters as much as the process. Ask what you will receive at the end: a list of candidate processes ranked by effort and impact, the data and access needed, the risks, and a recommendation on what to do first and what to leave alone. If you are not sure yourself where to begin, our guide on what to automate first in a small business gives you a way to form your own opinion, which makes it easier to tell whether the agency's recommendation makes sense. A vendor who tells you to leave something manual for now is usually more trustworthy than one who says everything should be automated.
Which tools and integrations should you check?
Ask the agency to name, specifically, how the solution will connect to the tools your team already uses. A chatbot or agent that lives in isolation creates a second inbox to monitor, and leads or orders end up copied by hand. The useful question is not "do you integrate?" but "how, with which of my systems, and what happens when a sync fails?". For example, if you run HubSpot, Shopify, Slack, Notion or Zapier, the agency should be able to explain what data moves in each direction and who gets notified when something breaks.
Check a few concrete points in the proposal:
- Which of your current systems are read from, and which are written to.
- What channels are covered (web, WhatsApp, Instagram) and what the platform's own rules require, such as WhatsApp's business messaging policies.
- Where conversations and customer data are stored, and who has access.
- What the fallback is when the agent cannot answer: a handoff to a person, a ticket, a notification.
If a CRM is part of the picture, read how connecting an AI chatbot to your CRM without losing leads works, so you can judge whether the proposed setup covers duplicates, field mapping and failed syncs. Those details are where many projects quietly leak.
Who owns the work once it is built?
You should know, in writing, who owns the prompts, workflows, configurations, knowledge base content and customer data. This is the question most buyers skip and later regret. Some agencies build everything inside their own accounts, so if you part ways you lose the system. Others build inside your accounts, with you as the administrator, which is usually the safer arrangement. Neither is automatically wrong, but you need to know which one you are buying and what leaving would involve.
Ask for specifics: will the accounts for your CRM, messaging platform and automation tools be in your name? Will you receive documentation of the flows and instructions the agent follows? Can you export your data and conversation history? Also ask about third-party model providers such as OpenAI or Anthropic, because their terms and data handling affect your own privacy obligations. Their official documentation is the right place to verify what happens to data sent through their APIs. For a wider view of the risks, security and ethics in autonomous agent deployment lists what to be careful about.
How will success be measured?
Agree on a small number of measurable outcomes before the build starts, tied to the problem you are solving. If the goal is faster replies, measure first response time and the share of conversations resolved without a person. If it is lead handling, measure how many leads are contacted and how many reach a booking or a sales call. The baseline matters: without today's numbers, any later improvement is just an impression. An agency that cannot say how it will know the project worked probably does not plan to find out.
Also ask how quality is checked, not only volume. An agent that answers quickly but wrongly is worse than a slow one. A serious team reviews real conversations, tracks failures and adjusts, which is the kind of approach described in evaluating AI agents beyond simple accuracy. For the business side, the ROI of AI agent automation explains how to connect those metrics to time saved or revenue, without pretending the numbers are guaranteed.
How much autonomy and human oversight is planned?
A reliable agency tells you where the AI acts alone and where a person approves or takes over. Sending an order status is low risk; issuing a refund, changing a booking or giving advice that could create liability is not. Ask which actions the agent can take without approval, how it recognises that it is out of its depth, and how a customer reaches a human. The honest answer is usually "start with narrow permissions and widen them as we see how it behaves".
This is a design choice, not a technical detail, and it should be explained in plain language. The article on autonomous versus semi-autonomous agents helps you think through which level fits each task. Be wary of anyone who promises a fully autonomous system on day one for a customer-facing process. Gradual expansion is slower to sell but much safer to run.
What does a sensible engagement look like, and which red flags matter?
A sensible engagement runs in stages: diagnosis, design, implementation, then measuring and scaling what works. Starting with one well-defined process, running it live with oversight, and expanding only after results are visible protects you from paying for a large build that does not fit. Ask what the first stage delivers, what you can see before the next one begins, and whether you can stop after it. Pricing and timelines should be agreed per project, in writing, and tied to a defined scope.
| Question | A good answer | A red flag |
|---|---|---|
| How do you start? | Diagnosis of our process, with a written output | A demo and a quote in the first call |
| What do you connect to? | Named systems, data flow and failure handling | "We integrate with everything" |
| Who owns it? | Accounts and flows in our name, documented | Vague or built only in their accounts |
| How do we measure? | Baseline, a few metrics, regular review | "You will see the results" |
| What if it fails? | Human handoff and a monitoring routine | Guarantees that it never will |
Other red flags are guaranteed results, pressure to sign quickly, no mention of maintenance, and no one able to explain what the agent actually does in simple terms. Maintenance is real: policies change, products change, and the agent needs updated information. Ask who handles that and how. If you want a sense of how the knowledge side works, see how to train an AI chatbot on your own business data.
What should you prepare before contacting an agency?
Prepare a short brief: the process you want to improve, how it works today, the tools involved, volume per week, and what a good result would look like. You do not need technical language. Even a page of notes and a few real examples of conversations or tasks lets the agency give a more useful first response and lets you compare vendors on equal terms. Contact two or three agencies with the same brief and compare how they respond, not only the price.
If you want to see how this kind of conversation goes, you can reach Kuan AI through the contact form; the team replies within 24 hours. Kuan AI designs and builds AI agents, chatbots, automations, websites and AI-assisted social media, and works through diagnosis, design, implementation, then measurement and scaling. Scope, price and timing are agreed per project, so use the checklist above on us as well.
Questions about hiring an AI agency
How do I know if an AI agency is credible?
It asks about your process before proposing anything, names the systems it will connect to, explains ownership and measurement in writing, and admits what should stay manual. Vague promises and demo-only sales are warning signs.
Should I hire an agency or a freelancer?
It depends on scope and ongoing needs. A freelancer can fit a single, simple automation. An agency tends to make sense when several channels, integrations and maintenance are involved. Either way, ask the same questions about diagnosis, ownership and measurement.
Who should own the accounts and workflows?
Ideally you, with the agency working as an administrator inside your accounts, and with documentation of the flows. If they build in their own accounts, ask what happens to your data and system if you stop working together.
How much does an AI agent project cost?
It varies a lot with scope, channels and integrations, so be careful with any fixed number quoted before a diagnosis. Ask for a written proposal tied to a defined scope. Kuan AI agrees pricing and timing per project.
What should a first project be?
One well-defined process with clear volume and a measurable result, such as answering repeated customer questions or following up leads. Run it live with human oversight, then expand once results are visible.
How do I compare two proposals?
Give both agencies the same brief, then compare the diagnosis approach, named integrations, ownership terms, metrics, handoff to humans and maintenance. Price matters less than whether the scope is clear.
Choosing well comes down to a few habits: ask for the diagnosis, name the integrations, get ownership in writing, agree on the numbers and start small. None of that requires technical expertise, only the willingness to ask direct questions and to notice when answers stay vague. An agency worth hiring will welcome these questions, because they make the project easier to deliver.
