Most AI projects do not fail because the technology is weak. They fail because nobody agreed what the agent should do, how anyone would know it worked, or what happens on the day it gets something wrong.
These are the questions to ask any agency, consultant or platform before you pay them. Each one comes with why it matters and what a good answer sounds like. A good partner answers all fifteen without hesitating. If you get vague answers to more than two or three, keep looking.
Why it matters: "Automate customer service" is a wish, not a project. Without one named task and a number to beat, you cannot tell success from a demo.
A good answer: "We'll start with one workflow, for example answering booking questions on your website. Today your team handles about 300 a month and takes four hours to reply. The target is 80 percent answered without a person, in under two minutes."
Why it matters: The most expensive mistakes come from an agent doing something nobody expected it to do, such as issuing a refund, promising a date or changing a record.
A good answer: A written list of actions that always need a person's approval, set before anything is built.
Why it matters: An agent that looks great on ten hand-picked examples can fall apart on your actual customers, documents and edge cases. This is the question most vendors struggle to answer.
A good answer: "We'll take a set of your real past cases, agree a pass mark with you, and the agent has to reach it before it touches a live customer."
Why it matters: A new platform means migration, retraining and another login nobody opens after month two.
A good answer: "It works through your existing CRM, inbox, calendar and internal tools, within the permissions you set. Nothing is migrated."
Why it matters: An agent is only as good as what it knows. If your prices, policies or services change and the agent does not, it will confidently say the wrong thing.
A good answer: One place that holds your services, prices, policies and tone, that every agent reads from, and that your team can update without a developer.
Why it matters: Every agent eventually meets a question it cannot answer. Guessing is what damages trust.
A good answer: "It hands the conversation to a person, with the full context, so your customer never has to repeat themselves."
Why it matters: When something goes wrong, "the AI did it" is not an explanation your customer or your team will accept.
A good answer: A log of every conversation and action, readable by your team, not only by engineers.
Why it matters: Switching everything over on day one is how small problems become public ones.
A good answer: "First it runs alongside your team without acting, so you can compare its answers to theirs. Then it takes a small share of the real work. It only takes full volume once the numbers hold."
Why it matters: Your customer conversations and documents are business data. You should know exactly where they go.
A good answer: A clear yes or no in writing, plus where data is stored and who can access it.
Why it matters: Per-conversation and per-task pricing can look cheap in a pilot and become your biggest software bill once it works.
A good answer: A price you can predict at two and three times today's volume, in writing.
Why it matters: Models change, your business changes and customers find new questions. An agent nobody maintains gets worse every month.
A good answer: A named person or team, a monthly review against the agreed target, and the cost of that work stated up front.
Why it matters: The model underneath can change without notice. You should not be locked to one.
A good answer: "We can move the agent to another model without rebuilding it, and we test it against your cases again when we do."
Why it matters: Starting from agents that already work in production is faster and carries less risk than writing everything new for you.
A good answer: "We start from agents we already run in production and fit them to your workflow. We only build new where your process really is different."
Why it matters: Demos are rehearsed. Customers are not.
A good answer: A real reference, or a short paid pilot with a clear target if they are too new to have one.
Why it matters: You should not be held hostage by your own conversation history, knowledge and setup.
A good answer: "Your data, your knowledge base and the record of what the agent did are yours, and you can export them."
We wrote this guide because these are the questions we ask ourselves before we build anything.
- We start with one workflow and a number. Every project begins with an AI Opportunity Audit: we map how the work runs today, measure where time and money go, and agree the first build and its target.
- Agents run inside the software you already use, your CRM, inbox, calendar and internal tools. Nothing is migrated.
- Every agent is tested on your real past cases against a pass mark we agree with you, before it goes live.
- Go-live happens in stages: shadow mode first, then a small share of the work, then full volume once the targets hold.
- People stay in control. Agents get only the access they need, risky steps need approval, and every action is logged.
- Most builds start from agents already in production for customer service, sales, marketing and operations, so you are not paying for a project from scratch.
If you would like us to answer these fifteen for your business, book a 30-minute call. We will map one workflow with you and tell you honestly whether an agent is worth building for it.