“We want an AI agent” is something we hear a lot on first calls. Sometimes that really is the right answer. More often, once we’ve talked through the work, the business needs something simpler: an automation with one AI step in it, or an assistant that answers questions from the company’s own documents.
The three are different tools, with different costs and different ways of going wrong. Picking the right one is the difference between something that runs quietly in production and a great demo that never ships.
Automations with an AI step
An automation runs the same steps every time something happens: an email arrives, a form is submitted, a file lands in a folder. The AI step does one narrow job inside it, like reading an invoice, sorting a support ticket into the right queue or drafting a reply.
Our invoice automation is a good example. Every invoice email is picked up, AI pulls out the details and assigns a category, and the file is filed and recorded. It’s predictable, cheap to run and easy to test, which is why we reach for this pattern first.
Assistants that answer questions
An assistant (or chatbot, if you prefer) answers questions in a conversation. The useful ones answer from your own content, such as help articles, policies, product data or past tickets, using retrieval, usually called RAG, rather than from whatever the model happens to know.
A typical case is a support assistant that handles order and returns questions and passes the conversation to a person when it isn’t sure. It saves a lot of time without ever taking an action on its own. We cover how we build these on our RAG and vector search page.
Agents that choose their own steps
An agent works toward a goal and decides which steps to take, using tools you give it: look something up in the CRM, check an order, open a ticket, send a draft for approval. Those tools are often connected through an MCP server.
Picture an operations agent that works out why an order is late by checking three systems, then drafts an update for the customer. Agents are good at that kind of varied, look-it-up-and-reason work. The catch is more moving parts, more testing, and the need for firm limits on what the agent may do without a person signing off.
How we decide
If the steps are the same every time, it’s an automation. If people mainly need answers, it’s an assistant grounded in your content. If the job needs judgment across several systems, an agent may be worth it, and we’d start it read-only. And if nobody can describe what a good result looks like, none of the three is ready yet; the first step is writing down the process.
What each one costs to run
Automations are the cheapest, because the model does one small job per run. Assistants cost more, since every conversation sends context to the model. Agents can cost the most, because they may call the model many times to finish one task. Whatever you build, set usage limits and track cost per task from the first day.
How we roll AI out safely
We start with one job that happens often and is easy to check. We collect real examples and agree with the client what a good answer looks like, test models against those examples, then launch to a small group with a person reviewing the output. Access widens as the accuracy holds up. It’s the same approach behind our AI automation and AI agent work, whatever the end result looks like.
Signs a project is too ambitious
The brief says “an AI that handles customer service” with no list of the questions it should answer. Nobody can say what the AI must never do. There’s no documentation for it to learn from yet. Or success is judged by how impressive the demo looks rather than by the hours it saves. Any one of these is a reason to start smaller.
Questions we often hear
Are AI agents ready for real business use?
Yes, for well-defined jobs with clear limits, good tools and human approval for risky actions. Fully hands-off agents running a whole department are still rare, and rarely the best first project.
Which model should we use?
It depends on the task. We test Claude, OpenAI and Gemini models on your real examples and choose on quality, speed and cost.
Can we start small?
Yes, and we’d recommend it. Most clients start with one automation or one assistant, prove the value, then build on it.
