What an AI agency does
An AI agency builds agents and automations for other companies. The work usually has four parts:
- Understanding the process: what comes in, who decides, what goes out and where AI really helps.
- Building the agent: choosing the AI model, connecting the agent to the company's tools and writing the instructions.
- Teaching it with the company's material: documents, prices, policies and the answers the team already gives.
- Maintaining it: fixing it when something changes, tracking the cost and updating the agent when the AI model changes.
Before hiring one, it is worth asking who answers when the agent gets something wrong with a client, what happens if the contract ends, and who keeps the agent, the instructions and the data after that.
Who builds AI agents
Four kinds of company build agents today: AI model makers, which launch their own agents; platforms where anyone can build their own; agencies and consultancies, which build to order; and companies that sell ready-made agents for one kind of work, such as customer service or management. The risks of depending on a single maker are covered in AI vendor lock-in.
Building your own agent: where and how
You can build an agent without writing code, on automation platforms that connect an AI model to the company's tools. The steps are always the same, because an agent has three parts: the model, the tools and the instructions (OpenAI).
- Define the goal: one clear task, such as answering pricing questions or booking appointments, not “help with everything”.
- Connect the tools: calendar, email, spreadsheets and the client records, with the minimum permission needed.
- Write the instructions: what it may do, what needs approval and what it never does.
- Test with real cases: the questions clients actually ask, including the hard ones, before opening it to everyone.
Training an agent, in most cases, means giving it clear instructions and the company's documents, not training the AI model from scratch. That is what makes it answer with the right information.
What stays with the builder is the work that does not show in the demo: protecting the agent against orders hidden in emails and web pages, tracking a cost that varies with use, and knowing what to do when the AI provider goes down. That is the subject of in the demo, everything works.
Buying a ready-made agent
A ready-made agent comes with the tools already connected and the security rules in place, and another company takes care of maintaining it. It starts the same day. In return, it works the way it was designed: if the company's process is very different from others', it may not fit. Before subscribing, it is worth asking for a trial with your own number and your clients' real questions.
The three paths side by side
- To start: the agency charges for the build; your own costs little money and a lot of time; the ready-made one starts with the monthly fee, often after a free trial.
- To maintain: with the agency, the maintenance contract; with your own, the builder's time; with the ready-made one, it is already in the monthly fee.
- Until it works: with the agency, the project time; with your own, the time to build and test; with the ready-made one, the initial setup.
- When it breaks: the agency fixes it, if the contract covers it; with your own, whoever built it; with the ready-made one, the vendor.
- Customization: the agency does what is asked; your own does what you manage to build; the ready-made one does what it was designed to do.
What it costs
Prices vary so much from one vendor to another that a single number says little. What helps is knowing the parts of the bill:
- Build or implementation, when there is one.
- Monthly fee for the service or the platform.
- AI billed by usage, when the agent runs on a model paid for separately: the more it talks, the more it costs.
- Messages on the official WhatsApp API: Meta charges per message, and since October 1, 2026, service replies are free only up to 1,000 a month on each number (Meta).
- Maintenance hours, from whoever built it or someone on the team.
A warning that applies to all three paths: Gartner predicts that over 40% of agentic AI projects will be canceled by the end of 2027 (Gartner). That is why it pays to start with one clear task, at a cost you can track.
Five questions to decide
- Is the process similar to other companies'? If so, a ready-made agent usually does the job. If it is very particular, a custom build is worth it.
- Who will maintain the agent six months from now? If the answer is “nobody”, the path is a ready-made agent, or an agency with a maintenance contract.
- What will it be able to do on its own? The more it does (sending messages, changing the calendar, answering clients), the more the rule on what needs approval matters.
- Who keeps the instructions and the data if you switch vendors? Ask before you start, not when you leave.
- Can I try it first? With your own number and your clients' real questions, not just a demo.
Where Orion fits
Orion is a ready-made AI agent that works as a manager: it serves clients on WhatsApp, uses the calendar, email and documents, organizes the team's deliverables and follows what was agreed. The connection to the systems the company already uses is built to measure, together with the company.
For agencies, consultancies and accountants that already serve small and mid-sized businesses, there is the Partner Program: the partner refers Orion to clients, and each referral turns into recurring commission for the partner; Orion signs up, bills and supports the product. Plans and the trial period are on the pricing page.
Frequently asked questions
- What does an AI agency do? It builds agents and automations for other companies: it understands the process, connects the agent to the tools, teaches it with the company's material and takes care of maintenance.
- Who builds AI agents? AI model makers, platforms where you build your own, agencies, and companies that sell ready-made agents for one kind of work.
- Where can I build an AI agent? On automation platforms that connect an AI model to the company's tools, many of them without code. Maintaining and protecting the agent stays with whoever builds it.
- How do you train an AI agent? In most cases, with clear instructions and the company's documents, not by training the model from scratch. Then by testing it with clients' real questions.
- How much does it cost to build an AI agent? It depends on the path: build and monthly fee with an agency; the builder's time and AI usage with your own; the monthly fee with a ready-made one. Always add the official WhatsApp API messages and the maintenance hours.
Sources
- The parts of an agent: OpenAI, A practical guide to building agents, accessed Nov 20, 2026.
- Fees: Meta, Pricing on the WhatsApp Business Platform, with the rates in effect since Oct 1, 2026, accessed Nov 17, 2026.
- Agentic AI projects: Gartner, Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027 (Jun 25, 2025).
