What Is a Custom AI Agent? (And How It's Different From a Chatbot or Custom GPT)

Ask ten vendors what an 'AI agent' is and you will get ten answers, most of them a chatbot wearing a new sticker. Here is the honest version. A custom AI agent is an AI you configure for your specific business -- a name, a role, plain-language instructions -- that can actually *act* on your real data: book the appointment, draft the invoice, update the customer record, look up the order. Not describe how it would. Do it.

The load-bearing word is *act*. A chatbot answers questions. A custom GPT answers questions using knowledge you fed it. A custom AI agent takes real actions inside the software your business already runs on, under permissions you set for each one. That last clause -- under permissions you set -- is the whole game, and it is the part most explainers skip. This guide defines the category cleanly, shows exactly how an agent differs from the two things people confuse it with, and tells you when you actually need one.

A custom AI agent is a configurable AI worker that reaches into your business apps -- your calendar, your CRM, your invoicing -- through a controlled set of permissions, and finishes tasks on your behalf.

Unpack that and three ideas are doing the work. *Configurable*: you give it a role and instructions in plain English, the way you would brief a new hire, not by writing code. *Reaches into your apps*: it can read and change real records, not just chat about them. *Controlled permissions*: you decide, per tool, whether it can look, change, or must ask you first. Take away any one of those and you no longer have an agent. Take away the actions and you have a chatbot. Take away the permissions and you have a liability.

These three get sold under overlapping language, so pin them down. A chatbot -- the classic kind -- follows a script you built: buttons, keywords, canned replies. It is cheap and predictable and knows only what you scripted. A custom GPT (or a Claude Project) is a big step up in language: it holds instructions and documents you give it and talks fluently about your business. But it lives in a chat window with no hands. Ask a custom GPT to book Thursday at 2 and it will write you a lovely paragraph about booking Thursday at 2. It cannot open your calendar.

A custom AI agent is the one with hands. It understands the freeform request like a custom GPT does, and then it reaches into your CRM, your calendar, your invoicing, and completes the task -- because it is wired to the tools, not just the text. The difference is not intelligence. It is access, and the control layer that makes that access safe.

Abstract definitions blur, so watch one in motion. A customer texts your business at 9:40 p.m.: 'any chance you can take a look at my sink tomorrow afternoon?' A chatbot replies with your hours and a link to a booking page, and hopes. A custom GPT could write a warm, perfectly worded reply -- and still leave the actual booking to the customer.

A custom agent does the whole thing. It reads the freeform text, understands 'tomorrow afternoon,' checks your real calendar for open slots, offers 2:30, and when the customer says yes, it writes the appointment, texts a confirmation, and drops a new lead into your CRM tagged 'sink -- via text, after hours.' By the time you wake up, the job is on the books and the record is clean. Nobody typed on your end. That gap -- between a reply and a booked job -- is the entire reason the category exists, and every example in this cluster is a version of it.

You do not have to start from a blank agent. Deelo ships templates -- a Sales agent, a Support agent, a Customer Success agent -- and each one arrives pre-briefed: sensible instructions, a starter set of app access, and approval defaults already chosen for the job. You pick the closest one and adjust it to your business instead of building from zero.

That matters more than it sounds, because the hardest part of a good agent is not the writing, it is knowing which permissions are safe to grant for a given role. A template encodes those choices as a starting point -- a support agent does not come with the power to issue refunds on its own, for instance. Start from the template that matches the job, change the name, the tone, and the specifics, and you have skipped the part most people get wrong.

It helps to see an agent as three pieces you configure separately.

The brain is the role and instructions -- who this agent is, what it is for, how it should talk, what it must never do. In Deelo you write this as plain text in a form; no prompt engineering degree required. You can also start from a shipped template (a Sales agent, a Support agent, a Customer Success agent) that comes pre-briefed and adjust from there.

The hands are the tools it can reach. A Deelo agent is born connected to the same roughly 597 tools across 49 apps that the interactive AI Assistant uses -- bookings, invoices, CRM records, orders, marketing, and the rest. You do not wire APIs; the business already runs on the platform, so the hands are already attached.

The leash is the permission and guardrail layer. For every tool you choose off, allow, or allow-with-approval, and you set an overall autonomy level. This is what turns a powerful agent into a safe one, and it is different enough that it gets its own guide on how to give an agent safe access to your data.

'Custom' is not marketing here; it is the four dials you turn. You choose which of your installed apps the agent may touch, and no more. You set the per-tool permissions. You pick an autonomy level -- from confirm-everything to run-freely-within-limits. And you decide where it lives: inside your workspace, on a schedule, or on a customer-facing channel like phone, text, or a shareable web link.

Each agent also keeps its own isolated memory and can hand work to another agent, so a front-desk agent and a billing agent are genuinely separate coworkers, not one bot with a split personality. Two businesses using the same template end up with very different agents, because the dials are set for their apps, their risk tolerance, and their workflow.

A custom AI agent is not a general-purpose coding or infrastructure agent. It acts on the data in your business apps; it does not spin up servers or write your product's codebase. Point it at the job it is for.

A few more honest limits worth knowing up front. On the phone it is turn-based -- it listens, then responds, a natural back-and-forth rather than a sub-second, interrupt-me-mid-word call. Voice and SMS run on the Communications add-on; the web channel runs on your regular AI Assistant credits. And safety is deliberately conservative: an outside caller or an anonymous web visitor can never approve a high-risk action like moving money or deleting a record. The agent gathers what it can and defers that decision to a human on your team. Phone, SMS, and web are live; WhatsApp and Facebook Messenger are on the roadmap, labeled coming soon rather than sold as shipping.

You need a custom agent when a job has to *touch your systems*, *run without you sitting there*, or *live on a channel your customers use*. Answering the same booking question at 11 p.m. and actually booking it. Chasing overdue invoices every Monday morning. Qualifying inbound leads into the CRM while you are on a job. Those are agent jobs, because they require action, timing, or a public front door.

You do not need one if you just want to ask questions of your own data now and then -- that is what the interactive AI Assistant is for, and it inherits your own permissions the moment you log in. The line between the two is worth getting right, and it has its own guide. For the wider view of where agents are heading across a whole business, the autonomous agents guide zooms out, and if the chatbot-versus-agent line is still fuzzy, live chat vs. chatbot vs. AI agent draws it cleanly.