# AI agent development

> We build AI agents that answer customers on WhatsApp, web and phone, book appointments and update your CRM. Free first call.

URL: https://gradual.pro/en/services/ai-agent-development

Updated: 2026-09-29

Assistants that answer on WhatsApp, web or phone and act in your CRM, calendar or ERP. With a person always one step away.

- An AI agent does more than reply: it looks up data and runs actions in your tools.
- The highest-return uses are customer service, call answering and lead qualification.
- A pilot runs in 4-6 weeks, with a fixed budget per phase agreed before we start.
- It can always hand the conversation to a person on your team.

## What is the difference between a chatbot and an AI agent?

A classic chatbot follows a tree of fixed answers: when the customer goes off script, it gets stuck. An AI agent uses a language model (such as GPT, Claude or Gemini) to understand questions phrased any way, and it has access to tools: it can check availability in your calendar, look up an order status or log a lead in your CRM.

In practice, a chatbot answers and an agent resolves. That is why agents take over whole tasks and not only the first reply.

## What can an AI agent do in your business?

These are the agents we deploy most, ordered by how fast they pay back:

- **AI receptionist**: answers calls 24/7, handles frequent questions and books appointments. See [AI receptionist](/en/use-cases/ai-receptionist).
- **WhatsApp agent**: replies to customers, sends quotes and follows up. See [WhatsApp AI agent](/en/use-cases/whatsapp-ai-agent).
- **Customer service agent**: resolves issues from your knowledge base and escalates complex cases. See [AI customer service](/en/use-cases/ai-customer-service).
- **Sales agent**: qualifies leads, books demos and updates the CRM.
- **Internal assistant**: answers your team from company documentation (procedures, price lists, contracts).

## How do you stop an agent from making up answers?

The real risk of an AI agent is that it confidently says something false. We reduce it with three measures: the agent answers only from your company's information (a technique called RAG), it has explicit instructions to hand off to a person when unsure, and we review a sample of conversations every week during the pilot.

Sensitive actions, such as issuing a refund or confirming a special price, can also require human approval before they run.

## What do we need from your company?

For the pilot we need the information the agent will work with (FAQs, price lists, policies, catalog), access to the tools it has to use, and one person on your team to review its answers during the first few weeks.

Your information doesn't need to be tidy: we structure it as part of the project.

## Process

1. **Define the job**: Which conversations it handles, what it can do, and when a case goes to a person.
2. **Prepare the knowledge**: We organize your information into a base the agent checks before it answers.
3. **Connect the tools**: Calendar, CRM, ERP, email or online store, with permissions limited to what is needed.
4. **Supervised pilot**: The agent works with real customers and we review its answers every week.
5. **Continuous improvement**: We measure resolution, response times and satisfaction, and tune monthly.

- of queries resolved without human involvement: 60-80%
- first response time, 24/7: < 10 s
- until the pilot is in production: 4-6 weeks

Typical ranges in customer service projects; they depend on the type of queries and the quality of the available information.

### Traditional chatbot, AI agent and human

|  | Traditional chatbot | AI agent | Human |

| --- | --- | --- | --- |

| Understands free-form questions | Poorly | Yes | Yes |

| Acts in your systems | No | Yes | Yes |

| Available 24/7 | Yes | Yes | No |

| Cost per conversation | Very low | Low (cents) | High |

| Complex or sensitive cases | No | Hands off to a person | Yes |

Tools: OpenAI, Anthropic Claude, Google Gemini, n8n, WhatsApp Business API, Twilio, Retell AI, HubSpot, Supabase

## FAQ

### How much does it cost to build an AI agent for a business?

It depends on the number of channels (for example WhatsApp, web or phone), the integrations, the volume, whether it uses voice, and ongoing maintenance. We budget per phase with a fixed price agreed before we start. AI model usage, telephony and WhatsApp fees are paid directly to the providers.

### Can the agent talk on the phone?

Yes. Current voice agents hold natural conversations with latency under a second. We connect them to your existing number or a new one, and they can transfer the call to a person when needed.

### Which AI model do you use?

We pick the model for the task: GPT, Claude or Gemini for conversation and reasoning, and smaller, cheaper models for classifying or extracting data. Because the agent doesn't depend on a single provider, we can switch models if a better or cheaper one appears.

### Is it legal to use AI agents with customers in the EU?

Yes, as long as you comply with the GDPR and the EU AI Act. The customer must know they are talking to an AI, the data needs a legal basis, and there must be a way to reach a person. We configure agents that way by default, and the same setup works for clients in the UK and US.

### Can I build an AI agent myself?

For simple cases, yes: tools like ChatGPT's custom GPTs or n8n's AI assistants let you build a prototype in an afternoon. The difference is connecting it securely to your systems, controlling its answers and maintaining it when something changes, which is what we do.