AI Agent Development Company

AI agent, chatbot or workflow automation - which one do you need?

The three are often sold under one name, but they do different jobs. A chatbot answers questions in a conversation. Workflow automation runs fixed steps: when a form arrives, create the ticket and send the email. An AI agent is given a goal, decides which tools to call and in what order, reads each result and picks the next step.

• Choose a chatbot when the job ends with an answer - product questions, policy lookups, order status. Our chatbot development team is the better fit.
• Choose workflow automation when the steps never change. It is cheaper, faster and easier to audit than any model.
• Choose an agent when the path depends on what it finds: a refund that needs an order lookup, a policy check and a carrier query, in an order that varies case by case.

What tools should an agent be allowed to use?

An agent is only as safe as the tools you hand it. Each tool is a small, named function - search_orders, create_refund, update_crm_contact - with a strict input schema, and the agent can do nothing outside that list. Miracuves starts every agent on read-only tools and adds write tools one at a time, once the read path has proved itself on real cases.

Permissions follow least privilege: the agent gets its own service account, scoped API keys, a rate limit per tool and a cap on how much it may change in one run. It never borrows an administrator's login. Every tool call is logged with its inputs and outputs, so any change in your systems can be traced to the run that made it.

Where should a human approve what the agent does?

Put the approval step wherever a mistake costs money, cannot be undone or reaches a customer. Typical approval points are refunds and credits above a set amount, messages sent to customers in the agent's own words, changes to prices, contracts or account status, and any deletion.

The agent prepares the action - the refund amount, the drafted reply, the reason - and pauses. A person approves, edits or rejects it in Slack, the helpdesk or an admin screen, and the run continues. Action types with a clean approval record can later be switched to automatic one at a time, on your decision, not ours.

How is an AI agent tested, and what happens when it fails?

Before the build starts we collect real cases from your history - tickets, requests, records - each with the outcome a good employee would reach. That set becomes the agent's exam. Every change to a prompt, a tool or the model is run against it, and a release ships only if task success holds or improves.

Failures are designed for rather than hoped away. A tool call that errors is retried once, then the run stops. An output that fails its schema check is rejected. A run that loops or passes its step or token budget is cut off. In every case the task goes to a person with the full trace attached, so nothing is dropped silently.

What does an AI agent cost to run each month?

The build is quoted once; the running cost arrives every month and grows with volume. An agent makes several model calls per task, because it plans, calls tools and checks what came back. The other lines are the vector database and hosting, the tracing and monitoring stack, and any paid APIs its tools call.

We measure cost per completed task on the evaluation set before launch, so you can multiply by your own volume instead of guessing. The usual levers are a small model for routing and classification, a large one only for the hard step, caching repeated lookups and capping steps per run. If a task costs more through the agent than through a person, we tell you before you commit.

Which systems can an agent connect to?

Anything with an API or a database the agent is permitted to read. Common connections are helpdesks such as Zendesk, CRMs such as Salesforce and HubSpot, Slack and email, Shopify, Snowflake and PostgreSQL, and your own internal services. Each connection becomes one or more tools, each with its own permissions.

Where a system already publishes an MCP (Model Context Protocol) server, the agent can use it instead of a custom connector, which shortens the build; we still review which tools that server exposes before switching it on. Systems with no API, such as an old desktop app or a supplier portal, can sometimes be driven through the screen, but that route is slow and breaks when the screen changes, so we treat it as a last resort.

When should you not use an AI agent?

When the task is identical every time, a script or workflow tool does it cheaper and with no chance of an inventive answer. When the job is answering questions from your documents, a retrieval chatbot is enough. When the acceptable error rate is zero and nobody can review the output, an agent is the wrong tool today.

Other warning signs: too few tasks a month to repay the build and testing, data the agent is not allowed to see, or nobody on your side to own the approval queue. If what you actually need is a model adapted to your own data rather than an agent calling tools, our LLM development team is the better place to start.