AI agents for business

A marketing agency we work with turned four manual steps into a routine of chained subagents, and in another area of the same client the team went from a handful of creative concepts per week to more than thirty per run. That’s the example this post on AI agents for business closes with: what they are, what types exist today and what one looks like running, with the tool it was built with and what changed once it went live.

What are AI agents for business

A chat assistant answers a question and waits for the next one. An AI agent receives a task, decides the steps needed to complete it, runs tools on its own (reading a file, calling an API, writing a document) and delivers a result, not a standalone reply. The difference is architectural: an agent has access to tools and can chain several calls to the model until the task is closed, without a person approving every intermediate step.

For Claude, that agentic capability is part of the base architecture. Claude Code applies it to the terminal and to code: it reads a repository, edits files, runs commands and fixes what breaks. Claude Cowork applies the same architecture to office work: it reads and writes documents, connects to Slack or Google Drive, and delivers finished spreadsheets or presentations. Underneath both sits the Agent SDK, the route for building custom agents on the same model when the use case doesn’t fit either off-the-shelf product.

What types of AI agents exist

It helps to separate agents by where they run and what they touch, rather than by what they claim to do:

• Coding agents: automate work on a repository (writing, debugging, reviewing pull requests). Claude Code is the fastest-growing example today.
• Office agents: automate work on documents and everyday business applications (reports, spreadsheets, presentations). Claude Cowork covers this case without a terminal.
• Custom vertical agents: built for one specific company process (qualifying leads, generating proposals, monitoring a sales pipeline), usually with the Agent SDK or MCP (the protocol that connects an agent to external tools and data) to reach internal systems.
• Specialized subagents: within a single workflow, different steps are split across agents with their own model and their own tools. A scraping step can run on a cheap, fast model while the final writing step runs on a more capable, more expensive one, all inside the same task.

There isn’t one type of agent that fits everything. The question isn’t which AI agent exists for a given case, but which part of the work repeats often enough to justify building one.

Example of AI agents built with Claude

The agency in the opening is one of our clients. Before working with us, their proposal pipeline ran through four separate manual steps: reviewing the client’s portal, enriching the data, applying the style guide and assembling the proposal. A person ran each step, one after the other.

We turned those four steps into a single chained routine using subagents: each step runs as an independent agent, with the model and tools it needs, and the output of one feeds directly into the next without anyone supervising in between. The full pipeline now runs on its own, start to finish.

The effect showed up elsewhere in the same client’s work: the team, once settled into the new workflow, went from a handful of creative concepts per week to more than thirty per run, with more than one hundred published across three markets within days. Running several agents in parallel over a process that used to be manual, step by step, explains that jump.

It’s one example among many of building AI agents for companies on top of an existing base, rather than building an agent product from scratch: the work was in identifying which steps always ran the same way, not in writing a new model.

What type of AI do companies use

In practice, almost no company uses just one type. The usual pattern is a mix: a general-purpose chat assistant for one-off questions (Claude, ChatGPT, Gemini), a coding agent for the technical team (Claude Code is the preferred choice among more experienced developers) and, when the volume of a specific process justifies it, a custom agent built with MCP or the Agent SDK for that case.

The gap between companies that get value from an agent and those that don’t usually has less to do with which product they picked and more with whether they correctly identified which process deserves a dedicated agent. A regular chat is still the fastest option for a single question; an agent only pays off once a task repeats and has several steps.

What are the 10 most used AI tools

There’s no single, stable ranking of “the 10 most used AI tools”: the numbers shift depending on whether you measure web traffic, active users, adoption inside companies or mobile app downloads, and each public source measures something different. What does hold across most studies is which assistants show up near the top: Claude from Anthropic, ChatGPT from OpenAI and Gemini from Google consistently lead the general-use rankings alongside DeepSeek, Grok, Copilot or Perplexity depending on the source, followed by more specialized tools depending on the industry.

That ranking measures which assistant more people use, a different question from which one handles a specific company process best. A company that needs an agent to qualify leads or generate proposals picks the model that runs that task well, at the cost and level of control it needs, regardless of where it sits in the ranking.

What is the best AI agent for work

There’s no “best” AI agent in the abstract, it depends on the work that needs solving. For coding tasks, an agent that runs in the terminal and understands a full repository (like Claude Code) makes sense. For office tasks on documents and applications already in use, an agent that runs on the desktop without a terminal (like Claude Cowork) fits better. For a very specific internal company process, neither off-the-shelf product works as-is: it takes a custom build with the Agent SDK, connected to internal systems through MCP.

In our experience with clients, the most common mistake is trying to automate a process that still isn’t well defined by hand, rather than picking the wrong agent. An agent executes fast whatever is already well understood, and executes an ambiguous process just as fast, without fixing it.

How to create AI agents for companies

The process we follow to create AI agents for companies always starts with the same step: mapping the manual process exactly as it runs today, step by step, before touching any tool. Only then does it make sense to decide the architecture.

1. Identify the repeatable process. A process that runs the same way every time, with clear steps, is a candidate for an agent. A process that changes depending on who’s doing it isn’t, yet.
2. Pick the right base. If the work is on code, build it on Claude Code. If it’s on documents and office applications, build it on Claude Cowork. If the process is specific to the company and doesn’t fit either, build it with the Agent SDK, connected to internal systems through MCP.
3. Split it into steps, not one agent that does everything. The proposal pipeline above works because each step (portal, enrichment, style, proposal) is a separate agent with its own model, not a single call trying to cover all of it.
4. Test with the client’s real data before rolling out to the team. An agent that works on test data and fails on real data isn’t ready, no matter how good the demo looks.
5. Let the process run on its own, with someone reviewing the result, not every intermediate step. The end goal is for the routine to run without constant supervision; human review moves to the end of the process, not the middle of each step.

Building AI agents for business doesn’t require an in-house AI engineering team. It requires being clear on which process repeats, what data it works with, and which part of that process can run on its own without losing control over the result.