AI agents for business: what’s worth building
AI agents for business are software systems where a language model decides which steps to take and which tools to call to finish a task – triaging a ticket, updating a CRM, fixing a failing test – rather than following a fixed script. They pay off on high-volume, well-bounded work with a cheap way to check the result.
Most companies that try agents start with the wrong use case: too open-ended, too risky, or impossible to measure. This guide ranks the realistic options, explains when a plain workflow beats an agent, and lays out a small first project you can evaluate in weeks rather than quarters.
What an AI agent actually is (and what it isn’t)
Strip away the marketing and an agent is a loop: the model gets a goal and a set of tools, picks a tool, sees the result, and decides what to do next until it’s done or gives up. The defining feature is that the model controls the control flow.
A workflow is different. You write the steps; the model fills in the parts that need language understanding – classify this email, extract these fields, draft this reply. The path is fixed in code.
That distinction matters because workflows are cheaper, faster, easier to test and easier to explain to an auditor. Agents earn their extra cost only when the path genuinely can’t be known in advance: debugging, research, multi-system support cases. A lot of what gets sold as an “agent” is a workflow with a chat box, and that’s often the right design. I go deeper on the patterns in agentic workflows.
Use cases ranked by ROI and risk
Here’s how I rank the common candidates, roughly from best first bet to “wait”. The pattern is simple: high volume, low cost of a wrong answer, and a fast way to verify the output.
- Internal knowledge lookup. Answering staff questions from your docs, runbooks and tickets. High volume, the reader is an employee who can sanity-check the answer, and citations make it verifiable. Usually a RAG workflow rather than a full agent.
- Support agent assist. Drafting replies and pulling account context for a human agent who sends the final message. The human is the guardrail, and time saved per ticket is easy to measure.
- Back-office data work. Extracting fields from invoices, contracts and forms into your systems, with low-confidence items routed to a person. Boring, high-volume and very measurable.
- Software development tasks. Coding agents working well-scoped tickets – dependency bumps, test coverage, small bug fixes – with every change going through normal code review. The review step is what makes this safe.
- Sales and ops research. Enriching leads, summarizing accounts before calls, preparing briefs. Wrong answers are cheap; usefulness depends on the quality of the sources you give it.
- Customer-facing conversational agents. Resolving customer requests end to end. Highest upside, but also the highest risk: the output reaches customers directly, and the agent needs write access to real systems. Do this after you’ve learned on internal use cases.
- Autonomous actions on money or production. Refunds, pricing changes, infrastructure changes without a human approval step. Don’t start here.
AI agents for small business vs SaaS companies
For a small business, buy before you build. Your helpdesk, CRM and accounting tools increasingly ship agent features, and a no-code automation tool connected to a model will handle a surprising amount of email triage and data entry. Custom development rarely pays back at small volumes. The useful work is choosing one process, measuring it before and after, and checking the vendor’s data handling.
For AI agents for SaaS, there are two separate questions. The internal one – agents that help your own team – follows the ranking above. The product one is whether your customers’ agents can use your product. That increasingly means exposing a well-designed API and an MCP server so agents in Claude, ChatGPT or an IDE can work inside your product with the user’s permissions. I explain how that works in what is an MCP server.
Build vs buy
Buy when the use case is generic and the vendor already integrates with your systems: support deflection on a mainstream helpdesk, meeting notes, coding assistants. You won’t out-build a funded team on their core product.
Build when the value comes from your proprietary data or workflows, when the agent needs to act inside your own product, or when vendor pricing per seat or per resolution gets expensive at your volume. Building doesn’t mean training models – it means orchestration code, tool definitions, retrieval over your data, evals and guardrails around a hosted model.
The hybrid is common and sensible: buy the commodity layer, build the thin piece that encodes what’s specific to you. See build vs buy software for the general framework.
How to start small
- Pick one task with a number attached. “Reduce time to first draft on billing tickets” beats “use AI in support”. Record the current baseline: volume, time per task, error rate.
- Collect 50–100 real examples with the correct outcome. This becomes your evaluation set and is the most valuable artifact of the whole project.
- Build the simplest version that could work – usually a workflow with one or two model calls, not an autonomous agent. Add agentic steps only where the fixed version demonstrably fails.
- Keep a human in the loop at the point of consequence: the agent drafts, a person approves. Measure how often they edit or reject.
- Give the agent the narrowest tools possible. Read-only first. Scoped write access later, per tool, with an audit log.
- Run it in shadow or on a slice of traffic for two to four weeks. Compare against the baseline, including the cost per task.
- Decide on evidence. Expand, adjust, or stop. Stopping is a legitimate result and much cheaper than a half-trusted system everyone works around.
What makes agent projects fail
- No evaluation set, so every prompt change is judged by vibes and regressions ship silently. See LLM evaluation.
- Too many tools. Models choose badly from a long list of overlapping tools. Ten clear ones beat fifty thin wrappers over every endpoint.
- Messy source data. An agent over contradictory documentation gives confident, contradictory answers.
- Write access before trust. One bad autonomous action can set an AI program back by months internally.
- Ignoring security. Agents that read untrusted content – emails, web pages, tickets – can be steered by instructions hidden in it. Read AI agent security before connecting anything with write access.
- Measuring activity, not outcomes. “The agent handled 10,000 conversations” means nothing if customers came back the next day with the same problem.
Where I fit
If you have a clear use case and engineers who can own it, you may not need outside help – start with the steps above. Where I’m useful is when the first project needs to ship into a real production system with auth, permissions, evals and cost controls, and your team hasn’t built one before. I do that as hands-on AI agent development, working inside your codebase. I also build agent-facing products myself: Craaft is a kanban board that people and AI agents work together through MCP and a REST API.
Questions
What are AI agents for business used for?
The most reliable uses today are internal knowledge lookup, drafting support replies for human review, extracting data from documents, well-scoped coding tasks, and research briefs. They work best on frequent, bounded tasks where a person or an automated check can verify the result quickly.
What is the difference between an AI agent and a chatbot?
A chatbot answers in conversation. An agent takes actions: it calls tools, reads results and decides the next step until a task is done. Many modern support “chatbots” are really conversational AI agents, because they can look up orders or update accounts rather than just reply with text.
Are AI agents worth it for a small business?
Often yes, if you buy rather than build. Agent features inside tools you already use – helpdesk, CRM, email – are cheap to try. Measure one process before and after. Custom-built agents rarely pay back unless the task is high volume or specific to how your business works.
How much does it cost to run an AI agent?
Running cost is mostly model usage: tokens per task multiplied by tasks per month, and agents use more tokens than single calls because they loop. Estimate it from a prototype on real examples, then check your provider’s current pricing. Build and maintenance time usually outweighs model spend early on.
Can AI agents replace employees?
In practice they replace tasks, not roles. The realistic outcome is a team that handles more volume with the same headcount, with people moving to review, exceptions and the work that needs judgement. Plans that assume full replacement tend to underestimate the cost of the mistakes.
Have a use case but not sure it’s the right first one?
Send me the process you’re thinking about. We’ll pick the smallest version worth testing and ship it in a two-week sprint with an eval set, so the decision to expand is based on numbers.
GMT+2 · Full EU overlap · US East mornings