← All articles
AI · 9 min read

AI agents for business: what they can really do today

Agents don't just answer questions — they use tools and complete tasks. Where they already save hours, where they still fail, and how to launch one without risk.

A chatbot answers. An AI agent acts: it reads a request, decides which tools to use — a search, a database query, an email, a ticket in your CRM — performs several steps and reports the result. That's what makes agents exciting, and also why they need more care than a simple chat widget.

How an agent works

RequestPlanUse toolsCheck resultAnswerLog
The agent loop: the model plans, calls tools, checks what came back and repeats until the task is done or it needs a human.

The language model is the “brain”, but the value comes from the tools you connect: read an order status, create a refund draft, search the knowledge base, book a meeting. Each tool is a small, well-defined API with clear permissions.

Where agents already pay off

TaskWhat the agent doesTypical effect
Customer supportLooks up orders, answers from policies, prepares refunds for approvalMost routine requests closed without waiting for an operator
Sales and lead handlingQualifies leads from forms, enriches data, drafts personalised repliesFaster first response, cleaner CRM
Internal operationsCollects data from several systems into a weekly reportHours of manual copy-paste saved
Document processingExtracts fields from invoices and contracts, flags anomaliesFewer manual checks, fewer typos
Development and QAWrites test drafts, triages failed runs, summarises logsEngineers spend time on decisions, not routine

Where they still fail

  • Long, open-ended tasks: the more steps, the higher the chance of a wrong turn
  • Ambiguous requests: without clear rules the agent guesses — and guesses confidently
  • Missing or messy data: an agent can't use what isn't in your systems
  • Irreversible actions: payments, deletions and messages to customers need a human checkpoint
Start with tasks that are frequent, well defined and easy to check. “Draft a reply and a refund for approval” is a great first agent; “run our customer support alone” is not.

Guardrails we always build in

  1. Least privilege: each tool can do only what it must — read-only where possible, limits on amounts and recipients.
  2. Human approval for anything that spends money, deletes data or contacts customers.
  3. Full logging: every step, tool call and result is recorded, so any decision can be explained.
  4. Evaluation before launch: a test set of real requests, including tricky and malicious ones such as prompt injection.
  5. Monitoring after launch: success rate, escalations, cost per task and user feedback.

How we launch an agent

We start with a two-week pilot on one process: map the steps, connect two or three tools, run the agent in “suggest” mode where people approve every action, and measure. When quality is stable, we let it act on its own for the safest steps and keep humans on the rest. See also how we choose the model and test AI products.

A good agent is like a good new colleague: it gets clear tasks, limited access at first and more responsibility as it earns trust.