Services

We tie strategy to implementation: clarify the problem, build the system and stay with the team until it works in real life.

All services
How we work together
Fixed-price offers
Service

Digital advisory

Strategy, audits and roadmaps for decisions that actually move the business.

from 1.800 €

See the service
← Back to journal

Put an AI Agent to Work in Your Firm Without Losing Control: 5 Monday Tips

AI agents now carry out whole tasks, not just answer questions. The smart part is keeping the last word for yourself. Five practical tips for a small business.

Put an AI Agent to Work in Your Firm Without Losing Control: 5 Monday Tips

This year everyone talks about "agents": AI tools that do not just answer a question, but carry out a task in several steps (Forbes). For a small firm, this can mean hours saved every week. But the difference between a real helper and a headache comes down to one thing: who has the last word. Here are five tips to put AI to work on a business task without losing control.

1. Pick a single narrow, repeatable task

Do not automate "everything to do with clients". Choose one clear activity that repeats often and has a result you can easily check. For example: preparing draft replies to frequent questions, turning some notes into a quote, or summarizing a long document into three points. The narrower the task, the less the AI gets wrong and the easier you verify it.

2. Keep a human in the loop: AI proposes, you approve

The safest way to work is to let the AI prepare, and you approve before anything reaches a client. This approach, called "human in the loop", lets the tool work fast while the decision stays with you (Bluetweak). In practice: the AI writes the draft email or quote, you read it, correct it, and send it. No message goes out automatically, without human eyes on it.

3. Match the level of control to the risk

Not every task needs the same attention. A simple, repetitive question, like working hours or the status of an order, can run almost automatically. A complaint, an invoice dispute, or a refund request should have the AI only propose, while you approve (Bluetweak). The simple rule: the more a mistake costs, the more human control.

4. Keep a record of what the agent does

Keep a trail of what it proposed and what you approved. You do not need a complex system: a folder with the generated drafts and the messages you sent is enough at the start. It helps with two things: you see where the AI often gets it wrong and correct it, and you have proof if a client later asks what you told them. This habit of documenting decisions is becoming more important because of new European rules (see below).

5. Start cheap and measure the time saved

You do not need an expensive subscription to test. Capable models have become very affordable: Anthropic launched Claude Sonnet 5 at an introductory price of 2 dollars per million units of text on the input and 10 dollars on the output, until August 31, 2026 (Anthropic). Take one task, use it for a week, and note how many minutes you save. If it works, you expand. If not, you lost an hour, not a budget.

Being careful with data and the new European rules

When an AI agent works with client messages, the text reaches the provider's servers. That is why you should not put there data you would not want to see outside your firm: full name and ID number, addresses, signed contracts, passwords, bank details. Anonymize first. Instead of "the client John Smith", write "a client". Instead of the exact amount in a contract, use a made-up value with the same shape.

This care is no longer just common sense, it is also law. GDPR has required you for years to protect personal data, and the new European regulation on artificial intelligence (2024/1689) comes into force in August 2026 and requires, among other things, human oversight of high-risk AI systems (AI Governance Institute). Even if your small firm does not fall under "high risk", the direction is clear: the human stays responsible. The five tips above put you exactly on the right side of this rule.

One concrete step for today

Pick a single repetitive task in your firm and set a simple rule: the AI proposes, you approve, and your clients' sensitive data never leaves without being anonymized. Test it for a week. It is the cheapest way to find out whether AI really helps you, without risking anything important.

Sources

  1. Forbes, small business tech roundup (July 12, 2026)
  2. Bluetweak, human in the loop in customer support
  3. Anthropic, Claude Sonnet 5 launch
  4. AI Governance Institute, regulation weekly (July 3, 2026).
Share

NEWSLETTER · WEEKLY

The week in AI,
distilled.

Every Friday, one short e-mail on what happened in AI and tech, translated into what it means for your business. No jargon, no spam.

  • The news that matters, already filtered
  • One practical tip you can apply on Monday
  • Unsubscribe in one click, anytime

Have a project in mind?

Book a free 30 minute strategy call.

Let's talk