What should your first AI employee do?

Update, 2 October 2026: the agent builder, approvals and the rupee cost of every task described below are live. Early access is no longer free: every new workspace starts with ₹25 of free credits and no card, then you pay as you go, in rupees plus GST, with a tax invoice for every top-up. See pricing.

The most common mistake with AI employees is starting too big. "Handle all our customer messages" sounds efficient, but it gives the agent a job with no edges, you no way to check it, and your customers a worse experience than before.

Start as you would with a new hire: one role, one channel, and a close eye on the first few weeks.

What makes a good first job

A good first job has five things going for it:

1. It repeats. The same questions or steps come up many times a week.
2. The answers are written down. Timings, prices and policies already exist in a document, or could be written in an afternoon.
3. Mistakes are cheap to fix. A wrong answer about opening hours is easy to correct; a wrong refund is not.
4. It is easy to check. You can read the work and quickly tell whether it is right.
5. The hand-off is clear. You can say exactly when a person should take over.

Jobs that usually fit:

• answering routine questions about products, fees, timings and delivery;
• replying to new enquiries within minutes and collecting the details your team needs;
• sending appointment reminders and handling simple rescheduling;
• sending polite reminders for invoices that are already due;
• preparing a daily summary of orders, enquiries and open issues.

Hold these back for now: complaints, refunds and discounts, negotiations, anything needing medical, legal or financial judgement, and anything involving sensitive information such as health or bank details.

Write the job description

Your AI employee is only as good as its brief. Write down:

• The role, in one sentence: "Answers WhatsApp questions about our courses, fees and batch timings."
• Its sources: the FAQ, price list and policies it may rely on. If an answer is not there, it should say so.
• Its limits: for example, never promise a delivery date, quote an unlisted price or discuss other customers.
• When to hand over (more on this below).
• Tone and language: formal or friendly; English, Hindi or Hinglish, following the customer's lead.

Start with approvals on

At first, have a person approve every message before it goes out. It feels slow, but every draft you edit reveals a gap in the brief or the sources, and closing those gaps is what makes the agent good.

Once drafts of one type are consistently right, stop requiring approval for that type. Keep approvals permanently for anything that commits money, makes a promise or touches a sensitive subject.

Set spending limits before you go live

Most AI tasks cost little, but a confused agent stuck in a loop can run up a bill. Set limits at two levels:

• Per task, so one runaway conversation stops early. Watch what normal tasks cost during your trial and set the cap comfortably above that.
• Per month for the whole workspace, with an alert well before you reach it.

Decide what happens at the limit. Usually, the agent should stop and tell a person.

Tell customers they are talking to an AI

Say so at the start of the conversation, in the customer's language. For example:

Hi! I'm the AI assistant for [your business]. I can help with timings, fees and bookings. Type "agent" at any time to reach a person.

नमस्ते! मैं [आपके बिज़नेस] का AI सहायक हूँ। मैं समय, फ़ीस और बुकिंग में आपकी मदद कर सकता हूँ। किसी व्यक्ति से बात करने के लिए कभी भी "agent" लिखें।

Do not give the agent a human name and backstory that suggest it is a person. Customers who feel tricked stop trusting the business. India's rules on AI-generated content have also been getting stricter, so disclosure is the safer position.

Get consent right before you message anyone first

Replying to a customer is different from messaging someone first. Before an AI employee does that, check three sets of rules. This is a general overview, not legal advice, so speak to a lawyer about your own situation.

The Digital Personal Data Protection Act, 2023 (DPDP Act). Names, phone numbers and message histories are personal data. The usual basis for using them is consent that is free, specific, informed and clear, given after a notice that explains what you collect and why. People must be able to withdraw consent as easily as they gave it. Use the data only for the stated purpose, keep it secure, and delete it when you no longer need it. The Act's obligations are coming into force in phases, so check which apply to you today.

TRAI's Telecom Commercial Communications Customer Preference Regulations (TCCCPR). These cover commercial SMS and voice calls. Senders must register on a DLT (distributed ledger) platform and get their sender headers and message templates approved. Promotional messages may only go to people whose preferences allow them or who have given recorded consent, and opting out must be easy. Service and transactional messages follow different rules, so never label a promotion as a service message.

WhatsApp's business rules. TCCCPR is written for SMS and calls, but Meta's own policies require a person's opt-in before a business messages them on WhatsApp, and pre-approved templates for conversations the business starts.

In practice:

• record when and how each person agreed, and to what;
• keep service messages ("your order has shipped") separate from promotions ("our sale starts today");
• honour "stop" at once, and make sure your AI employee does too;
• never buy or scrape contact lists.

Know when to keep a human in the loop

Even a well-trained AI employee should hand over when:

• a customer is upset, complaining, mentions legal action or asks for a person;
• money is involved beyond the limits you set;
• the question needs medical, legal or financial judgement;
• the customer seems vulnerable or the matter is sensitive;
• the answer is not in its sources, or it is unsure.

A good hand-off passes the conversation and a short summary to your team, so the customer does not have to repeat themselves.

Your first-week checklist

1. Pick one job and one channel.
2. Write the role, sources, limits and hand-off rules.
3. Turn on approvals for every outgoing message.
4. Set per-task and monthly spending limits.
5. Write the AI disclosure greeting in each language you use.
6. Check your consent records for anyone you will message first.
7. Run test conversations, including awkward ones.
8. Review tasks daily, improve the brief, and relax approvals one message type at a time.

Approvals, spending limits and the rupee cost of every task are all part of the Dhanur AI agent builder we are shipping next. Create a free workspace today and be ready when it opens.