AI Agent for Trade Show Lead Follow-Up: From Cards to Reviewed Messages

An AI agent for trade show and event leads reads business cards, badge photos and notes taken during the event, organizes each contact into a standard record, and drafts a follow-up message that refers to the actual conversation. The salesperson reviews every draft before anything is sent, and only contacts with a valid legal basis and proper permission receive messages. The gain comes from turning messy material into clear actions, with a trail and human accountability, not from sending contacts automatically.

Key takeaways

• The agent organizes raw material into fixed fields and keeps track of where each piece of information came from.
• The follow-up draft cites only what was said or noted, without inventing interest.
• No message goes out without the salesperson's approval, and the channel must respect the LGPD and WhatsApp rules.
• Missing or doubtful data becomes an open item, never a guess.
• A pilot at a real event, with an agreed metric and a decision log, comes before any expansion.

Why trade show material usually gets lost

Trade shows concentrate many contacts into a few days. Cards end up in bowls, notes stay in notebooks, badge photos sit on phones, and conversations at the booth turn into loose reminders. When the team returns to the office, much of the context is already gone, and the follow-up either goes out generic or never goes out at all.

The problem is rarely a shortage of contacts. It is a shortage of context: who talked about which product, what pain was mentioned, what deadline came up, who asked for material, and who simply walked past the booth. A well-configured agent works precisely on the link between what was recorded and what the salesperson needs to do next.

Step by step: setting up the workflow

1. Map the sources. List the real inputs: business cards, photos, voice notes with annotations, QR code scanner exports and phone notes. Define who feeds each one.
2. Define the record. Create a fixed template with fields such as name, company, role, stated interest, date and place of contact, suggested next action, and source. Fields without information stay empty, never filled by guesswork.
3. Limit access. The agent reads only the material from the event in question, with no access to the whole commercial database. Remove what does not serve the purpose, such as personal documents that appear in photos.
4. Generate drafts. For each contact, the agent proposes a message that cites what was actually noted and suggests a channel.
5. Review. The salesperson approves, edits or discards each draft. Record who approved it and when.
6. Send and log. Only after approval does the message go out through the permitted channel, and the result returns to the CRM.

Handling personal data under the LGPD

A business card contains personal data. Before any automation, the company needs to know why it collected that data and whether using it for later commercial contact fits that purpose. This is not a technical detail. It is the first decision of the project.

In practice, separate three things: what the person explicitly authorized, what was exchanged in a business context, and what should not be used at all. Sensitive information, such as health details mentioned in conversation, should not enter the commercial record. When the agent influences a decision about a person, such as prioritizing or excluding a contact, the company must be able to explain the criteria and allow human review. For the right to review, read LGPD automated decision-making and the right to review under Article 20.

You also need to map where the data is processed. If the tool sends information to services outside Brazil, that must be documented before the pilot, as explained in international data transfer under the LGPD when using AI providers. For official guidance, the ANPD, Brazil's data protection authority, is the reference.

How the agent should write the follow-up

The most common mistake is asking the model for a personalized message without supplying what was said. The result sounds warm but empty, or worse, invents an interest that never existed. The rule is simple: a draft may only use facts present in the record, with the source available for the salesperson to check.

A good message has three short parts: a reference to the meeting, one useful piece of information tied to what the person mentioned, and a simple question to continue the conversation. It should not include promises about deadlines or prices, because those depend on information the agent does not have. Anthropic's guide on building effective agents recommends starting with simple workflows and keeping human checkpoints in place, which fits this approach.

When the contact came through WhatsApp, observe the customer service window and the rules for outbound messages. See the 24 hour window on WhatsApp API and how an AI agent should act and WhatsApp opt-in and how to collect permission for outbound messages.

Only messages the salesperson has approved are sent.

Illustrative scenario: an industrial equipment fair

Illustrative scenario: an industrial equipment company takes two salespeople to a multi-day trade fair. At the end of each day, they send the agent photos of the business cards and a voice note with their notes. The agent returns a spreadsheet with the fields filled in, marks as pending the contacts without a role or stated interest, and proposes a draft for each one. The salesperson discards two drafts that sounded generic, adjusts another because the conversation had been about a specific deadline, and approves the rest. The person who asked for material receives the message the next day, and the person who only walked past the booth does not enter any active sequence. At the end, the team reviews what worked and what needs to change in the record template.

None of this depends on a specific tool. It depends on clear rules and on someone responsible for the review.

How to apply the method in a pilot

In how we work with clients, the starting point is mapping the current process: how contacts arrive, how long they take to become action, and who decides. Then we limit data and actions, so the agent reads only the fair material and never sends anything on its own. The pilot runs at a real event, with a log of every draft, approval and send, and with a metric defined before it starts, such as the share of contacts that receive a follow-up within the window the team set. At the end, we hand over the workflow and decide whether to expand, adjust or stop. The NIST AI Risk Management Framework helps organize this risk control from the start.

This sequence keeps a company from contracting a large automation before knowing whether the sales process can support the change.

Mistakes that most often hurt results

• Treating the trade show list as a marketing database without checking the purpose of collection.
• Letting the agent send messages without human approval.
• Accepting fields filled by assumption, without a source mark.
• Measuring only message volume instead of the quality of the conversations that follow.
• Keeping the fair material accessible indefinitely, instead of setting a retention period and access limits.

Frequently asked questions

Can I use a business card photo as the basis for the agent? Yes, as long as the purpose of contact is clear and the processing respects the LGPD. The photo should be treated as source material, with restricted access and a retention period set by the company. The agent should extract only the fields needed for the commercial record.

Do I need consent to message someone after the fair? It depends on the context and the channel, and the answer should not be improvised. On WhatsApp, outbound messages follow their own rules for permission and templates, so check them before building the sequence. When the question is legal, the company's compliance lead should validate it.

Can the agent decide which contacts to prioritize? It can suggest an order based on explicit criteria, such as stated interest and urgency mentioned. The final decision and the exclusion of contacts should stay with the salesperson, with the criteria recorded so they can be reviewed.

How long does it take to put this into production? It depends on event volume, data sources and CRM maturity. The safest path is to start with one event, measure the quality of the drafts, and only then decide what to expand. If you already have a CRM update workflow, you can build on the structure described here, adapting the fields to your process.

Sources

• ANPD: Autoridade Nacional de Proteção de Dados
• NIST AI Risk Management Framework
• Anthropic: Building effective agents