TCPA Exposure in Automated Outreach: What an AI System Must Log to Be Defensible

If an AI system calls or texts your customers, a TCPA claim comes down to what your logs can prove. They need to show who consented, to what, when, and what the system knew when it sent each message. If your records can't rebuild a single contact, you are arguing from memory. The plaintiff is holding a screenshot.

That gap matters more every year. TCPAWorld's year-end count found 2,788 TCPA cases filed in 2024, up 67% from 2023. In November 2024, 95.5% of those filings were class actions. Plaintiffs' firms are building volume practices around automated outreach. They look for companies whose records won't hold up.

This is a checklist, not legal advice. Your counsel decides which rules apply to your business. Our job is to make sure your system can produce the evidence when they ask for it.

Why does AI change the exposure?

A traditional dialer or texting platform follows a script. Each message is a template, the send list is fixed, and the timing is scheduled. You can rebuild what happened from the campaign settings.

An AI system makes decisions. It writes each message on the spot. It decides when to follow up and reads customer replies to choose what to do next. Each of those decisions creates a new kind of exposure.

• ▸The content changes from message to message. An appointment reminder that adds "ask about our whitening package" now reads as marketing. Your consent records may not cover marketing.
• ▸Opt-outs come in many forms. Customers rarely type STOP. They write "please stop texting me," "wrong number," or "not interested, take me off." A keyword filter misses most of these. The AI may treat them as conversation and reply.
• ▸The system decides timing. A follow-up agent can choose to send a message at a time your rules never approved.

If the system speaks rather than texts, the same applies. The FCC confirmed on February 8, 2024 that AI-generated voices are "artificial" voices under the TCPA.

Damages under the TCPA are counted per call or message. Every one of those decisions is therefore its own potential violation. Volume works against you. The statute, 47 U.S.C. 227(b)(3), sets $500 per violation, and a court may treble that for willful or knowing violations. The FCC's compiled TCPA rules add the calling-hour and consent requirements the system also has to enforce.

What is the exposure worth, in plain arithmetic?

Here is an illustrative example. It is not drawn from any real company. Plug in your own numbers.

Exposure = messages per month × miss rate × months × D

Here D is the damages amount per message, which your counsel can tell you.

Say you send 20,000 automated texts a month. Suppose 0.5% go to people who had already opted out in words your filter didn't catch. That is 20,000 × 0.005 = 100 messages a month. Over 12 months, that is 1,200 messages, so your exposure is 1,200 × D.

The miss rate is the number to watch. You won't know it unless your logs record every inbound reply, how the system classified it, and what it sent next. Without those records, you can't measure the rate or show that it was low.

What must the system log for every contact?

This is the core of the post. Each row is a question opposing counsel will ask. The middle column is the record that answers it. The last column is where most systems fall short.

Rows 2 and 9 are the ones most teams skip. They are also the ones that let you rebuild a decision months later. Without row 2, you know only what the CRM says today, not what the system saw that day. Without row 9, you can't explain why the system acted as it did. The logic may have changed four times since.

Where should the compliance check sit?

Logging shows what happened. A gate stops the bad message before it is sent. You need both, and the order matters.

The right design puts a deterministic check between the AI and the outside world. The AI drafts the message. A separate rule layer then checks it: consent scope, purpose label, suppression status, local time, and opt-out history. Only then does the message go out. Every pass or block is logged with the reason. The inbound version of that gate, who is allowed to say what, enforced by the system, has the same shape.

Don't ask the AI to police itself. A line in the prompt such as "never text people who opted out" is a request, not a control. Models follow those instructions most of the time, and "most of the time" multiplied by D is still a large number. A rule outside the model either blocks the send or it doesn't. Its decision can be logged and replayed.

What does the opt-out path need to look like?

This is where AI systems fail most visibly, so it gets its own section.

Every inbound message should be classified before the conversation continues. One of the categories must be "revocation or possible revocation". Unclear messages, such as "who is this" or "wrong number", should stop automated sending and route to a human agent. Treating them as an invitation to chat is a mistake. The log should record what the customer wrote, how it was classified, the confidence level, and the suppression timestamp.

Then test it. Write a set of 200 to 500 realistic opt-out phrasings, including typos, slang, and other languages if you serve bilingual customers. Run them through the classifier before launch and after every prompt or model change. The miss rate you measure is the miss rate in the formula above. It is far better to learn it from your own test set than from a complaint.

FAQ

Q: Does using AI change which TCPA rules apply to us? Your counsel answers that question, not your engineers. What AI changes is how much you can prove. Generated content, flexible timing, and free-text replies all create decisions a template system never made. Each one needs a record.

Q: Isn't a STOP keyword filter enough? No. It catches customers who use the keyword and misses those who write in plain language. You need a classifier that reads intent, sends unclear replies to a person, and is tested against a written set of opt-out phrasings. Each result should be logged.

Q: How long should we keep these logs? Set the retention period with your counsel, based on the limitation periods that apply to you. Expect years, not months. The more urgent check is whether your vendor's default retention is shorter than that. Many dashboards keep detailed records for only weeks. If yours does, start exporting to storage you control this quarter. Twilio documents a default of up to 13 months (400 days) for messages and media, after which media is permanently deleted.

This week: run the reconstruction drill

Pick five contacts at random from 60 days ago, including at least one that ended in an opt-out. For each one, try to produce all ten records above within an hour, without asking an engineer to search raw server logs.

Expect to fail this drill the first time. That isn't a scandal. It is how pilots become production: the demo proved the AI could hold a conversation, and nobody scoped the evidence trail. This is part of why AI proofs of concept fail in production. Legal reviews the system after launch, and no one can answer its questions.

If you can produce all ten records for all five contacts, you are in good shape. If you can't, the gaps tell you exactly what to build first. Rows 2, 3, and 9 usually come first. Adding logging to a running system only covers the future, so design the evidence layer before volume grows, not after the first demand letter.