AI Chatbot with SMS Follow-Up That Keeps the Conversation Going

What happens after a visitor closes the chat window

A website chat is a short, fragile conversation. The visitor asks about availability, gets a helpful answer, types “let me check with my wife,” and closes the tab. If the only record is a transcript sitting in a dashboard, that lead is effectively gone. Either nobody reaches back out, or someone does days later with a cold call the visitor no longer connects to your website.

SMS follow-up changes the ending. Before the visitor leaves, the chatbot offers to continue by text: “Want me to text you these details so you have them when you’re ready?” If they agree, the conversation moves to their phone, where people tend to read messages promptly and can answer whenever it suits them. The same AI that answered on the website keeps answering in the text thread, and a person on your team can step into that thread at any point.

The follow-up exists to finish something specific that started on the site: a booking the visitor didn’t complete, a question they wanted to think over, a document they asked for, a quote they were waiting on. Every text should point back to that unfinished thing. When it’s done, or when the customer says they are no longer interested, the follow-up ends.

What separates a working setup from an irritating one comes down to three questions: did the person agree to receive texts, does the thread know what was said on the website, and does the automation step aside when a human or the customer wants it to?

Asking for the phone number the right way

Texting a customer who typed a number into a chat box is not the same as texting a customer who agreed to be texted. The chatbot has to collect both the number and a clear permission, and your records have to be able to prove it later.

How the ask is worded

The permission request appears as its own step, never buried inside a longer form. It names your business, says what kind of messages will follow (for instance, answers about this request and scheduling updates), notes that message and data rates may apply, and explains that replying STOP ends the texts. Your attorney approves the exact language. We build the flow so that approved wording is shown verbatim to every visitor; the model is not allowed to paraphrase it.

What gets stored

• The phone number, normalized to one standard format so duplicate contacts are caught.
• The timestamp, the page the visitor was on, and the version of the consent wording they saw.
• The scope of what they agreed to. Permission to follow up on a quote request is not permission to join a monthly promotions list, and the system keeps those as separate flags.

When the visitor says no

Declining must be easy and must not cost the visitor anything. The chatbot offers email instead, or simply wraps up the conversation politely. A bot that keeps pressing for a phone number after a refusal spends goodwill your business worked hard to earn.

Carrying the website conversation into the text thread

The weakest SMS follow-up tools send something generic: “Thanks for visiting! Reply with any questions.” The visitor can’t tell which conversation this refers to, and the AI on the text side knows nothing about what was already discussed. A properly built system passes a structured handoff record from the web chat to the SMS side.

That record usually holds:

• A short summary of what the visitor wanted, drafted by the model and checked against the transcript.
• The facts they supplied: name, service needed, location, preferred timing, and anything else your intake requires.
• What is still open, such as a missing address, an unanswered question or an unconfirmed slot.
• The reason for the follow-up and which sequence should run.

With that record in hand, the first text can be specific and immediately useful:

The business and customer above are illustrative. The pattern is what matters: the text refers to the visitor’s actual request, offers a concrete next step and carries the opt-out instruction. Notice too that the AI holds a time rather than promising one, because in this example the business wants a coordinator to confirm. Whether the AI may book outright or only propose is a policy decision we make with you.

A follow-up sequence with built-in exits

The sequence is where most text follow-up goes wrong. A typical drip campaign keeps firing on schedule no matter what the customer said. Ours are designed from triggers and exit conditions first, and timing second.

A few rules apply across every sequence. Texts go out only inside permitted hours in the customer’s own time zone, inferred from the area code and service address and easy for staff to correct. No sequence runs longer than your team agrees to, and we usually recommend fewer messages than owners first propose. Any reply at all, even “not now,” pauses the automation so the next message answers the person rather than following a timer. And as soon as a staff member writes in the thread, every scheduled message for that contact is cancelled.

Scheduled messages are editable templates written with your team and filled from the handoff record. The AI composes freely only when replying to something the customer wrote, and then only from the knowledge and policies it has been given.

What the AI does when someone texts back

Text replies are shorter and messier than website chat. People write “ok,” “how much,” “wrong number,” “can u do sat,” or send a photo of the problem. The reply handler sorts each incoming message before deciding what to do with it, in this order:

1. Opt-out words first. STOP, UNSUBSCRIBE, CANCEL, END and QUIT are caught before the message ever reaches the model. The contact is suppressed immediately, one confirmation goes back, and no further automated messages are sent to that number. Plain-language requests such as “please stop texting me” are treated the same way even though they aren’t the exact keyword.
2. Wrong numbers and confusion. “Who is this?” gets a direct answer naming your business and the website request. “Wrong number” suppresses the contact and flags the record so staff can check whether the visitor mistyped their digits.
3. Answerable questions. Hours, service areas, what’s included, how the process works. The AI answers from your approved knowledge base, the same one the website chatbot uses, so the two channels never contradict each other.
4. Simple actions. Picking one of the offered slots, confirming an address, asking for a call. The AI completes these through the connected calendar or CRM when your rules permit, and otherwise queues them for staff approval.
5. Anything sensitive or unclear. Complaints, price negotiations, billing disputes, messages the AI can’t classify confidently, and photos that need a professional’s judgment go to a person with the full thread attached.

Pricing deserves its own mention. The AI repeats only pricing information you have written down and approved for customers. It will not estimate a figure from a photo, offer a discount, or accept a price the customer proposes. Those messages reach a human every time.

Which number sends the texts, and who sees the thread

Choosing the sending number

Customers are more likely to trust a number they recognize. There are usually two workable routes. Many existing business landlines and VoIP numbers can be text-enabled, so messages come from the number already printed on your vehicles and invoices while calls keep ringing as they do today. Alternatively, a new local or toll-free number can be dedicated to messaging. Either way, US carriers expect business senders to register their use case, through A2P 10DLC for local numbers or toll-free verification for toll-free ones, and unregistered traffic is frequently filtered. We prepare that registration with you and hold off on live sending until it is approved.

Giving staff a real inbox

Where possible, text threads appear in a place your team already works, such as the conversation view in your CRM; otherwise we set up a simple shared web inbox. Each thread opens with the website chat summary, marks the AI’s messages as automated, and offers a clear control to take over. Once someone takes over, the AI stops replying in that thread until it is handed back. Staff can also leave internal notes the customer never sees.

Alerting the right person

Pinging the whole team for every “thanks” teaches people to ignore alerts. Routing rules decide who gets notified: a new booking goes to the scheduler, a complaint to a manager, and a warm lead asking for a call to whoever covers sales that day.

How text follow-up turns into spam

We are regularly asked to repair follow-up setups that customers complained about or carriers began filtering. The causes repeat:

• Treating a typed number as consent. A phone field on a chat form with no clear permission language is both a legal exposure and a deliverability problem.
• Long automated sequences. Daily texts for weeks to someone who never answered reads as harassment and drives opt-outs and spam reports.
• Blindness to human activity. The automation sends “Just checking in!” an hour after the owner personally closed the deal on the phone.
• Suspicious-looking links. Generic URL shorteners are a common trigger for carrier filtering, so we link through your own domain.
• Invented urgency. A model left to its own devices may write “only one slot left” when that isn’t true. Ours are instructed never to manufacture scarcity, and scheduled templates are fixed text.
• Two systems texting one person. A marketing platform and the chatbot both message the same lead because neither knows about the other. During discovery we look for existing senders and either connect them or retire one.

Plenty of off-the-shelf chat widgets list SMS as a feature, and some handle simple cases well. Where they tend to fall short is the list above: consent records tied to a wording version, exits that react to staff activity, one knowledge base shared across channels, and routing that matches how your team actually divides the work. A custom build needs more planning up front, and it is only worth it when follow-up is a real part of how you win business. If your main problem is enquiries that never got answered at all, an AI chatbot for missed leads may be a better first project.

What to measure, and how EVOTECH builds it

Numbers worth tracking

• Opt-in rate in chat: of the visitors offered text follow-up, how many accept. A low figure usually points at the wording or the moment of the ask.
• Reply rate by sequence: which follow-up reasons start real conversations and which get ignored.
• Opt-out rate by message: a template that prompts STOP more often than the others gets rewritten or dropped.
• Outcome per thread: booked, quote accepted, handed to staff, declined or went quiet, recorded in your CRM so follow-up can be tied to revenue.
• Takeover rate and reasons: which questions the AI passes to people, so the knowledge base can close the most frequent gaps.

Our build process

1. A free phone or video consultation. Together we trace what happens to a website lead today, from your chat and forms through your phone setup and CRM.
2. Written, fixed-scope quote. It lists the sequences, integrations, number setup and takeover rules we will deliver, so what’s included is never ambiguous.
3. Consent and registration. Your attorney signs off on the consent wording, and we submit carrier registration and wait for approval before live sending.
4. Build and connect. Website chat, SMS gateway, knowledge base, CRM or calendar, and the staff inbox are wired together, with the handoff record flowing between them.
5. Test on real phones. Every sequence runs end to end, including STOP, wrong numbers, time zones, staff takeover and a lead who books halfway through a sequence.
6. Launch and review. After launch we read threads with you, adjust templates and tighten routing based on what customers actually write.

EVOTECH IT LLC has more than 20 years behind it and a 5.0 Google rating; the people doing the work are US-based, near Houston, and serve clients nationwide by remote delivery. If the chatbot also needs to write into your pipeline, see our AI integration services, or start with the broader AI chatbot development overview.

Related services

Frequently asked questions