Bet on automation to turn one-way marketing texts into conversations

Demand was already there. Clients were churning to competitors over conversational gaps, and leadership moved two-way texting up the roadmap. At the same time, we found 20 clients using an unpromoted keyword-autoresponse feature. It wasn’t self-serve: each one needed technical support to set up. That clients pushed through that friction at all was the signal.

A survey of our clients showed what they wanted from a conversation: help a shopper find the right product and buy. It also showed the catch. A few had a customer-service team on call, but most had no one to answer texts in real time.

Competitors hadn’t cracked it either. Email platforms were moving into texting fast but had no real conversations yet. Texting competitors either offered a few keyword replies or, like Tone, staffed human agents. Automated tools were simple or robust, never both.

The core question: how do clients hold conversations when they don’t have anyone to hold them? I led workshops with product, engineering, customer success and marketing. One debate took over: automate replies, or staff real people to text subscribers for our clients. One of our founders was interested in a premium human concierge tier.

The debate ate a whole meeting. Regrouping with my PM, I proposed we ground the debate in trade-offs, so we scored them. We weighted speed highest, since churn was rising and competitors were closing in.

Before going to leadership, we pressure-tested the winner with a quick confidence, impact and risk check. It held: concierge might earn more per conversation, but not enough to justify a slower launch and more unknowns. We framed it for leadership as automation first, not automation forever. The best automated use cases would become the baseline to test human agents against later.

Automation raised the next question. As I sketched, my ideas kept splitting on how guided setup should be. Full control assumed nothing about what clients wanted to test. Strong guidance prevented errors and kept texts easy for subscribers to answer.

Guidance, not control. I tested prototypes with 5 customers: 4 from our core persona of non-technical marketers, and 1 power user. Only the power user wanted full control. The other 4 wanted guardrails and a place to start.

I think you all have a better idea on best practices and standards for these texts, so I’d want to be helped along the way, maybe like tips or templates.

— Marketing manager, prototype testing

Scoping to the fastest safe MVP. I ran an affinity mapping session with my PM and engineers to turn research into scope. I framed it around one question: what gets us to launch fastest with the least harm to the subscriber’s experience? Writing the message came first. What happens when subscribers reply unexpectedly became smart defaults, watched by opt-out rate.

Guided replies, not a blank field. I designed setup as mad-libs: write a question, add numbered replies, and watch a live phone preview. It was a little less flexible than one free-form box, but it kept texts formatted like top performers. It also turned replies into clean data we could pattern-match, built to support suggesting better questions by industry and goal. Every tester rated ease of use 4 or 5 out of 5.

Planning for unexpected replies. Customers trusted us to handle replies we didn’t expect, but they were anxious about opt-outs. So I put that load on the system, not the marketer:

• A reply we don’t recognize gets one “try again” text, then exits to the brand’s Help message.
• No marketing texts go out while a customer-service conversation is open.
• Opt-out rate became the beta’s guardrail metric.

Settling a debate with a test. My PM wanted the 2-day no-reply exit drawn into every journey, worried clients wouldn’t know what happens when a subscriber goes quiet. Rather than debate it, I added it to the prototype test. Most customers found it behind the settings icon on their own, so it stayed there. If beta clients got confused, we’d surface it.

What testing changed.

• Users weren’t sure whether they’d already set up their replies. Now, when a step saves, the map shifts the replies into view, with a one-time tooltip.

The beta hit its adoption goal and its guardrail, and missed its biggest bet.

The win came from a different direction. Texts that asked subscribers to reply, instead of clicking through to a site, did just as well: no drop in revenue and no rise in opt-outs. And every reply was free first-party data. A home goods brand, for example, now knows which subscribers knit.

One client built segments from those replies and retargeted them. Compared with messages that didn’t use reply data, clickthrough rose 23% and revenue rose 6%.

Next steps. I left Attentive soon after the beta launched. Here’s how I’d have phased it:

1. Widen the search: test more use cases with a small set of clients to find one that hits 2x.
2. Confirm the win scales: roll out the top use cases as templates, including the reply-for-data pattern.
3. Test against humans: run the best automated use case against human agents, and settle whether a concierge tier is worth building.

Reflecting, with AI in the picture. This worked because it was constrained. Guided replies meant clean, predictable data, which is why automation could beat people in the first place. Freeform AI replies would bring back the unpredictability we designed out. Today I’d let AI propose journeys, with the client as editor and final checkpoint before anything reaches a subscriber.