The setup was simple enough. A model with access to someone’s calendar and task list, and the goal of being useful without being asked. Notice things. Offer things.

I got stuck on how to organize the suggestions. There seemed to be two obvious ways to slice it.

By what the suggestion does: Some are just information. Some hand off into another app. Some group nearby errands together.

By what the suggestion is about: Health and wellness. Family and relationships. Errands and chores. Travel and commute.

I started with the second one, because it felt cleaner and it maps to how people describe their own lives. It took a while to notice it was the wrong foundation.

Why themes fell apart for me

Themes describe the subject of a suggestion. They turn out to say nothing about whether the suggestion is any good.

Here are two things that both sit squarely in “health and wellness”:

• You haven’t scheduled any exercise this week.

• Your physical therapy appointment is at 4pm, and you need to leave by 3:20 because of traffic.

The first is a lecture. The second saves your afternoon. Same theme, wildly different value, and the theme label gives you no way to tell them apart.

What made this concrete for me was watching what the structure encouraged. When themes are the backbone, the natural thing to build is coverage: one decent suggestion per theme per day. And once you’re optimizing for coverage, most of what you generate is the first kind, because the first kind is always available. You can always tell someone they haven’t exercised.

There’s a softer version of this too. Nobody wakes up wanting a family-and-relationships suggestion. They want to not forget their kid’s recital. Organizing around the category put me at one remove from the thing people actually care about.

Why action types weren’t much better

I tried flipping to action types next. Better, but it has a different hole in it.

Action types describe the output, not the reason. “Deep link the user into a shopping app” is a mechanism. It says nothing about when doing that is a good idea, and when-is-this-a-good-idea turned out to be the entire problem I was actually trying to solve.

So I’d swapped a taxonomy that ignored value for one that ignored timing.

What actually worked: stop looking for one taxonomy

The unlock, eventually, was realizing these aren’t competing answers to one question. They’re answers to three different questions, and I’d been mashing them together.

Layer 1: the trigger. Why are we speaking at all?

This is the layer I’d skipped entirely, and it turned out to be the foundation.

A trigger is a specific pattern in the data that justifies opening your mouth. Not a topic. A pattern. The list I ended up with looks something like:

- Two events overlap

- Not enough time to travel between two locations

- A deadline approaching with no time set aside for it

- A recurring task that’s overdue

- A free gap big enough to fit a pending task

- A task whose location is near somewhere you’re already going

- Five meetings back to back with no break

- A meeting that just ended and implied a follow-up

The thing that surprised me is that this list is finite. I can write it down. For calendar and task data it seems to land somewhere between fifteen and twenty-five patterns, and I got a rough version down in an afternoon.

That property is what makes everything else possible. A list you can write down is a list you can test, count, rank, and delete from. The version I’d been building before, where you hand the model the calendar and ask it to suggest something helpful, has no list. So there was nothing to test, and nothing to remove when something felt off. I could only tune the prompt and hope.

Layer 2: the action shape. What can they do about it?

Once a trigger fires, there’s a separate question about what you’re offering. This list roughly sorts itself by value:

1. One tap and it’s done. The assistant completes it.

2. Change my schedule. It rewrites the calendar or task list for you.

3. Take me there. Hands off into an app or service, ideally carrying context so you don’t start over.

4. Bundle several things. Groups related items into one offer. Three errands near one appointment.

5. Just telling you. Nothing to do. Pure information.

That last one is where I think most of these products quietly die. It’s the cheapest kind to generate and the least valuable to receive, and there’s no natural brake on it. I’ve come to treat a hard cap on it as non-negotiable, though I don’t have a principled way to pick the number yet.

Layer 3: the packaging. How does it show up?

This is where grouping by distance actually belongs, which was the other thing I’d been placing wrong. Proximity grouping isn’t a category of suggestion. It’s an optimization that runs after you’ve already decided a set of errands is worth surfacing.

Same story for time-of-day grouping, or collapsing several small things into one card. Presentation decisions, made last, on whatever survived.

Themes came back, just somewhere else

I didn’t end up throwing themes out. They just moved off the critical path. Four jobs where they earn their keep:

Limiting repetition: Without them there’s no way to notice you’ve fired four health nudges today.

Keeping variety: If five suggestions are on screen, themes stop all five being about work.

Consent: This is the strongest reason to keep them, and the one I underrated. Health isn’t errands. Family isn’t commute. People should be able to say yes to one and no to another, and that setting needs a label to attach to.

Finding blind spots. Breaking results down by theme is how you notice what you’re failing at completely.

So: settings screen and metrics dashboard, not ranking logic. Which in hindsight is a fine place for a taxonomy to live. I’d just given it a promotion it couldn’t handle.

What I found when I looked at how others did it

Once I had the layered version, I went back through what shipped systems actually do, and the shape is pretty consistent:

something changes in the data

↓

detect which triggers fired

↓

add context (location, travel time, history)

↓

drop anything ineligible (too soon, too often, muted, low confidence)

↓

rank what’s left

↓

bundle and package

↓

pick the moment and the channel

↓

watch what happens, learn

The detail I keep coming back to: The decision to speak is made by rules. The words are written by the model.

Models seem to be unreliable at judging when to interject, the ability doesn’t track model size, and prompting for more proactivity barely moves it. [ProactiveBench], [a multimodal proactiveness study].

Four things models are genuinely very good at:

• Reading messy human text.

• Deciding which service or tool could actually fulfill something.

• Writing the one sentence a person reads.

• Checking tone, and catching the suggestions that are technically correct but socially wrong.

That last one I’d underestimated. There’s a whole class of suggestion that is factually right and still lands badly, and rules are hopeless at spotting it.

The rule I’d keep above all the others

Every suggestion has to be able to show its work.

I ended up putting this in the data structure rather than the prompt, so a suggestion can’t exist without a pointer back to the calendar entry or task that produced it.

The difference reads like this:

• You might want to pick up your prescription.

• Your pharmacy is a two-minute detour from your 3:30 dentist appointment, and “refill prescription” has been on your list since Monday.

The second is checkable. If it’s wrong, you can see exactly where it went wrong, and it stays trustworthy even when it misses. The first is a guess wearing a confident face, and a few bad ones seem to poison the whole feature.

Timing, which I got wrong twice

Every suggestion has a window where it’s useful, and it took me two passes to start carrying that window in the data.

The buckets I settled on:

• Right now: Under fifteen minutes. Earns the right to interrupt.

• Today: Belongs on an ambient surface, not in your ear.

• This week: Belongs in a morning summary.

• Standing: Never urgent. Belongs somewhere browsable.

The window is what picks the channel. And the same correct piece of information, delivered as a voice interrupt instead of a quiet card, is the difference between helpful and irritating. Nothing about the content changes.

The related thing I keep relearning: when confidence is low, ask instead of acting. A guess phrased as a question is charming. The same guess executed as an action feels like a betrayal.

If I were starting over

I’d spend the first two weeks on a spreadsheet instead of the integration.

1. Write down the trigger list. Fifteen to twenty-five patterns, plain English, no code.

2. Collect a couple hundred real calendar and task snapshots.

3. Label them by hand. Which triggers fired, and would I have actually wanted to hear about it?

No infrastructure required. What comes out of it is the knowledge that five or six triggers produce nearly all of the value, which settles the taxonomy question with evidence rather than argument, and tells you where not to spend the quarter.

I did this in the wrong order and built twenty triggers of roughly equal polish before I had any idea which ones were carrying the thing.

What I’m watching

Acceptance rate and time-to-action are the obvious numbers, and they’re also the flattering ones.

The honest number is how often people turn it off. Mutes and disables. A good acceptance rate alongside a rising mute rate is a system that’s failing, and the acceptance rate will tell you the opposite right up until the end.

The other one worth tracking is which triggers get dismissed repeatedly. That’s the delete list, and mine was shorter and more obvious than I expected.

Still plenty I don’t have answers for. How to set the interruption budget without just guessing. Whether the trigger list stays finite once you add email and messages, or whether that’s where the whole approach starts to strain. How much of the ranking can be learned rather than hand-tuned.

But the trigger list was the thing I should have started with, and I didn’t.

But after spending a significant amount of time doing this, one thought keeps crossing my mind: Was I healthier when I didn’t know this much? Is my body’s ecosystem already optimized? Am I overdoing it—and possibly making things worse?

A similar question comes up when I think about the health metrics provided by fitness trackers.

When I first started using one, it felt like a novelty. Over time, though, it slowly became a reason for me to run, work out, cycle, or swim. That behavior can be interpreted in two very different ways: motivational or obsessive. And depending on the context, I can convincingly argue either side.

This tension became even more real when I started building in the same space - integrating health metrics with Alexa to make smart homes even smarter. The product manager in me immediately asked: How do I segment users for a product like this when the user in me isn’t fully convinced of its utility?

I’ve come across countless articles, research papers, and personal testimonials with strong, and often opposing views. Usually, I can predict where a friend will land on a debate like this. But this one is harder. The trade-offs are nuanced, and the stakes so personal.

I wanted to understand it better.

Motivated by that curiosity, this piece covers learnings from cutting through the noise around health and fitness trackers.

Who benefits the most from them? Who should avoid them altogether? And in which scenarios should people be especially cautious about their use?

Positive effects of wearables

Wearables deliver measurable improvements in physical activity, metabolic health, sleep quality, and cardiac monitoring when matched to the right population and paired with behavioral support. A study published in The Lancet shows evidence base spanning over 160 randomized controlled trials and multiple systematic reviews consistently showing that wearables increase daily movement, support weight management, and improve disease-specific outcomes such as glycemic control in diabetes.

Types of wearables, their positive effects, and best customer segments

1. Activity trackers and fitness bands (Fitbit, Garmin, Xiaomi, Mi Band): The Lancet study found that users wearing activity trackers averaged 1,800 additional steps per day compared to control groups, equivalent to approximately 40 extra minutes of walking daily. The same review documented an increase of 6 minutes per day of moderate-to-vigorous physical activity, with effects most pronounced with goal-setting, coaching, or feedback mechanisms.
   Overweight and obese users experience substantially larger benefits. A 2024 systematic umbrella review in NPJ found that health intervention through tech in this population achieved step increases of 2,185 per day - more than double the 1,007 steps seen in general populations.
   
   Best Customer Segment: Activity trackers are most effective for individuals who start with a significant deviation from baseline health or activity norms. Overweight and obese adults, as well as highly inactive individuals, experience the greatest gains because the feedback provided by trackers represents a meaningful signal, not marginal noise.
   For these users, step counts, streaks, and visual progress indicators translate into large, observable changes in daily movement and modest but tangible weight loss. The gap between their starting point and recommended activity levels makes the data salient and motivating, reinforcing behavior change.
   The larger the initial deviation from the mean, the more actionable and rewarding the feedback becomes.
   

2. Smartwatches with coaching programs (Apple Watch, Samsung Galaxy Watch, Garmin): Smartwatches paired with structured coaching dramatically amplify adherence and long-term behavior change. A 2025 study found that participants recently diagnosed with Type 2 Diabetes using smartwatch-enhanced coaching programs were 10 times more likely to begin regular exercise compared to control groups. At six months, they remained seven times more likely to stay active, and at one year, three times more likely to sustain activity levels. 52% percent of smartwatch-coached users met recommended physical activity guidelines versus only 17% in the control group.
   
   Best customer segment: Smartwatches excel for general adults seeking habit formation through integrated coaching, reminders, and adaptive goals. They are particularly valuable for busy professionals who want an all-in-one device that combines health tracking with communication and productivity features. The coaching element is critical - without it, smartwatches function as expensive activity trackers with marginal additional benefit.
   

3. Smart rings (Oura Ring, WHOOP 4.0, Samsung Galaxy Ring): Smart rings offer discrete 24/7 monitoring with a focus on sleep, recovery, and readiness metrics. A 12‑month randomized controlled trial using a biometric smart ring plus guided feedback showed large, clinically relevant gains: sleep onset latency dropped from 25 to 14 minutes at three months; VO2max improved by 11.8% at three months; body fat fell by ~3.8 percentage points; HRV increased by ~40%; and daily steps rose by ~60%. Most improvements persisted even after coaching stopped, with 99%+ device adherence over a year.

   

   Best-fit segments are: (1) adults seeking sustained lifestyle change (sleep, fitness, body composition); (2) competitive and serious recreational athletes optimizing training load and recovery; and (3) people with insomnia or sleep-architecture concerns wanting non-pharmacologic improvement;

Negative effects: studies and worst-affected customer segments

While wearables offer clear benefits for many populations, they can worsen motivation, trigger anxiety, and reinforce disordered behaviors in specific groups. The harms are most pronounced among adolescents, individuals with eating disorders or body image concerns, cardiac patients prone to health anxiety, and users who interpret data through a perfectionistic or threat-focused lens.

1. Adolescents in school physical education programs (ages 13–14): A randomized controlled trial by University of Birmingham examined the motivational impact of Fitbit activity trackers on 84 adolescents aged 13–14 years enrolled in mandatory physical education classes over eight weeks. Rather than increasing intrinsic motivation for physical activity, the intervention produced significant declines across multiple dimensions of self-determination theory. Autonomous motivation - the internal drive to be active for its own sake—decreased by 0.22 points. More troubling, amotivation - a state of disengagement and lack of any motivation to be active - increased by 0.35 points, a large effect size . Perceived competence, and autonomy significantly decreased, indicating that the wearable undermined students' confidence, sense of control, and social connection around physical activity.
   
   The mechanisms driving these negative effects centered on non-personalized step targets, peer comparison, and external pressure. Adolescents reported feeling judged when they failed to meet arbitrary targets. Instead of fostering joy, the Fitbit became a source of guilt, replacing intrinsic enjoyment with judgement. The study highlights that wearables in youth settings can backfire when they emphasize competition and standardized goals rather than personal progress and autonomy-supportive coaching.
   
   Wearables that impose external benchmarks and make performance visible to others can undermine the very motivation they are intended to build, particularly in teens who are not already intrinsically motivated to exercise.
   

2. Young adult women with body image vulnerability (ages 17–35): Multiple studies in university and community samples of young women document associations between fitness and diet app use and higher levels of disordered eating, body dissatisfaction, and compulsive exercise. Women who use diet tracking apps show significantly more eating pathology than those who use fitness-only apps or no tracking apps. Those who combine diet and fitness tracking exhibit the highest scores on validated measures of disordered eating, drive for thinness, and obligatory exercise.​
   
   When fitness becomes about optimizing numbers and achieving aesthetic ideals rather than enjoyment, health, or performance, the psychological costs can outweigh any physical benefits. "Fit-ideal internalization" acts as a moderator: women who strongly endorse cultural ideals of leanness are most vulnerable to the negative effects of tracking.
   
   This finding is particularly concerning given the marketing of fitness wearables and apps to young women through appearance-focused messaging (e.g., "get bikini ready," "tone and sculpt"). The data suggest that these tools can nudge vulnerable users toward disordered eating and exercise behaviors

3. Older adults with limited technology literacy and physical impairments: Usability studies of wearables in older adults (65+) document frustration and disengagement related to four primary factors: inaccurate or confusing data displays, difficult setup instructions, device malfunctions, and physical discomfort during wear. Older users also report challenges with small touchscreens, gesture-based interactions, and lack of clear explanations for complex metrics such as HRV.
   
   Age-related changes in fine motor control, vision, and cognition make interactions that are trivial for younger users—such as navigating nested menus or interpreting trend graphs—genuinely difficult for many older adults. When wearables are mandated by healthcare providers or family members without adequate training and ongoing support, they can become sources of stress, incompetence feelings, and technology abandonment.

What This Means for You

The research reveals a clear pattern: wearables amplify existing trajectories.

If you’re significantly below baseline health (overweight, sedentary, uncontrolled diabetes), the feedback is signal - loud, clear, actionable. The ~2,000 extra steps for obese users aren’t marginal, they’re meaningful.

If you’re already health-optimized, the same device produces noise. Those 1,800 steps matter less when you’re starting from 8,000. The juice may not be worth the squeeze.

Three decision rules:

1. High benefit, low risk: Overweight/obese adults, newly diagnosed Type 2 diabetics, serious athletes, people with sleep disorders—all paired with coaching or clinical oversight

2. High risk, uncertain benefit: Adolescents 13-16, individuals with eating disorder history, young women with body image concerns, patients with health anxiety, older adults with limited tech literacy

3. Everyone else: Marginal gains require active engagement. If you ignore the data, don’t wear the device. If you obsess over the data, don’t wear the device. The sweet spot is narrow.

The question isn’t whether wearables work. It’s whether they work for you, right now, with the support structures needed to make them helpful rather than harmful.

1. Introduction: Why Voice AI Is Becoming Workflow Infrastructure

2. Revenue Capture & Transaction Intake

   1. Case Studies: Wendy’s, Apollo Hospitals (Apollo 24|7)

3. Service Delivery & Customer Support

   1. Case Studies: Meesho, Smartness

4. Revenue Recovery & Accounts Receivable

   1. Case Study: Southwest Recovery Services

5. Trust, Risk & Compliance

   1. Case Studies: HSBC UK, Barclays

6. Conclusion

Introduction: Why Voice AI Is Becoming Workflow Infrastructure

My fascination with voice AI deepened during my summer internship at Amazon, where I worked at the intersection of Alexa voice AI and the smart home ecosystem. Since then, I’ve contributed to projects spanning multiple industries focusing on applying voice AI to real business workflows - not just to save time and effort, but to meaningfully scale operations and elevate the quality of customer experience.

While the technology itself has scaled rapidly and advanced significantly, I believe its real value and true validation comes from how effectively it is applied within business workflows. This has been the core problem statement I’ve been working on over the past few months, and it made me curious to draw inspiration from how other industries are leveraging similar capabilities. This piece is a synthesis of those learnings and observations.

Voice AI has moved well beyond “press 1 for billing.” As per Grand View Research, the global voice AI generators market is growing at a CAGR of 29.6% and expected to reach USD 21.8 Bn by 2030. Understanding the trends in such a hyper-growth sector is crucial if you are in the industry and want to stay abreast with what’s happening around.

I’ve structured this article around the different stages in the lifecycle of business workflows that are most disrupted by voice AI, along with case studies for each stage to illustrate how these applications have materialized in the real world.

1. Revenue capture & transaction intake

2. Service delivery & customer support

3. Revenue recovery & accounts receivable

4. Trust, risk & compliance

Revenue capture & transaction intake

Across industries, leaders are increasingly treating voice as a revenue and transaction channel, not just a support cost - because a meaningful share of high-intent demand still shows up as real-time conversations (phone calls, voice kiosks, in-car voice, voice assistants). The workflow they’re optimizing is essentially:

Demand → capture → structure the order/reservation → confirm → route to fulfillment (POS/KDS/reservations system) → upsell where appropriate

A few common mental models show up across deployments:

• Front-door transaction layer (not just an IVR): Voice AI isn’t there to deflect calls; it’s there to complete the transaction and push clean data into POS/KDS/reservation systems (often with a human fallback).

• Throughput engine for peak hours: The goal is to remove the ordering bottleneck - reduce service time, increase order completion, and keep staff focused on prep and handoff. It can also be deployed during after-hours to capture any missed opportunities.

• Guardrails + integration > pure conversation: The best results usually come when the system is tightly constrained around offerings/menu/reservation rules and integrated with POS/reservations - rather than being a totally open-ended chatbot.

Case studies:

Wendy’s

Wendy’s scaled its FreshAI, a voice AI system that automates drive-thru ordering, (alongside digital menu boards) as a top-of-funnel transaction intake play - automating order taking, improving the in-restaurant operating rhythm, and using AI-driven suggestions to lift check size. The rollout is positioned as part of a broader customer experience strategy focused on personalization, convenience, and hospitality. The core impact:

• Target rollout to 500+ restaurants by end of 2025 (FreshAI + digital menu boards).

• Already deployed to 160+ U.S. restaurants in Q1 2025

• Broader digital performance cited alongside this strategy: digital mix reached over 20% of total sales in Q1; conversion rate hit an all-time high (not solely attributed to FreshAI).

• Improved order accuracy and efficiency as staff can focus on speed of service and correct order delivery

What did Wendy’s implement in the customer journey:

1. They inserted FreshAI at the earliest conversion choke point - the drive-thru speaker: FreshAI is deployed as the order-taker at the drive-thru, automating the conversation customers normally have with a crew member. The intent is to make ordering faster and more consistent while keeping humans focused on execution and hospitality.

2. The experience is designed to handle real drive-thru ordering complexity (not a rigid script): Wendy’s explicitly frames FreshAI as solving problems traditional rule-based chatbots struggle with - casual conversation, slang, and heavy customization - by using generative AI that adapts in real time instead of following a narrow decision tree.

3. Orders flow into restaurant systems (POS + hardware), with guardrails: In the Google Cloud partnership announcement, Wendy’s described the system as being powered by Google’s foundational LLMs using Wendy’s menu data, business rules / conversation guardrails, and integration with restaurant hardware and the POS - so the outcome of the conversation becomes an executable order, not just text.

4. “Human-in-the-loop” is built into the workflow: Wendy’s positions FreshAI as an assistant that empowers the crew, not a replacement. Success is measured in “orders submitted without human intervention,” but the operating model expects crew to stay available for exceptions and to focus on speed-of-service and getting the order out accurately.

5. They paired voice automation with adjacent accuracy tooling downstream: Alongside the drive-thru voice AI, Wendy’s is rolling out tools like menu item label printers (to ensure customizations are executed correctly) and smart delivery scales (to confirm all items are included), reinforcing the perfect order promise after the order is captured.

[Source: 1]

Apollo Hospitals (via Apollo 24|7)

Apollo 24|7 built an omnichannel booking and engagement stack where patients can discover services and complete bookings inside familiar channels - most notably WhatsApp, and (for appointment handling) voice integrated into their customer care workflows. The core impact:

• +49% increase in bookings using WhatsApp Flows vs their usual chatbot booking approach

• +72% increase in average revenue per order using WhatsApp Flows vs usual chatbot booking approach

• 20% decrease in calls to the customer care center after integrating Voice API (with quicker query resolution)

How Apollo 24|7 is using voice + conversational automation to improve top-of-funnel and booking:

1. They redesigned booking as a native in-chat transaction (WhatsApp Flows): Apollo 24|7 moved diagnostic test booking into a single, end-to-end flow that lives entirely inside WhatsApp. Instead of sending users across links and forms, the conversation itself becomes the booking interface: patients can select the relevant test, request home collection or a kit where applicable, pick a date and time, and apply a coupon - all without leaving the chat.

   This reduces friction at the exact moment of conversion. The less context-switching a patient has to do, the more likely they are to complete the booking (and the more likely they are to add higher-value services).

2. They used acquisition + re-engagement channels that drop users directly into the booking flow: To drive bookings, Apollo 24|7 shortened the path from “interest” to “confirmed appointment” by pushing traffic straight into WhatsApp. Their campaign combined click-to-WhatsApp ads (to acquire or retarget), offline QR codes (to convert physical touchpoints like leaflets and pharmacy posters), and WhatsApp messages (to re-engage existing users).

   This is essentially lowest-friction conversion: marketing → WhatsApp → Flow → booking, with fewer steps where users typically drop off.

3. Where “voice” comes in - recovering drop-offs and routing patients: On the appointment side, the setup also uses voice as a support channel - routing patients to the right consultant and calling back users who drop off or get disconnected mid-booking. That’s a simple but high-leverage conversion recovery loop in healthcare, where abandoned booking attempts are often high intent.

[Sources: 1, 2]

Service delivery & customer support

As we all have seen and experienced, in service delivery, voice AI isn’t a nice-to-have anymore - it’s becoming the front door to resolution for customers who still default to calling when something is urgent, confusing, or high-stakes. The north star is shifting from deflect calls to resolve intent fast, safely, and empathetically, while keeping a seamless human handoff for edge cases. That means leaders increasingly evaluate voice AI on whether it can operate reliably under real constraints like peak spikes, noisy audio, multilingual callers, regulated policy language, safety routing, etc.

Common mental models show up across voice AI deployments in service delivery & customer support:

• Resolution engine (not just deflection): The goal isn’t to make calls go away, it’s to solve the customer’s problem end-to-end—answer, authenticate if needed, take the action, and only escalate when the issue truly requires a human.

• Triage + routing layer (get to the right place fast): Voice AI needs to act as the smart front desk - identify the right intent, gather the minimum necessary context, and route the customer to the right queue or specialist. The win is fewer transfers, less repetition, and lower customer effort.

• Surge absorber for peak demand: In many industries, call volume is spiky (outages, travel disruption, payment failures). Voice AI is treated as elastic capacity - handle the surge, keep wait times sane, and protect CSAT - while human agents focus on the hardest cases.

• Guardrails + system integration > free-form chat: The best deployments are tightly grounded in approved policies, knowledge sources, and backend systems (CRM, billing, scheduling). They prioritize reliable actions and compliant answers over open-ended conversation - especially in regulated or high-stakes contexts.

Case studies:

Meesho - India (in collaboration with ElevenLabs)

Meesho used ElevenLabs voice AI tools to deploy a real-time AI voice agent to automate high-volume customer support calls - especially “where is my order?” style queries like delivery status, delays, cancellations, and refunds - in Hindi and English, with a focus on keeping the experience empathetic and human-like at massive scale. The core impact:

• Handles 60,000+ customer calls per day

• Reported 95% resolution rate, significantly reducing the need for human intervention/escalations

• ~50% improvement in average handle time (AHT)

• +10% higher CSAT after deploying the voice bot experience

• ~75% lower per-call cost vs human-operated calls

What did Meesho implement?

1. A real-time voice agent at the highest-volume support entry point (voice calls): Meesho deployed a voicebot that answers customer calls and handles the most common “where is my order?” flows - so customers get instant support in natural speech without waiting for an agent.

2. Multilingual support: The voice agent supports both Hindi and English from day one, enabling Meesho to serve a broad base of users in India in the languages they naturally speak on calls.

3. Engineered for real Indian calling conditions (noise, interruptions, basic devices): The bot is described as handling peak seasonal demand reliably and includes interruption intelligence to distinguish affirmations (ji/okay/yes) from true interruptions - keeping conversations fluid instead of stopping mid-sentence. It’s also framed as robust for noisy environments and accessible to users on basic smartphones.

4. Designed the workflow to maximize containment while reducing handle time: By focusing on high-frequency intents and resolving many queries without human intervention, the implementation is explicitly positioned around reducing AHT and freeing human agents for complex cases.

[Sources: 1, 2]

Smartness (in collaboration with ElevenLabs)

Smartness (Italian SaaS serving hospitality operators globally) built two internal voice agents using ElevenLabs Conversational AI: a Support Agent that captures customer inquiries during off-hours and turns them into prioritized Zendesk tickets, and an AI SDR Agent that engages qualified leads and schedules product demos - so the company can offer more continuous service while reducing time spent on low-priority sales outreach. The core impact:

• Continuous customer support coverage during off-hours by capturing inquiries and converting them into prioritized Zendesk tickets

• Sales team time reclaimed: the AI SDR agent reduces time human SDRs spend on lower-priority accounts by handling initial outreach and booking demos

Implementation:

1. Inserted an AI Support Agent at the after-hours support choke point: it collects customer requests when the team is offline and automatically creates prioritized tickets in Zendesk.

2. They operationalized the handoff by auto-creating prioritized Zendesk tickets: Instead of just collecting a voicemail or transcript, the agent turns the interaction into a structured Zendesk ticket that’s already prioritized - so the support queue starts the next day with clearer triage.

3. They used a helpdesk integration pattern that makes the agent ticket-native: ElevenLabs’ Zendesk integration is designed to let voice agents create tickets and use historical ticket data/knowledge base to resolve issues faster. The Smartness story explicitly confirms ticket creation + prioritization; the broader Zendesk integration capabilities show the typical way this is implemented technically.

4. Added an AI SDR Agent at the top-of-funnel qualification step: it behaves like a virtual sales rep - engages qualified leads and schedules product demonstrations, reducing manual SDR effort on low-priority accounts.

[Sources: 1]

Revenue recovery & accounts receivable

Voice AI can be leveraged wherever revenue is leaking because a human couldn’t get to the call: unanswered inbound traffic, after-hours calls, low-connectivity outbound campaigns, or agents spending time on repeatable verification + payment flows. The winning pattern is not to replace collectors - it’s to expand coverage and tighten execution: automate the repeatable steps (reach, verify, negotiate within guardrails, take payment, document outcomes), and escalate exceptions to humans fast.

Common mental models show up across deployments:

• Coverage > headcount: use voice agents to answer every inbound call (including after-hours) and run large outbound campaigns without staffing spikes.

• RPC/PTP as the funnel: collections teams treat Right Party Contact (RPC) and Promise-to-Pay (PTP) like conversion metrics, and voice AI is optimized against those.

• Compliance-first conversation design: constrain what the agent can say/do (disclosures, permitted offers, escalation rules), then scale.

• Payment as the end-state: best systems don’t chat - they drive to an executable outcome (on-call payment, payment portal handoff, or a documented arrangement).

Case study:

Southwest Recovery Services

Southwest Recovery Services adopted Skit.ai’s Voice AI to automate a portion of inbound debt-collection calls - especially to handle call volume they couldn’t cover and to answer outside business hours - while maintaining a live-agent transfer path when needed. The core impact:

• 10× ROI reported within a few weeks of going live.

• 50% Right-Party Contact (RPC) rate and 10% Promise-to-Pay (PTP) rate reported

• Enabled 24/7 assistance for consumers and improved agent productivity (agents focus on transfers + exceptions)

Implementation:

• Inserted voice AI at the inbound front door: Skit.ai answers inbound consumer calls first, then transfers to a live agent when requested.

• Automated core collections intents with guardrails: the deployment targets common inbound intents (questions, next steps), then routes to RPC / PTP flows and escalation paths.

• Enabled end-to-end resolution options: the solution is described as supporting on-call payment processing and negotiation capabilities, pushing conversations toward closure rather than deflection.

• Expanded beyond inbound once stable: the CEO quote also references automation across outbound calls (scale + cost efficiency) once the model proved out.

[Source: 1]

Trust, risk & compliance

Trust and compliance is where voice AI stops being convenient and starts being critical. When real money, identity, safety, or sensitive data is on the line, the voice channel becomes a prime target for fraud, and at the same time, customers have near-zero patience for clunky security rituals. The north star here is shifting from more authentication steps to smarter, lower-friction verification. In practice, leaders evaluate voice AI in this workflow on whether it can hold up under hard constraints - spoofing and deepfakes, regulatory auditability, bias across accents and languages, privacy and consent requirements - all while keeping the experience fast enough that customers don’t abandon the interaction.

Common mental models show up across deployments:

• Silent security layer (verify while you talk): Instead of interrogating customers with knowledge-based questions, teams use passive voice biometrics to confirm identity during normal conversation - reducing friction while improving security.

• Risk engine + step-up: Voice becomes one signal in a broader risk model. If risk is low, let the user proceed quickly; if risk is high (new device, unusual behavior, suspected spoof), trigger step-up authentication or route to a specialist.

• Deepfake era readiness (liveness + multi-layer defense): As synthetic voice improves, leaders increasingly assume voice can be spoofed and plan for layered defenses (liveness detection, multi-signal scoring, and continuous monitoring).

Case Studies:

HSBC UK

HSBC UK embedded Voice ID (voice biometrics) directly into its telephone banking authentication flow, so callers can be verified by matching their live voice to a stored voiceprint, instead of relying primarily on knowledge-based security questions that fraudsters can sometimes steal or guess. The bank positions Voice ID as a practical way to make phone banking both easier for legitimate customers and harder for impersonation scams, and it pairs the “verify the customer” use case with a second line of defense. The core impact (as of 2021):

• ~£249M prevented from being lost to telephone fraudsters

• Attempted telephone banking fraud down 50% year-on-year.

• 43,000+ fraudulent phone calls identified since the technology was introduced in the UK

• £981M+ of customers’ money protected since introduction

• Voice ID was used by 2.8M+ active customers, and HSBC said it was enrolling ~14,000 customers per week at the time

Implementation:

1. Inserted Voice ID at the authentication choke point: during a normal phone banking call, the system checks whether the caller’s voice matches the voiceprint held on file, allowing faster, lower-friction verification than repeated Q&A security steps.

2. Added a fraudster-voice cross-check capability: beyond verifying legitimate customers, HSBC describes building a library of fraudsters’ voice prints to cross-check new incoming calls, turning Voice ID into both an authentication tool and a fraud detection layer.

3. Positioned it as part of a broader channel strategy: the announcement situates Voice ID inside a wider push to manage high contact volumes (e.g., new voice response system handling 450k+ calls/week) while steering general queries to chat - suggesting HSBC was thinking about risk + operational load together, not in isolation.

[Source: 1]

Other examples: Barclays Customer Service Solutions

Conclusion

I’ve come to believe the real value of voice AI is validated only when it is embedded into workflows in a way that drives measurable outcomes. That has been the core problem statement I’ve been working on over the past few months and it’s what made me curious to study how other teams are deploying voice AI in the wild. This article is a synthesis of those observations.

Across the case studies, one pattern keeps repeating: the best deployments don’t treat voice as a standalone bot. They treat it as a workflow primitive. That means voice AI is judged less by how human it sounds, and more by whether it can reliably do three things:

1. Capture structured intent from messy, real-world speech,

2. Push that intent into systems of record (POS/CRM/helpdesk/payment rails), and

3. Operate under constraints -peak demand, multilingual accents, regulatory policy, fraud risk - while handing off gracefully when needed.

If there’s one meta-takeaway, it’s this: voice AI creates value when it changes the shape of a workflow. It either (a) removes a bottleneck at the moment of intent, (b) increases resolution capacity during spikes, (c) drives recovery where humans can’t reach at scale, or (d) embeds security without breaking experience.

I believe that the next wave of differentiation won’t come from voice sounding more human - it will come from voice being more operational: deeper integrations, tighter guardrails, clearer measurement, and faster iteration loops. The teams that win will be the ones that treat voice AI not as a channel experiment, but as workflow infrastructure.