The AI Communications Market Report

01 · Executive summary

Seven findings for busy readers

• Speed is the cheapest conversion lever most teams have not pulled. The two studies that measured it put qualifying odds around twenty one times higher at five minutes than at thirty, and roughly seven times higher inside the first hour than an hour later. Both were funded by companies selling software for the problem, which is worth knowing and does not make them wrong.
• Voice quality stopped being the differentiator. Every serious platform now delivers sub second, natural, interruptible conversation. Choosing by demo call is choosing by the part that no longer varies.
• Every deployment runs a six layer stack. The conversation, the runtime, telephony, orchestration, integrations, and operations. The demo shows one layer; the other five decide the outcome and are owned by someone, on purpose or by accident.
• The market sorts into five categories, not fifty vendors. Voice runtimes, no code builders, enterprise conversation platforms, point solutions, and managed layers. Naming your category first shrinks the shortlist to two or three real candidates.
• Most of the statistics in this category do not survive checking. The famous SMS and follow up numbers trace back to a retracted whitepaper and a 1942 survey of forty people. Section three shows the working, because a business case built on those numbers collapses the moment somebody checks.
• The operational layer, not the model, is where pilots die. A widely cited MIT study found ninety five percent of generative AI pilots show no measurable financial return within six months. The consistent cause: nobody owns tuning, monitoring, and upkeep after launch.
• Compliance is a market entry ticket, not paperwork. The United States prices consent mistakes per call, Europe prices them as a share of revenue, and India decides whether your calls connect at all. The rules are learnable, and section nine maps them.

02 · The problem

The revenue leak, in numbers

Most teams do not lose deals in the negotiation. They lose them in the gaps: the form filled on Tuesday that gets a reply on Friday, the missed call that never gets a callback, the warm prospect whose follow up never went out. Each gap looks small. Together they are usually the largest untapped line in the funnel.

Two things drive this, and neither is fixed by hiring another person. The first is that buyer interest decays on the order of minutes. Someone filling in a form is comparing vendors at that moment, and every study above measures the same effect from a different angle: the gap between reaching them now and reaching them later is not incremental.

The second is that demand ignores office hours. A buyer researching at 9pm on a Sunday gets whoever answers, and voicemail loses to the next name on the list. Neither problem is a motivation problem. Both are scheduling problems, which is a much easier thing to solve.

03 · The evidence

The numbers this market repeats

Before we go further, a caution about the statistics in this category, including some we have quoted ourselves in the past. While preparing this report we traced the most repeated figures in AI sales and messaging back to their primary sources. Several do not survive the trip.

This is not pedantry. If you are building a business case on a number that turns out to be a 1942 survey of forty people, the case falls over in the meeting where someone checks.

The pattern behind these is worth understanding, because it will keep producing new ones. A vendor publishes a figure that flatters its product. A blog quotes the vendor. A second blog quotes the blog and rounds the number up. Within a few years the original is unreachable and the figure has become something everybody knows. Speed to lead research is the clearest case: the two studies below are real and worth reading, but both were funded by companies selling software to fix the problem they measured, and the older one is now nineteen years old.

None of this means speed and follow up do not matter. Our own operating experience says they matter enormously. It means the honest version of the argument rests on mechanism rather than on a decorated statistic, and a vendor who cannot tell you where a number came from is quoting someone who could not either.

04 · The technology

What changed in AI voice, and why the demo is the wrong test

You could always tell. There was a beat before every reply, the delivery was flat, and anything off script produced the same polite non-answer. That has mostly stopped being true. Agents now answer inside a second, cope with being cut off halfway through a word, and stay with a buyer who switches language or doubles back on something said a minute ago.

That creates an awkward problem for anyone running an evaluation. If every credible vendor sounds good, the demo has stopped being a test. It shows you what the category can do. It tells you very little about what this particular vendor will do for you in a year. The questions that still separate them are the ones a demo never reaches:

• Nobody picks up. Does it recognise voicemail, leave something worth returning, text straight away, and try again at an hour when the person might actually be free?
• The gap between touches. Do SMS, WhatsApp and email share one memory with the call, or are they four tools that happen to be aimed at the same person?
• The record. Does the CRM update itself while the conversation is happening, or does it wait for notes that nobody writes?
• Month six. Is somebody reading transcripts and adjusting, or has the agent been running unattended since launch while your offer moved on without it?

Most of what follows is really about those four questions.

05 · The framework

The operating stack: six layers, whoever runs them

Underneath the branding, every production deployment runs the same six layers. Which makes the useful evaluation question quite a narrow one: which of these does the vendor run, and which quietly become your job?

Three of them cause most of the trouble. Telephony breaks without telling you: nothing throws an error, connect rates just sag as your numbers age and carriers start treating them differently. Orchestration is the biggest piece of invisible engineering on the list, and the one teams consistently underestimate, because queues and retry logic and shared memory are unglamorous work that only shows up when it is missing. Then there is operations, which is not a task at all. Prompts drift as your offer changes, models update underneath you, and the rules move in whichever markets you call.

Whoever owns those three owns the result. Every category in the next section is a different answer to that one question, dressed up in different marketing. There is a longer version of this framework in the production voice AI stack.

06 · The market map

Five categories, not fifty vendors

The landscape looks crowded until you sort it by operating model, at which point it collapses into five groups. Each one is a different split of those six layers between vendor and customer. Work out which group you are actually shopping in and the shortlist tends to pick itself.

How to use the map: pick the category that matches who will own layers three through six of the stack. Only then compare vendors inside it. Cross category comparisons are ownership decisions wearing a feature list.

07 · The players

Who is who, and what each one leaves you

The names a shortlist usually contains, sorted by the categories above. Every row describes a published product model, not a verdict: these are all real products with real deployments, and the last column is the honest question a demo will not answer.

Two layers sit underneath all of them and are worth naming, because they are often mistaken for competitors: telephony and messaging carriers such as Twilio, and realtime media infrastructure such as LiveKit. Almost every product above rides on one or both. If a vendor cannot tell you which carrier and which media layer they run on, that is a useful question to keep asking.

08 · The positioning map

Where each player sits

Two axes, both structural. Across: who operates the system once it is live, from a platform you run yourself to a service run for you. Up: how much of the motion the product covers, from the phone call alone to voice and messaging on one shared memory.

How to read it

The top right is not a winner’s circle. It is the highest service, lowest control corner of the market, and plenty of teams belong in the bottom left instead. A company building a voice product should be suspicious of anything that takes the runtime out of its hands; a ten person sales team with no engineers should be equally suspicious of a toolkit. The map tells you which trade you are making, not which vendor is better.

• Toolkits, bottom left. Voice runtimes and in house builds. The product is the call, and every layer around it is yours. Cheapest per minute, highest total ownership.
• Assembled stacks, top left. Multi channel motions that exist because a team built the orchestration. Powerful and specific to you, and the maintenance never ends.
• Managed voice, bottom right. Enterprise platforms and point solutions where someone else runs a scoped job well. Strong inside the scope, quiet outside it.
• Managed motions, top right. Operators who run the whole stack across channels. You buy an outcome and give up direct control of the components.

Three honest caveats. Vendors move: several runtimes are adding managed tiers, and several enterprise platforms are adding outbound. Placement is by primary product model, so a vendor with a services arm sits where the bulk of its customers actually land. And a dot near an axis is a judgment call about emphasis, not a measurement.

09 · The playbook

What converts: speed, then sequence

Across industries, the deployments that pay back share two mechanics. First, instant response: the agent calls or replies within a minute of the trigger, while intent is at its peak. Second, a sequenced motion: the call is one touch in a coordinated cadence across text, voice, and email, all sharing one memory of the buyer.

Section three is the reason this one is thinner on statistics than you might expect. The famous channel numbers, the ninety eight percent SMS open rate and the four times pickup lift from a pre call text, did not survive checking, so we are not going to repeat them here to make the case look stronger. What follows rests on mechanism and on what we see running these systems.

The mechanism is straightforward. A buyer who has been told a call is coming answers a number they would otherwise ignore, because the call arrives with context instead of as an interruption. The nearest rigorous evidence for that effect comes from survey research rather than sales: a randomised trial found pre notification texts raised response rates by roughly sixty percent. Real, useful, and a long way short of the four times a vendor blog claimed.

A reference cadence

Adapt days and channels to your market; in India, WhatsApp typically takes the lead role.

• One memory across every touch. The call references the text; the email references the voicemail. Disconnected tools reintroduce the leak the system exists to close.
• Every touch lands in the CRM as it happens, with outcome, objections, and next step, so a human can take over any conversation with the full thread in front of them.
• Judgment escalates to people. The system carries volume; pricing exceptions, relationships, and negotiations route to a human with context attached.

10 · Industry lenses

The same leak, four different disguises

The mechanics travel, but they show up wearing local clothes, and the version you recognise is usually the one from your own industry. Four from markets where the conversation is most of the job:

If your industry is not listed, the transfer test is simple: find where interest decays fastest in your funnel, and put the always on layer exactly there first.

11 · The decision

Build or buy: compare capacity, then ownership

Most build or buy cases get argued on cost per seat against cost per licence, which compares two things that do not produce the same output. Capacity is the more useful axis: what each option lets you do in a day, and how quickly you can turn it up. Then the second question, which is who ends up operating the stack.

The teams getting the strongest returns are not replacing people. They split the work: software carries first pass volume and follow up; people keep the conversations where judgment moves the number. Then the build or buy question reduces to four questions about the stack.

12 · Compliance snapshot

The rules that decide whether your calls connect

Every market runs on the same skeleton: get permission, respect the do not call lists, call at humane hours, say who you are, and stop when asked. The details differ, and the details are what regulators fine. Treat this as a working map rather than legal advice, and have counsel sign off before your first campaign.

• Disclosure is now table stakes for AI calls in major markets: the agent says what it is at the start, every time.
• Consent has a shape: written, naming the business, tied to the number, never a condition of purchase. A clear checkbox does it; a bought list never does.
• Rules move. Whoever owns your operations layer owns tracking them, per market, as part of the standing job.

The market by market detail, including primary sources, is in our voice AI compliance guide.

13 · Getting started

The ninety day blueprint

Build it, buy a platform, or hire an operator: the deployments that pay back tend to move in the same order anyway. Find the leak. Build for that one leak. Put it in front of real volume while somebody is still watching. Then keep tuning it, which is the part most teams stop doing around week five.

The mapping phase is the one people want to skip, and it is the one that decides whether the rest means anything. Without a baseline from your own funnel you end up comparing results to a vendor’s slide, which is not a comparison. Start wherever interest decays fastest, usually speed to lead, because that is where the difference between doing nothing and doing something is most visible. And whoever ends up running the system, ask how often it gets tuned. That cadence predicts month six better than anything in the original build.

14 · About the authors

Where Kaigen Labs sits in the map we drew

We should say plainly where we sit. Kaigen Labs is a category five vendor, a managed layer. We build on the runtimes from category one, pick between them per deployment, and then run the rest of the stack for mid market revenue teams: voice, SMS, WhatsApp and email as one motion, CRM write back included, monitored and tuned every week, with most pilots taking calls inside two to three weeks.

None of this report is written from the outside. The stack in section five is our own architecture map, the cadence in section nine is a motion we actually run, and the four questions in section eleven are the ones we tell prospects to put to us.

• If you are building: our operator guides compare the runtimes honestly, including Retell AI vs Vapi, Retell AI vs Bland AI, and the full production stack.
• If you are evaluating the managed path: we publish side by side comparisons against every major alternative, half discs and all.
• If you want the map drawn on your funnel: a twenty minute call with the Kaigen team. You bring how you sell; we sketch the layers for your case and say honestly which ones you could run yourself.

15 · Method and sources

How this report was put together

If a report is going to place other companies on a chart, it owes you its method. Here is ours, including the things we chose not to do.

• None of the statistics are our own research. Every figure is a widely cited industry finding, and each one carries its attribution in the sentence or the chart source line. Where a number is directionally useful but loosely sourced, we say so rather than dressing it up.
• The positioning map plots product model, not performance. The axes are who operates the system and how much of the motion the product covers, both readable from what each vendor publishes about itself. Nobody was scored, surveyed, ranked, or charged to appear, and no quadrant is labelled as the good one.
• We are on the map. Kaigen Labs publishes this report and sits in the managed motions quadrant, marked as ours in the legend. Read our placement with the scepticism you would apply to any vendor mapping itself.
• We publish no market size figure on purpose. Published estimates for this category differ by an order of magnitude depending on whether they count contact centre software, speech APIs, or agent platforms. A number that swings that far with its definition tells a buyer nothing useful.
• Vendor descriptions come from two places: what each company publishes about its own product, and our experience running several of these platforms in production. Where we have not operated something, we describe the category rather than inventing detail.
• Corrections are welcome. If you work at a company on this map and we have described your product wrongly, tell us and we will fix it in the next edition.

16 · References

Where to check everything above

Every external claim in this report, with somewhere to verify it. Where a source has a commercial interest in its own finding, or a limitation worth knowing, that is noted rather than buried.

1. 1Speed to lead: 21x qualifying odds at five minutes versus thirty, and 100x contact oddsInsideSales.com with Dr James Oldroyd, Lead Response Management Study, 2007. Vendor funded, never peer reviewed.content.marketingsherpa.com
2. 2Response time distribution and the 7x and 60x qualifying findingsOldroyd, McElheran and Elkington, The Short Life of Online Sales Leads, Harvard Business Review, March 2011hbr.org
3. 395% of generative AI pilots showing no measurable profit and loss impactMIT Media Lab Project NANDA, The GenAI Divide: State of AI in Business 2025, July 2025. Preliminary, not peer reviewed.mlq.ai
4. 4Enterprise AI projects abandoned after proof of conceptGartner press release, 30% of generative AI projects will be abandoned after proof of concept, July 2024gartner.com
5. 5Companies struggling to scale value from AIBoston Consulting Group, Where’s the Value in AI?, October 2024bcg.com
6. 6The SMS open rate claim, and its retraction by the original researchersMobilesquared, SMS engagement dropping but still outperforms other channels, August 2022mobilesquared.co.uk
7. 7The origin of the five to twelve touches claim: a 1942 survey of fewer than forty peopleSales and Marketing Executives International, on the provenance of widely quoted sales statisticssmei.org
8. 8Pre notification messages and response rates, the nearest peer reviewed evidenceBull et al., Pre-Survey Text Messages Improve Participation Rate in an Australian Mobile Telephone Survey, PLOS One, 2016journals.plos.org
9. 9WhatsApp reach in IndiaDataReportal with Kepios, Digital 2025: Indiadatareportal.com
10. 10Email open rate benchmarks, and the effect of Apple Mail Privacy ProtectionMailchimp email marketing benchmarks, and Litmus research on Apple Mail Privacy Protectionmailchimp.com
11. 11AI generated voices count as artificial voices under the TCPAFederal Communications Commission, Declaratory Ruling FCC 24-17, CG Docket 23-362, February 2024docs.fcc.gov
12. 12TCPA statutory damages of $500 per violation, trebled for willful violations47 U.S. Code section 227law.cornell.edu
13. 13The one to one consent rule vacated, a separate action from the AI voice rulingInsurance Marketing Coalition v. FCC, Eleventh Circuit, January 2025media.ca11.uscourts.gov
14. 14GDPR fine tiers: up to 4% or 20 million euro at the higher tier, 2% or 10 million at the lowerRegulation (EU) 2016/679, Article 83gdpr-info.eu
15. 15UK marketing call rules and the Telephone Preference ServiceInformation Commissioner’s Office guidance on direct marketing and PECRico.org.uk
16. 16India commercial communication rules, the 140 and 1600 series, and the do not disturb registryTelecom Regulatory Authority of India, TCCCPR 2018 and subsequent directionstrai.gov.in
17. 17Canadian rules: CASL covers electronic messages, while voice calls fall under the National Do Not Call ListCASL section 6(8), and CRTC Unsolicited Telecommunications Ruleslaws-lois.justice.gc.ca
18. 18Australian telemarketing calling hoursTelecommunications (Telemarketing and Research Calls) Industry Standard 2017, ACMAdonotcall.gov.au
19. 19WhatsApp 24 hour customer service window, template approval, and opt inMeta, WhatsApp Business Platform documentation and messaging policydevelopers.facebook.com
20. 20Market size estimates spanning roughly five times in the same year, by definitionCompiled from published estimates by MarketsandMarkets, Precedence Research, and IMARC Groupmarketsandmarkets.com

The AI Communications Market Report, 2026 edition. Written by the Kaigen team, verified in August 2026. Guidance, not legal or financial advice. You are welcome to share it in full.