🎙️ Key Insight from this Week’s “AI to Go” Podcast

Chris Carter, Chairman & CEO of Approyo, Founder of Mugatu AI - the man who built one of the first SAP HANA clouds

Christopher Carter has been embedded in the SAP ecosystem since the 1980s. So when I asked him what AI founders should be embedding into right now, his answer surprised me. He didn’t name a model lab. He didn’t name a cloud. He named the ecosystem - and told me, point blank, that the founders he meets at AI events are getting this exact decision wrong.

“I take slivers of where I see gaps in ecosystems - and that’s what I focus on.”

That’s the part that stuck with me as a buyer. From inside a F500, the question I’m answering on every AI vendor evaluation is: where does the data sit, who owns the integration, and whose ecosystem governance does this fall under? Founders who haven’t picked a clear ecosystem stance can’t answer that. And that is usually where their pilot dies.

“If you don’t know how the enterprise works and how they integrate - you’re going to be left behind.”

🧠 The Buyer’s Brief

How Enterprise buyers actually evaluate your ecosystem choice - across compute, model, middleware, and app layers

Chris built Approyo, Inc. by going deep on a single ecosystem when everyone else went broad. SAP - nothing else. He turned down >90% of the enterprise software market on purpose: “nobody can be an expert at your solutions except for you.” That bet held for 13 years because SAP held - per SAP’s corporate fact sheet, 77% of the world’s transaction revenue touches an SAP system. AI founders face the same decision in 2026. The stack now has four layers, each with distinct ecosystem dynamics: compute (chips, data centers - inference capacity allocation, sovereign placement, portability), model (frontier and open-weight LLMs - swap risk, sovereignty, unit economics), middleware (orchestration, evals, agent frameworks - lock-in, observability), and application (SAP Joule, Salesforce Agentforce, Microsoft Copilot - integration depth, audit trail, residency).

Most AI founders pick a model-layer ecosystem and assume the app layer comes free. It does not. The DSAG Investment Report 2026 found only 3% of SAP customers run SAP Business AI in production, while 77% of AI-active enterprises use non-SAP tools like Microsoft Copilot. Even the deepest enterprise app ecosystem is losing AI workloads to the layer above.

That’s a real opportunity for founders who pick well.

The 5 buyer-side criteria for ecosystem evaluation

Data gravity. Where do the workflows you want to touch already live? Pick the ecosystem with the gravity.

Integration surface. Documented, governed APIs (or emerging standards like MCP and A2A)? API politics will slow your pilot.

Compliance posture. SOC 2, ISO 27001, EU AI Act readiness, sovereign cloud options? Inheriting compliance is faster than building solo.

Co-sell economics. Does the ecosystem actively bring you into deals (Sapphire, Dreamforce, AWS Marketplace)? If you’re building in Enterprise, co-sell motions are key.

Replacement velocity. How fast can the ecosystem itself be disrupted? The hardest one - and why the next section matters.

7 Triggers that can break the Ecosystem you bet on

Chris’s bet held because SAP held. In AI, the substrate moves under your feet. Here are seven triggers that signal an Ecosystem (layer) may shift in 2026:

#1 New model architectures. In March 2026, Yann LeCun left Meta to found AMI Labs with a $1B seed at a $3.5B valuation, betting on world models as a successor to LLMs. If world models work, the model layer reshuffles. Founders on transformer-only assumptions need a contingency thesis.

#2 Unit-economics collapse in the model layer. Per DeepSeek’s API documentation, R1 prices are roughly 96% cheaper than comparable OpenAI o-series rates. The moat is the system, not the model. If your ecosystem can’t accommodate model swap, your gross margins erode while your buyer’s procurement team runs the math.

#3 Cybersecurity capability shocks. In April 2026, Anthropic announced Claude Mythos Preview alongside Project Glasswing. Per Anthropic, Mythos has identified thousands of zero-day vulnerabilities in every major operating system and web browser. The UK AI Security Institute confirmed Mythos can autonomously execute multi-stage attacks on vulnerable enterprise networks. If your ecosystem can’t show a credible response, your pilot stalls in CISO review.

#4 Compute supply constraints. Dominion Energy disclosed 40.2 GW of data center power requests by February 2025, up from 21.4 GW seven months earlier (per S&P Global). Sightline Climate finds ~50% of global 2026 data center projects face delays from grid and equipment shortages. The Uptime Institute names power as the single defining constraint on data center growth. Founders depending on a single provider sit downstream of that provider’s allocation queue.

#5 Hyperscaler / frontier lab interdependency. In March 2026, Microsoft and OpenAI restructured their partnership; Microsoft’s license is now non-exclusive. In April 2026, Anthropic and Amazon expanded their collaboration for up to 5 GW of new compute. On April 24, Google committed up to $40B to Anthropic with 5 GW of TPU capacity. When compute and distribution alliances reshuffle, the ecosystem your tool sits on can change overnight. Map your second-source path now.

#6 Regulatory and sovereignty shifts. EU AI Act, U.S. state-level AI laws, sovereign cloud requirements reset which ecosystems serve which geographies. A founder anchored to a U.S.-only regulatory regime cannot serve a German Global 2000 company.

#7 Agent interoperability standards. MCP and A2A are being adopted across SAP, Anthropic, OpenAI, and Google. Fast adopters become the default substrate. Laggards become legacy faster than anyone expects.

Ecosystem choice is not a one-time, single-layer decision. Treat it as a checkbox and you lose.

📌 One Thing to Do This Week

Pick one ecosystem layer - compute, model, middleware, or app - and write a one-page memo for your team answering five questions about it: data gravity, integration surface, compliance posture, co-sell economics, and replacement velocity. If you can’t credibly answer all five, you do not have an ecosystem strategy. You have a hope. Enterprise buyers will tell the difference inside the first 15 minutes of a vendor call.

“Inside Enterprise AI” is written by Johannes Deubener - AI Implementation Lead at Deutsche Telekom, ex-Founder, and host of the AI to Go podcast. Every issue is drawn from real vendor evaluations, real enterprise pilots, and the messy enterprise reality that AI founders building for large Enterprises need to understand.

If this was useful, forward it to one AI founder who builds in the Enterprise segment.

→ Watch the full episode with Chris Carter:

→ Subscribe to “AI to Go” on YouTube, Spotify, Apple Podcasts

Newsletter for Enterprise AI founders. Every issue: one key insight, one practitioner deep-dive, one thing to do this week.

🚨 Key Insight

In any Enterprise setting, a text chatbot that’s 80% accurate can be useful. A voice agent that’s 80% accurate is a different story - because the 20% happens live, with no retry button. Every failure is a refund, an escalation, or a compliance incident.

That’s why voice AI is the 2026 story for Enterprise AI founders specifically. Text-based AI has commoditized - every SaaS vendor ships a copilot, and buyers treat the LLM as a substitutable layer. Voice breaks that pattern. The stack is multi-model, latency-sensitive, regulation-heavy, and failure-mode-rich. Which means the margins, the moats, and the founder-scale opportunities are back. Three data points explain why this is the year:

• The latency wall broke. Best-in-class voice stacks now hit sub-500ms end-to-end. Industry median is still 1.4–1.7 seconds per Hamming’s analysis of 4M+ production calls - which is exactly why the gap between leaders and the field matters. The leaders make Voice AI conversations feel natural and human-like.

• The economics inverted. Enterprise voice AI now runs at roughly $0.05–$0.15 per minute all-in (per Retell, Klariqo, and multiple 2026 platform comparisons). That’s consistently 5–10x cheaper per minute at production scale.

• The market went industrial. The voice AI agent segment is projected at ~35% CAGR through the early 2030s (per Market.us; estimates vary across research firms), and enterprise production deployments have scaled materially in the last 12 months - from isolated pilots to departmental infrastructure.

🧠 The Buyer’s Brief: The 5 Strategic Gaps in the Enterprise Voice AI Stack

That’s the bullish case. Here’s what no vendor deck mentions: the enterprise-readiness layer around the voice stack is mostly missing. Five gaps. Each one a credibly billion-dollar company.

Gap 1: Languages and Accent Coverage

This is the gap I see most clearly from my seat at Deutsche Telekom - and the one Silicon Valley founders most consistently underestimate. Every voice AI platform ships with great English and reasonable Spanish. Almost none of them ship with production-grade coverage of the 23 other official EU languages, the 22+ languages of India, or the major African languages. The leading multilingual players cluster around 30–40 languages - and quality degrades sharply outside the top five.

The accent problem inside English alone is worse than founders realize. A 2020 study by Stanford audited five major ASR systems - Amazon, Apple, Google, IBM, and Microsoft - and found an average word error rate of 35% for African American English speakers versus 19% for white American speakers. A 2024 study by Washington University audited 2,700+ speakers from 171 countries and found statistically significant performance disparities for non-native English accents across all major commercial ASR services. The absolute numbers have improved since publication with newer models, but the disparity persists as a structural problem tied to training data composition. For an enterprise deployment in Mumbai, Lagos, Berlin, or São Paulo, those numbers mean voice AI that literally doesn’t work on a material fraction of customer calls.

And this isn’t a “nice to have” - it’s a procurement blocker. Every European buying team starts the voice AI evaluation with: “Does this work in all our markets at parity?” If the answer is “English is great, Portuguese is okay, Polish and Czech are a year away,” the deal easily stalls at the regional level before the global contract is ever signed.

The Founder Opportunity: Of the ~7,000 living languages, only about 20 are genuinely “high-resource” for voice AI - meaning they have the training data depth to build production-grade Voice AI at parity with English. That leaves ~3 billion speakers of “low-resource” languages - concentrated in Asia and Africa - materially underserved by today’s voice stack. And here’s the part that matters for enterprise: even widely-spoken languages like Hindi (600M speakers), Bengali (230M), and Swahili (200M) fall in the low-resource tier. This isn’t a small-markets problem. It’s a structural gap in what the industry has built, creating a - structural - advantage for founders who are Non-English native speakers and work outside the SF/NYC fundraising corridor. I see three credible wedges building Voice AI infrastructure as language-first, not English-first:

(I) region-specific platforms that dominate one linguistic market (a voice AI company built for Indian languages, or the Nordic markets, or MENA, or West Africa) with data depth competitors can’t match.

(II) an accent-adaptive layer that plugs into existing voice stacks and demonstrably closes the WER gap on non-American English - a real-time translation equivalent for accent conversion.

(III) a multilingual evaluation and benchmarking layer that exposes the language/accent performance gap publicly and forces the horizontal platforms to compete on it. Linguistic specificity is also the hardest moat to copy because the data sets that create it take years to build.

Gap 2: Voice Monitoring and real-time Failure Handling

A buyer question no founder has answered cleanly: “A customer complained in call #4,827 this morning. Can you tell me what happened?”

In text AI, you have logs. In voice AI, you have audio - and usually no trace of which ASR hypothesis won, which LLM tokens routed the call, or which TTS chunk was playing when the customer interrupted. The stack’s state is locked inside the call-scoped object, and most vendors don’t expose it. This observability gap has a live-call twin: the runtime failure problem. Industry-leading containment is 60%+, meaning 40% of calls still hand off to humans, and those handoffs are where customer trust lives or dies. The demo shows a smooth transfer; production shows the customer repeating themselves to a confused human with no context. Worse: A 2024 study by Cornell University found that roughly 1% of Whisper transcriptions contained entirely fabricated phrases or sentences that never appeared in the source audio. Of those hallucinations, 38% contained what the authors classify as “explicit harms” - fabricated violence, inaccurate personal associations, or invented authority (fake URLs, credentials, or citations). On a live call, an undetected hallucination is a liability event, not a bug.

The Founder Opportunity: A unified voice runtime layer that covers both in-call failure handling and post-call debuggability: real-time hallucination detection, sentiment-based escalation triggers, and context-preserving handoff that passes the full conversation summary and the specific escalation reason to the human agent - AND every conversation treated as a distributed trace (ASR hypotheses, LLM prompts and tool calls, TTS timing, barge-in events, per-stage latency), searchable and exportable for the customer. These two capabilities get bought together because they answer the same buyer question - “can I trust this in production?” - and vendors that unbundle them leave Enterprise buyers unconvinced.

Gap 3: The “Voice to Structured Data” Extraction Layer

The gap buyers feel most acutely but vendors talk about least: what happens to the transcript after the call?

Every voice deployment produces two outputs - a resolved interaction and a raw transcript. The transcript, in most deployments, is dead weight. The question after “does the agent work?” is “what structured data comes out and where does it land?” Claim details into the claims system. Lead qualification into Salesforce. Sentiment into CX dashboards. Compliance flags into audit. The transcript is the raw material; the structured extraction is the product. Very few vendors build the second half.

The second dimension most founders miss: real-time extraction enables dynamic routing, agent assist, mid-call interventions. Post-call extraction enables analytics, coaching, workflow automation. Different architectures, different buyers, both underserved.

The Founder Opportunity: A voice-native extraction layer between the voice agent and enterprise systems of record. Closest analog: the conversation intelligence category (Gong, Chorus, Observe AI) - but those were built for sales coaching and QA, not for the agentic voice era. The 2026 version is built for voice-first workflows, not call recordings. Bonus moat: the structured output schema becomes the defensibility. Own the last mile of the voice call, own the data every downstream system depends on.

Gap 4: The Compliance Layer Nobody Built

Voice creates regulatory risk text never did:

• TCPA: FCC classifies AI voices as “artificial.” Violations run $500–$1,500 per call with no cap - a 10,000-call campaign is $15M of potential exposure.

• HIPAA: Voice recordings of PHI sit inside the full weight of HIPAA enforcement - Security Rule requirements, breach notification, and state-level retention rules. HIPAA penalties range from ~$137 to ~$2.1M per violation tier, and call recording systems that don’t enforce access controls, encryption, and state-specific retention are a common finding in OCR audits. 2025 OCR enforcement actions against healthcare-adjacent firms (Warby Parker $1.5M; Solara Medical Supplies $3M in 2024) show the enforcement regime is active and material.

• BIPA, GDPR, Colorado’s biometric + AI stack (2026): Illinois BIPA imposes $1,000–$5,000 per violation for biometric voiceprint handling without consent. GDPR treats voiceprints as special-category data with fines up to 4% of global revenue. Colorado’s Biometric Amendment (HB 24-1130) classifies voiceprints as sensitive data requiring prior consent; Colorado’s AI Act (SB 24-205, effective June 30, 2026) adds “high-risk AI” obligations when voice AI is used in consequential decisions (employment, insurance, credit, healthcare). Enterprises deploying outbound voice AI in regulated verticals now have to clear all three simultaneously.

The Founder Opportunity: Compliance-as-a-platform for voice. Consent capture, state-aware recording rules, PII redaction at the ASR layer (not post-hoc), retention enforcement, audit-ready exports. Wedge strategy: own one regulated vertical (telco, healthcare, financial services, or insurance), then expand.

Gap 5: Agent Provenance in Outbound Voice

Here’s a buyer-side question that’s come up more and more recently, and that no incumbent cleanly answers: when our voice AI calls a customer, how does the customer know it’s actually us?

This is the outbound trust problem, and it’s the mirror image of what Pindrop, Resemble, and a few others, already solve well on the inbound side. The existing defenses assume a caller reaches your contact center and you need to decide if they’re real. But 2026’s dominant threat vector is the reverse: an attacker clones your brand, your CFO, or your customer service line and calls your customers, suppliers, or employees. The Arup $25.6M deepfake incident (January 2024, Hong Kong office) shows the outbound attack vector at scale - a video-and-voice conference call with a deepfaked CFO and colleagues convinced a finance employee to wire $25.6M across 15 transactions. Pure voice vishing operates on the same principle without the video component, and it’s accelerating fast: deepfake-enabled vishing surged over 1,600% in Q1 2025 vs Q4 2024 per Right-Hand Security, and Pindrop’s 2025 Voice Intelligence Report separately documented a +1,300% rise in deepfake fraud attempts across 1.2 billion customer calls. Voice cloning now takes 3–10 seconds of sample audio scraped from any podcast, earnings call, or LinkedIn video.

STIR/SHAKEN doesn’t solve this. It authenticates the phone number at the network layer - not the content, not the AI agent behind the call, not the brand the call claims to represent. Branded calling (e.g. Hiya) is a real step forward, but it’s built around phone number identity for human outbound calling, not for verifying that a specific AI agent is legitimately authorized to act on behalf of a specific brand in a specific conversation.

The Founder Opportunity: The “verified agent” layer for voice - the DKIM/DMARC analog that email got in 2007 and voice still doesn’t have. Cryptographically signed agent provenance at the call level: who is the AI, which enterprise authorized it, what scope is it permitted to act within, and how can the receiving party (human or another AI agent) verify all of that in real time? Adjacent play: agent-to-agent trust protocols for when your voice agent calls another company’s voice agent (an emergent category with no incumbent). The founders who establish the standard here own a category, not a product. Email security built its multi-billion-dollar layer on exactly this foundation. The standard-setter captures disproportionate value because trust layers compound via network effects, not feature races.

Where this leaves us

Voice AI in 2026 is the rare category where the technology is finally ready and the enterprise infrastructure around it isn’t. The demos work. Production is where the gaps still live. And every gap - language/accent coverage, observability and failure handling, structured data extraction, compliance, outbound agent trust - is a billion-dollar-plus company waiting to be built.

📌 One Thing to Do This Week

If you’re not building voice-native but want to add voice to your product: Resist the urge to ship a voice feature this quarter. Instead, map exactly one high-frequency, high-value moment in your existing customer workflow where a voice interaction would remove real friction - then pressure-test it against the five gaps above. If your answer to “how does the enterprise buyer evaluate this on compliance, observability, and agent provenance?” is “we’ll figure that out later,” you’re building a demo, not a feature. The founders who add voice successfully in 2026 don’t ship voice everywhere. They ship it at the single workflow moment where it creates 10x value, with the enterprise layer designed in from day one.

If you’re building voice-native: Audit your product against one buyer-side question: “Can you give me a full call-level trace of a failed conversation - with PII redaction, signed agent provenance, and structured output extracted - exported to my data base within 60 seconds?” If the answer involves a support ticket or a screen-share, you’re not production-ready at Enterprise scale. That single capability moves you past 90% of the field before your next RFP.

About this Newsletter

“Inside Enterprise AI” is written by Johannes Deubener - AI Implementation Lead at Deutsche Telekom, ex-Founder, and host of the AI to Go podcast. Every issue is drawn from real vendor evaluations, real enterprise pilots, and the messy enterprise reality that AI founders building for large Enterprises need to understand.

If this edition was useful, forward it to one AI founder who builds in the Enterprise segment.

Newsletter for Enterprise AI founders. Every issue: one key insight, one practitioner deep-dive, one thing to do this week.

🎙️ Key Insight from this Week’s “AI to Go” Podcast

Guest: Tim Daines, Founder of Synaptix Solutions - AI enablement company, 22-year enterprise expert, building in Dubai.

Tim’s quote that struck me:

“The demo looks brilliant. The board gets excited. And then six months later, the thing is sitting in a corner - like an unused Peloton bike from COVID.”

I’ve been that enterprise buyer. I’ve watched brilliant demos. I’ve signed off on pilots. And I’ve watched too many of them never reach production.

What Tim named the “ROI Gap” is something I see every week from the inside.

The real cost of enterprise AI is not the LLM. It’s everything that nobody budgeted for: integrations, governance, workflow redesign, behavior change, exception handling, and data management.

Most founders are fundraising for the user interface. Few are building the guts.

Tim’s diagnosis: VCs are getting nervous not because AI doesn’t work - but because their portfolio companies are burning cash on beautiful demos that can’t survive contact with a real enterprise. The factories are being built. The toys are already in them. But the production-ready environment? Not there yet.

My buyer’s reaction: This is exactly what it looks like from inside the enterprise. We don’t kill pilots because the technology failed. We kill them because the vendor didn’t design for what comes after the demo.

🧠 The Buyer’s Brief: The Six Reasons Enterprise AI Demos Never Reach Production

What I’ve learned evaluating 100s of AI vendors inside a global F500 enterprise - and what Tim confirmed from the other side of the table.

Most enterprise AI conversations focus on the wrong problem. Founders optimize the demo. Buyers evaluate the demo. And then both sides are surprised when the demo doesn’t translate.

Here are the six reasons enterprise AI demos die - and what winning founders do differently at each stage.

1. You Sold the Demo. You Didn’t Design for the Day After.

The demo is the easiest part of an enterprise AI deployment. The hard part is everything that comes after: integrations with legacy systems, exception handling when the model is wrong, human oversight workflows, security reviews, access permissions, and change management for the team who has to use it.

What winning founders do: They walk into the first meeting asking: “What does your production environment look like?” Not “Can I show you what this can do?” The shift from demo-first to production-first thinking is the single biggest differentiator Tim sees between founders who win and founders who die in procurement.

2. You Conflated POC, Pilot, and MVP.

These are not the same thing. A POC proves the technology works. A pilot tests it in a controlled slice of the real environment. An MVP is something you put in front of users. Founders who conflate all three - and from there, they jump straight to “production” - create the most common enterprise AI failure pattern.

What winning founders do: They sequence deliberately. POC → Pilot → MVP → Production, with explicit success criteria at each gate. They don’t declare victory at pilot.

3. You Started With the Data You Had - Not the Data You Need.

This is where Tim’s concept of Minimum Viable Data becomes critical. Most founders assume data availability. They don’t ask: what is the minimum data signal needed to make this specific workflow function reliably? The result: beautiful AI built on incomplete, inconsistent, or siloed data that produces unreliable outputs at scale.

What winning founders do: Before writing a single line of code, they map the workflow. They identify exactly where the minimum viable data lives - and whether it’s actually capturable. If the answer is no, they fix the data problem first.

4. You Ignored the Knowledge Bleed.

Here’s the big problem underneath the data problem. You can’t identify your Minimum Viable Data until you’ve mapped how to capture relevant workflow knowledge and, even more importantly, where institutional knowledge is leaving the business. Tim calls this Knowledge Bleed - and it’s accelerating. Baby boomers are retiring. Zero-hour contract workers are quitting after three months. Senior engineers move on. Every time someone leaves or changes the job, they take with them undocumented workflow knowledge that was never captured, never structured, never made machine-readable.

The enterprises that are hardest to deploy AI into aren’t the ones with bad technology. They’re the ones whose operational knowledge lives entirely in people’s heads.

What winning founders do: They start by asking “Where is your institutional knowledge currently living, and what happens when the person who holds it leaves?”. That question opens a conversation that most vendors never start - and it immediately differentiates you from everyone selling chatbots.

5. You Sold Technology. The Buyer Was Buying Trust.

This is Tim’s most important observation - and it matches exactly what I see from the buy side. Enterprise buyers are not buying your technology. They are buying when you can be trusted with their governance, their data, their compliance posture, and their political exposure.

What winning founders do: The founders who win do not lead with the shiny demo. They lead with sovereignty, security, and governance. They speak the language of the people who can kill a deal: procurement, IT, legal, compliance. And they frame their product not as a technology purchase - but as a risk reduction decision.

Before the first meeting, they prepare answers to the questions that will end the deal if unanswered: How does this behave when the model is wrong? Who has access to what data? How does this fit within our data sovereignty requirements? What’s the handoff when this goes into production?

6. You Measure Productivity. The CFO Measures Margins.

Most founders pitch ROI on Enterprise AI programs with productivity metrics: hours saved, tasks automated, speed improved. Enterprise CFOs - who are increasingly owning the AI investment decision - are asking a completely different question: what is the lifecycle cost of all this AI investment, and where is it actually impacting margins?

A co-pilot that saves 2 hours per week per employee while quietly inflating data center costs is not a win. It’s a liability that the CFO will eventually notice.

What winning founders do: They speak in CFO language from day one. Not productivity metrics. Revenue recovered. Lifecycle cost expected. Margin impact quantified.

The Pattern That Wins

Tim’s clearest summary of what winning founders have in common:

They sell a system that learns the business, remembers the business, and builds a strategic moat from the inside out - using institutional knowledge as the key ingridient.

That is not a chatbot. That is not a co-pilot. That is a Knowledge Management System designed to become the intelligence integration layer of the enterprise. The founders who build this category create switching costs, defensibility, and genuine ROI that survives the CFO’s quarterly review.

The founders selling “another LLM with great evals” - won’t be on the shortlist in 12 months.

📌 One Thing to Do This Week

Map the Knowledge Bleed in your next prospect’s org.

Before your next enterprise sales conversation, ask one question: “What happens to the workflow knowledge when a senior team member leaves?” Listen for hesitation. That hesitation is your product opportunity. If they don’t have an answer - you’ve just found your Minimum Viable Data starting point, and a reason for them to buy that no chatbot vendor can replicate.

Do this in the next three conversations. It will change every pitch that follows.

“Inside Enterprise AI” is written by Johannes Deubener - AI Implementation Lead at Deutsche Telekom, ex-Founder, and host of the AI to Go podcast. Every issue is drawn from real vendor evaluations, real enterprise pilots, and the messy enterprise reality that AI founders building for large Enterprises need to understand.

If this was useful, forward it to one AI founder who builds in the Enterprise segment.

→ Watch the full “AI to Go” episode with Tim Daines:

In this conversation, Alex Sherman, CEO at Bluefish AI, and host Johannes Deubener explore the seismic shift happening in marketing right now, and why 2026 is make-or-break for every CMO and marketer.

Chapters

00:00 Introduction: AI Marketing Channel Revolution

00:59 Welcome and Introduction of Alex Sherman

01:23 Alex's Origin Story

03:30 The Evolution of Technology

04:34 From European History to Enterprise Software

05:20 An "Accidental" Entrepreneur: His Path to Startups

06:47 Alex's "Pirate" Philosophy: Building Against the Odds

08:54 Flying Through the Storm: Leadership in Uncertainty

11:45 Leadership in the AI Age: Absorbing Stress and Filtering Noise

12:47 AI Marketing: The New Content Paradigm

16:41 The Wake-Up Call: 66% of Shoppers Use AI

17:03 From Novel Idea to Essential Channel

20:20 How Alex used Deep Customer Discovery to get started

23:17 The Bluefish Platform: Visibility and Influence Across the AI Ecosystem

26:13 The Attribution Problem: Tracing Marketing Impact in AI

29:47 Alex's Message to CMOs

33:21 Early Mover Advantage: Data Quality and Learning Curves

36:47 The Attention Economy vs. Agent Economy

40:19 The Future of Brand Storytelling in an AI World

45:09 Regaining Control: Putting Brands Back in the Driver's Seat

47:57 Real-Time Marketing: From Planned Campaigns to Continuous Optimization

51:59 The End of Paid Ads?

55:13 Alex's Advice for AI Founders

57:23 Closing Remarks

--------------------------------------------------------------

Don't be shy and come say hi:

Connect with Johannes: https://www.linkedin.com/in/johannesdeubener/

Follow Alex: https://www.linkedin.com/in/alsherman/

Check out Bluefish AI: https://www.bluefishai.com/

--------------------------------------------------------------

More about AI To Go:

✨ Watch more Episodes on YouTube: https://www.youtube.com/channel/UCiL3_cC9nmXALQiaqlYs2Kw
✨ Listen on Apple Podcasts: https://podcasts.apple.com/us/podcast/ai-to-go-podcast/id1866997458

✨ Listen on Spotify: https://open.spotify.com/episode/5X9XgIv1qtCYtExziNMyXK
✨ Receive all Updates on LinkedIn: https://www.linkedin.com/company/ai-to-go-podcast/

Tags: #AIMarketing #EnterpriseAI #MarketingStrategy #CMO #ChatGPT #Podcast #Leadership #Innovation #BusinessStrategy

Karin shares her journey from a music career to entrepreneurship, emphasizing the importance of creativity, community, and mental health. She discusses the transformative power of AI in mental health solutions, the need for breaks to recharge creativity, and the role of community support in entrepreneurship. Karin also highlights the significance of data-driven approaches in mental health and the importance of personalizing mental health solutions to enhance wellbeing.

Our conversation delves into the complexities of navigating mental health knowledge and real-life implementation, emphasizing that having access to information and wisdom is not sufficient for overcoming personal struggles. Karin highlights the importance of support systems in dealing with trauma and mental health issues, as well as the role of AI in providing assistance. If you have big plans in 2026, building healthy habits while managing stress and anxiety are critical. Karin shares her personal best practices and explains how Earkick makes a real difference in the lives of their users.

Chapters

00:00 Karin's Journey from Musician to Entrepreneur

03:24 Creativity and the Importance of Breaks

05:48 The Role of Community in Entrepreneurship

10:29 Mental Health as a Business Opportunity

13:40 The Evolution of Mental Health Awareness

17:16 Introducing Earkick A New Approach to Mental Health 20:13 The Science Behind Earkick's Effectiveness

23:28 Personalization in AI Therapy

26:38 Real-Life Impact of Earkick

29:54 The Role of Employers in Mental Health

33:14 Building Trust in Mental Health Solutions

36:55 The Future of AI in Mental Health

--------------------------------------------------------------

Don't be shy and come say hi:

Connect with Johannes: https://www.linkedin.com/in/johannesdeubener/

Follow Karin: https://www.linkedin.com/in/karinstephan/ Check out Earkick: https://earkick.com/

--------------------------------------------------------------

More about AI To Go:

✨ Watch more Episodes on YouTube: https://www.youtube.com/channel/UCiL3_cC9nmXALQiaqlYs2Kw

✨ Listen on Apple Podcasts: https://podcasts.apple.com/us/podcast/ai-to-go-podcast/id1866997458

✨ Listen on Spotify: https://open.spotify.com/episode/5X9XgIv1qtCYtExziNMyXK
✨ Receive all Updates on LinkedIn: https://www.linkedin.com/company/ai-to-go-podcast/