Sam Bobo — AI Product Manager

Building Products at the Intersection of AI & Human Experience

I close the gap between what AI can do and what an organization will do with it, and that gap is where revenue is lost or won.

Work With Me Explore My Work
Sam Bobo — AI Product Manager
Microsoft & IBM Watson
Harvard & Emory
10+ Years in AI

My Approach to AI Product Management

At the intersection of artificial intelligence and product management lies a unique challenge: translating cutting-edge technology into products that people actually want to use.

My AI Journey

How the field evolved, and how I moved through it — not always in the same order

Chapter 01

Conversational AI

2014 – 2019

When AI found its voice — NLU, intent classification, and dialogue management transformed how people interacted with technology, and how I thought about product design.

Industry Milestone
February 2011

Watson Wins Jeopardy!

IBM's Watson defeats champions Ken Jennings and Brad Rutter, demonstrating that machines can parse natural language, reason over vast knowledge, and respond.

August 2014

BlueSpark

I join IBM Watson as the third cohort of the BlueSpark program, a rotational-based, leadership development program.

October 2014

IBM Watson Headquarter Launch

IBM opens its Watson headquarters in New York City at 51 Astor Place, signaling a major commitment to conversational AI.

Early 2015

Industry Intelligence

Acquired, operationalized, and productized scholarly industry-specific content into corpora to augment natural-language models with domain-based ground-truth, improving accuracy in industry AI applications by 18%.

Work done after being selected to shadow under the Chief Marketing Officer and head of the Watson Ecosystem

Model Adaptation
2015-2016

Watson Ecosystem

Consulted 40 leading technology start-ups in the Education, IoT, Healthcare, and Data Analytics industries from product design and implementation through revenue maximization, to solve new and evolving customer needs with natural language technology.

AI Consultation
2016

Watson Starter Packs

Uncovered AI service usage patterns through cluster analysis, analyzing 250k+ IBM Cloud developer environments to inform the creation of use-case based Starter Kits, reducing developer time-to-”hello world” for Watson-powered applications by 85%.

IBM Cloud was called Bluemix at the time

Developer Relations
Chapter 02

Rules-Based AI

2019 – 2022

When AI meant meticulously crafted logic trees, expert systems, and deterministic decision engines — I learned that intelligence begins with structure.

2019

Nuance Speech Suite

Started owning the Speech Suite product roadmap, Nuance's highest-grossing on-premise voice AI platform, shipping 24 releases to sustain $125M+ in peak annual revenue across 1,800+ worldwide enterprise customers

Voice AI
2020

Microservice Decomposition

Championed the redesign of Nuance's speech engines into cloud-native microservices, directly laying the infrastructure foundation for Dynamics 365 Contact Center’s debut and achieving 99.999% availability and 99.99% SLA.

Cloud Native
Industry Milestone
March 2022

Microsoft Acquires Nuance Communications

The Nuance Enterprise business starts being incorporated into the Business Apps organization

OpenAILarge Language Models
Chapter 03

Generative AI

2022 – Present

When AI stopped following scripts and started writing them — large language models, multimodal systems, and agentic architectures are redefining what's possible in every product category.

Industry Milestone
November 2022

OpenAI debuted ChatGPT

OpenAI launches ChatGPT, reaching 100M users in 60 days — the fastest product adoption in history. Generative AI shifts from research artifact to mainstream conversation, and every product team suddenly has an AI mandate.

OpenAILarge Language Models
2023

Code Generation Tooling

Built three AI code-generation tools that automated Nuance-to-Microsoft bot migration for systems integrators, cutting migration effort and time by 75% at 96.3% accuracy.

CodeGen
2023

Constrained Speech Recognition

Drove adoption of constrained speech recognition for enterprise Copilot development, reducing alphanumeric misinterpretation rates by 60% and cutting ASR word error rate by 50%, boosting self-service containment.

ASR
Microsoft Milestone
July 1, 2024

Dynamics 365 Contact Center

Microsoft officially launches its premier Contact Center offering

VoiceAI
2025

Empathetic Voices

Grew the D365 Contact Center voice gallery by 24%, adding expressive, customizable branded and standard voices with cross-lingual and local dialect coverage to improve CSAT across global markets, supporting 100M+ calls annually per customer.

TTS
2025

Self-Learning Systems

Explored agentic self-learning, memory systems, and behavioral fine-tuning for Contact Center AI agents using human-to-agent feedback loops that informed Microsoft's agentic AI roadmap.

Learning Loops
Now

Shaping the Next Wave

Focused on model fine-tuning and agent creation

Agents

Core Competencies

Voice AI

Speech Recognition (ASR), Text-to-Speech (TTS), Natural Language Understanding, Dialogue Management

In practice →

In practice

Owned voice AI products processing 100M+ calls annually — tuning ASR accuracy, architecting NLU dialogue flows, and launching empathetic voice personas across Nuance and Microsoft Copilot. Latency, accent coverage, and domain vocabulary compound in ways text-first PMs rarely anticipate.

Frameworks

Enterprise Strategy, Mental Models, Execution Blueprints, Cross-functional Alignment, Decision Architecture

In practice →

In practice

I build frameworks when an organization has the right pieces but can't see how they connect. At IBM it was Application Development, Nuance it was engine decomposition, and at Microsoft it was Adaptation and Learning Loops.

Responsible AI

Model Cards, Harms Assessment, Transparency, Fairness & Bias Mitigation, Governance Frameworks, Regulatory Compliance

In practice →

In practice

I've delayed launches when bias audits surfaced real issues and built governance processes that treat responsible AI as an upstream engineering constraint. The goal is always the same: make the risk visible early enough to do something about it.

High Fidelity Design

Figma Make, Interactive Prototypes, Product Mockups, Design Systems, UX Flows

In practice →

In practice

I produce high-fidelity prototypes in Figma that communicate product intent with the precision of a shipped feature. Stakeholders align faster, user feedback is real, and engineering rework drops — because decisions are made in pixels, not assumptions.

Product Evangelism

Video Vignettes, Live Demos, Conference Speaking, Developer Advocacy, Narrative Storytelling

In practice →

In practice

I've produced 20+ product video narratives and presented on-stage at 5 major tech conferences. I translate model capability into a story that a non-technical buyer can feel in under two minutes, directly accelerating enterprise deals.

Systems Thinking

Feedback Loops, Dependency Mapping, Emergent Behavior, Complexity Navigation, Root Cause Analysis

In practice →

In practice

Enterprise AI products sit inside workflows, incentive structures, and organizational habits that can reject a technically sound product. I map dependencies before scoping — it's how I catch failure modes before launch instead of after.

The Glossary

A field guide to the AI primitives I have hands-on depth in. Each entry pairs the concept with a story of where I've shipped it.

Showing all capabilities.

Voice · TTS Lexicon Adaptation noun
The mapping of words to their phonetic spellings, used to teach text-to-speech systems how to pronounce industry-specific jargon and proper nouns.
How I use it → At Microsoft, I led the unification of phonetic-rule adaptation across both deterministic and Generative AI voice models — translating problematic words into IPA, X-SAMPA, and inline phonetic forms so administrators don’t have to be speech scientists. Hit 96.3% pronunciation accuracy on a healthcare drug-name benchmark.
Voice · TTS Custom Voice noun
A production-grade voice replica trained on samples from a specific speaker — surfaced in Microsoft Dynamics 365 to verified customers under a sensitive-use review process with responsible-AI guardrails.
How I use it → I drove the customer-facing release of Custom Neural Voices for D365 Contact Center, partnering with legal to land a deepfake-mitigation: Custom Voices are blocked from outbound-calling scenarios so the cloning use case can’t be weaponized against unsuspecting recipients.
Recognition · Speech Constrained Speech Recognition noun · technique
A grammar-driven recognition approach that constrains spoken-input output to a predefined list (per the W3C SRGS spec), far outperforming open-ended STT on accuracy, latency, and cost for closed-set tasks like account numbers, zip codes, or stock tickers.
How I use it → I owned the constrained-grammar system at Nuance Speech Suite and led its productization into Microsoft Copilot Studio — letting customers who’d spent hundreds of thousands tuning grammars carry that investment onto the modern stack instead of starting over.
Recognition · Entity Confirmation Mode & Threshold design pattern
A bot-design pattern that re-confirms captured information (account numbers, SSN, etc.) when recognition confidence falls below a tunable threshold — balancing transactional accuracy against caller friction.
How I use it → I designed the confirmation-strategy controls (never / always / threshold-gated) for high-stakes self-service flows. Below threshold the bot reads the value back: “to confirm, your account number is 8921782 — is that correct?”
Learning Loops Autonomous Agent Feedback Loop system pattern
A closed-loop system where conversation analysis identifies repeated topics, clusters them, and uses generative AI to author new automation routes — turning the contact center from a cost center into an “idea center.”
How I use it → I led Project Blackbird in early 2023 — pre-agentic-AI common parlance — building a proof of concept that performed cluster analysis on customer-service call transcripts and generated Copilot Studio YAML topics for human review.
Affective Computing Emotion AI noun · field
Voice and speech models that infer emotional state — disappointment vs. disgust, calm vs. frustrated — informing both agent assistance and AI-voice tone modulation in real time.
How I use it → At the Microsoft 2025 Hackathon I led an integration of Hume AI’s models into Dynamics 365 Contact Center, surfacing inferred caller emotion in the agent dashboard. The project demonstrated value but didn’t ship — bumping into responsible-AI sensitivity around inferred-emotion use plus the friction of onboarding a third-party vendor.
Voice · TTS AI Clone noun · application
Voice replicas of specific individuals trained on a controlled set of reading-prompt samples — used by leaders for executive demos and as a benchmark of realism against competitors like ElevenLabs.
How I use it → I produced a 15-hour-trained replica of our VP from 500 prosody-targeted prompts. He used it in 10–15 executive briefings; the realism reached the uncanny valley enough to influence enterprise deal momentum. Made replicas of two other senior leaders after.
Voice · TTS Saudi Voice project
Locale-specific neural voices tuned to a regional accent or dialect — addressing identity-tied gaps where general-purpose voices feel “off” to local audiences.
How I use it → To unblock $11M in Saudi/Emirati pipeline (12 companies losing deals because Saudi Arabic voices sounded too Egyptian), I shipped a Najd-dialect Custom Voice — first proving feasibility with an internal employee, then contracting an SI to broker a local voice actress with defined accent-fidelity acceptance criteria.
Voice · TTS AI Voice Portfolio PM responsibility
The full slate of TTS voices a contact-center provider ships across genders, ages, languages, and styles — managed as a portfolio against competitor offerings, evolving capabilities, and customer demand.
How I use it → I lead the TTS portfolio for Microsoft Dynamics 365 Contact Center — picking voices to ship, deprecating older ones, and balancing GenAI/HD voices against deterministic neural voices across 100+ locales.
Voice · TTS HD Voice noun · platform
GPU-accelerated text-to-speech that adds emotional, stylistic, and contextual variation to synthesized speech — closing the experience gap between neutral-sounding bots and frustrated callers.
How I use it → I led the rollout of HD Voices from Azure AI Speech into D365 Contact Center, addressing the common pain point where a frustrated caller hears a neutrally-cheerful bot and ends up more frustrated.
Voice · Infra HD Voice Routing infrastructure
Intra-region inference routing that keeps voice TTS data-residency compliant while sending requests to whichever sub-region has GPU capacity — bridging the gap between voice-AI demand and uneven hardware availability.
How I use it → When telemetry showed mounting customer frustration in regions where HD voices were unreachable due to GPU scarcity, I worked with engineering to design intra-region routing that preserved data residency while unblocking the experience.
Voice · TTS Multilingual Voices noun · feature
Neural voices that fluently speak multiple languages from a single voice persona — letting a single bot persona handle truly global, multi-language conversations without persona switching.
How I use it → I unblocked multilingual access in Dynamics 365 Contact Center by expanding the data contract with Azure AI Speech — exposing JennyML’s 76+ secondary languages so bots could use one consistent voice across English / French / Japanese in a single conversation.
Voice · Localization Automatic Language Localization workflow
AI-driven translation of bot authorship into target languages, applied in parallel to development rather than as a post-build bottleneck — with human-in-the-loop approval from local language experts.
How I use it → I introduced Automatic Language Localization in Microsoft Copilot Studio so multilingual bot authoring no longer required dumping the bot, hiring translators, and re-uploading. Translation runs in parallel with authoring; local experts approve translated text inline.
Recognition · Speech Multi-Type Recognition design pattern
A speech-recognition design pattern allowing one bot prompt to accept multiple valid entity types — e.g., “say your account number OR last 4 of your SSN” — with conditional routing based on what was captured.
How I use it → I shipped Multi-Type Recognition into Microsoft Copilot Studio so bot designers no longer had to write complex PowerFx logic across multiple authoring nodes for “this OR that” entity capture. Reduced bot authorship time by 10%.
Recognition · Entity One of Multiple Entities recognition pattern
Recognition of multiple instances of the same entity type within a single user utterance — e.g., a caller listing prescription drug names in one breath rather than one-at-a-time.
How I use it → I exposed multi-instance entity extraction (already supported by Copilot Studio’s NLU engine but not surfaced to designers) on the bot authoring canvas. Cut authoring time and improved user experience by 15% — refilling four prescriptions becomes one utterance instead of four turns.
Learning Loops CollabLLM technique · RL
A reinforcement-learning technique that optimizes LLMs to balance intrinsic rewards (enhancing the user experience by asking clarifying questions) against external rewards (achieving the conversation’s stated goal).
How I use it → At the Microsoft 2025 Hackathon, I partnered with research to fine-tune Qwen 2.5 7B Instruct using CollabLLM — modest lift (84.4% → 85.2%) constrained by training-data volume, but the demonstration moved fine-tuning onto the contact-center roadmap.
Learning Loops System Memory capability
Per-customer persistent memory layered onto evaluation/coaching agents — letting an off-the-shelf agent remember client-specific policies and idiosyncrasies that aren’t in its training.
How I use it → I led the System Memory capability for Microsoft’s Quality Evaluation Agent — letting QEA learn client-specific quality standards (e.g., “always cite the publicly-available policy URL”) rather than treating every customer deployment as generic.
Responsible AI Transparency Notes documentation
Public-facing documentation that accompanies every AI feature/model at Microsoft — covering training data, intended use, guardrails, expected inputs and outputs.
How I use it → I authored the Transparency Note for Constrained Speech Recognition and championed creation of others for NLU and TTS engines in the Dynamics 365 Contact Center portfolio. An underrated form of customer trust-building when AI features ship.
LLM Tuning Constraint in LLMs technique
Steering large-language-model output to adhere to a predefined format or grammar — using techniques like Context-Free Grammars for post-processing bias rather than fine-tuning the model itself.
How I use it → With the shift to sequence-to-sequence ASR architectures (trained on 1M+ hours of audio), accuracy improved but grammar-based bias became hard. I collaborated with research on combining seq-to-seq models with constrained output via post-processing — recovering the precision needed for alphanumeric and list-based recognition.
TTS · Generation Temperature Slider UX control
A user-facing control that exposes the temperature parameter (output randomness/creativity) of generative voice models — letting customers tune voice output from “more static” to “more animated” without engineering involvement.
How I use it → I shipped temperature as a configurable global setting for HD voice bots in both Dynamics 365 Contact Center and Microsoft Copilot Studio admin — exposing what was previously a developer-only knob through an intuitive admin slider.

Writing, Speaking & Press

Sharing ideas on AI product development across stages, screens, and publications

Writing

Dec 16, 2025 · 5 min

Layered Adaptation — The Specificity of Fine-Tuning Models

Read on Medium →
Jan 4, 2024 · 10 min

The Council of Experts Leader-Agent Model — A Proposed Framework for General Intelligence

Read on Medium →
Aug 15, 2019 · 8 min

LabOS

Read on Medium →

Speaking

Panelist

Sarasota.tech Tech Summit

Sarasota, FL · January 9, 2025

"AI — The Good, The Bad, and The Ugly"

View Summit →
Guest Lecturer

Harvard Extension School

Cambridge, MA · December 2023

Guest lecture for the course "Innovation & Entrepreneurship"

Press

Sarasota Magazine

"Smart City | Meet the Innovators and Professionals Shaping Our AI Ecosystem"

Featured Interview · January 2026

Read Article →
Berkeley Prep Magazine

"Decoding the Future | How Berkeley Preparatory School and its Alumni are Embracing AI"

Featured Interview · June 2025

Read Article →

Let's Connect

Interested in collaborating, speaking opportunities, or just want to chat about AI products?

I'm always interested in connecting with fellow product managers, AI researchers, startup founders, and anyone passionate about building responsible AI products.