From Unscripted with Jeff Pedowitz, featuring Matthew Guggemos, speech-language pathologist and CTO behind InnerVoice
Everyone in AI is obsessed with getting machines to talk. Matthew Guggemos spent more than 20 years teaching humans to talk, specifically children for whom communication doesn't come free. Today he builds AI at scale: an AAC app with over 50,000 users and an evaluation tool that cuts assessment time by up to 90%.
That path, from speech pathology to AI leadership, holds lessons for anyone deploying AI in their organization. Here are the five that matter most.
Ask why people leave special education and Guggemos doesn't hesitate.
"The top three reasons are paperwork number one, paperwork number two, and paperwork number three."
A single special education assistance request can run 50 pages per state. The people leaving aren't the ones who can't teach. They're subject matter experts who also mastered pedagogy, and they're walking away over administrative load.
The business translation: your best people rarely quit the work. They quit the friction around the work. That's exactly where AI belongs first. Guggemos built his evaluation tool, Easy, because language sample analysis took him 6 hours per student. Counting morphemes by hand, notating pronunciation in the International Phonetic Alphabet, transcribing 50 utterances. The AI now handles transcription and analysis. The clinician reviews and confirms. Same expert judgment, 90% less grind.
Some educators embrace AI. Others want nothing to do with it. Guggemos is blunt about the resisters: most can't define the thing they're rejecting.
Even the name works against adoption. "Artificial" loses to "real" every time (nobody picks saccharin over sugar), and we still can't agree on what "intelligence" means. The lesson for leaders: resistance is usually an AI literacy gap, not a character flaw. You close it with education and hands-on exposure, not mandates.
The deepest insight in the episode comes from linguistics. Language isn't a physiological process like growing teeth. People say things because saying them works. The research base is Michael Tomasello's usage-based linguistics: you learn what's useful and you say what you hear around you.
Guggemos has watched what happens when a non-speaking child uses a communication device and the adults around them ignore the output. The child asks for help 3 times, nobody responds, so they stand on a chair and get it themselves. The device gets abandoned. Not because it failed technically. Because nobody honored what it said.
"If what you're saying is not understood, not recognized, or not honored, the value of saying words drops. And the role of actions increases."
Swap "communication device" for any AI tool in your stack. If your team's AI outputs go unread, unused, or overridden without explanation, usage will collapse for the same reason. Adoption isn't a training problem. It's a response problem. InnerVoice, his AAC app, bakes this in: its AI avatar behaves like a trained communication partner that always responds to what the user says, asks only questions the user has the words to answer, and never asks them to repeat themselves into a void.
Guggemos played in Michigan State's top jazz band and still gigs regularly. He argues music is one of the best backgrounds for AI work, and the reason is combinatorics: a finite set with nearly infinite combinations.
A vocabulary of 300 combinable words generates billions of possible utterances. A drummer's rudiments produce infinite solos. A jazz combo improvises freely, but inside a key, a tempo, and a form. This conversation itself was improvised inside constraints: English, one hour, AI as the topic.
That's the design principle for agents and chatbots. Don't give your AI everything. Give it the right finite set and clear rules for combination. InnerVoice only asks questions that can be answered with the words on the user's board. Your customer-facing agent should operate the same way: constrained enough to be reliable, combinable enough to be useful.
Guggemos is a CTO with no computer science degree. His secret for learning to build full stack: whenever he hit something he didn't understand, he asked AI to explain it in music terms. Python explained as a snare drum etude. Cloud architecture as a score.
He calls the resulting role "systems conductor." Beethoven couldn't show the timpanist how to hold the sticks, but he understood every instrument's parameters well enough to compose for them. That's the bar for AI leadership: not mastering every instrument, but knowing each one's range, role, and limits. Domain experts who understand language, workflow, and people can absolutely lead AI work. The analogy is the on-ramp.
Guggemos tested Claude, Suno, and Udio on real musician instructions: play in 19/16 like Vinnie Colaiuta, double a figure in C harmonic minor, use quintuplet grids. His review: "AI is a pretty bad musician." The training data for that depth of craft simply isn't there yet.
And his answer to the question every Unscripted guest gets, what humans must consciously hold on to as AI advances: multi-sensory communicative intent, read in real time. The handshake from a friend you haven't seen in years. A robot version might spike dopamine once. It won't mean the same thing.
Who is Matthew Guggemos? A speech-language pathologist with 20+ years of clinical experience, a jazz-trained drummer, and the CTO behind InnerVoice, an AI-powered augmentative and alternative communication (AAC) app from iTherapy with more than 50,000 users. He got his start in AI through a Microsoft AI for Accessibility grant.
What is AAC? Augmentative and alternative communication: tools and devices that let non-speaking people, including many autistic children, communicate through buttons, phrases, and speech output.
What does InnerVoice do? It uses computer vision and generative AI to turn a photo of the user's surroundings into contextual vocabulary, and pairs users with an avatar trained to behave like an ideal communication partner. It honors every response and only asks questions the user has the words to answer.
What is Easy? An AI-powered speech and language evaluation tool that cuts assessment time by up to 90%. It uses speech diarization to separate the child's voice from the adult's in a single-microphone recording, then keeps the clinician in the loop to confirm every attribution.
What's the main business takeaway? Point AI at the friction, not the craft. Automate the paperwork that drives experts away, honor the outputs so usage sticks, and design systems around constrained, combinable sets rather than unlimited options.
Unscripted with Jeff Pedowitz features honest conversations with remarkable people from security and science, business and medicine, technology and the arts. Subscribe wherever you get your podcasts.
The Pedowitz Group | pedowitzgroup.com