Elon, ChuckGPT, and the Future of Work According to Andrew Amann
with Andrew Amann, 923.ai
Elon, ChuckGPT, and the Future of Work According to Andrew Amann
Show Notes
Andrew Amann is the co-founder and CEO of 923.ai, an AI studio that has been building custom AI and machine learning solutions since 2016 - well before generative AI became mainstream. Clients include FanDuel, Consumer Reports, and a range of mid-market and enterprise companies. Before 923, Andrew was a nuclear submarine engineer, then spent the early years of his career helping manufacturing and government organizations transition from paper to software during the early internet era.
923.ai's most visible recent project: Chuck GPT - the AI version of Charles Barkley that appears in FanDuel's commercials at chuck.fanduel.com. The concept took three to four days to build. Making Charles Barkley brand-safe took four to five months. That ratio is Andrew's core lesson about enterprise AI deployment: the fun part is fast; the responsible part is where the real work is.
Why Chatbots Are Not the AI Solution Companies Think They Are
The first thing Andrew wants founders and executives to understand: deploying a chatbot is not an AI strategy. Klarna famously fired 700 employees to install an AI chatbot - and then quietly hired those employees back. The fundamental problem is that large language models are non-deterministic systems. They do not always give the same answer to the same question, and they do not always give the right answer. 923.ai's content team was able to purchase a Chevy Tahoe truck for a dollar through a chatbot that had been deployed without proper guardrails. That is a brand and financial liability, not an efficiency gain.
The solution is evaluation suites - systematic processes for predicting and testing what responses an AI system will produce, checking them for accuracy and brand safety before the system goes live. 923 has been building these for nine years. Most internal teams deploying LLM-based chatbots have not thought about edge cases at all. The lesson: the upside of AI deployment is real, but it requires the same rigor you would apply to any production software system, and more.
The Right Way to Think About AI in Your Business: Eight Hands
Andrew's mental model for AI ROI: instead of asking “which employees do we replace?”, ask “where can we give each employee eight hands?” The goal is not headcount reduction - it is output expansion. An employee with two hands and an AI agent gets the output of eight hands. The consulting firm case study: $500K in revenue became $1M with half the overhead. Nobody was fired. Costs were reduced by eliminating waste and excess time. Revenue increased because the people already there were freed to do revenue-generating work instead of administrative tasks.
The tactical question to ask in your business: which tasks are repetitive and should never have been done by a human in the first place? Data entry. Reformatting invoices. Passing information between departments. Communications overhead. These are the AI entry points - not because they are glamorous, but because they are genuinely wasteful of human capacity. When you eliminate them, you do not reduce headcount; you promote everyone to work that actually requires judgment.
AI Adoption Is 80% Human, 20% Technology
The technology part of deploying AI successfully is, in Andrew's experience, the smaller part of the challenge. The larger part is human: getting people to believe in the system, training them to initiate and review AI outputs, building the managerial skills to run AI-assisted workflows, and earning buy-in from teams who are anxious about what the change means for them.
Andrew's model for how this should work: human passes to agent → agent returns to human for review → human approves before any output goes live. This loop is deliberate. The humans are not being removed from the process; they are being repositioned within it. The part that requires the most investment is training people on that repositioning: what does it mean to be “at the front” of an AI process? How do you write the outline? How do you review an AI output critically? How do you know when to override? These are new skills that take time to build, and they are where most AI deployments succeed or fail.
He draws the parallel to the early internet era - the same fear (“will this replace our jobs?”) accompanied the shift from paper to Excel to SaaS. It never did. It created entirely new categories of work. AI will do the same.
How to Find Your First AI Use Case: Start with the Constraint
Andrew's consulting methodology is straightforward: map the entire business from end to end, identify every bottleneck, and then ask whether removing that bottleneck requires AI or not. If a simpler, deterministic software solution works, use that - 923 has 15 years of non-AI software experience and will use it when it is the right tool. AI is introduced only when the bottleneck genuinely requires it.
The wheel company case study illustrates this. A Minnesota company that makes specialty wheels for fighter jets and fire trucks came to 923 asking for an iPhone app that could photograph a wheel, match it to a database, and automatically ship a replacement. The technical problem was genuinely hard (all wheels are black and round; depth cannot be captured from a photo). But when 923 mapped the whole business, all the red sticky notes - the most painful problems - clustered around sales. The salespeople didn't have enough information to close deals on the phone. The solution: a HubSpot integration that put all the relevant customer, product, and order data on screen during a sales call, so reps could answer questions and take action without calling back in three days. No AI. No iPhone app. Immediate revenue impact.
The lesson: founders who come in asking for AI often need something simpler. The companies that come in not knowing what they need often have a clear constraint hiding behind the request - and finding that constraint is where the real value is.
Why Humanoid Robots Are the Wrong Design Target
Andrew's challenge to the robot industry: stop trying to make robots human-shaped. The bipedal, two-armed human form is extraordinarily difficult to engineer - balancing, adjusting for unexpected weight shifts, managing the moment-to-moment physics of upright locomotion - and it is not necessarily the right form factor for most robot tasks. Roomba is his model: singular purpose, executes perfectly, never needs thought. Coffee maker, same principle. The question for every robotic application should be: what is the actual task, and what physical form optimizes for that task?
The same principle applies to AI agents: a SQL-writing agent does not need to be humanoid or conversational. It just needs to write SQL code accurately when called. The anthropomorphization of AI - giving it a face, a voice, a human personality - is an interface convention we inherited from Hollywood sci-fi, not a design requirement. In some cases (like Chuck GPT, where the entire point is Barkley's personality) it is exactly right. In most operational contexts, it adds complexity without adding value.
Apple, Google, and the Data Gravity Problem
Andrew's diagnosis of why Apple is struggling with AI is simple: they don't have the data. Google and Microsoft have your email, your documents, your search history, your calendar - the context needed to make AI personalization meaningful. Apple has your device. When you strip out iCloud and look at what Apple actually owns versus what you could easily recreate on an Android, the answer is: not much. The blue-bubble/green-bubble dynamic is brand loyalty built on social pressure, not data moats.
His recommendation: if your company operates on Google Workspace, commit to Gemini. If it operates on Microsoft 365, commit to Copilot. Stop copying and pasting between ChatGPT and your actual work tools. The AI value is in personalization, and personalization requires your data to live in the AI's context. The AI tools integrated into your existing workflow will outperform any standalone chatbot you paste content into - and they will compound over time as they learn more about you and your work.
Tools & Resources Mentioned
- 923.ai - Andrew's AI studio; builds custom AI solutions for enterprise clients including FanDuel and Consumer Reports; founded 2016
- Chuck GPT (chuck.fanduel.com) - Charles Barkley AI chatbot built by 923.ai for FanDuel; three days to build the concept, four to five months to make it brand-safe
- Microsoft Copilot - Andrew's top recommendation for Microsoft 365 shops; use the AI built into your existing tools rather than standalone chatbots
- Google Gemini - recommended for Google Workspace users; same reasoning as Copilot
- HubSpot - used in the wheel company case study; integrating product and customer data into the sales CRM view eliminated the need for callback delays and increased close rates
- The Matrix (film) - Andrew references it as surprisingly prescient about AI; notes that the VR goggles / simulation risk is real when split across income demographics
- Benjamin Franklin biography - Andrew recently reread it and draws parallels between Franklin's combination of invention, business, and political engagement and Elon Musk's trajectory
Frameworks
Eight Hands Per Employee
The right mental model for AI ROI is not headcount reduction - it is output expansion. Every employee currently has two hands. The right question is: where can we give each person eight hands through AI agents? The goal is not to eliminate a role but to expand what that role can produce. When you frame AI deployment this way, the entire organizational conversation changes: instead of 'who do we replace?' you ask 'where is each person constrained by repetitive tasks that shouldn't require human judgment?' Eliminate those constraints, and you get more output from the same team - without the friction of a restructuring.
Human-AI-Human Loop
Every production AI deployment should follow the same basic structure: human initiates (writes the brief, defines the goal, provides context) → AI agent executes (drafts, codes, analyzes, communicates) → human reviews and approves before output goes live. This loop is not temporary scaffolding for an immature technology; it is the correct architecture for AI work. It maintains human accountability for quality, prevents brand and compliance failures, and keeps the human skill of critical evaluation actively exercised. Deployments that skip the final human review step are where the Chevy-for-a-dollar types of failures happen.
Find the Constraint First
The most common mistake when bringing in an AI consultant: arriving with a solution in mind. The right process is constraint mapping. Document the entire business end to end, identify every bottleneck, and sort by severity. Then ask: is this a problem that needs AI, or will a simpler deterministic solution work? If AI is warranted, what is the minimum viable AI system that removes this specific constraint? The wheel company case study is the archetype: the client asked for AI image recognition; the actual constraint was sales information access; the solution was a database integration. Start with the constraint, not the technology.
Evaluation Suites Are Not Optional
Enterprise AI deployments fail not because the AI is incapable but because they were deployed without systematic testing of edge cases. An evaluation suite is a structured test library that predicts what the AI will say across the range of possible inputs - including the adversarial, the embarrassing, and the off-brand - and validates that every output is accurate and brand-safe before launch. Building a Chuck GPT that sounds like Charles Barkley takes days. Building an evaluation suite that ensures it never says something FanDuel would regret took months. Any AI deployment at scale should budget accordingly.
AI Adoption Is 80% Human Change Management
The technology portion of an AI deployment is typically the smaller challenge. The larger challenge is human: securing belief in the system, training people to work in the new loop, building managerial capacity to oversee AI workflows, and addressing the fear and resistance that any major operational change produces. Andrew draws the explicit parallel to the early internet: the same fears accompanied the shift from paper to Excel to SaaS, and every wave of those fears was resolved by education, training, and demonstrated value - not by ignoring the human side of the equation. Budget at least as much for change management as for the technology itself.
FAQ
What does 923.ai do and what makes it different from other AI agencies?
923.ai is an AI studio that has been building custom AI solutions for enterprise clients since 2016 - before generative AI existed as a commercial category. They work on both the machine learning side (predictive, analytical systems) and the generative AI side (LLM-based applications and agents). What distinguishes 923 is nine years of experience building evaluation suites: systematic test libraries that predict and validate AI outputs before deployment. Most internal teams deploying chatbots and LLM tools do not build these, which is why so many enterprise AI deployments produce embarrassing or brand-unsafe results at scale. 923's core value is building the responsible infrastructure around AI, not just the AI itself.
What is Chuck GPT and what did building it teach 923 about enterprise AI?
Chuck GPT is an AI version of Charles Barkley built for FanDuel, accessible at chuck.fanduel.com. You can have a full conversation with it. The concept took three to four days to build - getting the model to replicate Barkley's personality and voice required extensive YouTube research and prompt engineering. Making it brand-safe - ensuring it couldn't discuss DraftKings, couldn't give betting advice, couldn't say anything that would embarrass FanDuel - took four to five months. The ratio is Andrew's core lesson: the capability is fast and cheap; the responsibility is where the real work and real value is.
Why does Andrew say AI adoption is 80% human?
In Andrew's experience, the technology component of a successful AI deployment is straightforward - pick the right model, build the right agents, connect the right data. The harder, more time-consuming challenge is the human side: getting people to trust the system enough to use it, training them on the human-AI-human loop (human initiates, AI executes, human reviews), building the managerial skills to oversee AI workflows, and working through the fear and resistance that any significant operational change produces. The parallel he draws is to the early internet era, when the fear of 'will this replace us?' accompanied every SaaS adoption. It never replaced people - it changed what they did. AI will follow the same pattern, but only if organizations invest in the human change management side with as much rigor as the technical side.
How should a founder find their first AI use case?
Andrew's methodology: map the entire business end to end and identify every constraint - every bottleneck, every place where work piles up or quality degrades. Sort those constraints by severity. Then ask, for each one: is this genuinely a problem that requires AI, or is there a simpler deterministic solution? If it requires AI, what is the minimum system that removes this specific constraint? Often the constraint that looks like it requires AI (image recognition, natural language processing) is actually a data accessibility problem that needs a database integration. Start with the constraint, not the technology, and let the constraint drive the solution.
What does Andrew think about the future of humanoid robots?
Andrew challenges the assumption that the bipedal, two-armed human form is the right design target for robots. The human body is extraordinarily difficult to engineer - balancing, weight-shift adjustment, the physics of upright locomotion - and for most tasks it is not the optimal form factor. His model: the Roomba. Singular purpose, perfect execution, no unnecessary complexity. He recommends asking 'what is the specific task, and what form factor optimizes for that task?' rather than starting from the assumption that it should look human. The same logic applies to AI agents: a SQL-writing agent does not need a face or a personality. It just needs to write accurate SQL when called.
What AI tools does Andrew recommend for founders and executives?
Andrew's primary recommendation: stop using standalone chatbots (including ChatGPT) as your main AI tool and commit instead to the AI integrated into your existing workflows. If your company uses Google Workspace, use Gemini. If it uses Microsoft 365, use Copilot. The AI value compounds through personalization - the more the AI knows about your work, your documents, your email, your colleagues, the better its outputs become. Copy-pasting between ChatGPT and your actual work tools prevents that compounding. The tools that live inside your existing systems will outperform external chatbots over time because they have context, and context is where AI value is generated.