AI Adoption for the Domain Expert Practitioner: What I Learned and How You Should Approach It
AI Adoption for the Domain Expert Practitioner: What I Learned and How You Should Approach It
- The domain expert advantage — why professional expertise is your most important AI asset
- Karpathy's framework: skill issue, December 2025, token throughput, jaggedness
- My approach since 2022: how domain expertise shapes AI adoption
- What my approach assumes — and what may be different for you
- Five pathways based on who you are
- The eight systems I use — and what each does best
- A structured approach to AI adoption for domain expert practitioners
- The token economy: how to think about value and cost
- Use case development: where AI generates real value in financial practice
- The frontier: agentic AI, AutoResearch, and what comes next for practitioners
- Glossary of AI terms for non-engineer practitioners
Why Domain Expertise Is Your Most Important AI Advantage as a Business Owner or CFO
I have been a CPA and CFO for 25 years. I have closed $700M+ in transactions, served as CFO for public companies navigating SEC reporting, secured $75.5M in debt financing for a multi-state cannabis operator, and managed nine-figure liquidity for a family office. That experience — accumulated across decades of high-stakes financial work — is not incidental to my AI practice. It is the foundation of it.
When ChatGPT became publicly available in late 2022, I adopted it immediately and applied it directly to my CFO advisory work. The reason I could do that from the start was not because I had AI training. It was because I had deep enough financial domain expertise to direct AI toward useful tasks, critically evaluate what it produced, and integrate the output into professional work that clients could rely on. That combination — domain expertise plus AI tools — is what creates genuine leverage. AI tools alone, without the professional judgment to direct and evaluate them, produce generic outputs. Professional judgment alone, without AI tools to accelerate and extend it, leaves significant productivity and analytical capability on the table.
This is the core insight I want to convey to business owners, operators, and senior practitioners reading this: your professional expertise is not a disadvantage on the AI learning curve. It is your primary advantage. The practitioners who will benefit most from AI are not those who understand how the models work — they are those who understand their own domain well enough to know when AI is right, when it is wrong, and exactly what to do with the difference.
AI tools amplify domain expertise. They do not replace it and they do not substitute for it. A senior practitioner with strong domain expertise and developing AI skills will consistently outperform a technical AI user with shallow domain knowledge on any professional task that requires judgment, context, and accountability.
Karpathy's AI Framework: Skill Issue, the December 2025 Inflection Point, and What It Means for Business Owners
Andrej Karpathy — co-founder of OpenAI, former director of AI at Tesla, and as of May 2026 a researcher at Anthropic — is one of the most technically credible AI practitioners speaking publicly. In a March 2026 podcast conversation on No Priors with Sarah Guo, he articulated the current state of AI adoption with a precision that every practitioner should understand. The podcast was recorded when Karpathy was an independent researcher; his views reflect his own and do not represent Anthropic or any other organization.
1. Everything Is a Skill Issue
This is the most important statement in the conversation for any practitioner approaching AI adoption. The bottleneck is not the AI's capability. It is the practitioner's skill in directing, sequencing, and integrating that capability into real work. For domain experts — senior finance professionals, operators, attorneys, physicians — this framing is clarifying rather than daunting. The skill required is not technical. It is the skill of knowing what to ask, how to specify it, and how to evaluate the result. Those are skills that experienced practitioners already exercise constantly in their professional work. AI extends the reach of those skills; it does not require replacing them.
2. December 2025 — The Agentic Inflection Point
Karpathy marks December 2025 as the moment when agentic AI crossed a qualitative threshold — from tools that were impressive in demos but frustrating in sustained use, to agents that could complete long-horizon tasks with genuine coherence and reliability. He describes having gone from writing most of his own code to delegating essentially all of it to agents, and not having typed a direct line of code since. For software engineers, this transition was immediate and dramatic. For financial practitioners, the equivalent transition is underway: research synthesis, financial model drafting, document preparation, and reporting workflows that were multi-day manual efforts are now structured as AI-assisted processes that run in hours. The December 2025 quality jump is the reason that distinction matters.
3. Token Throughput as the Productivity Metric
Karpathy describes feeling uncomfortable when he has subscription tokens left unused — because it signals that he is not maximizing his leverage. At the advanced practitioner level, the goal is to consume tokens as productively as possible. The bottleneck is not the AI's capacity; it is the practitioner's ability to direct that capacity toward high-value work continuously. This framing reframes the economics of AI adoption: the cost of a subscription is not the question. The question is whether the work you are doing with those tokens generates proportionally more value than what you would produce without them.
4. The Jaggedness Problem — Why Human Judgment Remains Irreplaceable
Karpathy's description of current AI models as simultaneously a "brilliant PhD student" and a "10-year-old" captures the current state with precision. Models are extraordinarily capable at tasks with clear, verifiable criteria — drafting structured documents, synthesizing research, running quantitative analysis. They are unreliable on tasks that require nuanced professional judgment, contextual awareness built from years of domain experience, and the kind of accountability that comes from a practitioner's name being on the work. For financial practitioners: AI accelerates the gathering and structuring of financial information. It cannot replicate the judgment that determines what that information means for a specific business, in a specific situation, with a specific set of stakeholders. That judgment is what clients are engaging when they retain a fractional CFO. AI makes that judgment more efficient — it does not substitute for it.
Karpathy's four-part framework — skill is the bottleneck, December 2025 was the inflection point, token throughput is the productivity measure, human judgment is irreplaceable for jagged tasks — maps directly onto the experience of a senior financial practitioner applying AI to professional work. The capability is there. The leverage comes from the domain expertise to use it well.
How I Applied AI to CFO and Financial Advisory Work Since 2022
My AI journey since 2022 has been shaped by one consistent priority: applying AI to financial domain work where I have deep enough expertise to direct it effectively and evaluate its outputs rigorously. The phases below describe how that application evolved — not as a story of a beginner slowly mastering technology, but as a senior practitioner systematically expanding the role of AI tools within an established professional practice.
2023
Phase 1 — Immediate Application: Financial Research and Analysis
Adopted ChatGPT immediately upon public release and applied it directly to financial research, earnings analysis, market research, and document drafting. My 25 years of financial domain expertise meant I could evaluate AI outputs critically from day one — catching errors, redirecting poorly structured responses, and identifying exactly where AI added value versus where professional judgment was required. Simultaneously began using Gemini, Grok, Perplexity, Copilot, DeepSeek, and Amazon's model to develop a comparative understanding of each system's strengths across financial use cases. All systems used on available tiers to maximize comparative learning.
Application — Year 12024
Phase 2 — Query Development and Use Case Expansion
Systematically developed query quality across all eight systems — sourcing professional prompts from LinkedIn, YouTube, and practitioner communities and adapting them to financial advisory contexts. Expanded use cases beyond research to include financial modeling support, client document drafting, valuation framework development, and investment analysis. The discipline of daily use across multiple systems produced a clear picture of which systems performed best for specific financial tasks. Query quality — the precision of instructions given to AI — became an explicit skill development priority, because domain expertise alone does not translate to effective AI direction without it.
Expansion — Years 2–3Phase 3 — Deep Professional Integration and Coding Capability
AI became fully integrated into every area of my CFO advisory practice: financial modeling acceleration, research synthesis, client deliverable drafting, website and content development, capital markets analysis, and investment research. Began using AI for coding assistance in Java and Python — applying AI to tasks that would previously have required a technical hire or outsourcing. Also developed investment analysis workflows where AI compresses research and synthesis time substantially, with professional judgment applied to every output before it informs any decision. The December 2025 quality jump in agentic AI tools marked a clear step change in the capability available for financial workflow automation.
Deep Integration — Year 4Forward
Phase 4 — Agentic Workflows and Personal Knowledge Infrastructure
Current focus: building agentic financial workflows where structured high-level inputs produce comprehensive outputs with minimal per-step intervention, and developing a personal WIKI-based knowledge system that serves as AI-searchable context for all professional and investment work. The goal is the principle Karpathy articulates: put in few tokens once in a while, and have a large amount of high-quality work happen on your behalf. For a financial advisory practice, this means AI-augmented research, analysis, and drafting workflows that produce deliverable-quality outputs at a fraction of the previous time cost — with professional judgment applied at the review and integration stage rather than throughout the production process.
Agentic — CurrentThis is a senior practitioner's four-year application of AI within an established financial advisory practice — not a beginner's learning journey. The phases describe expanding capability applied to professional work, not a progression from ignorance to competence. The foundation throughout was financial domain expertise. The AI tools extended the reach and speed of that expertise.
Is This AI Adoption Approach Right for You? What to Consider
My approach was shaped by specific circumstances: deep financial domain expertise built over 25 years, a professional practice where AI application generates direct client value, and the discipline to apply AI tools daily across multiple systems from the moment they became available. Several aspects of that context may not apply to your situation.
Deep domain expertise as the starting point. My ability to evaluate and direct AI outputs in financial contexts from day one is a direct function of decades of professional experience in finance, accounting, and economics. A practitioner earlier in their career, or applying AI in a domain where their expertise is shallower, will face a different learning curve. The AI tools are the same; the ability to direct them effectively depends heavily on what you bring to the interaction.
Daily professional application as the primary learning mechanism. I built AI skill through daily professional use — not through courses, not through tutorials, and not through experimentation disconnected from real work. That approach works when you have a clear professional context for applying AI immediately. If your use cases are less defined, a more structured learning approach may produce better results initially.
Multi-system comparative use. Using eight systems across multiple years produced a comprehensive understanding of the AI landscape that informs which system I choose for which task. For practitioners with limited time, going deeper on one system may produce better near-term results than spreading across many. The comparative understanding is valuable — but it has a time cost that not everyone can absorb.
Karpathy himself is explicit about uncertainty at the frontier: even as one of the world's most capable AI practitioners, he describes constantly feeling behind. If that is his experience, the appropriate posture for any practitioner is humility about the path and continuous learning — not adherence to any single approach, including the one described here.
My approach is a reference point, not a prescription. Read it as one framework developed under specific conditions — deep domain expertise, daily professional application, multi-system comparative use from 2022. The five pathways that follow are designed to address a broader range of starting points and situations.
Five AI Adoption Pathways for Business Owners, CFOs, and Professional Practitioners
The right AI adoption approach depends on your professional context, your domain expertise, and your available time. The following pathways are illustrative frameworks — not rigid prescriptions. Most practitioners will draw from several and adapt them to their specific situation.
The common thread across all pathways is the same: consistent professional application, deliberate use case development, and the discipline to keep learning as the frontier moves. The specific path depends on who you are and what you bring to the tools. Domain expertise is the differentiating variable — not AI familiarity.
The Eight AI Systems I Use in My CFO Practice — Compared and Reviewed
I use eight AI systems regularly across my CFO advisory practice and investment work. The following is my practitioner's assessment based on sustained professional use — not a technical evaluation, not a vendor review, and not a ranking. Different use cases will point to different systems as the most valuable. The comparative understanding I have developed across all eight is itself a professional asset — knowing which tool to use for which task is part of the skill.
No single system is optimal for every professional task. The comparative understanding I have developed across eight systems is itself a professional capability — knowing which tool to deploy for which task is part of effective AI practice. Karpathy's prediction of model speciation — specialized models for specific domains — suggests that the landscape will become more differentiated over time, making comparative system knowledge increasingly valuable.
A Step-by-Step AI Adoption Framework for CFOs, Business Owners, and Senior Professionals
The following sequence reflects how I would advise a senior practitioner approaching AI adoption today — informed by four years of professional application across eight systems and shaped by Karpathy's framework. It is specifically designed for domain experts: practitioners who bring deep professional knowledge and need AI to extend the reach and speed of that knowledge, not replace it.
Step 1 — Start with Your Highest-Value Professional Use Case
Step 2 — Select Your Primary System Based on Your Use Case
Step 3 — Build Query Quality Deliberately
Step 4 — Expand to Comparative Multi-System Use
Step 5 — Develop Professional Use Case Architecture
Step 6 — Maintain Continuous Learning as a Professional Discipline
This six-step approach is designed specifically for domain expert practitioners — those who bring deep professional knowledge and need AI to extend it, not those building expertise from scratch. The sequence prioritizes professional application from the first step, because domain experts learn AI most effectively through real work, not through general experimentation.
AI Subscription Costs and Value: How to Think About Token Economics as a Business Owner
When you use a paid AI system, you are paying for token consumption — the basic unit of text that AI models process. The economics of AI adoption for a professional practitioner are straightforward: does the value generated by AI-assisted work exceed the subscription cost? For a senior financial practitioner with clear professional use cases, the answer is typically yes from the first engagement.
Karpathy frames this from the productivity perspective: at the advanced practitioner level, unused tokens signal wasted leverage. The goal is not to minimize token consumption — it is to consume tokens as productively as possible. For a financial practitioner, that means structuring professional work so that AI is producing research, analysis, and drafts continuously, with professional judgment applied at review and integration rather than at every production step.
| Practitioner Profile | Primary Value Driver | Subscription Timing | Key ROI Metric |
|---|---|---|---|
| Senior Domain Expert | Domain expertise × AI speed — professional outputs produced in hours vs. days | Immediate — use cases are clear and value is direct | Professional output quality and time savings per engagement |
| Business Owner | Research, analysis, and drafting tasks removed from owner's plate | Early — once a single high-value use case is identified | Owner hours recovered and redirected to higher-value activities |
| Earlier-Career Professional | Learning acceleration and output quality improvement | When specific professional use cases justify the cost | Quality improvement in professional outputs relative to non-AI baseline |
| Technical Practitioner | Coding productivity — December 2025 quality jump makes ROI immediate in most cases | Immediate — Karpathy's transition applies directly | Development velocity and code quality per unit of time |
For a senior domain expert practitioner, the question of AI subscription economics is simpler than it appears: if your professional use cases are clear, the cost of a quality AI subscription is a professional development investment that generates returns from the first engagement. The more relevant question is not whether to subscribe but how to structure your professional workflows to maximize the value of every token consumed.
AI Use Cases for CFOs and Controllers: Where the Technology Generates Real Financial Value
The most common complaint from practitioners who have tried AI and found it disappointing is some version of: the answer wasn't good enough. This is almost always a use case or query quality problem, not an AI capability problem. Generic queries produce generic outputs. Professionally structured queries, designed around specific high-value tasks and evaluated with domain expertise, produce outputs that are directly useful in client engagements.
Karpathy's AutoResearch is the frontier expression of this principle at scale: identify the workflow, design the delegation, structure the input so that valuable output is produced with minimal ongoing intervention. For a CFO advisory practice, that principle operates at a smaller but equally meaningful scale: a well-designed research workflow, a structured financial modeling assistant, a client report drafting process — each built around a specific professional use case and refined through repeated application.
CFO Advisory Practice Use Cases
AI is integrated into every substantive area of my practice: capital raise preparation and financial model drafting, market research synthesis for client engagements, client deliverable drafting and refinement, SEC filing analysis and comparison, earnings and valuation analysis, industry research for specific transactions, and the website and article content you are reading now. In each case, the AI produces a research synthesis, a first draft, or a structured analysis — and professional judgment determines what to use, what to revise, and what to discard. The combination produces professional outputs at a fraction of the previous time cost.
Controller Function Use Cases
AI has material application across the Controller function — the area of financial practice most immediately transformed by the technology. In my Controller engagements, AI accelerates month-end close research and documentation, supports technical accounting research across ASC 606 revenue recognition, ASC 842 lease accounting, and other complex GAAP topics, assists in drafting financial statement footnotes and management reporting narratives, and compresses the time required to prepare audit support documentation. AI also supports accounting system evaluation and implementation work — synthesizing vendor capabilities, drafting requirements documentation, and producing comparison frameworks that would previously require days of manual research. The critical discipline in all Controller applications is the same as everywhere else: AI produces the research, the draft, and the structured analysis. Professional accounting judgment — built through years of public accounting experience and client engagement — determines what is correct, what requires adjustment, and what cannot be delegated. A Controller who uses AI to produce faster financial statements without applying rigorous professional review is not more efficient. They are faster at producing errors. The combination of AI speed and Controller judgment is the value. Neither alone is sufficient.
Investment Analysis Use Cases
I apply AI extensively to my personal investment practice: earnings transcript analysis, comparable company research, DCF model development, sector and industry research synthesis, and portfolio monitoring workflows. AI compresses the time required to gather and synthesize the information that informs investment analysis. Professional judgment — built over 25 years of financial analysis across public and private companies — determines what that information means and what to do with it. AI does not make investment decisions. It makes the information base for those decisions more comprehensive and faster to develop.
| Task Type | AI Leverage | What AI Does Well | What Requires Professional Judgment |
|---|---|---|---|
| Research Synthesis | Very High | Aggregating large information volumes, identifying patterns, summarizing sources with citations | Evaluating source quality, applying business context, determining relevance to specific situation |
| Financial Modeling Support | High | Formula development, model structure, sensitivity analysis construction, error checking | Assumption setting, business judgment, model interpretation — the CFO function |
| Document Drafting | High | Structure, first drafts, formatting consistency, tone calibration at scale | Professional judgment, accuracy of specific claims, client-specific context and relationship knowledge |
| Valuation Analysis | High | DCF framework construction, comparable company research, precedent transaction synthesis | Assumption validation, business quality assessment, negotiation context — the CFO's analytical judgment |
| Capital Raise Preparation | High | Market research, comparable transaction research, financial model structure, investor materials drafting | Lender/investor relationship knowledge, deal structuring judgment, negotiation strategy |
| Earnings and Investment Analysis | High | Transcript synthesis, financial ratio analysis, sector comparison, model updating | Investment thesis development, risk judgment, portfolio decision-making |
| Strategic Financial Decisions | Medium | Framework application, scenario generation, analytical structure | All final judgment — AI is an analytical accelerant, not a decision-maker |
| Controller Function | High | Technical accounting research, footnote drafting, audit support preparation, close documentation, system evaluation frameworks | All GAAP technical judgments — revenue recognition, lease accounting, complex accruals, and management reporting context require professional accounting expertise |
| Client Relationships | Low | Communication drafting, meeting preparation materials | All direct client interaction — trust, judgment, and relationship management are irreplaceable |
The use cases where AI generates the most value in financial practice are those where the task is information-intensive, the output is structured, and the quality evaluation requires professional judgment. That is precisely the intersection where a senior financial practitioner with domain expertise and AI tools is most effective — the AI accelerates the production, and the domain expertise determines the quality.
Agentic AI and the Next Frontier: What Business Owners and CFOs Need to Know Now
Karpathy describes the current moment as one of genuine psychosis — a constant state of trying to figure out what is possible, feeling behind, pushing further. He is not being dramatic. The frontier is advancing faster than any individual can fully track, and the practitioners who were actively engaged with AI tools before December 2025 are experiencing the quality jump from an established position. Those who were not are starting from behind.
The Agentic Transition for Financial Practitioners
December 2025 was the inflection point for coding agents. The equivalent transition for financial workflow agents is underway. Research agents that synthesize market information on a defined schedule, reporting agents that produce management reporting from structured data inputs, analysis agents that update financial models with new earnings data and flag material changes — these are not science fiction. They are the near-term application of the agentic capability that Karpathy describes for software engineering contexts, translated to financial practice. The practitioners building these workflows now will have a structural productivity advantage over those who wait.
AutoResearch — The Principle Applied to Financial Practice
For a CFO advisory practice, the AutoResearch principle translates to: design your financial workflows so that a structured high-level input — a company name, a transaction type, a market question — triggers a comprehensive research and analysis process that produces a deliverable-quality output, with professional judgment applied at review and integration rather than at each production step. The technology to build simple versions of this workflow exists today through Claude's Projects, structured prompt chains, and emerging no-code agentic tools. Building toward it is a current professional development priority.
The Jobs Question — What Karpathy Actually Says
On the question of AI's impact on professional employment, Karpathy is explicitly uncertain — and his uncertainty is more credible than confident predictions in either direction. He invokes the ATM/bank teller paradox: ATMs did not reduce bank tellers. They reduced the cost of running a branch, which created more branches and more teller demand. The analogous question for financial professionals is whether AI makes financial services cheaper and therefore more broadly demanded, or whether it directly substitutes for the professionals who provide them. Karpathy's honest answer: it is too early to know. The tasks will change. The demand economics are genuinely uncertain. The practitioners who build AI capability now are better positioned regardless of how that uncertainty resolves.
My Current Frontier — Personal Knowledge Infrastructure
My current development priority: building a personal WIKI-based knowledge system that serves as AI-searchable context for all professional and investment work — a structured repository of financial frameworks, transaction experience, market research, and analytical methodologies that AI can retrieve and synthesize on demand. The technology to build this exists today (Obsidian, Claude's Projects, RAG-based pipelines). The discipline of consistent curation over time is the hard part — and it is a professional development investment with a long compounding horizon.
The frontier is a moving target that Karpathy himself cannot fully keep up with. The appropriate response is sustained engagement: maintain active AI use in your professional practice, stay current with the practitioner community, and build toward more sophisticated workflow automation as the tools mature. The practitioners who will have the most leverage from AI in 2028 are those building the foundational capability and workflows now.
AI Adoption for CFOs and Business Owners — Common Questions
These are the questions practitioners most commonly ask when approaching AI adoption for the first time or evaluating whether to expand their current AI practice.
AI Glossary for Business Owners and Financial Practitioners — Plain-Language Definitions
Simplified definitions for general educational purposes. Includes Karpathy's framework terms from the No Priors Podcast, March 20, 2026. Definitions reflect usage as of May 2026.
| Term | Definition (Simplified) |
|---|---|
| Agentic AI | AI systems that autonomously plan and execute multi-step tasks toward a goal, without requiring human input at each step. Karpathy's "claw-like entities" that do work on your behalf. The next major phase of AI capability development for practitioners at every level. |
| AutoResearch | Karpathy's term for autonomous AI research loops — agents that design experiments, collect data, optimize, and self-improve without human involvement at each step. The principle: structure your workflows so that significant work happens on your behalf from a single high-quality input. |
| Context Window | The amount of text an AI model can process in a single interaction. Larger context windows allow work with longer documents and more complex instructions without losing earlier content — a key practical capability for financial work involving large documents and extended analytical tasks. |
| Foundation Model | A large AI model trained on broad data that serves as the base for many applications. GPT-4, Claude, and Gemini are foundation models. Most AI tools practitioners use daily are built on top of foundation models. |
| Hallucination | When an AI model generates factually incorrect information with apparent confidence. A known limitation of current AI systems. AI outputs must be verified with professional judgment before being relied upon in professional, financial, or legal contexts. Part of Karpathy's jaggedness problem. |
| Jaggedness | Karpathy's term for the uneven capability profile of current AI models — simultaneously a "brilliant PhD student" and a "10-year-old." Extraordinary at verifiable, structured tasks; unreliable on tasks requiring nuanced professional judgment. The reason domain expertise remains essential: professionals can identify and correct jagged outputs that a less experienced user would not catch. |
| Model Speciation | Karpathy's prediction that the current monoculture of large general-purpose models will give way to a more diverse ecosystem of specialized models optimized for specific professional domains. Implication: the best AI system for financial advisory work in two years may not be the best general-purpose model available today. |
| Prompt / Query | The instruction or question given to an AI system. Prompt quality — the precision, structure, and context of the instruction — is the primary variable under the practitioner's control. Well-structured professional prompts produce dramatically better outputs than vague queries. |
| RAG | Retrieval-Augmented Generation. A technique that enhances AI responses by retrieving relevant information from a specific knowledge base before generating a response. The technical foundation of personal WIKI and "second brain" AI systems — allows AI to answer questions from your own documents and professional knowledge base. |
| Second Brain | A personal knowledge management system — typically a structured note repository or WIKI — designed to serve as AI-searchable external memory. Combined with RAG-based AI, functions as an AI-augmented long-term professional knowledge base. A current development priority for advanced AI practitioners. |
| Skill Issue | Karpathy's framing of the current AI adoption bottleneck: the constraint is not the AI's capability — it is the practitioner's skill in directing and integrating that capability. "Everything is skill issue." For domain experts, this framing is clarifying: the skills required to use AI effectively are adjacent to skills experienced practitioners already have. |
| Token | The basic unit of text that AI models process — roughly 0.75 words. Paid AI subscriptions are priced based on token consumption. For professional practitioners, token economics are straightforward: does the professional value generated by AI-assisted work exceed the subscription cost? For clear professional use cases, the answer is typically yes. |
| Token Throughput | Karpathy's productivity metric: the rate at which you consume tokens productively. At the advanced practitioner level, the goal is to maximize token throughput — to have AI doing as much professional work on your behalf as productively as possible. The transition from worrying about token cost to maximizing token productivity marks a significant development in AI practice. |
Fractional CFO AI adoption guide: 8 systems, Karpathy skill issue framework, agentic AI explained, and how 25 years of CFO expertise makes AI work in professional practice.
