Enterprise AI Adoption Hinges on Infrastructure, Trust, and Tailored Solutions
TL;DR
- Enterprise AI adoption lags significantly behind model performance and consumer use due to complex data infrastructure, workflow redesign, and trust requirements, necessitating a decade-long integration journey.
- Externally driven AI builds are twice as effective as internal efforts, challenging the notion that large enterprises can solely rely on internal teams for successful AI implementation.
- The AI talent market is constrained, with top engineers primarily working in startups and large tech firms, creating a significant risk for enterprises attempting to build from scratch.
- Pricing models for AI solutions are shifting towards "pay as it works" and proof-of-concept-driven sales, reflecting a move away from traditional SaaS models to prove value upfront.
- Specialized, niche data requirements are becoming critical for AI training, moving beyond generic labeling to complex, domain-specific expertise that requires sophisticated sourcing and validation.
- The future of enterprise AI lies in hyper-personalization, moving beyond off-the-shelf SaaS to solutions tailored to specific customer data and workflows, enabled by modular architectures.
- While model benchmarks show progress, enterprise AI adoption hinges on hyper-specific performance and trust, not generalizability, requiring rigorous fine-tuning and validation processes.
Deep Dive
Enterprises face a significant chasm between the impressive performance of AI models and their actual adoption, with only a small fraction of AI projects successfully reaching production. This gap is not due to a lack of AI capability but rather the complex realities of enterprise deployment, which demand robust data infrastructure, workflow redesign, and critical elements like trust and observability--processes that are only in their early stages and will likely take a decade to mature. The key to bridging this divide lies in shifting from a "science project" mindset to focused, business-led initiatives with clear KPIs, often leveraging external expertise rather than solely relying on internal builds, which can be two times less effective.
The current enterprise AI landscape is characterized by a proliferation of vendors, many of whom offer solutions that fail to deliver on their promises, evidenced by low accuracy rates in out-of-the-box agents. This necessitates a shift towards a "pay when it works" model, where vendors must prove value through proof-of-concepts or solution sprints before receiving payment. Invisible Technologies, for instance, operates on this principle, offering free eight-week engagements to demonstrate the efficacy of its modular AI software platform. This approach emphasizes hyper-personalization over standardized SaaS solutions, integrating diverse data sources and custom workflows to create tailored business solutions. The core challenge for enterprises is not technical expertise but rather the discipline to define clear operational metrics and ownership, anchoring AI initiatives in tangible business outcomes rather than broad, unfocused experimentation.
The future of AI adoption hinges on a more integrated, "for-deployed engineer" model that drives workflow embedding and customization, a realization that Palantir pioneered a decade ago. This contrasts with the traditional SaaS model, where successful companies relied on quick out-of-the-box setups and external system integrators. In the AI era, the value lies in adapting and fine-tuning models to specific customer contexts, which requires ongoing effort and specialized expertise. While many believe synthetic data will replace human feedback, the inherent complexity and contextual nature of many real-world problems, particularly in fields like legal services and multimodal analysis, necessitate continued reliance on human expertise for data generation and validation. This nuanced approach to AI implementation, focusing on tailored solutions and proven value, will ultimately determine which companies thrive in this rapidly evolving market.
Action Items
- Audit enterprise AI adoption: Identify 3-5 key barriers to successful implementation beyond model performance (e.g., data infrastructure, workflow redesign, trust).
- Develop a framework for evaluating AI vendors: Define 3-4 critical operational metrics for a specific use case (e.g., contact center) to measure ROI.
- Implement a "pay-as-it-works" pilot program: Offer 2-3 solution sprints for new AI initiatives, deferring payment until user acceptance testing and validation are met.
- Establish a cross-functional AI initiative team: Assign 3-4 operational leaders (not solely tech) to champion and drive AI adoption based on clear business KPIs.
- Create a runbook template for AI deployments: Define 5 required sections (setup, common failures, rollback, monitoring, ownership) to prevent knowledge silos and ensure maintainability.
Key Quotes
"I think the cognitive distance that has occurred over the last couple of years is model performance has increased exponentially... but the enterprise has not... I think the reason for that is deployment in the enterprise is a lot more than just models themselves it's the data infrastructure to support those models it's the redesign of workflows it's the process of figuring out which operational leader takes accountability for that and most importantly it's trust it's observability it's all the things that you know I spent a decade building things like credit models in in banking and in those cases you need to go through model risk management testing training validation and so I think that whole process is in the first inning in the enterprise."
Matt Fitzpatrick highlights a significant gap between the rapid advancement of AI model performance and its actual adoption within enterprises. He explains that enterprise adoption is complex, involving not just the models themselves but also the supporting infrastructure, workflow redesign, and crucially, trust and accountability. Fitzpatrick believes this entire process is in its early stages within the enterprise, suggesting it will take a decade to mature.
"I think the perception from external or from general kind of tech crowds is that internal teams for I don't know you name your boring large enterprise it's just really low quality you're not getting the top tier ai engineers you're not getting top tier devs is that true look I think the amount of talent that knows how to do this well is not large and so that that finite group mostly works in ai startups of various forms right and and large tech companies."
Fitzpatrick addresses the common perception of lower quality talent within internal enterprise teams compared to AI startups and large tech companies. He acknowledges that the pool of highly skilled AI engineers and developers is limited. Fitzpatrick explains that this finite group is predominantly found working in AI startups and major tech firms, implying that enterprises may struggle to attract and retain top-tier talent for their internal AI initiatives.
"I think the simplest advice I give and by the way this is how we sell quote unquote is start with proof of concepts start with we call solution sprints don't pay a dollar until you prove the tech works so like we don't actually sell anything we meet a customer who we say we will we will do it for free for eight weeks and prove to you the tech works and that's a very simple way if your tech works you'll show it it's an expensive way to do business it is and it's not but look so let me give an example like how one of our deployments works because I think it fair enough if the answer is that you know it takes you two years to build anything but like I'll give you an example."
Fitzpatrick advocates for a proof-of-concept-driven sales approach, emphasizing that clients should not pay until the technology is demonstrably effective. He explains that his company offers free eight-week "solution sprints" to prove the technology's value. Fitzpatrick acknowledges this is an expensive business model but asserts it's a simple and effective way to build trust and showcase the technology's capabilities before any financial commitment.
"I think the biggest one is just the view that when I first started this job the main pushback I always got was that synthetic data will take over and you just will not need human feedback two three years from now and it's interesting i don't from first principles that actually doesn't make very much sense if you think through it right if you think about the diversity of tasks that exist in the world and then how long it would take you to get comfortable with the accuracy it doesn't make any sense."
Fitzpatrick challenges the notion that synthetic data will entirely replace human feedback in AI development. He explains that this belief doesn't align with the fundamental reality of the diverse tasks AI needs to perform. Fitzpatrick argues that the complexity and contextual nuances of many real-world tasks make it impractical to rely solely on synthetic data for achieving the necessary accuracy and understanding.
"I think the way our educational system will shift will function will shift material we're a talent assessment company an enormous amount of people we bring in did not go to college and we assess them on cognitive aptitude and skill and so I think the really positive note I would leave on is I think the way that people learn the topics they learn the way we look at resumes and how to screen and assess people will move in a really positive direction and I think a very different one than we've had for the last 100 years."
Fitzpatrick expresses optimism about the future of education and talent assessment, particularly for those without traditional college degrees. He highlights that his company assesses individuals based on cognitive aptitude and skills, noting that many successful hires did not attend college. Fitzpatrick believes this approach signifies a positive shift in how learning, resumes, and candidate assessment are viewed, moving away from a century-old paradigm.
Resources
External Resources
Books
- "Seven Powers" by Hamilton Helm - Mentioned as a framework for understanding pricing premiums and competitive advantages.
Articles & Papers
- MIT report on GenAI deployments - Cited to indicate that only 5% of GenAI deployments are currently working in enterprises.
- Gartner report on enterprise GenAI projects - Cited to suggest that 40% of enterprise GenAI projects may be canceled by 2027.
- KPMG report on consumer GenAI usage - Cited to show that 60% of consumers use GenAI weekly.
- Salesforce AI Research report on out-of-the-box agents - Cited to highlight the accuracy limitations of current enterprise agents on single-turn (58%) and multi-turn (33%) workflows.
- World Economic Forum report on AI's environmental impact - Cited to suggest that AI will be a net positive for the environment.
- Johns Hopkins study on avoidable medical errors - Cited to report that 250,000 deaths per year are due to avoidable errors.
People
- Matt Fitzpatrick - CEO of Invisible Technologies.
- Harry Stebbings - Host of the podcast "20VC".
- Turing McCourt - Mentioned as a company that has reached several hundred million dollars in revenue.
- Francis - Founder of Invisible Technologies.
- Smash Khanna - Senior partner at McKinsey and a mentor to Matt Fitzpatrick.
- Pat Grady - Head of Sequoia Capital.
- Claudia - Girlfriend of Matt Fitzpatrick.
- Jason Lampkin - Co-host of a weekly news show on the Invisible Technologies business.
- Rory - Co-host of a weekly news show on the Invisible Technologies business.
- Sam Altman - Mentioned in relation to "war mode" discussions.
- Nick Saban - Mentioned as an example of successful recruiting in sports.
- Nick - Mentioned as a CEO with a "brutal culture" but focused on winning.
- Jamon - Mentioned in relation to a discussion about benchmarks.
Organizations & Institutions
- Invisible Technologies - Company focused on AI training and enterprise deployment.
- McKinsey - Consulting firm where Matt Fitzpatrick previously worked.
- Quantum Black Labs - A group within McKinsey focused on tech development.
- Sequoia Capital - Venture capital firm.
- Goldman Sachs - Financial institution developing its own AI tools.
- MIT - Massachusetts Institute of Technology.
- Gartner - Research and advisory company.
- KPMG - Professional services network.
- Accenture - Consulting firm.
- PFF (Pro Football Focus) - Mentioned as a data source for player grading.
- NFL (National Football League) - Professional American football league.
- New England Patriots - Professional football team.
- Palantir - Company known for its work in enterprise AI and forward-deployed engineers.
- Databricks - Company providing AI and data infrastructure.
- McCord - Mentioned as a player in the AI training space.
- Surge - Mentioned as a player in the AI training space.
- Chering - Mentioned as a player in the AI training space.
- SAIC Vantor - Partnered with Invisible Technologies on a US Navy project.
- US Navy - Collaborated on a project involving underwater drone swarms.
- Toyota - Mentioned for its production system and institutional memory.
- Uber - Used as an analogy for sourcing talent and price discovery.
- Lyft - Used as an analogy for sourcing talent.
- Airbnb - Used as an analogy for marketplace dynamics.
- VRO - Mentioned in the context of marketplace dynamics.
- Revolut - Mentioned as an example of a company with a potentially "brutal culture" but focused on winning, and as an AI-native business in the physical world.
- Y Combinator - Accelerator program mentioned for its recent class's revenue performance.
- Microsoft (MCP) - Mentioned as an example of an open architecture component.
- Pinterest - Mentioned as the former occupant of Invisible Technologies' San Francisco office space.
Tools & Software
- ChatGPT - AI tool for writing.
- Notion - Productivity software for documents.
- Gmail - Email service.
- Slack - Communication platform.
- Superhuman - AI productivity suite.
- Alphasense - Research platform for decision-making.
- Tegus - Acquired by Alphasense, a resource for expert insights.
- Daily Body Coach - Service for personalized nutrition, training, and psychology-based coaching.
- Aura Ring - Wearable device mentioned in the context of patient data.
- Google Sheets - Spreadsheet software, mentioned in the context of early commodity labeling.
Other Resources
- AI Productivity Tax - A concept describing the loss of context and time when using multiple AI tools.
- Reinforcement Learning from Human Feedback (RLHF) - A method used by Invisible Technologies to train large language models.
- Model Risk Management - A process used in banking for testing and validation of models.
- Synthetic Data - A type of data used for training AI models.
- Institutional Memory - A concept from Hamilton Helm's "Seven Powers" framework, exemplified by the Toyota Production System.
- Jevons Paradox - An economic concept related to increased consumption with advanced technology.
- FDE (Forward-Deployed Engineer) - A role focused on executing specific workflow builds.
- SAS (Software as a Service) - A software distribution model.
- GMV (Gross Merchandise Volume) - A metric related to sales volume.
- KYC (Know Your Customer) - A process mentioned in the context of banking operations.
- GLP-1s - A class of drugs mentioned in the context of healthcare.
- LBO (Leveraged Buyout) model - A financial model used for investment analysis.
- FICO scores - A credit score used in financial contexts.
- MS (Multiple Sclerosis) - A medical condition mentioned in the context of fundraising.