The Only Good AI Strategy: Modern AI Platforms, Not DIY Builds
At COLLIDE 2023, Graft CEO Adam Oliner said a good AI strategy must use all of a company's data, support experimentation, lift the whole organization and empower everyone. His conclusion, from a vendor's perspective, was to adopt a modern AI platform rather than build one.
By The AI Enterprise Magazine Team3 min read
Key takeaways
- Strategy is the top barrier. Oliner said the lack of a clear strategy is the number one reason companies are not adopting AI.
- Four properties. A good strategy is multimodal, supports experimentation, builds collaboratively and empowers everyone.
- Beyond LLMs. Text-only generative AI is a small subset of the broader foundation-model trend, he said.
- Building is costly. He cited AI talent at roughly a quarter million dollars per person per year, plus ongoing infrastructure upkeep.
- Not differentiating. Because successful AI companies build near-identical platforms, he argued companies should buy rather than build.
At COLLIDE 2023 in Atlanta, Adam Oliner, CEO and founder of Graft, argued that companies need exactly one AI strategy: adopt a modern AI platform that every employee can use, and do not build it in-house. In a 15-minute talk, he laid out the properties he said a good strategy must have, then made the case against building the platform from scratch. Oliner, who previously led machine learning at Slack, was also pitching his own company's product, so his conclusion should be read in that light.
Why do companies need an AI strategy?
Oliner listed the outcomes he associated with AI: higher productivity, lower costs, better decisions, new revenue from services or product features, more successful customers and a competitive moat built from data. Companies also face the risk of missing the technology wave entirely.
So why isn't everyone doing it? Oliner said the number one barrier to adoption is the lack of a clear strategy, with infrastructure and expertise also high on the list. His thesis: "In five years there will be two types of companies: those with an AI strategy and those that are dying."
What does "modern AI" mean?
Oliner defined modern AI as the way top AI companies had been working for years, before ChatGPT reached the mainstream. It is built on foundation models, which are large pre-trained models, and takes a data-centric approach that abstracts away much of the underlying machinery. It also covers the full life cycle, so teams are not left building in a notebook and then struggling to reach production.
He stressed that large language models and generative text are only a small subset of that broader trend. At Slack, he said, his team shipped dozens of use cases built on these methods, which he called a proof point for doing modern AI outside of Google, Meta and Amazon.
What are the properties of a good AI strategy?
Oliner offered four. First, it should be multimodal. He said 80% of data is unstructured, so a strategy should reach beyond tables to text, images, audio, video and graphs.
Second, it should support experimentation. He asked the audience whether anyone knew every AI feature they would ship in the next two years, and no hands went up. A platform should let teams test across many use cases, because a strategy built only around LLMs will leave companies feeling behind again when the industry moves to the next thing.
Third, it should be collaborative. Investments should compound, so the next project is easier than the last, and work done by one team should benefit others. Data silos should be intentional, he said, not accidental.
Fourth, it should empower everyone. Asked what share of a team should get AI "superpowers," Oliner said the only sane answer is 100%.
Why not build the platform yourself?
Oliner described the build path as daunting: identify dozens of tools across orchestration, storage and security, write glue code to assemble them, package them behind a user interface, deploy them reliably and maintain them for life. "Infrastructure is like a child," he said. "It needs care and feeding."
He estimated AI talent at about a quarter of a million dollars per person per year, and said a team would likely need several people, putting the people cost alone in the millions. He also noted that component choices cannot be made independently, because the pieces must work together as one system.
He cited data showing only 4% of enterprises had shipped any generative AI use case to production, and that most of those had shipped just one or two, with some having to withdraw them. Risks of getting it wrong, he said, include cybersecurity and reputational damage from leaking private information or making bad recommendations.
His strongest argument was strategic: successful AI companies end up building nearly the same system from similar components. "It's not differentiating," he said, and he called it a distraction from the core business. His conclusion was to use a modern AI platform, such as the one Graft built, so teams can focus on their own business.
Where to learn more
COLLIDE 2027 returns to Sandy Springs Performing Arts Center in Atlanta on Wednesday, October 13, 2027, and The AI Enterprise Conference 2027 takes place Tuesday, August 31, 2027, at Pier Sixty in New York City. Details are at /events/collide/ and /events/ai-enterprise-conference/. Registration is coming soon.
Frequently asked questions
What is a modern AI platform?
Oliner defined modern AI as a foundation-model-based, data-centric approach that covers the full life cycle from experimentation to production, and said a platform should be accessible to everyone in the organization.
What makes a good AI strategy?
He said it should use all of a company's data (multimodal), support experimentation across future use cases, build collaboratively so investments compound, and empower the entire team.
Should companies build their own AI platform?
Oliner advised against it, citing cost, risk, ongoing maintenance and talent needs, and arguing that the platform is not a differentiator. He represented Graft, which sells such a platform.
How many enterprises had shipped generative AI to production?
Oliner cited data showing only 4% of enterprises had shipped any generative AI use case to production at the time, most of them only one or two.
This article is based on remarks by Adam Oliner, CEO & Founder, Graft at COLLIDE Data + AI Conference 2023 on October 3, 2023. More coverage: COLLIDE AI Conference.
Filed under: AI Strategy & ROI, Data & AI Infrastructure
