AI Doesn’t Have To Be That Hard, Fivetran CEO Says

Complexity seems to be part and parcel of the AI game these days. New technologies demand new tools and new platforms, with a host of new skills to bring it all together. New business models are springing up around AI, with new ways of measuring success. AI can seem so overwhelming, but it doesn’t have to be, says Fivetran CEO and Co-Founder George Fraser.

Fraser co-founded Fivetran back in 2013 to address the complexity around data integration, specifically the extract, transform, and load (ETL) process of taking data from operational systems and putting it into a data warehouse (or a data lake). Fraser acknowledges that everybody hates ETL because data pipelines are brittle and prone to breaking, but he insists that Fivetran is different.

“It’s funny to be in the business of selling something that people sort of despise. They don’t despise us, but they despise the need to do it,” he says. “[ETL] is a thing that’s been around forever. It’s not going anywhere, and it can be a pain–although if you use Fivetran, it’s a pain for us, but it’s not a pain for you.”

As companies embark upon AI, they’re rediscovering the joys of technological complexity. Fivetran has a front-row seat into many of these initiatives, and it’s not always a pretty sight.

“Sometimes I think people want this to be more complicated than it has to be,” Fraser tells BigDATAwire in an interview this week. “I’m not saying it’s just like super easy, in which case, why has not everyone done it? But I think one of the reasons sometimes why do people struggle is sometimes they have these mega initiatives with everything in the world. I’m like, well, that project is not going to succeed.”

Fivetran co-founder and CEO George Fraser

Gartner recently predicted that 40% of current AI projects will fail by the end of 2027. Just like with the big data wave before it, companies often get infatuated with new technology, which makes them susceptible to mission creep. The devil lives in the details, and he thrives when there are lots of them.

“Sometimes they go out of their way to make it more complicated because it’s kind of some sort of Skunkworks thing,” Fraser adds. “And they’re really more interested in using new technologies than they are in solving a problem.”

If you’re thinking about developing your own LLM, training an LLM, or even fine-tuning an existing one, you’re probably doing it wrong, Fraser says. “My opinion is there’s very few companies in the world that should be training their own language models,” he says.

Most companies should just be users of AI, not developers of it, he says. In fact, most companies already have many of the tools that they will need to build a basic AI application, such as a chatbot or agent that accesses a company’s knowledgebase, Fraser says. There’s no need to go out and buy more.

“What I’ve seen be super successful with that is leverage your existing data stack. Use Fivetran, use your data warehouse, or your data lake if that’s the direction you’ve gone,” he says. “If you leverage the tools you already have, it makes it a lot easier. You can get this up and running pretty fast, if you’re trying to do this enterprise knowledge base thing.”

The basic pattern is this: Get all your data together in one place, such as the data warehouse or the data lake, which you probably already did, Fraser says. Use your ETL tool to transform it into a shape that’s ready for AI. That shape is usually a pretty simple one.

Creating AI knowledgebase apps doesn’t have to be hard, Fraser says (Jacek-Kita/Shutterstock)

“It’s like a very tall, skinny table with not a lot of columns, and one of them is a text column, and that’s the thing you’re searching,” Fraser says. “It’s almost disappointing to people. They want it to be more complicated. And I’m like, guys, a really great tool for data management is SQL. And you take your existing data warehouse or data lake and you write like a big freaking union query that pulls it all together. And that’s the thing that’s going to feed your AI pipeline.”

You don’t need anything fancy to store the data that’s going to become the knowledge base, which is primarily text data. Fivetran is moving a lot of data into data lakes and lakehouses these days, and transforming data into Apache Iceberg table format. But there’s nothing stopping you from using your good old pre-existing database to house text data as a blob, or a binary large object.

“Relational databases are very good at storing text blobs like, since like Oracle v3. This is not a new function,” Fraser says. “I deny the supposed contradiction between relational and text data. Text data lives just fine in a relational schema. And then you plop your search application down on top of that, and it works super well. We have it at Fivetran. People love it.”

That doesn’t mean things can’t go wrong. Fraser saw one company build an elaborate data pipeline to shuttle PDF documents into a data warehouse that was serving as a knowledge base for an AI search application. “The project was a big success, but guess what? At the end there were 300 PDFs,” Fraser says. “There were so few [PDFs] and then there was tons of data in Salesforce and their support system.”

Most of the data that companies want to feed into AI already exists as text in the systems of record apps, Fraser says. That data can be replicated just as easily as tabular data residing in databases, or data pulled over a SaaS application’s API, he says.

Many companies are building AI apps using the retrieval augmented generation (RAG) pattern, but that pattern is going by the wayside, Fraser says. Instead of creating embeddings from existing knowledge and then “comparing the sort of approximate semantic content of the two documents” and hoping for “some kind of overlap in this abstract high dimensional space,” companies are finding success with the “self-talk” pattern, i.e. reasoning models such as OpenAI o3.

Use AI tools; don’t develop them (Ole.CNX/Shutterstock)

“There’s a better thing to do, which is you have the language model do this self-talk pattern where it goes and it says, ‘The user asked this question. What should I do to answer this question?’” Fraser says. “Not only can you search all the text documents, but if you want to, you can search specific text documents. You can search our documentation. You can search our internal wiki. You can search our opportunity notes in Salesforce. Then it can be more precise about the searches it’s doing right, so I think that’s sort of where things are headed.”

The number one thing that companies can do to succeed with AI is to get software engineers to use AI tools, says Fraser, who is a 2023 BigDATAwire Person to Watch.

“That is probably the single most important thing for any company that writes software to be to be doing with AI right now, is just internally using the AI tools that are available,” he says. “Don’t build your own. Just go adopt the tools from the most popular providers.”

As a software tool provider, Fivetran is also on the road to AI adoption. But since it has more than 5,000 paying customers, the company needs to be sure its code is bug-free.

“It hasn’t worked yet, but we’re trying to use them more,” he says. “It’s like having an infinite supply of software engineers who are super hardworking and will do whatever you tell them. And they type really fast, but they’re kind of dumb so you’ve still got to do the architecture piece and you’ve got to constrain them. That’s how you make them succeed.”

Eventually, we’ll get to the point where Fivetran’s connector code is all AI written. “But it has to live within this platform that constrains them and makes sure that everything follows these key best practices,” Fraser says. “So that’s the future we’re trying to build towards.”

Fivetran Raises $565 Million, Buys CDC Vendor HVR

Can We Stop Doing ETL Yet?

The post AI Doesn’t Have To Be That Hard, Fivetran CEO Says appeared first on BigDATAwire.