Does Vendor Lock-In Belong in an Enterprise AI Data Platform?

An acquisition is just about to close this Friday. That means, by Monday, finance wants to know where the group is losing margins. The problem is the new enterprise functions on its own systems, and onboarding a single data source is quoted at weeks or months. Meanwhile, the foundation model your board approved last quarter sits on top of all of it, waiting for data it cannot yet read. 

That is the position one of Advancing Analytics' hospitality customers, a global food and beverage group built through acquisitions, was in before it changed how it brings data on board.

In this episode of the Don't Panic It's Just Data podcast, recorded live on the show floor at Big Data London, host Shubhangi Dua, Podcast Host and B2B Tech Journalist, speaks with Andrew Phethean, UKI Industry Lead, Enterprise Growth at Databricks, and Terry McCann, CEO of Advancing Analytics.

This podcast episode is especially relevant to CDOs, CEOs, and CEOs who are now beyond AI experimentation and are now pursuing plugging agentic systems into production. The speakers cover why the foundation of the platform’s model is the key factor when it comes to the outcome, what Databricks calls the four C's, and why the cost of doing nothing rarely makes it into the business case.

"I think a lot of the time enterprises are experimenting with the big vendors of the world and putting a lot of investment in that, but not necessarily putting that into production properly,” Phethean tells Dua.

How Should the Enterprise Invest in Scaling AI From Experiment to Production?

Phethean tells Dua that the latest models are not always what an enterprise requires. "We don't necessarily need them for all different business operations," he says.

“For us, it's about how you scale that motion where you can leverage, whether it be open source models or lower cost models, or machine learning,” he says. The main thing is to “leverage your foundational data at the bottom to bring out more value in what the enterprise is aiming to achieve.”

The problem is that many enterprises are experimenting and investing in big vendors but not necessarily putting that into production adequately. The question really is about how an enterprise should invest in the revenue-generating areas of their goals and not just focus on the cost savings or the efficiency driver aspect.

Alluding to an enterprise buying access to Anthropic or OpenAI, McCann says to look at its data. "You become unstuck very quickly," he says, without a governed platform that has been modelled in the language of the business. His customers, he says, often skip the plumbing entirely.

“It's like step one, buy the thing, step three, profit. And people generally forget about step two,” McCann tells Dua, alluding to a boardroom example.

Step two is where projects fall apart, the CEO of Advacing Analytics emphasises. It focuses on building the platform, followed by hosting and serving models to the business. The solution he says for instance he says be could be the Databricks' AI Gateway as one way to build a model and serve it out. 

Where Enterprise Investment Should Be Directed

Investment has to go into both the data platform and the foundation models according to the CEO. 

The theme that ran through Big Data London this year, according to both guests and the host, was context. Dua notes that her first visit centred on generative AI, the second on agentic AI, and this year on context. 

McCann further ties it to governance and how data used to be governed for the people in a business. Now, he says, it is governed "more for the agents who are working with us." Ontology, and tools such as Databricks Genie, add the layer of business meaning that keeps a model from returning "garbage in, garbage out."

What Are The Four C’s in Big Data?

Dua asks Phethean for the biggest challenge he faces. He says that the shift from experimentation to execution is the major challenge. He alludes to how Databricks frames the obstacles as four C's:

1. Context: Models need to understand the business they are answering questions about.
2. Control: Governance has to run across every layer of the stack.
3. Cost: Spend must stay manageable. "You're not token maxing anymore," Phethean says. The aim is to "value max."
4. Choice: Across vendors, clouds and LLMs. "Vendor locking is not a strategy," he says.

Dua adds a fifth from her own recent episode on the topic: keeping data current. 

Phethean then points to Databricks' newly released product, codenamed Reyden, which targets how quickly context can be drawn from data as it arrives. That means the data is current. Ultimately, all speakers agree that how fast an enterprise turns inputs into outputs will decide where it finds value.

LakeForge & Genie Code Cut Data Onboarding to Minutes

McCann says Advancing Analytics has been a Databricks partner for about seven years. Recently it became a Databricks Ventures portfolio company, with Databricks co-investing alongside the firm's primary investors. The shared aim, he says, is delivering value to customers in a way that is cost-effective.

Enterprises showcasing Big Data LDN are using many tools to solve one key problem, the CEO of Advancing Analytics tells Dua. Meanwhile, Databricks is forming the data ecosystem.

He exemplifies this by alluding to entity resolution for customer 360, security and SIEM, and retiring OLTP databases to run everything in one place. Customers, he says, do not want to buy fifteen different tools. They want one platform they know is continually invested in.

Working with Databricks and Genie Code, Advancing Analytics built an agentic platform called LakeForge to produce Databricks patterns quickly. According to the website, Databricks LakeFlow framework is powered by Advancing Analytics’ team of LakeForge Agents. “It gives you the speed of vibe-coding, with the guardrails of traditional engineering,” the website says.

It is now deployed with a hospitality customer McCann describes as one of the largest food and beverage companies in the world. The customer has global coverage, a long record of acquisitions and a spread of disparate sources. It needed data onboarding that was fast, reliable, and repeatable.

The starting point was "weeks or months to onboard a data source." McCann says LakeForge and Genie Code have brought that down to minutes. The business units can now see where they are selling and, as he puts it, "where they're hemorrhaging margin."

What CEOs Need to Pay Attention to?

The host asks what CEOs should pay attention to even when it does not move the numbers this quarter. Phethean points to untapped value in structured and unstructured data, and to getting foundations in place before attempting more advanced work. 

He also argues that investment should go to revenue-generating areas, not just cost savings and efficiency. UKI Industry Lead at Databricks calls it an innovator's dilemma. "You can't just improve that with AI. You've got to redefine it."

While McCann says, "It's a lot about trying to understand the cost of inaction.” Businesses that attend BDL, talk, and never commit watch competitors reach the market faster and better understand their customers. He adds that customers leave, and employees eventually stop wanting to work for a company that is not investing in the right areas.

"We often talk a lot about justifying the ROI of an investment into data or an investment into AI,” he tells Dua. “Part of the problem isn't necessarily justifying the ROI. It’s really about actually moving and committing.

On what the tech industry gets wrong about AI, McCann returns to the failure of project — step two, where boardrooms commit to a foundation model company and assume the outcomes will follow. Phethean adds the organisational side. Roles such as the CDO and the chief AI officer need a voice at board level, he says, along with cross-functional teams, which he calls tiger teams, pointed at the largest tasks in the business.

Buying a model is the easy part. Both guests argue that AI projects stall because the governed, well-modelled data platform underneath was never built, and that boards treat inaction as free when it is not. Databricks' four C's of context, control, cost, and choice describe what has to be in place. 

Key Takeaways

• Buying a foundation model without a governed data platform stalls projects.
• Newer models aren't needed for every business operation, says Phethean.
• Open source, lower-cost models and machine learning can carry production work.
• Databricks frames AI obstacles as four C's: context, control, cost, choice.
• Cost discipline now means maximising value, not tokens.
• Vendor lock-in isn't a strategy: choice spans vendors, clouds and LLMs.
• Data governance now serves AI agents as well as people.
• LakeForge cut data source onboarding from weeks or months to minutes.
• The cost of inaction can outweigh the ROI case for investing.
• CDOs and chief AI officers need a seat at board level.
• McCann's CDO advice is speed to market; Phethean's is translating IT to business.

Chapters

• 00:07 – Welcome and introductions
• 00:48 – How has Big Data London been?
• 02:32 – From generative AI to agents to context
• 04:33 – The biggest challenge: getting the plumbing right
• 05:49 – The four C's (and a fifth)
• 07:31 – The Advancing Analytics and Databricks partnership
• 09:52 – Customer story: LakeForge and Genie Code
• 11:22 – The shift CEOs aren't watching
• 15:08 – ROI versus the cost of inaction
• 16:44 – What the tech industry gets wrong about AI
• 18:38 – If your AI agent had a personality
• 20:02 – One takeaway for CDOs
• 21:30 – Closing and where to find the guests

For more information, visit: advancinganalystics.co.uk