Your AI strategy has a data problem

Why bad data is the biggest obstacle to enterprise AI adoption

Luke: Brad, when you’re starting an engagement on AI readiness, how do you approach it?

Brad: The first question I ask is: does your team believe in AI yet? Because that determines the strategy. If they don’t, you don’t start with architecture. You start with a small win. Find something you can demonstrate quickly that makes people say, oh, this actually works. That’s the catalyst.

If they’re already believers, then you architect toward your end state and start eating the elephant. But in either case, there’s a people component that doesn’t get enough attention. Identify the person or people on your team who know the data best. They’re often buried in operations or finance, not IT. Pull them out of their day job. Make this a priority. Give them real ownership. They have to be invested and incentivized for success. Without that internal champion, even the best external help stalls out.

AI governance and data security: a risk-based approach

Luke: Miles, how did you decide who gets access to AI inside your organization?

Miles: Deliver it to who’s asking for it. Those are your champions. The faster you let them run, the more momentum you build across the organization. You need guardrails depending on your regulatory environment, but the instinct should be to supercharge the people who want to use it, not slow them down. For roles that are heavily governed or regulated, you move more deliberately, test for accuracy, and co-author the process. But don’t let caution in one area hold back progress everywhere else.

Luke: What about risk and governance more broadly?

Miles: We evaluate data for AI based on blast radius. In a breach scenario: what’s the impact if the data fed to our models becomes public? We focus governance on the highest blast radius items first. PII, API keys, passwords, credentials: those have strict controls. CRM data is a different conversation. The question isn’t whether to govern, it’s where to govern hardest. You can’t treat everything like it’s ITAR-controlled or you’ll never ship anything.

Brad: On the heavily regulated end, we helped a government defense contractor build their AI strategy. Data classification and segmentation was everything. For ITAR-controlled data in an Azure CMMC High environment, Microsoft Purview is a powerful tool to perform the required classification before you connect any models. Once the data is controlled and protected, there are frontier models available within GCC FedRamp High environments. The sequence matters: classify first, deploy second.

The highest-value AI use cases in finance and operations

Luke: Brad, what are the highest-value AI use cases you’re seeing?

Brad: Back-office automation is table stakes at this point. Reconciliations, Excel hell in finance, budgeting, data integrations—everyone should be looking at this. You can realistically eliminate most manual Excel workbook in your accounting and finance function. The logic is almost always reverse-engineerable. Significant time savings, and frankly, better accuracy.

But the use case I’d put at the top of the underutilized list: connecting AI to your customer contracts. Most companies have years of contracts sitting in PDFs, completely unindexed and unsearchable in any meaningful way. Extract the key provisions. Make them indexable. Enrich with external data sources. That accelerates future contract creation, improves market segmentation, and generates decision-making insights that used to require a lawyer or a long weekend. A huge percentage of your downstream rules for accounting, finance, and operations can be derived from that data. It’s a great place to start.

On the sales side, AI is supercharging top-of-funnel activity, identifying opportunities earlier and improving conversion rates. Finance teams using AI for FP&A and forecasting are getting time back and better outputs.

The SPAC mechanics themselves weren’t the problem. What broke was the discipline around what went into the structure and how sponsor incentives aligned (or didn’t) with shareholder outcomes.

In 2021, a significant portion of SPAC targets were pre-revenue companies with ambitious five-year projections and very limited operating histories. A liability safe harbor for forward-looking statements meant those projections could appear in deal documents without meaningful legal exposure. Sponsors had powerful economic incentives to close deals because their economics were tied to closing, not to how the combined entity performed afterward.

And then came the redemptions. SPAC shareholders have the right to take their investment back from the trust account rather than remaining in the combined entity, even if they vote in favor of the deal. In 2021 and 2022, redemption rates in some transactions exceeded 90% of trust capital. Companies had built their post-combination operating plans around a certain capital base and closed the deal to find barely enough cash to fund near-term operations. By late 2022, the average de-SPAC transaction was down approximately 40%.

It exposed the fundamental misalignment at the core of the structure as it was being used at the time, and the SEC had started paying very serious attention.

How to build a centralized AI data infrastructure

Luke: Miles, you mentioned arming your champions with what they need. What does that look like at your organization?

Miles: The goal is ensuring they have the tools and the data for their specific use cases, whether that’s internal data or external subscriptions. We’re aggregating a lot into Snowflake as our data warehouse. It gives people a one-stop shop. Our business applications are still the system of record, but now we only have to integrate once, with controls, and we can turn on the right data for each AI use case.

The prep work is iterative. You understand the source data, figure out which attributes matter, and understand how it’s used in practice. There’s a tagging process by data domain that helps AI better understand unstructured content. And then there’s the feedback loop: you test for accuracy and inference quality, and you keep improving.

Luke: Can you give a concrete example of what your teams are actually doing with this?

Miles: We’re automating a lot of market intelligence work, aggregating what’s happening across the industry, what competitors are doing, tracking customer behavior signals. That used to require significant manual effort. Now it runs continuously.

One that caught us by surprise: we used Claude to recover tax refunds. We found out one county we operate in was still charging us for abandoned pipe that had been out of service for years. To figure out exactly how far back the clerical errors went, we put Claude to work on our historical documentation for the area. It pieced together the entire story and calculated a highly accurate figure for overpayments. A real dollar win from unstructured data nobody had looked at in years.

Minimum viable AI governance: what every company needs in place

Luke: What’s the minimum required setup for AI governance for a less regulated company?

Brad: For most companies, you don’t need much to get started. Here’s what we recommend:

Luke: What are they building?

Miles: Everything from project tracking dashboards to a CRM we just launched to replace Salesforce.

Why data readiness is a compounding AI advantage

Luke: Brad, make the case for urgency. Why should companies be organizing their data right now?

Brad: Here’s the cool part: once the foundation is in place, it works like a flywheel. Every refinement takes a little less effort than the last. The upstream processes, the downstream processes—each iteration just needs an increment. Your systems get smarter every cycle, not because the AI magically improved, but because the data is getting continuously refined through the governance and feedback loop.

There’s never been a better time to build. Start owning your data. The companies that invest in this foundation now are the ones who will compound the advantage over time, and the gap between them and organizations that wait is going to be significant.

10 actionable steps to get your organization AI-ready

Here’s the distilled advice from the session for leaders who want to move from conversation to action.

The bottom line on AI data readiness

The AI opportunity is real. But the companies that capture it aren’t the ones with the most aggressive AI strategy. They’re the ones with the cleanest data foundation underneath it.

What Miles and Brad described isn’t glamorous work. Building a data inventory, establishing a business glossary, deduplicating your ERP: none of that makes for exciting board slides. But it’s exactly what separates companies that get transformative AI outcomes from those that get expensive hallucinations.

The good news: you don’t need a perfect data environment to start. You need a starting inventory, a champion who knows the data, and enough governance to protect the blast-radius items. Then you start. Because the flywheel only works if you give it a push.

If you’re ready to get your data house in order, Embark’s AI consulting practice helps organizations assess their readiness, build the foundation, and identify the use cases worth moving on first.