AI: Make or Build

Crafting AI strategy with Limited Resources

Chances are, if you are in charge of setting your company’s Data & AI strategy, the one question your CEO wants answered right now is the Return on AI (ROAI).

Here are some angles you might consider in preparation for your pitch to the senior leadership.

Depth vs. Breadth

The board and/or the senior execs will inevitably ask whether they should go for depth or breadth—i.e. should you focus on 1-2 core projects, or spread out and initiate numerous projects simultaneously?

Especially in the beginning, you want to focus on a handful of high-impact areas that you believe AI can provide strategic value and align with the core company objectives.

You might be tempted for the opposite—to pursue a multi-armed bandit strategy, where you hedge your bets on a larger surface area. You might even think Pareto—cast a wide net at first until you identify the impactful 20%, and then drop the rest of the projects.

Logical, but I disagree. You should be able to pick your battles. Not all AI initiatives are created equal. You know what your company sells, and what you core value proposition is. Block out the rest and bring AI to your core business first and foremost.

AI Your Core Value Proposition

You are probably familiar with the value vs. complexity crosstab, which is a great cheat sheet as a prescriptive aid for strategy leaders. You can use it as-is with AI, but what I find more helpful is switching out complexity in favour of core value proposition.

In this variant, the Y axis is unchanged—business impact or strategic value—but the X axis represents the spectrum of core value proposition for your company. For example, if you are a SMB SaaS company, it makes sense that AI can help with initiatives on lead scoring, churn prediction, or employee satisfaction (eNPS).

However, such initiatives are generic; all companies need and benefit from them. But they don’t influence the calculus of what your company is trying to accomplish.

I subscribe to the notion that the products we pay for in business are necessary evils. To steal Product lingo, the customers have a job to be done (JTBD); and when they pay for your product, it means currently, you are the best worst option they know about.

Let’s take the concrete example of lead scoring. Your company probably generates a bunch of leads every month, of which a much smaller percentage is actually meaningful. And if there has been no or limited data science initiatives in-house, your RevOps & Marketing colleagues might be either i) doing it manually, or ii) using a third-party vendor to score the leads using limited data from Salesforce.

Jumping to the outcome, let’s say using AI, your team improved the lead scoring process by x%, leading to some commercial efficiencies and perhaps positive, demonstrable ROAI.

This is not bad at all. Most data science/AI initiatives don’t see the light of day in production—so congratulations!

However, it is unlikely that you achieve strategic value by improving lead scoring efficiency (unless, you are in the business of lead scoring). Yes, it makes a difference, and yes, it is easy to quantify the ROIA. It is also a great quick-win scenario, when you need to establish trust and credibility after starting a new position.

In contrast, let’s look at a scenario where you apply AI to your core value proposition. This is your reply to your prospective customers’ JTBD; it is your best worst option offering to their problem. Again, let’s jump to the outcome that we have scored some improvement using AI on your core value prop; now you do it y% better, faster, bigger etc.

This is a much likely scenario where you achieve strategic value, because you improved the underlying process instead of a surface-level commercial interaction.

How far can you push lead scoring by optimising lead scoring? Without touching the core product, you are making any meaningful improvements to what you are selling. If you are getting 1,000 leads/month, and now you can qualify 110 with AI rather than 100 without. It’s a nice improvement, but it doesn’t affect the big picture.

OTOH, if you improve your core offering, then you can see significant changes in your lead numbers—that’s strategic value delivered and ROAI well-deserved.

As much as possible, try to constrain yourself to only build if something has high strategic value and it is a core value prop for your company. This is where your competitive advantage lies. If you have any resemblance of a moat, this is it—the data and processes you have been capturing for years. Of course, all your competitors are in the same situation. But potential != certainty. We know that not all who have it will be able to use it, so press your advantage (while you have it) and start building.

If something is a core value prop but has low strategic value, keep an eye out on potential interactions and cross-collaborations across the company. You want to build a coalition around the perceived business impact of such initiatives, before you take on the cost of building them out.

For anything that is not a core value prop—don’t build. If they have high strategic value, consider the option of buying them—either for their people or for their data.

Finally—and this could be your busiest quadrant—when things are of low strategic value and not a core value prop, you want to outsource and manage costs. Look for affordable third-party vendor plugins that offer ‘AI-powered’ solutions to your problems. Caveat emptor—it could be difficult to ascertain whether the AI powering the solution is actually AI; but hey—if it solves your problem, do you even care?

Amp Up Your Data & Analytics Game

Seriously. Data & Analytics (D&A) is fast becoming that neglected sibling in the family. Of course, data professionals know that without proper D&A, there is no way you will see positive ROAI. But senior leadership may not share that view, which tends to manifest as being comfortable in throwing a lot of money if it’s AI, and a lot scrutiny for any D&A related budget item.

If this is you, you need to revert this narrative, fast. Without high data quality, robust and resilient data infrastructure, enforced data governance, and a non-zero-sum data culture, you will find it difficult to show any ROAI. For more on this topic, check out my previous article on the Four Horsemen of Data Apocalypse:

TL;DR: Good data is the fuel for AI.

Foster an AI-driven Culture

The AI inflection point is real. This is much worse than the data science hype cycle of 2010; which was primarily limited to industry job seekers and companies that has no business hiring data scientist employing them in droves.

Now, AI is everywhere, for everyone. Up until several years ago, some of us were still correcting people who call any run-off-the-mill data science or machine learning ‘AI’; not the battle is lost, forever (Skynet is now AGI).

Unless the senior management enforces a policy of ‘no AI’, expect everyone to use AI tools, from the most mundane task all the way up to their actual one job (cue in company time-series data thrown into ChatGPT).

Given this expectation, embrace it. Yes, it is tiresome to read an article that regurgitates a summary of another that was also written by AI. See this as an opportunity for being the change you want to see.

If you are a data professional, write a jargon-buster and share with everyone in the company. To many non-data peeps, the entirety of data science/machine learning/AI is one big sameness. Provide examples on types of data—numeric, text, image, audio—and how embeddings change the game by enabling similarity search for unstructured data. Quickly demonstrate supervised learning—inputs and outputs will suffice. Explain autoregressive LLMs—how they predict the next token given some input—so that others may intuit which business cases are more inducive to LLMs. Highlight the power of data and emergent characteristics of powerful models; emphasise that the models are not causal but correlational.

You don’t want AI to be perceived as magic, and yourself a magician. To make your job easier, start breaking barriers and lower entry requirements. Language is important! When a non-data person jokingly blurts that they are too stupid to understand, rebuke them immediately—don’t let them get away so easy!