Council Post: Six Lessons From Two Years Of Building An AI Adoption Program
Ankur Pal is Head of AI at Aplazo, an AI-native buy-now-pay-later fintech platform serving millions of customers across North America.
getty
At nearly every conference I attended in 2022, board members, cofounders and AI leaders asked the same question: What's the ROI on AI?
Nobody had a convincing answer, my company included. We had always made investment decisions through a strict ROI lens, and that discipline became a bottleneck applied to a technology whose value was still forming.
We began thinking differently about AI ROI after an investor arranged a conference, in collaboration with a frontier AI company, for his portfolio's growth-stage companies, bringing together more than 50 founders and AI leaders. One insight reframed the problem: Staying at the innovation frontier matters more than a clean ROI case. When it comes to AI adoption, focusing on ROI can often matter less than losing ground while you wait to figure it out.
I brought that thinking back to our leadership working session and won sponsorship to invest aggressively in AI org-wide. But here's the lesson for other leaders facing the same problem: Getting buy-in from the top was the easy part.
Diagnose before you standardize.
Before building any plan, you need an actual starting point.
Our starting point was that people were already using AI tools, some paid out of pocket and others provided by us, with no org-wide visibility. Knowing this, we ran quarterly surveys, designed by the CTO and me, covering tool awareness, usage, value and support needs.
Three things came out of the survey: There was a good deal of shadow spend; there was a wide gap between power users and everyone else, telling us to raise the floor; and usage mapping revealed heavy tool duplication, which was fixed by choosing department-specific platforms.
By identifying the key issues, you can begin to shape your overall AI strategy. For us, this meant identifying a small set of high-leverage platforms (Copilot, OpenAI and Claude) and then rolling out licenses org-wide without over-optimizing first.
Give it structure.
Momentum without structure fades. The best way to create structure is through an AI steering committee. Our steering committee consisted of the cofounder, CTO, CPO, head of people and myself, which met biweekly to plan and track direction.
To give the committee guidance, I suggest conducting surveys that are blunt about capability. We found that people didn't know how to use what they'd been given. To solve this issue, we onboarded a third-party training platform combining self-serve courses with live trainer support, setting a six-month target for core completion, expecting completion to translate into real use.
Open doors create risk, so build for them early.
Broad access creates real exposure: data privacy, IP, hallucination and production risk. Building a dedicated AI productivity function, therefore, can be essential for guiding end-to-end implementation for engineering and product.
With that safety net in place, let the organization experiment. We started quarterly, theme-based generative AI and agentic AI hackathons, with every department fielding a small cross-functional team spanning business, product, AI and engineering. Every cycle surfaced three to four strong ideas, several shipping into production.
Biweekly town halls are also useful for sharing AI success stories. Often, a shipped peer project can move the organization's baseline more than any policy.
Setbacks can teach more than success.
Three honest mistakes shaped our understanding of AI as much as anything that worked.
First, a few enthusiastic teams shipped generative AI and agentic AI features fast and superficially, creating real second-order risk—which is why the AI productivity function exists today.
We also didn't monitor token and credit usage closely enough at first, until tiered usage monitoring and a clearer optimization framework corrected it.
Finally, urgency drove duplication. Business teams grew impatient waiting on in-house proofs of concept, so they independently adopted third-party platforms solving problems we were already building toward.
A democratized build-buy-hybrid decision framework can help solve this final problem, as it gives every team a shared way to choose between building, buying or blending both.
Building the muscle, not just the habit.
To ensure consistency across the organization, make sure no team has to rebuild a capability from scratch.
Our AI operations team built the platform and infrastructure powering production generative AI and agentic AI, which spanned PromptOps, RAGOps, SafetyOps, evals, AgentOps and agent governance.
It can help to create a mentor-mentee program for the team building this infrastructure, matching technical experts with engineers on dedicated three-to-six-month projects.
Look at the numbers.
Finally, I suggest tracking adoption in three tiers rather than one number:
• Level 1 is access and basic onboarding.
• Level 2 is regular weekly use.
• Level 3 is embedded, value-generating use with a measurable output.
By measuring productivity through half-year performance cycles, you should look to see an increase of usage at each level. With the strategies shared above, 90% of our team reached Level 2 in the first year, and 40% reached Level 3. By year two, 100% had reached Level 2, and 78% had progressed to Level 3.
What I'd Tell Another AI Leader Or Founder
None of this came from one decision, but from a dozen smaller ones, several of which we got wrong before we got right.
If your board is asking for the ROI number today, my advice: Don't wait for one clean answer. Build the diagnostic surveys, the governance structure, the training, the risk function and the tiered measurement first.
The number becomes real, and defensible, only once that system exists.
Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?