You’ve got to have it, but you’re not sure exactly why.

It’s a familiar feeling in the early stages of most technology innovations. We had to have a computer on our desk long before they really made us more productive. We had to get “on the internet” before we figured out how to conduct business over it. We needed a portable phone years before cell phones changed our lives — business and social — forever.

Artificial intelligence (AI) feels a bit like that to me. While we are making incredible progress, many companies are still at the “we have to have it” stage but haven’t answered the critical “why?” Agentic AI raises the stakes even higher, because it doesn’t just generate suggestions; its agents can act autonomously inside your workflows, for better or worse.

I read some recent McKinsey research that finds about 88% of organizations now use AI in at least one business function, yet roughly two‑thirds are still stuck in pilots and only a small minority report meaningful, enterprise‑level financial impact. Not surprising, as it tracks with my experience.

At Calix, we’ve spent the last year moving AI from experimentation into everyday operations. We face the same basic challenge: How do you help agentic experiments make it into production so they can drive real business results? Our people enthusiastically built about 700 AI agents on their own, but most didn’t make it out of the experiment phase. Forty were different: They solved a real business need in a workflow context, even if they weren’t the most technically impressive.

The biggest barrier we’re facing has nothing to do with the AI models. Agentic AI will support a true reinvention of how work flows across a company — not how we improve individual pieces of our operations. I think of tapping AI for incremental improvements as “little AI,” and it yields little returns. The real opportunity is what I call “big AI,” where we use agentic AI to create an entirely new way of working that involves cross-functional workflows or lifecycles. That’s what unlocks real value. That’s the reason we have to have AI.

Here are four questions learning and development (L&D) leaders must answer before AI can deliver business impact.

Question 1: Where, exactly, should AI change the work?

“In general” or “everywhere” are not answers. But neither is finding small pockets of a process that could use an efficiency boost. You can’t just drop AI into whatever tools we already have and hope something good will happen. Wiring a model into a ticketing system or attaching a bot to a learning platform without ever asking how the entire workstream flows won’t change operational models. And, in reality, very little work runs in a neat line from A to B to C. Experienced people adapt. They may skip steps, loop back, jump ahead based on context, and make judgment calls that never show up in the official process map.

For AI to really change the work, you need to be able to examine how work flows across functions, from initial customer interest to sales, to marketing, to product development and fulfillment — all with an eye on driving specific outcomes. Agents can reroute steps, parallelize tasks and move decisions closer to where the information lives, but only if you’re willing to see where agents might improve a workflow by redesigning it.

That’s agentic AI’s biggest potential: changing how work flows in the first place. And herein lies the biggest challenge, and the biggest opportunity for L&D professionals: There are very few people in companies today who have the deep insight across workflows to enable cross-functional transformation. Many operations experts have too narrow a purview. They are masters of incremental improvements to the part of the workflow they know inside out. But step outside the piece of the workflow they know, and they can’t grasp the potential of the larger transformation agentic AI can support.

We’ve seen this at Calix. We have a leader who moved from sales operations into IT. Because he understood the sales workflow end-to-end and the technology, he was able to identify rather than the smaller, local savings that teams doing narrower, function-by-function optimization surfaced.

This combination — owning the business workflow end-to-end and knowing the tools — is rare, and it’s the actual skill L&D should be building. AI will reward leaders who can define end‑to‑end customer or employee lifecycles and redesign flows. When L&D designs programs that give leaders a cross‑functional view of entire workflows and the systems that support them, you’re training the people who can put “big AI” to work for the company.

Question 2: Who owns the outcomes, and the accountability?

This gets harder when you recognize that agents are different from other technologies. Think of them as a non‑human agent workforce embedded in real workflows: They hold goals, interpret context, make decisions inside systems, and coordinate with people and other agents. They are much more than another productivity tool for humans to wield.

Recent research shows, though, that treating AI like a teammate can have unwanted side effects. In a large‑scale experiment, managers reviewing documents did something subtle but important: When the drafter was described as an “AI employee” instead of an AI tool, they shifted blame toward the AI, took less personal responsibility, escalated more often and missed more errors, even though the underlying system hadn’t changed at all.

That gives us two opposite, equally bad patterns. One is the classic “moral crumple zone,” where humans absorb blame for failures in complex systems they barely control. The other is the mirror image we’re seeing now, where “the AI” takes the blame and no one is truly accountable. Neither of those outcomes works.

Agents only create durable value when every meaningful workflow has a clear outcome owner who knows three things:

• Which outcomes they are on the hook for
• What the agent is allowed to do autonomously in that workflow
• When they must step in, approve or override

And the bigger the workflow, the more that ownership has to span functions. For “big AI” and true transformation, you need someone whose role is to own change across flows, not just within one department — a transformation lead who can hold that cross‑functional accountability.

For L&D leaders, that means teaching people how to use agents. But it also means developing them into transformation leaders, or people who can hold end‑to‑end accountability and negotiate changes across functions.

Question 3: How will we measure real business impact, not just activity?

Do not measure AI like you’d measure a new app. Usage, logins or similar measures won’t tell you whether agents are improving performance or learning outcomes in a way that matters to the business. Recent analyses of AI programs show that most companies track technical performance and basic adoption, but very few connect AI all the way to operational key performance indicators (KPIs) and financial impact.

For example, at Calix, the agents that we’ve picked to move into production have one thing in common: We can point to concrete changes in how work was getting done and the results we were seeing. A team built an agent to handle first‑pass triage on incoming requests. That’s a perfect use of agentic AI because it targets an area of repetitive judgment that consumes so much time for a person but is a snap for AI. We saw productivity across that workflow jump 20%.

We also had our supply chain team build an agent to look for the initial symptoms of an imminent customer complaint. Agent anticipation really worked: We were notified in time to act before the customer logged an issue, which cut downstream escalations and avoided costly remediation. Those are the big AI metrics we care about: fewer downstream incidents and lower remediation costs across a named and transformed workflow. Tracking these higher level metrics will keep us focused on progress against our goal of . We now treat that supply chain flow project as a booster trial for the rest of the business.

Many teams measure “little AI” adoption metrics — agent runs, active users, prompts — because they’re easy to count. And they should be tracked. But for real impact (big AI), you need a different class of metrics: workflow performance, role capacity and business outcomes that matter to your chief financial officer and operating leaders.

• Workflow metrics:cycle time, throughput, error rates and escalation volumes in named, end-to-end workflows with embedded agents. These measures help determine whether agents are improving workflow performance, reducing friction or accelerating outcomes across the end-to-end process.
• Role metrics:capacity and scope changes for human roles that signal transformation, not just tool use, such as hours shifted from low‑value tasks to redesign and negotiation work, number of cross‑functional workflows a single leader now owns end‑to‑end, and reductions in handoff friction or rework across the flows they lead.
• Business outcome metrics:the hard numbers your CFO already cares about, including total time‑recovery value (hours × fully loaded cost), cost or margin improvements tied to specific workflows (fewer expediting fees, overtime, or remediation costs), and revenue or volume growth handled without proportional increases in headcount or operating expenses.

For L&D leaders, this means another opportunity for training innovation. Just as big AI requires new skills in cross-functional workflow understanding, it also requires people who know how to measure the effectiveness of transformed workflows. Leaders need to think in terms of cycle times, throughputs, escalations and anticipatory signals as a whole, not just generic piece-part productivity scores. They’ll need to connect flow metrics to changes in their roles and, ultimately, to the business outcomes their CFO cares about.

In other words, measurement becomes a transformation skill: The ability to design metrics that show how agents are changing the work, not just how often agents get used.

Question 4: How will we build trust and fluency at scale?

It’s tempting to make AI feel friendlier by treating agents like colleagues. Some companies have given AI agents names, listed them on org charts and refer to them as “AI employees.” Recent HBR research suggests that move can backfire. Framing AI as an employee reduced individual accountability, increased unnecessary escalation, lowered review quality and made managers more uncertain about their own roles, without improving adoption. That’s a tough trade: less clarity and less trust, with no real gain.

At Calix, we’ve taken a different path. We do think of agents as part of a workforce embedded in workflows, but we treat them as systems with very clear guardrails. Each agent has an identity, acts on behalf of a specific person or team, and is given a narrow, well‑defined scope of work. Just as important, we log every meaningful action so we can see what the agent did, why it did it and undo it if necessary.

For L&D leaders, that trust model yields a concrete teaching agenda. People need to learn the tool and how to supervise it. That includes knowing when to rely on agents, when to question their outputs, how to escalate decisions that don’t feel right and how their own accountability fits into an AI‑enabled workflow. It also means updating role descriptions and performance expectations. Then employees can understand that success requires good oversight — catching issues, challenging recommendations and using agents to improve decisions, not just speed or volume.

Of course, we’re still in the early days of figuring out what it means to work alongside an agentic workforce. There isn’t a playbook yet for how to build expertise across functions and insight into the nuances of workflows across a company. Agentic AI isn’t like opening a box of shrink-wrapped software and learning how to use it. Agents become a de-facto part of the workforce.

That uncertainty is exactly where L&D can lead. The opportunity is big AI, transformation at the workflow level that knows no functional boundaries. L&D leaders can create the training and role design for it. They can equip tomorrow’s leaders with the new skills they’ll need to oversee and collaborate with AI in ways that none of us were taught in school.