Salesforce is changing the game with Large Action Models (LAMs), creating AI that doesn’t just analyze but actually does work for businesses. These smart systems can now run on devices and handle complex tasks without needing constant internet connection, making workflow automation faster and easier. The models are lightweight yet powerful, designed to solve multi-step problems with remarkable efficiency. By making advanced AI more accessible and practical, Salesforce is helping companies of all sizes transform how they work. This isn’t just a tech upgrade—it’s a whole new way of thinking about artificial intelligence in the workplace.
What Are Salesforce’s Large Action Models (LAMs) and How Will They Transform Enterprise AI?
Salesforce’s Large Action Models (LAMs) are advanced AI systems designed to autonomously execute complex enterprise workflows, moving beyond traditional analysis to proactive task completion. These lightweight, on-device models enable organizations to automate multi-step processes with unprecedented efficiency and precision.
A New Chapter in Enterprise AI
If you’d told me five years ago that AI would soon be capable not just of suggesting what to do, but actually rolling up its digital sleeves and getting things done for you—well, I’d have snorted into my Americano. And yet, here we are. Salesforce, never one for half-measures, has just flung open the doors to a new breed of AI: agentic models designed not to merely analyze, but to autonomously act within enterprise workflows. For any organization riding the turbulent currents of digital transformation, this is no incremental tweak; it’s a tectonic shift. (You can read the details in this CIODive article.)
As a top manager at Customertimes—a company that’s spent the last eighteen years navigating clients through windswept fields of CRM and automation—I can feel the static in the air. It smells like freshly ground coffee and opportunity. But I digress. The announcement of Salesforce’s new large action models (LAMs), purpose-built for agentic AI, is more than just another press release. It’s an inflection point, reminiscent of when the first graphical UIs transformed computing from cryptic runes to intuitive canvases.
I had to pause and ask myself: are we truly ready for AI that not only thinks alongside us, but acts on our behalf? It’s thrilling, sure—but a little unnerving, too.
From Passive Insights to Proactive Agents
For years, artificial intelligence has played a supporting role: the insightful friend in the corner, always ready with a pie chart or a predictive metric. Those days, it seems, are numbered. Salesforce’s LAMs empower AI to leap from the sidelines to the center stage, orchestrating entire workflows, fielding customer requests, and even juggling multi-step supply chain maneuvers. It’s as if the Roomba suddenly decided to redecorate the living room. Hyperbolic? Maybe, but the comparison stands.
What’s particularly arresting—yes, I’ll use that word—is how these models can now run on-device. No more being shackled to the caprices of cloud latency or unreliable Wi-Fi. This is a lifeline for our clients whose operations stretch from the glass towers of Manhattan to remote outposts where a Zoom call sputters like an old Lada in January. The tactile benefit here? Teams can execute sophisticated AI workloads without waiting for the broadband gods to smile upon them.
I remember a client in agribusiness—let’s call them “Illustrative Farms”—whose field technicians spent half their time cursing at spinning loading screens. When we piloted a rudimentary offline AI tool, their relief was almost comical. “It just works,” one of them said, thumping the device like a Soviet-era radio. The human texture of that moment still sticks with me.
Technical Elegance Meets Real-World Grit
Let’s address the elephant in the server room: bigger isn’t always better. Many AI vendors chase jaw-dropping parameter counts—hundreds of billions!—as if size alone conferred wisdom. In practice, enterprise infrastructure is a palimpsest of new and legacy tech. Not everyone’s running hyperspectral GPUs cooled by liquid nitrogen.
Salesforce’s introduction of lightweight xLAM models, starting at a modest yet mighty 1 billion parameters, is a pragmatic move. These models pack robust planning, reasoning, and function execution into a svelte package. According to Salesforce’s own benchmarks, they’re even outpacing GPT-4o and the GPT-4.5 preview in certain enterprise agent tasks. I’ll admit, I was skeptical at first (who wouldn’t be?)—but the numbers don’t lie. Sometimes, leaner models cut to the chase where bulkier ones flounder.
And then there’s TACO—no, not the edible kind, though I confess that’s where my brain went first. Salesforce’s new multimodal action model family is engineered for multi-step problem solving. It doesn’t just answer; it decomposes, reasons, and executes in a chain-of-thought-and-action structure. On the MMVet benchmark, TACO reportedly scored a whopping 20% higher than previous top performers. Bam! That’s not a mere incremental bump; it’s a quantum leap.
The Agentic Maturity Model: A Practical Roadmap
Of course, technological wizardry is only half the battle. Most clients don’t know where to start with agentic AI, let alone how to scale it. Enter Salesforce’s Agentic Maturity Model—a structured guide to help organizations crawl, walk, and eventually run with autonomous agents. It’s like a good recipe: start simple, iterate, and resist the urge to toss in every spice at once.
I’ve seen this in action. Alpine Intel (a real client, not illustrative this time) adopted Salesforce’s agentic framework and shaved hours off their insurance claims processing. They were skeptical at first. So was I. But the results spoke for themselves, and I felt a twinge of pride—mixed with a pinch of envy, if I’m honest. The process wasn’t flawless, but the learning curve can be deliciously steep.
Wiley, another early adopter, echoed similar sentiments. I recall their team describing the shift from manual to agentic processes as “going from a rusty tricycle to a Tesla.” Hyperbole? Perhaps. But emotionally apt.
Responsible Automation and the Democratization of AI
There’s a whiff of caution in the air, too. Agentic AI is not a toy; it’s a force multiplier, and with great power comes—you know the rest. Governance, security, and clear ethical guardrails aren’t just afterthoughts. They’re prerequisites. At Customertimes, we’ve always insisted that technology must empower, not imperil. I once rushed a project that, in hindsight, needed tighter oversight. Lesson learned: trust, but verify. Salesforce’s insistence on responsible deployment reassures me we’re not alone in this conviction.
What makes this era truly remarkable is accessibility. By optimizing models for diverse hardware, Salesforce is casting a wider net. It’s not just the Fortune 500 getting the AI edge; it’s the mid-market, the public sector, the upstart with more ambition than budget. That democratization feels as crisp and invigorating as autumn air. You can almost hear the gears of innovation whirring—softly, persistently—in the background.
The historian in me wonders: will we look back on this as the moment when AI moved from the hands of the few to the many? I’d wager yes. Or, at least, I hope so.
The Road Ahead: Imperfect, Electric, Unfinished
So, where does that leave us? Somewhere between exhilaration and humility. The future is arriving in fits and starts—action by action, half-baked idea by prototype by breakthrough. We’re all learning, sometimes stumbling, occasionally sprinting.
In the end, agentic AI represents more than a new toolset; it’s a new partnership between humans and machines. The lines blur, the possibilities multiply, and yes, sometimes the coffee goes cold before you figure it all out. But what a ride.
Ready?
Because I’m not entirely sure I am. And yet, who could resist?