Most enterprise AI deployments peak on day one. You ship the model, it works reasonably well, and then — slowly, imperceptibly — it starts to drift. The language patterns your customers use evolve. Your internal workflows change. The business context shifts. But the model doesn't know any of this. It's still frozen at the moment you deployed it.
This is the silent killer of enterprise AI ROI. Not the flashy failures. Not the hallucinations. The slow decay of models that can't learn from the very production environment they're running in.
Datadog's acquisition of Adaptive ML, announced yesterday, is a direct answer to this problem. And it's one of the most strategically significant enterprise AI moves of 2026.
What Adaptive ML Actually Built
Adaptive ML didn't build another AI model. They built the infrastructure that lets enterprises make their models better — continuously, from their own production data.
They call it Reinforcement Learning Operations, or RLOps. Think of it as the evolutionary successor to MLOps. Where MLOps focused on deploying and monitoring models, RLOps focuses on teaching models to improve based on real-world feedback signals.
The core insight: the algorithm was never the bottleneck. Every major foundation model provider — OpenAI, Anthropic, Google — has world-class algorithms. What enterprises lacked was the production-scale infrastructure to apply those algorithms to their own proprietary data streams, continuously.
Adaptive ML built that infrastructure. And Datadog just acquired it.
AT&T's Results Tell the Real Story
Before this acquisition, Adaptive ML's highest-profile deployment was with AT&T — and the numbers from that partnership reveal exactly what's possible.
AT&T had the same problem most large enterprises do: they were running AI on generic foundation models that didn't understand their specific business context. Fraud patterns at a telecommunications company are different from fraud patterns at a bank. Customer service conversations at AT&T involve specific billing terminology, product names, and technical jargon that general-purpose models handle poorly.
Using Adaptive ML's RLOps platform, AT&T's data science teams specialized their models on their own data. The results were measurable and significant.
For call summarization, AT&T fine-tuned a Gemma 12B model that delivers 30% faster performance over general-purpose APIs when generating daily transcripts for customer service calls.
For fraud detection, the impact was even sharper. Specialized reasoning models reduced fraud case review time from six minutes to just 30 seconds — a 12x throughput improvement for every fraud analyst on the team.
Mark Austin, Vice President of Data Science at AT&T, described the strategic shift: "To get the most value out of agentic AI applications, we need to move beyond generic prompts and specialize our models on our own data and workflows."
That's not a technology statement. That's a business strategy.
Why Datadog Is the Perfect Acquirer
Here's what makes this acquisition more interesting than a typical acqui-hire.
Datadog operates at a scale that gives Adaptive ML's technology a unique advantage. With Q1 2026 revenues of $1.006 billion — up 32% year over year — and guidance of $4.3-4.34 billion for the full year, Datadog is embedded in the production infrastructure of thousands of enterprises globally. That means they have access to something Adaptive ML's algorithms desperately need: continuous, real-world signals at unmatched scale.
Every trace, every log, every metric, every anomaly that Datadog captures is a potential training signal for AI systems that need to improve over time. When your AI agent fails to resolve a production incident, Datadog knows exactly what happened. When it succeeds in diagnosing a security alert in seconds, Datadog logged the entire reasoning chain.
That's the data flywheel Adaptive ML's CEO Julien Launay pointed to when describing the rationale: "The missing piece was never the algorithm, the hardest part was production scale. With Datadog, and the continuous stream of real-world signals that only a platform operating at this unique reach can provide, we will work directly from the foundation that intelligent agents need to drive exponential productivity gains, reliably and consistently."
Ameet Talwalkar, Datadog's Chief Scientist, framed it in terms of first-party intelligence: "Our lab is focused on leveraging our data and domain expertise to build specialized agents and models, and to effectively turn our data into first-party intelligence."
First-party intelligence. That phrase deserves attention from every enterprise CIO and CFO.
What This Means for Enterprise Technology Leaders
For CIOs and CTOs evaluating their AI strategy, this acquisition signals something important: the next phase of enterprise AI isn't about which foundation model you license. It's about who owns the improvement loop.
Right now, most enterprises are renting intelligence. They pay OpenAI, Anthropic, or Google for models that get smarter — but get smarter for everyone, not specifically for you. The model improvements that result from your enterprise usage accrue to the model provider, not to you.
The RLOps model inverts this. Your production data, your feedback signals, and your business-specific outcomes become the training inputs that make your AI systems better over time. Competitor data doesn't improve your models. Your data doesn't improve competitors' models. The intelligence compounds in your organization.
Datadog has already been building toward this with its Bits AI suite. Bits Investigation — an autonomous AI agent for troubleshooting production incidents — has already conducted hundreds of thousands of investigations on behalf of customers. Bits Code and Bits Security Analyst extend the same autonomous reasoning to development and security workflows. At DASH 2026, Datadog launched over 100 new AI-powered capabilities, including AI Guard for agentic security.
Adding Adaptive ML's RLOps capabilities to this suite means those Bits agents will no longer just be smart on day one. They'll get smarter as they run on your specific infrastructure, learn from your specific incident patterns, and adapt to your specific security posture.
What This Means for Business Leaders
For CFOs, COOs, and business unit leaders, the financial logic is straightforward.
Generic foundation model APIs charge per token. Every call costs money. And because generic models aren't specialized for your business, they often require more calls — more back-and-forth — to reach a reliable answer. Specialized models trained on your data get to the right answer faster with fewer tokens. AT&T proved this: 30% faster summarization, 12x fraud detection throughput. Those translate directly to cost reduction and capacity expansion.
But the larger strategic implication is competitive differentiation. In a world where every enterprise can access the same foundation models, the companies that win will be those that turn their proprietary operational data into a defensible AI advantage. The company that has spent three years improving its fraud detection model on its own transaction patterns will consistently outperform a competitor running a generic model — even if the competitor uses the same base model.
The Adaptive ML CEO put it plainly: "AT&T is ensuring their domain-specific data becomes a compounding strategic asset." The word "compounding" is key. This is not a one-time advantage. It's an advantage that grows over time.
Where Enterprises Should Focus Now
The Datadog-Adaptive ML combination is unlikely to be the last move in this space. Expect enterprise software vendors across observability, security, CRM, and ERP to make similar acquisitions as RLOps becomes the standard infrastructure for self-improving AI systems.
Conversations with technology leaders across industries suggest this is already top of mind. The shift from "deploy a model" to "build a model that continuously improves from your data" is becoming a defining capability question for enterprise AI programs.
For organizations thinking about their AI roadmap, a few questions are worth prioritizing now:
What production signals does your organization generate that could train domain-specific models? Customer interactions, support tickets, transaction patterns, incident logs — these are all potential training inputs that most enterprises currently discard.
How are you instrumenting your AI agents today? You can't improve what you can't measure. Observability isn't just for infrastructure anymore. Every AI agent interaction should generate structured feedback that enables continuous learning.
Are you building intelligence or renting it? The answer doesn't have to be binary — hybrid approaches using foundation models fine-tuned on proprietary data are already proving effective in production at scale. But the question of ownership should be explicit in your AI strategy.
The Bottom Line
Datadog's acquisition of Adaptive ML is more than an R&D investment. It's a statement about where enterprise AI value creation is heading.
The race for general intelligence was always going to be won by a handful of hyperscalers. But the race for specialized, continuously improving enterprise AI — AI that knows your business better than any generic model ever could — is still wide open. And it runs through RLOps.
Most enterprise AI programs today are building models that peak on day one. The winners will be those that build AI that gets smarter every day they run it.
Rajesh Beri is the founder of THE D*AI*LY BRIEF, an enterprise AI newsletter for technical and business leaders. Connect on LinkedIn or X/Twitter.
