AI agents are revolutionizing marketing departments across industries, moving beyond simple automation tools to become intelligent partners in campaign strategy and execution. Today's marketing teams are integrating these sophisticated systems into nearly every stage of their processes, from creative generation to data analysis.

This transformation represents a fundamental shift in how marketing operates. Rather than replacing human creativity and strategic thinking, AI agents amplify these capabilities while handling the heavy lifting of execution, optimization, and analysis. The result is faster, more personalized, and data-driven marketing that would be impossible with human resources alone.

This article explores the role of AI agents in five core areas of the marketing workflow.

1. Creative Content Generation: From Weeks to Hours

The creative process—producing ads, copy, and visuals for campaigns—has been dramatically accelerated by generative AI agents. Tasks that once required teams of copywriters and designers working for weeks can now be completed in hours.

Coca-Cola's "Create Real Magic" Success Story

In 2023, Coca-Cola launched its "Create Real Magic" platform, enabling users to co-create artwork using GPT-4 and DALL·E. The results were staggering: over 120,000 pieces of unique content created in just a few months. But the impact extended beyond external campaigns.

Internally, Coca-Cola's marketers now use generative AI tools for design and ideation. As Selman Careaga, President of ASEAN & South Pacific category at  Coca-Cola, explained: "Whether it's research or analytics that can create better insights... or using AI as a tool together with our human insights team to come up with more relevant ideas." The company's human creatives partner with AI agents to generate ad copy, social posts, and even flavor concepts—as seen in its AI-developed Y3000 limited-edition drink.

Practical Efficiency Gains

At a more operational level, e-commerce retailer Adore Me automated product description creation with AI. This mundane but necessary task previously consumed 30-40 hours per month. An AI writing agent reduced this to just one hour—a 97% time savings that freed up human resources for strategic work.

These examples demonstrate that across industries, AI content agents help teams generate and iterate on creative materials at unprecedented speed, while humans provide brand guidelines and final approval.

2. Customer Research & Insights: Mining Insights at Scale

The very foundation of effective marketing is a deep understanding of customers: their needs, preferences, and opinions. AI agents are revolutionizing this process by analyzing vast data sources and surfacing actionable insights in record time.

Instead of human analysts manually sifting through survey responses or social media comments, AI research agents use natural language processing to aggregate and summarize sentiments from thousands of data points. Early implementations across industries include:

• Consumer goods companies analyzing call center transcripts and online reviews to identify product issues and opportunities
• B2B marketers researching target accounts by examining public financial reports, news, and LinkedIn data
• Market research firms experimenting with AI moderators that autonomously conduct customer interviews via chatbot and compile findings

This shift enables marketing teams to base strategies on comprehensive data analysis rather than limited sample sizes or intuition alone.

3. Ad Campaign Execution: Real-Time Optimization at Scale

Once marketing creative and plans are finalized, executing campaigns across multiple channels becomes a complex operational challenge. AI agents excel as tireless coordinators, automating routine steps, optimizing performance in real-time, and making proactive adjustments.

Google's Performance Max: AI in Action

Google's Performance Max essentially functions as an AI agent for advertisers. It autonomously manages campaigns across Search, YouTube, Gmail, Maps, and other platforms, allocating budget and adjusting placements to meet specified goals. Marketers simply provide objectives and creative assets, while AI handles granular execution, continuously learning which audiences and messages perform best.

This cross-channel optimization reacts in real-time to shifts in consumer behavior, something manual optimization would accomplish far more slowly.

Enterprise Implementation Examples

Enterprise marketing teams are embedding agents throughout their campaign operations. Accenture uses autonomous agents to run large campaigns with minimal human intervention, achieving notable efficiency gains. Retail giant Carrefour's AI marketing studio not only personalizes content but automatically adapts creatives for different social platforms, accelerating campaign rollout across Facebook, Instagram, and other channels.

The common thread is real-time decision-making: these agents can pause underperforming ads, redistribute budget to top-performing channels, or recommend email send time changes by analyzing engagement data instantly. Humans set strategy and guardrails while AI agents handle rapid response tactics, creating campaigns that are "always on" with 24/7 optimization.

4. Lifecycle Marketing: User-Level Personalization

Lifecycle marketing focuses on engaging customers with the right message at each stage of their journey—onboarding, retention, win-back, and beyond. AI agents excel at analyzing customer data and orchestrating personalized touchpoints at scale.

Starbucks' Deep Brew Success

Starbucks relies on an AI engine called Deep Brew to power hyper-personalized offers for Rewards members. Deep Brew functions as a behind-the-scenes marketing agent, mining loyalty and app data to recommend products and promotions tailored to individual customers. It might suggest a new drink based on past orders or time a discount when a lapsed customer is due for a repeat visit.

The system considers contextual factors like weather, local events, and time of day to optimize engagement. The impact has been substantial: personalized recommendations drove higher visit frequency and larger ticket sizes, contributing to mobile orders now representing 30%+ of U.S. transactions.

Advanced Personalization and Targeting

AI agents like Auxia's Decision Agent deploy adaptive, ultra-personalized campaigns across channels. These systems automatically rank, score, and predict the optimal form of content (e.g. email, SMS, etc) for each user based on their behavior or preferences to serve the best journey for that person. This ensures each individual receives far more timely and relevant outreach, which results in far better performance than what manual segmentation could achieve.

5. Data Analysis & Decision Support: 24/7 Marketing Intelligence

Modern marketing generates enormous amounts of data—campaign metrics, web analytics, CRM data, sales figures, and more. AI agents are being deployed to process this information and support faster, better decision-making.

These agents function as intelligent marketing analysts that work around the clock. An AI analyst agent continuously monitors campaign KPIs across channels, alerting teams when something notable occurs like a sudden drop in conversion rates or an ad trending below benchmarks. Rather than waiting for monthly analytics reviews, marketing leaders receive real-time insights, produce executive-ready reports, and highlight trends or anomalies as they happen.

Workflow Transformation: Before and After AI Agents

To illustrate the practical impact, consider how AI agents transform a typical email campaign workflow:

Traditional Approach

A human marketing manager relies on a data scientist or engineer to create segments of their customer base using database queries, marketing automation rules, or (in the most advanced teams) predictive models. They collaborate with copywriters to develop email content and design, set up A/B tests for the subject lines and body, then schedule and send the campaign. Over the following days, they manually monitor performance metrics, export data to spreadsheets for analysis, and eventually prepare reports to guide follow-up actions. This process involves multiple handoffs and considerable time investment.

AI-Agent Enhanced Approach

A "Content" agent generates multiple email variations based on all the hypotheses your team has on what will drive a customer to convert. After a human marketer reviews and approves these drafts, a “Decision” agent automatically analyzes customer data, leverages machine learning to discover behavioral patterns for each individual customer, and serves the optimal variation for each person based on what’s predicted to have the highest impact.

As the campaign runs, it monitors performance in real-time, proactively adjusting the emails that are distributed for each individual within pre-approved guardrails. Simultaneously, an "Analyst" agent autonomously processes live data, integrating email statistics with web traffic and sales conversions to identify hidden trends.

When specific clusters of customers outperform others, the “Analyst” agent alerts the team and recommends specific re-engagement strategies for the underperforming group. Post-campaign, it auto-generates comprehensive reports with key insights and actionable recommendations.

In this scenario, the marketer's role shifts to strategic oversight: reviewing AI-generated content, approving agent suggestions, and providing direction. The heavy lifting of execution and analysis is automated, compressing timelines from weeks to days while enabling rapid, data-driven adjustments that human teams might miss or implement too slowly.

Overcoming Implementation Challenges

Despite their promise, AI agents face several key adoption barriers:

Trust and Oversight

As AI agents take on more autonomous roles, companies need robust guardrails to ensure brand safety and accuracy. Marketing involves creativity and nuance that AI can mishandle—a tone-deaf automated social post can create PR disasters. Building trust requires implementing review checkpoints, maintaining "human veto" power, and creating transparent logs of agent decisions. Until marketers trust AI systems, they'll remain reluctant to grant autonomous authority.

Data Integration and Quality

AI agents are only as effective as the data they receive. Many marketing teams still operate with data silos—separate CRM, web analytics, and event collection streams—along with data quality issues. Autonomous workflows require unified, clean data streams for effective AI reasoning. Setting up these integrations while ensuring accuracy represents a significant technical hurdle.

Real-time data access is crucial for proactive agents. Any latency or batch processing limitations can severely restrict an AI's ability to react quickly to changing conditions.

Skills and Change Management

Adopting AI agents requires marketers to develop new competencies, from prompt engineering to interpreting AI outputs and managing exceptions when systems escalate issues to humans. Cultural resistance and anxiety must be addressed, as 43% of marketers who haven't embraced AI worry about becoming too reliant on these tools.

Successful companies approach this in a number of ways. Some embed AI within every team and role at the company, while others create internal "AI task forces" or centers of excellence to coordinate initiatives and provide training. Both approaches can work, but it typically depends on the stage of the organization, how AI-native the workforce is, and what your enterprise’s goals are.

Strategic Implementation Focus

With significant AI hype, many teams experiment with agents without clear plans, leading to scattered pilots that never scale. Best practices emphasize starting with specific workflow pain points—"producing weekly reports is slow" or "social media monitoring is overwhelming"—then deploying targeted agents to solve these problems.

Early wins, like Adore Me's 97% time reduction in product copy generation, build momentum for broader implementation. However, not every marketing task benefits from AI automation; some creative and strategic functions still require human-led approaches.

The Future of Human-AI Marketing Collaboration

AI agents are transforming marketing from a primarily human-intensive craft into a sophisticated collaboration between human creativity and machine efficiency. Today's marketing teams may use AI writing assistants, analytics bots, and automated schedulers—but this represents only the beginning.

As enterprises integrate these agents into unified systems, marketing workflows will become increasingly intelligent and autonomous. We're witnessing marketing departments evolve into augmented teams where humans provide strategic direction and ensure brand integrity, while AI agents handle execution, data processing, and routine decision-making.

The industry sits at a crucial turning point. The foundational pieces—generative AI, integration APIs, preliminary trust frameworks—are in place. Forward-thinking companies are assembling them to reinvent campaign management entirely. While most marketers remain in early adoption phases, trailblazers have demonstrated the potential.

Over the next 1-3 years, expect more marketing functions to incorporate proactive AI assistants and limited autonomy. Each success will build confidence to push boundaries further, likely creating hybrid human-AI teams capable of achievements in personalization, scale, and speed that purely human teams could never accomplish.

Conclusion: AI Doesn’t Replace Marketers, It Augments Them

The adoption of AI agents in marketing isn't about replacing human creativity or strategic insight—it's about amplifying these uniquely human qualities with machine efficiency and intelligence. In marketing, embracing AI agents means reimagining workflows and roles so that the combination of humans plus AI exceeds what either could achieve alone. Organizations that master this human-in-the-loop, multi-agent model will lead marketing innovation, delivering more personalized and impactful campaigns while competitors struggle to keep pace.

The era of AI-agent augmented marketing has arrived, promising a future where creative strategy and algorithmic execution work hand in hand to drive sustainable growth. The question isn't whether AI agents will transform marketing—it's how quickly your organization will adapt to harness their full potential.

Ready to learn more about Auxia? Schedule a demo.

‍

The marketing landscape is experiencing a fundamental shift. What began as rule-based automation—triggered emails, scheduled social posts, basic segmentation—is evolving into something far more sophisticated. Today's AI agents don't just execute predefined workflows; they observe, learn, and make decisions that were once the exclusive domain of human marketers.

This transformation represents more than technological advancement; it's a reimagining of how marketing teams operate. Where marketers once spent hours analyzing campaign performance, AI agents now continuously optimize in real time. Where creative teams labored over endless A/B test variations, intelligent systems generate and test hundreds of permutations simultaneously. Where campaign managers juggled complex multi-channel orchestration, autonomous agents seamlessly coordinate across touchpoints without human intervention.

Yet this evolution raises profound questions about the future of marketing work itself. As AI agents move from helpful co-pilots to fully autonomous campaign managers, we must grapple with what this means for creativity, strategic thinking, and the fundamentally human aspects of brand building. The challenge isn't simply technical, it's organizational, creative, and philosophical.

This article explores the spectrum of AI agency in marketing, from today's sophisticated assistants to tomorrow's autonomous campaign orchestrators.

The Emergence of Multi-Agent Marketing Teams

Marketing teams are evolving into multi-agent systems where human marketers lead networks of AI-driven specialists. Rather than replacing humans, these AI agents serve as collaborative co-pilots embedded throughout the marketing lifecycle—from planning and creative generation to execution and analysis.

This creates a new organizational model: humans-in-the-loop guiding specialized AI assistants. Success requires rethinking workflows beyond simple one-to-one mappings of human job titles to AI tools. Instead, it's about designing intelligent information flows and determining when human oversight should intervene to ensure seamless AI-human collaboration.

In this model, human marketers become orchestrators and quality controllers, supervising agent "colleagues" that hand off tasks to one another in modular pipelines. This balanced approach maintains trust and contextual awareness while embedding AI agents deeper into marketing operations.

The Five Stages of Agentic Maturity in Marketing

Not all AI agent deployments are equal; they vary widely in sophistication and autonomy. We can categorize the maturity of agent integration in marketing teams into five stages (adapted from the general maturity model suggested by Scott Belsky):

Stage 1: Glorified Personalized Help: At this most basic level, the AI functions like a smart FAQ or assistant that provides personalized answers or content recommendations when asked. In marketing, this might be a chatbot that can retrieve information (“What was our Q3 ad spend on channel X?”) or a simple content suggestion tool. It’s helpful, but essentially reactive and one-step. Many early marketing AI tools (like FAQ chatbots on websites or basic personalization rules engines) fall into this category; they tailor outputs to the user, but don’t do any multi-step reasoning or proactive work.

Stage 2: Reactive Recommendations: Here, the agent can do some work on your behalf upon request. Marketers input a prompt or command, and the agent produces something useful, often assembling data or content. For example, an AI that generates an email draft or a social media calendar on demand based on parameters you give is at this stage. Most current "AI in marketing" tools operate at this stage. The agent doesn't act unprompted but can handle complex tasks when directed, delivering prepared assets or analyses that save significant marketer effort.

Stage 3: Proactive Recommendations: At this stage, the agent starts to anticipate needs and suggest actions without being explicitly asked each time. In a marketing context, an agent might observe your campaign setup and proactively suggest, “Based on previous campaigns, you may want to add a promotional message via notifications for audience X.” Or an Analyst agent might send a suggestion, “Key insight: users from segment Y are trending down, perhaps create a re-engagement campaign.” These recommendations are not user-prompted but context-triggered. Achieving this reliably often requires the agent to have a degree of situational awareness (e.g., access to live campaign data or content creation workflow state). While still relatively rare in 2025, we are starting to see early examples: some advanced systems will proactively learn and flag opportunities as agents gain better contextual awareness and predictive capabilities. This moves the AI from a passive tool to more of an active advisor on the marketing team.

Stage 4: Proactive Actions: Now the agent goes beyond suggesting; it takes actions autonomously (or semi-autonomously) when it sees an opportunity, while keeping a human in the loop. In marketing, this could mean an AI agent that doesn’t just recommend adding a channel or adjusting an ad bid, but actually executes the change: for instance, pausing a poorly performing ad on its own, or launching a test it thinks will improve results, then informing the team of what it did. At this stage, the AI behaves like a junior colleague who can “work side by side” with human marketers. We see early glimpses of this in areas like programmatic advertising (where the system allocates budget across creatives in real time) and some coding co-pilot agents have reached this stage (making code changes proactively). For broader marketing applications, Stage 4 action agents are just beginning to emerge.

Stage 5: Autonomous Workflows: The most advanced stage is a fully autonomous marketing workflow run by one or multiple agents with minimal human input beyond high-level goals. This is the vision of an AI agent (or a team of agents) that can carry out an entire marketing process end-to-end: for example, conceive a campaign for a new product, allocate budget, produce content, deploy across channels, and optimize – all autonomously, “negotiating with other agents on our behalf” if needed and only asking for human sign-off at key checkpoints. In theory, such an autonomous marketing agent could be given an objective (e.g., “Launch a campaign to increase subscriptions by 10% among Millennials in Q3”) and it would handle the rest, within guardrails. Currently, this level exists primarily in experimental forms, with the closest examples being Google's Performance Max and Meta's Advantage+ automated campaign systems. True autonomous marketing workflows remain largely aspirational.

Current State and Future Outlook

Most organizations in 2025 are transitioning from Stages 1 and 2 into Stage 3, with leading-edge cases exploring Stage 4. According to product leader Scott Belsky, most startups pitching agent-based functionality still operate in the "Reactive Recommendations" category.

Truly proactive marketing agents that reliably anticipate needs (Stage 3) are just beginning to appear, though Belsky notes, "I have not seen many great examples of this yet, but they are coming." Stage 4 proactive action agents are emerging in specialized applications like programmatic advertising, while Stage 5 full autonomy remains a near-future aspiration.

Navigating the Human-AI Partnership in Marketing's Future

The path forward requires more than simply deploying the most advanced AI tools available—it demands a fundamental reimagining of how marketing organizations structure themselves, allocate responsibilities, and maintain competitive advantage in an increasingly automated landscape.

Building AI-Native Marketing Teams

Organizations that succeed won't bolt AI agents onto existing processes, but will redesign their marketing operations around human-AI collaboration from the ground up. The most effective teams will operate as hybrid intelligence networks, where human creativity and strategic intuition guide AI execution and optimization. Human marketers will focus on high-level strategy, brand positioning, and emotional resonance, while AI agents handle execution, testing, and real-time optimization that scale beyond human capacity.

The Trust and Control Challenge

As AI agents become more autonomous, marketing leaders face a critical balancing act: capturing automation's efficiency gains while maintaining control over their brand's voice and actions. Success requires establishing robust governance frameworks early—defining clear boundaries for AI decision-making and creating feedback loops that preserve brand authenticity and customer relationships. The most sophisticated AI can optimize for engagement metrics, but human judgment ensures those optimizations align with long-term brand values.

The New Competitive Landscape

As AI capabilities democratize, sustainable advantage will shift from having access to tools to how thoughtfully organizations integrate them. The brands that thrive will use AI agents not to replace human creativity, but to amplify it—freeing marketers for the work that truly differentiates: understanding deep customer needs, crafting compelling narratives, and making strategic bets in crowded markets.

Organizations should focus on mastering Stages 2 and 3 before rushing toward full automation, investing in the infrastructure and cultural changes needed for effective human-AI collaboration. Most importantly, this means maintaining sight of marketing's ultimate purpose: connecting with human needs and emotions in ways that create lasting value.

The future belongs to organizations that embrace this partnership, leveraging AI agents as collaborative tools while preserving the human elements that create truly memorable brands. Success won't be measured by how much human work AI can replace, but by how much human potential AI can unlock.

Ready to learn more about Auxia? Schedule a demo.

‍

The Fragmentation Crisis

After speaking to dozens of marketing teams over the past year, the first clear theme that emerges from conversations with CMOs, CTOs, and CIOs is that they are tired of managing a plethora of point solutions.

The numbers tell the story: even mid-sized companies typically operate multiple different solutions in their marketing stack. Each tool was implemented to solve a specific pain point—email automation here, customer data platform there, A/B testing in another corner. But what marketing leaders are discovering is that all of these tools have their own silos, with absolutely no cognition, cohesion, or unified customer point of view.

The hidden costs extend far beyond software licensing. Companies invest millions in integration specialists, data engineers, and agency staff just to make these disparate systems work together. It’s not unheard of to spend more on the people managing a marketing stack than on the technology itself.

Most frustrating of all is the personalization paradox: despite having more marketing technology than ever before, most companies struggle to deliver truly personalized experiences. Rules-based systems and batch processing create experiences that feel generic and disconnected from real customer behavior.

The Convergence: Technical Enablers

Two fundamental technological shifts have converged to make marketing stack consolidation not only possible but inevitable.

The first is the maturation of cloud-native data warehouses. Since Snowflake’s emergence in 2015, enterprises have invested billions of dollars in getting all their data into centralized, accessible repositories. For the first time, companies have a true single source of truth for customer information: structured transaction data, unstructured behavioral signals, and everything in between, all sitting in one place.

The second enabler is the transformer architecture revolution that powers modern large language models. These advances allow AI systems to understand and process both structured and unstructured data at unprecedented scale and speed. What previously required extensive data science teams and weeks of analysis can now happen in real-time, enabling immediate decision-making based on the most current customer behavior.

Combined with improved API accessibility, these technologies create the foundation for a fundamentally different approach to marketing technology; one where intelligence, rather than data movement, becomes the primary value driver.

The New Architecture: Intelligence Layer

The future marketing stack looks dramatically different from today’s fragmented landscape. Instead of separate tools for data collection, storage, analysis, and activation, a new intelligence layer sits between your data warehouse and customer touchpoints.

This intelligence layer represents a complete paradigm shift. Rather than moving data between systems and applying rigid rules, AI agents make dynamic decisions in real-time based on complete customer context. A decision agent determines the optimal action for each individual customer. An analyst agent continuously evaluates performance and suggests improvements. Content agents create and optimize messaging and creative elements so teams can focus on more strategic priorities.

The result is continuous optimization rather than periodic campaign adjustments. Instead of running A/B tests for weeks to determine what works, the system learns and adapts continuously, testing hundreds of variations and automatically promoting the most effective approaches.

This architecture eliminates the need for most traditional point solutions. Why maintain separate tools for audience segmentation, experimentation, personalization, and analytics when a single intelligence layer can handle all these functions more effectively?

Real-World Results

Companies implementing unified AI-driven platforms are already seeing dramatic results. Take two recent examples from our work at Auxia: a leading international marketplace with billions in gross merchandise value achieved an 84% increase in customer lifetime value by replacing their fragmented, segment-based campaigns with Auxia’s user-level AI decisioning engine. Meanwhile, a global financial institution with over $500 billion in assets boosted onboarding completion rates by 50% while conducting 22 times more experiments than their previous system allowed.

The breakthrough in both cases came from moving beyond rules-based segmentation to AI systems that consider complete customer context—purchase history, real-time behavior and demographic signals—to make optimal decisions in milliseconds rather than the weeks required for traditional A/B testing cycles.

The Choice Ahead

What makes this consolidation wave different from previous enterprise software cycles is the capabilities it unlocks. After years of marketing technology breaking apart into specialized point solutions, AI is now driving them back together—but with fundamentally new possibilities.

Take real-time personalization as an example. Most systems today claim to be real-time, but they’re really not.

Here’s what true real-time looks like: when someone looks for a product in the fashion category on a digital marketplace, AI algorithms can immediately analyze what they just bought and determine the optimal next category to surface—maybe sports memorabilia—along with the specific subcategories and incentives most likely to drive a cross-sell. This decision happens in milliseconds, based on that individual’s complete purchase history, browsing behavior, and demographic signals.

That kind of in-session decision-making, based on complete customer context, simply wasn’t feasible with fragmented tools. Traditional systems might recommend similar products—more jeans if you bought jeans—but they can’t make the intelligent leap to complementary categories that actually increase lifetime value.

The companies getting breakthrough results today understand this goes beyond buying new technology. They’re reimagining how marketing teams operate when AI agents can handle the analysis that used to require teams of consultants.

Marketing leaders can either drive this consolidation proactively or be forced to react as competitive pressures mount. The enterprises that embrace this transformation early will set customer expectations for personalized experiences that fragmented systems simply can’t match.

‍

Auxia is an Agentic Customer Journey Orchestration platform that delivers personalized marketing recommendations to enterprise customers. Customers integrate with Auxia by calling our API to retrieve a recommended treatment—what we call a “Decision”—across in-app surfaces, emails, or other digital experiences.

When you’re serving enterprise customers, any infrastructure faces significant scaling challenges. We presently handle a peak rate of over 6,000 requests per second (RPS), aiming for a 99th percentile prediction latency of 100ms. Each request requires selecting from approximately 1,000 potential user-facing treatments.

This post details how we built a high-performance inference infrastructure that meets those demands.

Scaling Real-Time Personalization

Three critical requirements drive our architecture decisions:

• High-volume, low-latency processing - Supporting thousands of concurrent requests with sub-100ms response times
• Real-time contextual data integration - Incorporating fresh user context and behavioral data to deliver relevant recommendations
• Multi-tenant model support - Running simultaneous inference across different models for each customer and goal combination

This combination of requirements demanded a high-performance, real-time inference system with dynamic model loading capabilities across load-balanced service instances.

Architecture: Kotlin + TensorFlow Serving Sidecar

We implemented a co-located system where a Kotlin server (our control plane) sits alongside a TensorFlow Serving binary (inference engine) within each Kubernetes pod. This design gives us both high performance and maximum flexibility through clear separation of concerns:

• Kotlin server: Handles dynamic model loading, model metadata, input/output tensor transformation, and lifecycle orchestration.
• TensorFlow Serving: Efficiently loads and executes predictions on trained TensorFlow models using a high-throughput gRPC API.

This separation of concerns ensures that model logic stays isolated, while all orchestration and business logic live in Kotlin.

As a Kotlin + gRPC organization, this design leverages Kotlin's strengths—particularly coroutines for async programming—while abstracting away ML infrastructure complexity. The Kotlin layer handles:

• Infrastructure abstraction - Hiding complexities of different model types and future ML frameworks (ONNX, TorchServe, hosted services)
• Dynamic model management - Loading models from our artifact registry and managing TensorFlow Serving's ModelService
• Contract translation - Converting user and treatment features into model input tensors, and output tensors back into treatment scores

Why TensorFlow Serving?

We chose TensorFlow Serving over alternatives like TorchServe for several key benefits:

• gRPC API advantages: strongly-typed, programmatically defined interface with binary serialization that reduces message size and CPU parsing overhead, which is critical given our large feature sets.
• High performance: C++ implementation designed natively for multi-threaded usage to deliver the speed required for real-time decisioning.
• Production-ready features: built-in support for dynamic model loading and automatic inference batching.

However, Tensorflow Serving introduced several challenges that needed to be addressed by the Kotlin Server:

• Dynamic loading limitations - While technically supported, the binary is optimized for static model sets at startup
• Limited production testing - Fewer real-world deployments mean issues like incorrect CPU detection in containers
• Optimization requirements - Not all TensorFlow operations are performant, requiring careful model contract design

Inference Abstraction Layer

Our Prediction Service abstracts inference complexity into a simple, treatment-oriented API. This enables support for diverse model architectures ranging from Bandits to Tree-based Uplift models to even Deep Learning based recommender models.

Key Features

• Unified API: Single, consistent interface independent of underlying model implementation. The API contract specifies flexible input formats for user features (accepting either generic key-value pairs or a structured attribute object) and a standard key-value format for treatment features. It also defines the exact output structures that models must return, such as a single score or a list of scores per treatment.
• Dynamic Model Management: Load and serve different models on-the-fly without restarts, allowing for seamless updates and experimentation.
• Built-in Monitoring: Automatic collection of key performance metrics (like latency and error rates) for every prediction, ensuring system health and reliability.
• Developer-Friendly Tools: Model inspection capabilities and safe testing APIs for non-production validation.

Input Contract Design For Optimized Inference

TensorFlow Serving requires model inputs as a flat namespace of named tensors, essentially a flat dictionary mapping tensor names to data. Unlike TensorFlow Python’s support for complex nested structures (tuples, dictionaries, RaggedTensors), TensorFlow Serving imposes stricter requirements for serving models in production. This creates challenges when representing structured user and treatment features at scale.

Performance Optimization Journey

We experimented with several input designs to find the most efficient approach:

• Initial approach: One tensor per feature with feature names as keys. Simple to implement but highly inefficient; tens of thousands of tensors per request created significant serialization and name resolution overhead.
• tf.Example approach: Encoding features into TensorFlow's standard serialized format. Failed to improve latency due to costly proto parsing during inference.
• Final optimized design: Collapsed same-type features into single-typed tensors, distinguished by index positions. This approach minimized tensor count while aligning with TensorFlow's internal feature resolution.

Implementation Details

• User features are packed into typed tensors like [None, 4] for numerical features, where indices correspond to specific attributes (signup age, LTV, etc.).
• Treatment features use [None, None, N] format supporting multiple treatments per user, with an accompanying treatment_counts tensor indicating treatment boundaries per user.
• GPU optimization leverages TensorFlow Serving's auto-batching. The treatment_counts tensor handles padding removal and correct treatment boundary reconstruction.
• Metadata coordination ensures alignment between training and serving through embedded metadata files that map tensor indices to human-readable feature names.

This optimized design enables us to serve large-scale real-time inference at p99 latency under 100ms while scoring up to 1,000 treatments per request at 6,000 QPS.

Serving Model Validation Framework

Every trained model must be compatible with our Prediction Service API. We built a comprehensive local testing framework with three stages:

Environment Setup

Complete, self-contained production stack instance including Kotlin Prediction Service and TensorFlow Serving process. Programmatically launched locally by fixtures to test against actual service binaries, not mocks.

Test Orchestration

Managed by pytest and helper classes that handle model artifact placement and provide high-level client abstractions. Test authors work with pandas DataFrames while the framework handles serialization, gRPC requests, and result parsing.

Test Execution

Validates model compatibility through a file-based testing endpoint. The service reads model and feature files, performs inference, and writes scores to output files for validation. This workflow confirms that trained models can be loaded, served, and queried correctly before production deployment.

Dynamic Model Loading

Auxia’s dynamic model‐loading system allows customer requests to specify models by name and digest, then transparently fetches, validates, and serves those models without inference server restarts.  

Model Distribution

Models are published as OCI images in Google Cloud Artifact Registry. Each image contains a TensorFlow SavedModel at /data/tensorflow_serving_model/model and a metadata.json file describing input/output contracts. This enables ML Engineers and Data Scientists to push new versions frequently with floating labels (latest, canary) for rapid production deployment.

Runtime Architecture

Incoming gRPC Predict requests route through a ModelRegistry that dispatches to appropriate ModelLoaders based on model name prefixes. For container-based TensorFlow models:

1. DockerModelLoader resolves fully qualified image names, handles live and canary tags, fetches image manifests, and produces lightweight specs pointing to chosen digests
2. TensorflowModelLoader stages models on disk and orchestrates two coordinated state machines maintaining synchronization between our internal view and TensorFlow Serving's configuration

State Machine Management

ModelStateMachine manages individual model lifecycles:

• NEW → DOWNLOADED (files staged locally)
• DOWNLOADED → LOADED (reload-config sent to TensorFlow Serving)
• LOADED → AVAILABLE (GetModelStatus confirms serving readiness)
• Idle models automatically unloaded and deleted after configurable duration
• Read/write mutex guards with backoff timers for transient failure handling

TFServingStateMachine aggregates all loaded models into a single ModelServerConfig, pushing updates via ReloadConfigRequest API. This prevents race conditions between concurrent operations and handles known TensorFlow Serving bugs where unknown-status errors indicate successful reloads.

Both state machines run in Kotlin coroutines on dedicated dispatchers, ensuring asynchronous operation without blocking server I/O threads.

Canary deployments are first‐class features of our architecture. Models tagged with _canary trigger the Docker loader to read ModelCanaryConfig protobuf from image labels, specifying traffic fraction and monitoring parameters.

The system probabilistically routes specified percentages of requests to new versions while maintaining traffic to live versions. Canary models are pre-warmed to minimize latency spikes, and automated monitoring of latency, error rates, and output distributions drives promotion or rollback decisions.

This dynamic loading system provides robust, zero-downtime model lifecycle management enabling rapid experimentation and safe production rollouts.

Latency and Performance Improvement

• Our optimizations delivered significant latency improvements across the inference pipeline:
• Note: Latency scale is logarithmic. Figures based on load-tested results under heavy serving pod load. Production figures can be 10x better at 99th percentile. No 99% latency data available for Python Pandas baseline.
• The combination of optimized input contracts, efficient batching, and dynamic model management allows us to meet our ambitious performance targets while maintaining the flexibility needed for rapid ML experimentation and deployment.

‍

‍

‍

Today, I’m incredibly excited to share something that’s been a long time in the making: we’re launching a major upgrade to our Analyst Agent, a product that’s going to change the way marketing teams run analyses on their existing marketing initiatives.

If you’ve ever waited days—or weeks—for a data science team to answer questions like “What are the characteristics of our highest performing customers?” or, “Which email variations resonated most with customers from a specific acquisition channel?”, you know how frustrating it can be. We built the Analyst Agent to solve that exact problem. And now, it’s better, faster, and smarter than ever.

Ask a Question, Get the Why—Instantly

With the new Analyst Agent, you don’t need to be technical to understand what’s driving your performance with Auxia. Just ask a question in plain English:

• “Which cohort responded best to the onboarding email we refreshed last week?”
• “Did our upsell nudges in the app outperform control for high-income users?”
• “What are the characteristics of creative that perform well? What performs poorly?”

Our AI Decision Agent is already making hundreds of millions of decisions every day—deciding the optimal action, content, incentive, surface, and frequency to drive our customers’ objectives. Now, with Analyst Agent, your team can actually understand why those decisions worked—and how to make them even better.

Saying Goodbye to the “Black Box”

Let’s be real: marketers have been stuck in the dark for too long. AI systems can often perform better than rules-based logic, but without visibility into what’s driving results, teams are left guessing. That’s the “black box” problem—and we’re breaking it wide open.

So what’s new with our latest release? Here are the three big breakthroughs we’ve delivered:

• Move As Fast As Your Ideas: Weeks of analyst work? Gone. With a new, chat-based interface, the Analyst Agent delivers answers at the speed your campaigns move. That means faster experiments, faster learnings, and faster revenue impact.
• Intelligence Without the Overhead: The Analyst Agent is built for the way marketers think and doesn’t require a technical background. Simply ask a question—about customer cohorts, regions, or content variants—and the agent adapts to surface what matters most. It’s powerful enough to handle complex questions, but intuitive enough for anyone to use.
• Enterprise-Grade Security and Reliability by Design: Connect securely to your existing data with the same compliance, privacy, and performance standards you expect from any mission-critical tool. Auxia does not share data across companies, ensuring your critical insights and data remain private and protected by default.

From Insight to Advantage: How Analyst Agent Compounds Value

Unlike traditional analytics tools, the Analyst Agent is designed to uncover deeper, campaign-level insights. It’s not just reading data—it’s learning from it.

What makes the Analyst Agent unique is that it connects directly to Auxia’s proprietary treatment framework and decisioning data. This means the agent operates on top of proprietary data and intelligence that already understands who saw what, when, and why–unlocking real-time insights that a generic BI tool can’t replicate.

Because every Auxia-powered experience is already structured for measurement from the start, the Analyst Agent can automatically isolate causal effects, compare treatment variants, and synthesize learnings without manual setup. The result: precise, actionable insight—without SQL, delays, ambiguity, or guesswork.

The best part? The more you use it, the better it gets. Every interaction with the Analyst agent adds to a growing body of institutional knowledge for your organization. It logs your questions, tests, and outcomes into a structured knowledge base, remembering prior decisions and identifying crucial patterns. Over time, it starts to surface the right insights before you even ask—reducing repetitive work, accelerating learning, and making your team smarter with every session.

With the Analyst Agent, growth intelligence doesn’t just scale—it compounds.

Using the Analyst Agent to Drive Smarter Decisions

The Analyst Agent isn’t just a tool you query—it’s a partner in your thinking.

From the moment you engage, it guides you through a collaborative exploration process designed to unlock insight, even when you're not sure what to ask. It can:

• Propose instructions for exploration based on the analyses it can run
• Ask clarifying questions to refine the scope of your inquiry
• Adapt in real time based on your responses, feedback, and hypotheses

Whether you're pressure-testing a campaign strategy or chasing an unexpected spike in conversions, the agent helps frame the right questions and drives you toward the “why”.

What’s Next

The enhanced Analyst Agent is rolling out to select customers now, and we’ll be expanding access throughout the summer. If you're already using our AI Decision Agent, you're going to love what this unlocks. And if you're new to Auxia, this is the perfect time to explore what Agentic Customer Journey Orchestration can really do for your organization.

If you’re curious to see it in action, fill out this form for a demo.

Today, we’re excited to announce Auxia’s official launch in Japan as we expand the reach of our Agentic Journey Orchestration Platform. Across countless conversations with marketing and product teams, we’ve consistently heard a similar theme: teams want to deliver more relevant, personalized experiences—but face real constraints around time, resources, and complexity.

We believe AI can shift this paradigm. Auxia enables teams to move faster, experiment more intelligently, and scale personalization in ways that were previously out of reach. With this launch, we’re thrilled to bring our platform to one of the most sophisticated and quality-driven markets in the world—and to support Japanese organizations in shaping the next generation of customer experiences.

Why Japan?

Japan is globally recognized for its exceptional standards in quality, precision, and customer experience—values that deeply resonate with us at Auxia. From the beginning, Japan has been a key market in our global vision. Today, we’re proud to introduce our Agentic Journey Orchestration Platform more broadly to the region. We've already partnered with several forward-looking enterprises, and we see tremendous potential to support Japanese companies as they navigate the global shift toward more intelligent, AI-powered customer engagement.

Local Team and Leadership

Leading the Auxia’s Japanese operations is Hirotaka Yoshitsugu, who has been appointed as CEO of Auxia Japan K.K.

Yoshitsugu brings over 20 years of experience in building platform ecosystems. After managing i-mode partnerships at NTT Docomo, he helped launch the Tokyo office of AdMob. Following Google’s acquisition of AdMob, he joined the U.S. headquarters, where he built the international expansion of Google Play from the ground up. Later, he returned to Japan to lead Google Play’s partner business locally.

His deep roots in Japan’s tech industry make him an essential leader for Auxia’s expansion in the region.

Bringing Agentic Journey Orchestration to Japan

Today, Auxia powers customer experiences across a wide range of industries—including finance, retail, media & entertainment, and telecommunications. But our platform offers far more than simple automation.

At its core, Auxia helps teams reimagine what’s possible by removing the barriers that traditionally slow personalization down. Our platform:

• Activates all your first-party data, eliminating the need for months of engineering work by automatically handling the infrastructure required to deploy machine learning models into production.
• Enables rapid experimentation at scale—testing hundreds of hypotheses in parallel without relying on rigid, rules-based journeys or time-consuming manual A/B tests.
• Delivers the right message, action, offer, timing, and frequency for each individual customer. Just define your goal, set the guardrails, and Auxia intelligently handles the rest.
• Continuously adapts and optimizes in real time, learning from every interaction to improve outcomes—without manual tuning.
• Surfaces granular behavioral insights, highlighting subtle differences across customer segments that are difficult to identify with traditional tools.

And perhaps most importantly, Auxia puts all of this power directly in the hands of marketers—no technical expertise required.

What's Next

In the coming months, we’ll be establishing a physical presence in Japan and deepening our partnerships across a wide range of industries. As part of this commitment, we’re actively hiring for key roles in customer success, solutions engineering, research, and more.

Our priority is to build a team rooted in local expertise—professionals who deeply understand the unique dynamics of the Japanese market and share our ambition to transform how customer experiences are designed and delivered. By investing in talent, infrastructure, and long-term collaboration, we’re laying the foundation to support our customers in Japan with the precision, responsiveness, and care they deserve.

Looking Ahead: Building for Long-Term Impact in Japan

To all enterprises and organizations in Japan looking to harness AI to elevate your customer journeys — we invite you to join us in building the next layer of intelligent infrastructure.

We’re excited to collaborate with you and help shape the future of your business, together.