In baseball, “middle-middle” is used to describe a pitch thrown “right down the middle” of the strike zone. Usually, these pitches are crushed into the bleachers for home runs. On Twitter, it is common for baseball fans to caption highlights “middle-middle” with no further explanation. However, the implication is clear: The pitcher made the batter’s life infinitely easier, and it resulted in a home run.

In software, middleware is a distinct layer that connects and facilitates communication between two or more applications, acting as a “bridge.” Middleware is used when different applications, systems, or services need to interact but use different formats, protocols, or communication methods (e.g., a database, web application, or enterprise SaaS product). It translates, routes, or processes data to ensure compatibility between these systems.

Middleware is often passive and relatively deterministic, facilitating data flow and communication without actively initiating processes. As such, multiple pieces of middleware from different companies—each with its own distinct cost structure, customer service staff, and SLAs—can be attached to a single application or database to handle different functions.

For example, your Snowflake database might be connected to Fivetran for ETL (Extract, Transform, Load), DBT for data transformation, Apache Kafka for real-time event streaming, and Looker for data visualization and BI.

According to BetterCloud’s most recent annual State of SaaSOps report, the average company uses 112 SaaS tools. Among their 411 respondents, companies with under 200 employees had an average of 42 SaaS applications, a 1-to-5 SaaS app-to-employee ratio. However, for the first time in over a decade, the overall average number of SaaS apps per company declined, with 112 representing a 14% decrease from the prior year. This trend of software consolidation should persist as AI-driven applications fundamentally reshape how businesses manage their software ecosystems.

The backbone of AI consists of ever-changing models that can easily handle unstructured inputs but do not have deterministic outputs. On the one hand, this is excellent. It intuitively eliminates the need for middleware that “massages” data. You can tell an LLM what output structure to adhere to, and you don’t have to format the data perfectly before it is input. However, it seems like a nightmare for developers, who need to test ever-changing black boxes that respond completely differently to the same input.

Due to and powered by AI, there is a new kind of software being developed, which I like to call Middle-Middleware. In this case, an enterprise’s developers are the formidable slugger, and the Middle-Middleware is a program which makes previously-nightmarish feedback loops feel like hitting off a tee. Paradoxically, Middle-Middleware eliminates the need for software sprawl because it is AI-native and user friendly.

Middle-Middleware is like a pitcher who throws a “middle-middle” pitch: It translates a unique, unrepeatable, non-deterministic process (a pitcher’s windup; an LLM’s non-deterministic output) into a home run (a “middle-middle” pitch; a software bridge that consolidates AI models’ outputs into a consistent, user-friendly format). However, the implications of Middle-Middleware go beyond business intelligence.

Middle-Middleware sits in-between humans, not pieces of software. It does this by consolidating previously-separate SaaS products into one user interface. It makes deployment of an AI model—a job which some AI SaaS companies need months and numerous bespoke integrations to complete—take mere hours or days. A great example of Middle-Middleware is the company NLX.

NLX is the epitome of Middle-Middleware. Its tagline is “the application layer for conversational AI,” but that pithy statement undersells the company’s true capabilities. Trusted and deployed by several Fortune 100 companies—including United and Comcast—NLX truly handles everything sitting between a human developer and a human customer, adding ease and delight every step of the way. It is currently deployed worldwide in over 65 languages, helping real-world customers while other companies struggle to duct-tape together disjointed SaaS tools that aren’t AI native.

NLX’s Generative Journey is “the Wordpress for AI,” a low-code/no-code solution that sets boundaries and scripts for AI voice and chat agents, without developers having to handle every edge case. It generates dozens of test cases for you and shows you the output, ensuring your conversational agent is on-brand for all interactions. Its Voice+ (try the Showroom) allows you to make agents with multi-modal output. If your customer is engaging with AI over chat on their computer, it allows them to take a picture with their phone. It allows your voice agent to text a receipt or ticket to a customer while still on the call. This technology is patented and wholly unique. Finally, their Journey for Retail transforms the shopping experience, allowing you to type, text, or talk to a personal retail assistant. A customer can say, “I need an outfit for a black tie event. I’m a size 2 and want size 5 2” heels. Keep the entire outfit below $400. I need a jacket because it’s winter in New York City.” NLX will return a full outfit, ready to check out. This fundamentally changes a “chat bot” interaction from reactive to proactive and leads to a “one cart” customer experience akin to TikTok or Instagram Shopping. Your store can now have the UX of a trillion-dollar social media juggernaut.

Middle-Middleware is the “one ring to rule them all.” Many investors argue that foundation models will become commoditized while trying to invest in 2010s software infrastructure. This is a mistake. The superpower of AI models—their ability to handle myriad data formats, as long as they are told what to do—means that one platform should intuitively handle many operations that sit between a human developer and a human user. You should have one piece of software that communicates directly with your database, provides business intelligence, and allows you to self-service deployments of various commoditized AI models.

Middle-Middleware makes foundation models truly interchangeable. A developer’s workflow shouldn’t change if she is using Claude or Gemini or OpenAI. She should have one platform to deploy, test, and analyze AI for numerous use cases.

Too many AI companies are trying to convince customers they need costly and slow integrations, to add one more wobbly string of SaaS spaghetti to their existing software sprawl. Then, they try to argue that you need to pay “outcome-based pricing.” Great, let my vendor tell me how much they deserve to be paid based on metrics only they see the inputs for. It’s Facebook ads all over again.

Convincing a customer that AI is just another confusing SaaS tool that requires even more bespoke integrations and agreements is great for the disingenuous AI companies: Perpetual service contracts and “use it or lose it” NRE are high-margin, recurring revenue. However, it is the worst of both worlds for customers: It is the slow deployment of yesterday combined with the unpredictable performance of today. In reality, AI allows for a massive consolidation of software tools and should enable even better self-service for companies of all sizes. AI gives users lots of control. That control should accrue to developers, not SaaS companies. The answer is Middle-Middleware.