A study found that 68% of institutions have moved fewer than 30% of their AI experiments into full production, showing that most efforts remain at the pilot or proof-of-concept stage. Data shows that this has only gotten worse over time. A recent survey of 650 enterprise technology leaders found 78% have at least one AI agent pilot running — yet only 14% have scaled an agent to organization-wide production use. Meanwhile, S&P Global data shows the share of companies abandoning most of their AI initiatives jumped from 17% in 2024 to 42% in 2025. The cost is not just wasted sprint cycles. It is trust erosion with boards, demoralized engineering teams, and market position handed to competitors who figured out the lab problem first.
We’re finding that the real competition in AI is not happening in the product itself, but earlier, in the systems companies use to experiment, test, and decide what AI should do before customers ever see it. Enterprise AI has followed a familiar trend: companies are adding AI features to their existing platforms, workflows are getting copilots, and roadmaps promise small improvements in efficiency. This happens because adding AI to existing workflows is the easiest first step, but it is also where it’s hardest to stand out.
The market does not need more AI added to old, legacy systems. It needs new systems, built from scratch, with intelligence embedded into how work happens. This kind of change starts before the product even exists.
The Upstream AI Engine
To build AI that is ready for enterprise use, companies need something new. They need a fast-moving experimentation engine that works before the main platform. This is where ideas are quickly prototyped, tested in real workflows, and checked with customers before scaling up. At Xactly, we’ve built a lab that works upstream, where new ideas are created, tested, and proven with customers. It is a fast-paced environment where we design, prototype, and validate new AI features before they become part of the product.
The lab is built on a four tier upstream model:
- The ideation tier includes rapid prototyping with LLMs, designing in co-creation sessions, and workflow mapping.
- The experimentation tier includes a multi-LLM evaluation, RAG pipeline testing on our proprietary data corpus, and Model Context Protocol (MCP)-driven agent workflow construction.
- The governed validation tier includes end-to-end agentic workflow runs with human checkpoints, and compliance, audit-trail, and data residency reviews.
- The platform integration tier includes hardened feature graduates to the platform through a composable architecture.
The connective tissue of this lab is the MCP — originally developed by Anthropic and now an open standard adopted by OpenAI, Google, and Microsoft. MCP functions like a USB-C port for AI agents: a standardized interface that lets any compliant agent connect to APIs, platform functions, customer data sources, and third-party systems without rebuilding integrations for every new model we evaluate. Companies using MCP-based architectures report saving 25% of the time to build AI systems involving multiple models. More importantly, it means we are never locked into a single LLM vendor — a critical hedge as the model landscape continues to shift rapidly.
Instead of making changes after launch, the focus is on proving what really adds value before anything becomes part of the product. This difference is important because, in AI, being wrong creates operational debt, financial risk, and loss of trust. If a system gives results that cannot be explained, checked, or controlled, organizations will not use it.
From Experimentation to Differentiation
In the past, research and development in enterprise software was slow and not very transparent. The leading companies in this new phase will build and learn faster by moving quickly from prototype to customer feedback, then using those insights to improve.
At Xactly, this model is already producing results at pace. We’ve released products and features, like AI agents, that were not born from a product roadmap. They were born from the lab. That distinction matters enormously to the speed at which we can respond to what customers actually need, rather than what we assumed they needed six quarters ago.
My team builds prototypes quickly, tests them with design partners, and tries out new LLM patterns, MCP-driven agentic workflows, and retrieval systems to see what really adds value.
Rather than shipping isolated features, this approach focuses on building and testing agent workflows that can handle complex processes from start to finish. These systems are judged on how they work and whether they meet enterprise standards like governance, explainability, and compliance.
Additional Effects of the Development Model
There is a dimension of this model that is rarely discussed in public: what it does to engineers. The best engineers do not just want to build features. They want to work on problems that matter, on systems that are new, with enough autonomy to take technical risks. PwC’s 2025 Global Workforce Survey of nearly 50,000 workers found that employees with the highest levels of psychological safety — the freedom to experiment and learn from failure — are 72% more motivated than those who lack it. The Jellyfish State of Engineering Management Report found that organizations where leaders encourage grassroots AI experimentation see the strongest combination of productivity gains and engineer satisfaction. Our upstream lab model creates exactly this environment. Engineers at Xactly are not just implementing product tickets. They are co-designing the future of the sales performance management category — with direct access to real customer problems, real data, and real tools. Tenure, output quality, and internal NPS on our engineering teams have all reflected this shift. In a market where top AI engineering talent has more choices than at any point in the industry’s history, building a lab culture is also a talent strategy
Another big change is who takes part in the innovation process. It’s best to directly involve partners and customers in shaping use cases, testing workflows, and setting requirements. This way, the path to product-market fit is much shorter, and what gets built matches real needs instead of guesses. With Xactly’s AI Council and Customer Advisory Board, we involve customers in the invention process, shaping features based on real demand and real data from top sales teams. In a world where it is easy to make impressive demos that do not work in real life, this grounding is essential.
All of this shows a bigger change in how enterprise AI is built. The industry is moving from features to systems, from assistants to agents, and from outputs to real results. Enterprise AI will be won by those who build the best systems for figuring out what is worth building in the first place.
