The Permanent Bet: Why Your AI Tooling Due Diligence is Failing on a Decade Scale

The AI tooling decisions you make today are not merely software purchases; they are permanent strategic commitments that will determine your enterprise's agility, cost structure, and competitive viability for the next decade, if not longer. Most companies are getting this fundamentally wrong. They approach AI vendor selection with a transactional, short-term mindset, focused on immediate benchmarks and quarterly gains. At Junagal, where we build, own, and run technology companies with permanent capital and a decade horizon, we view every AI tooling vendor as a potential co-architect of our future. This perspective forces a radically different, far more rigorous due diligence process — one that exposes the hidden liabilities and strategic lock-ins that will otherwise suffocate innovation within 3-5 years.

Our Decade Horizon: The Junagal Imperative

At Junagal, we operate under a unique mandate: permanent capital. This isn't a venture fund with a 5-year exit clock ticking; it's capital deployed for perpetuity. This singular difference reshapes everything we do, especially how we select the foundational technologies our companies are built upon. When we evaluate an AI tooling vendor, we're not just looking at their current feature set or their immediate roadmap; we're assessing their potential to become a foundational pillar for businesses we intend to own and operate for a hundred years. This isn't hyperbole; it's our core operating principle. This means our due diligence isn't about minimizing short-term risk, but about maximizing long-term optionality and resilience. We don't just buy a tool; we are, in essence, integrating a strategic partner into our core technological DNA, a partner whose choices will directly impact the total cost of ownership, technical debt, and competitive advantage of our ventures for generations.

The Illusion of Agility: Why Most Enterprises Get It Wrong

The prevalent 'move fast and break things' mantra, while having its place in early-stage product development, is catastrophic when applied to core AI infrastructure. I see countless enterprises, seduced by slick demos and impressive model benchmarks, rushing to adopt AI platforms without truly understanding the long-term implications. They're prioritizing speed to market over foundational soundness, leading to an inevitable reckoning. We've witnessed companies, even Fortune 500s, adopting specific niche agent frameworks only to hit unforeseen scaling bottlenecks, data egress cost explosions, or model deprecation headaches within 18 months. The initial 'agility' quickly devolves into expensive re-platforming exercises or, worse, strategic paralysis. This isn't just about technical debt; it's about embedding brittle decision-making into the very fabric of an AI-native future.

The 12 Questions We Ask: A Framework for Permanent Relationships

Category 1: Foundational Architecture & Resilience

• Q1: What is your true multi-cloud/hybrid story, and where is the escape hatch?

  Many vendors claim multi-cloud, but dig deeper, and you find they're heavily optimized for a single provider, making true portability excruciatingly expensive. We need to understand the cost and complexity of migrating our data and workloads should we need to diversify our cloud strategy or pull data on-premises. This isn't about hypotheticals; it's about maintaining optionality against future geopolitical shifts, cost fluctuations, or vendor-specific outages. We need precise figures on egress fees, re-hydration costs, and API compatibility with alternatives. We look for tangible commitments to open standards like Kubernetes, Apache Arrow, or MLflow, rather than proprietary orchestration.

• Q2: How do you handle model evolution, and what's your plan for supporting previous versions and migrations?

  The pace of AI model innovation is relentless. A powerful model today can be obsolete tomorrow. We need to know how a vendor manages this lifecycle: Do they provide clear deprecation policies? What tools exist for automated fine-tuning migration? Will our proprietary fine-tunes on an older version remain performant, or will they degrade? A vendor with a robust plan for backward compatibility and assisted migration pathways (like some mature MLOps platforms from Databricks or Snowflake) is far more appealing than one that forces wholesale re-training every six months.

• Q3: What's your story on hardware abstraction? Are you tying me to a specific GPU/TPU generation or vendor?

  The 'AI factory' concept is becoming central to compute strategy [2]. While NVIDIA's advancements, for example, are undeniable, true long-term resilience requires a degree of hardware independence. We probe how vendors abstract away underlying compute infrastructure. Are they built on open standards like OpenCL or ROCm, or are they deeply embedded in proprietary stacks? The goal is to avoid situations where our AI compute becomes irrevocably tied to a single hardware vendor's release cycles and pricing models, limiting our ability to leverage future innovations or cost efficiencies from diverse silicon providers.

Category 2: Data Sovereignty, Lifecycle & Explainability

• Q4: Precisely where does my data live, and what are the specific guarantees around its use for model training or improvement?

  This is non-negotiable. Data is our core asset. We demand explicit, legally binding assurances that our proprietary data, including any fine-tuning datasets, is never used to train the vendor's foundation models or shared with third parties without our explicit, granular consent. This goes beyond standard privacy policies; it's about competitive advantage and intellectual property. We prefer vendors who offer true data isolation, like private endpoints and dedicated instances, ensuring our data remains ours and ours alone. Vendors like Palantir, focused on highly sensitive data environments, often have stronger stances here than those prioritizing broad network effects.

• Q5: What is your data governance and lineage story, from ingestion to model output?

  In complex AI systems, understanding how data transforms and impacts model outputs is paramount for auditing, compliance, and debugging. We expect a clear, auditable trail for every piece of data: where it came from, how it was processed, which models it trained, and what decisions it influenced. This is particularly crucial for regulated industries. If a vendor cannot articulate a robust, automated data lineage and governance framework (e.g., integration with tools like Apache Atlas or Collibra), they're a non-starter. This isn't a 'nice to have'; it's fundamental to responsible, long-term AI deployment.

• Q6: How do you facilitate debugging, explainability, and error handling for agentic workflows?

  As AI agents become more autonomous, their 'black box' nature becomes a critical risk. When an agent misbehaves or makes a suboptimal decision, we need to understand *why*. We push vendors on their tools for tracing agent thought processes, understanding model activations, and isolating error sources in complex, multi-step workflows. This is especially vital in applications like those being explored by AdventHealth with OpenAI [12], where explainability in healthcare decisions is paramount. If we can't debug an agentic system, we can't trust it at scale.

Category 3: Ecosystem, Interoperability & Openness

• Q7: What is your true commitment to open standards and open source? Where do you draw the line between proprietary core and community contribution?

  We actively seek vendors who contribute to and embrace the open-source community, seeing it as a long-term hedge against lock-in. Companies like Hugging Face, or even Warp's bet on building open source with GPT-5.5 [6], exemplify a healthier ecosystem approach. We examine their license structures, contribution models, and how easily we can extend, modify, or even replace components with community alternatives. A vendor that maintains a completely closed-source, proprietary ecosystem without clear extension points is a red flag for long-term viability and flexibility.

• Q8: How do you integrate with the broader AI ecosystem beyond your platform?

  No single vendor will own the entire AI stack. We need to understand their integration philosophy with other essential tools: our existing data warehouses (Snowflake, Databricks), MLOps platforms, vector databases, and domain-specific applications. We're wary of 'walled garden' approaches. True interoperability, often through robust APIs, SDKs, and commitment to industry standards (e.g., for model serving, data formats), is crucial. This ensures we can compose the best-of-breed solutions rather than being limited by one vendor's offerings.

• Q9: What's your API contract's stability and versioning strategy for major architectural shifts?

  Breaking API changes are the bane of long-term development. For a business built to last, API stability is paramount. We scrutinize their versioning policies, their commitment to backward compatibility, and the transparency of their deprecation roadmap. We want to see a history of thoughtful API evolution, not frequent, disruptive changes that force expensive refactoring on our teams. We benchmark against companies like Stripe, known for their developer-friendly APIs and clear versioning.

Category 4: Talent, Sustainability & Strategic Alignment

• Q10: What's your hiring pipeline for prompt engineers, ML architects, and data scientists, and how do you support my team's upskilling?

  The best tooling is useless without the talent to wield it. We evaluate vendors not just on their product, but on their commitment to ecosystem growth. Do they offer comprehensive training? Are certifications meaningful? Do they actively foster a community of practitioners? We look for evidence that they are investing in the human capital necessary to support their platform's long-term adoption, recognizing that a vendor's success is intertwined with our team's ability to maximize their technology.

• Q11: What's your business model's long-term sustainability, beyond the next VC round?

  For permanent capital, we need permanent partners. We conduct deep financial due diligence, looking beyond valuation to unit economics, customer churn, and runway. We need confidence that the vendor will be around for the long haul, investing in R&D, and not be forced into an acquisition or pivot that disrupts our operations. A heavily subsidized freemium model without a clear path to profitability is a major red flag, regardless of how innovative the tech.

• Q12: How do you handle ethical AI considerations, bias mitigation, and regulatory shifts in my specific industry?

  Responsible AI is not a checkbox; it's an ongoing commitment. We demand a clear articulation of their approach to fairness, transparency, and accountability. This includes their internal governance, their tools for bias detection and mitigation, and their strategy for adapting to evolving regulations (e.g., GDPR, sector-specific AI laws). For our portfolio companies operating in highly regulated sectors like healthcare or finance, a vendor's proactive stance on ethical AI is as critical as its performance metrics.

The Strongest Counter-Argument: Speed Over Rigor

I can already hear the counter-argument, loud and clear: "Anil, your process is too slow! In AI, speed is everything. The landscape changes weekly, and if we spend a year on due diligence, we'll be irrelevant before we even launch. You can't afford a year-long procurement process. Get something in production, iterate, course-correct. The cost of delay outweighs the cost of a suboptimal choice. This rigorous, permanent-capital approach is a relic of pre-AI enterprise software procurement."

This is a compelling argument, especially in the breathless pace of AI innovation. I acknowledge the imperative for speed. Nobody at Junagal advocates for paralysis by analysis. However, this argument fundamentally misunderstands the nature of AI tooling lock-in. It's not like swapping out a CRM; it's like performing organ transplant surgery on a living, breathing organism. When you deeply embed an AI model, its data pipelines, and its agentic logic into your core operations – say, automating inventory for a global retailer like Walmart or optimizing supply chains for a grocer like Kroger – extracting it becomes astronomically expensive. It’s not just code; it’s data gravity, specialized fine-tuned models, embedded knowledge, and the operational inertia of hundreds or thousands of workflows. A superficial 'fast' decision creates strategic debt that doesn't just slow you down, it can permanently cripple your ability to adapt. We've seen projects with initial 'speed' of 6 months take 2-3 years to unravel and replatform because the initial vendor choice created an unmanageable proprietary data format lock-in, forcing a complete restart. The true cost of that 'speed' was billions in lost market opportunity and wasted engineering cycles.

What We Got Wrong: The Human Integration Gap

Despite our meticulous technical due diligence, we've had our share of painful lessons. One particular failure mode stands out: underestimating the 'human API'. We once selected an AI data labeling and annotation vendor for a portfolio company in the logistics sector (think advanced robotics for warehouse management, akin to Ocado or JD.com's automation efforts). Their technology was cutting-edge, offering unparalleled accuracy and speed through novel active learning algorithms. Our 12 questions were rigorously applied, and they passed with flying colors on architecture, data security, and model lineage.

What we got wrong was the *velocity of human integration*. While their tech was brilliant, their support model was rigid, their documentation updates were slow, and their onboarding process was geared towards static, large-batch projects, not the iterative, fast-evolving needs of an AI-native startup. Our internal ML engineering team, operating at breakneck speed, frequently hit roadblocks that required custom workarounds because the vendor's support response times were measured in days, not hours. The friction points mounted. The 'seamless integration' we anticipated became a continuous battle against a bureaucratic support structure. We eventually had to pivot to a technically 'inferior' but operationally superior competitor who deeply understood the need for rapid iteration and offered a co-development model, even if their core algorithms were slightly less advanced. The lesson: the best technology, if it doesn't seamlessly integrate with your human operating rhythm, becomes a drag, not an accelerator.

The Road Ahead: Building for a Century, Not a Cycle

The era of treating AI tooling as a disposable commodity is rapidly drawing to a close. As AI moves from a specialized function to the central nervous system of every enterprise, the stakes of vendor selection become existential. Companies that fail to internalize AI due diligence as a core strategic capability will find themselves increasingly trapped in proprietary data-model silos, unable to adapt as the landscape shifts from general intelligence to specialized, composable intelligence.

My prediction is stark: within the next five years, the market will bifurcate dramatically. There will be the AI-native leaders, like the next Stripe or Shopify, who built their core on open, flexible, and strategically sound AI tooling partnerships, giving them unparalleled agility. And then there will be the laggards, drowning in AI technical debt, perpetually playing catch-up, their innovation cycles dictated by their vendors' roadmaps rather than their own strategic intent. Your call to action is immediate: reframe your procurement teams from cost centers to strategic architects. Bring your sharpest technologists, legal minds, and operational leaders to the table. Treat AI tooling selection as a permanent capital decision, because for your business, it will be.