Here’s what’s on our plate today:

• 🧪 Why AI budgets explode the moment you leave the pilot stage.

• 🗞️ Meta's face problem, OpenAI buys a brain, AI influencers crash Coachella.

• 🗳️ Wednesday Poll on where you actually stand on the AI cost curve.

• 🧰 Three tools worth having open this week: Leonardo AI, Modal, and QuillBot.

Let’s dive in. No floaties needed…

The Laboratory

TL;DR

• Cheap entry, expensive stay: $20/month subscriptions made experimentation accessible, with 88% of organizations now using AI in at least one function. But production deployment reveals hidden costs in data prep, integration, and change management that can exceed entire budgets.

• Pilots keep crashing on contact: Around 95% of generative AI pilots fail moving from test to deployment, with cost underestimates running 5x to 10x. Only about 5.5% of companies are seeing meaningful financial returns.

• Three uncertainties compound the problem: Usage-based pricing can turn a $5K/month chatbot into $50K/month at scale. Regulatory requirements add compliance costs. And most SMEs lack the tools to measure whether AI is delivering returns.

• Staged rollouts beat perfect planning: Fixed spending caps per team, a mix of open-source and commercial models matched to task complexity, and cost monitoring from day one. Treat AI budgets the way companies learned to manage runaway cloud spending.

• Waiting has its own price tag: Gartner projects $2.5T in worldwide AI spending by 2026. SMEs that wait for cost clarity risk finding their competitive landscape already reshaped by rivals who moved first.

Why small businesses can’t budget for their most important technology decision

Entrepreneurs are often described as people who jump off a cliff and build a plane on the way down. The idea captures the essence of building a business from scratch, where owners often make decisions first and figure things out later, learning and adapting under pressure rather than waiting for perfect conditions.

While these decisions can cover almost every aspect of the business, businesses often face one of their biggest challenges when adopting transformative technologies before the economics are fully legible.

Take the example of Cloud computing, SaaS, or mobile devices. In each case, companies that waited for perfect cost clarity often waited too long and risked losing out to the competition.

Today, AI presents the same bargain; however, the situation is a little more complex because, in AI, the cost structure is inherently unpredictable, shaped by variables that shift continuously and that most small and medium enterprises lack the tools or expertise to forecast.

Cheap to start, expensive to run

You see, for most of AI’s commercial history, meaningful adoption required data science teams, GPU clusters, and seven-figure budgets. All this changed in 2022 and 2023, when OpenAI, Anthropic, Google, and others began offering powerful models through simple APIs at what appeared to be accessible price points.

This meant a business could send a text prompt to an API endpoint and receive a response for fractions of a cent. Subscription tools like ChatGPT lowered the entry price to $20 per user per month, which, according to McKinsey’s 2025 State of AI survey, led to 88% of organizations regularly using AI in at least one business function, up from 78% a year earlier.

For SMEs, the accessibility was genuine: they could experiment with customer service automation, content generation, and document processing without hiring machine learning engineers.

However, despite the accessibility, the API price point that made experimentation easy also created a misleading impression of what production-grade deployment would actually cost.

The blind spot emerged in AI integration, data preparation, training, and ongoing maintenance, all of which become apparent only after a project moves past the pilot stage.

The hidden costs

For the entrepreneur figuring out how to build the AI system within their organization, data preparation is typically the first surprise. Most SME data sits in inconsistent formats across disconnected systems: spreadsheets, CRM tools, email inboxes, legacy databases.

Before AI can deliver reliable outputs, that data needs to be cleaned, structured, and connected. An estimated 96% of businesses lack the high-quality training data AI requires, resulting in $10K to $90K in unplanned data preparation costs. For a mid-sized company, this single line item can exceed the entire anticipated AI budget.

Then there is the cost of integration, which adds another layer: connecting AI to existing accounting, inventory, or customer management systems requires custom development whose complexity scales with the age and fragmentation of those systems.

CIO reported that data platforms are the top driver of unexpected AI costs, followed by network access to AI models. Then comes change management, training employees, redesigning workflows, and updating quality standards, none of which appear on a vendor’s pricing page, all of which vary from one business to the next.

When pilots break

These often-overlooked aspects of integrating AI into business functions are the best way to understand why 95% of generative AI pilots at companies fail.

Studies show that companies often underestimate the cost of AI projects by five to 10 times when moving from small tests to real-world deployment. A 2025 survey by McKinsey & Company found that only about 5.5% of companies were seeing meaningful financial gains from AI, while nearly two-thirds were still stuck in the testing stage and had not rolled it out across their business.

The visibility problem

The failure rate, however, does not justify inaction and should not be taken as grounds for not pursuing AI integration.

Gartner projected worldwide AI spending to reach $2.5T in 2026, driven by organizations that have correctly concluded that AI capabilities are becoming a baseline competitive requirement. AI-enabled competitors are resetting customer expectations around response speed, personalization, and service availability.

Similarly, supply chain optimization, fraud detection, and predictive maintenance are moving from differentiators to table stakes. An SME that waits for AI to become simple and cheap risks discovering that its competitive landscape has already moved on. The challenge, then, is not whether to invest, but how to invest without reliable cost visibility.

Currently, the cost calculation problem for SMEs is not a single gap. It is three distinct uncertainties that interact and amplify each other.

Three uncertainties

The first is pricing. Unlike older software with fixed subscription costs, AI pricing depends on usage. What starts as a $5k-a-month chatbot in testing can quickly turn into $50k a month when rolled out more widely, even if nothing about the tool itself has changed. More advanced systems that handle multi-step tasks make this worse, since they use far more computing power per request, often multiplying costs several times over.

The second is regulation. New rules, especially in the European Union, could require high upfront costs and ongoing compliance expenses. In the U.S., different state-level laws add further complexity.

The third is measuring returns. AI doesn’t deliver a fixed payoff; its impact changes over time as systems improve and expand into new uses. Most small and mid-sized businesses don’t have the tools to clearly track what AI is actually contributing, which makes it difficult to know whether the investment is paying off.

Together, these factors make AI spending far less predictable than it first appears, especially as companies move from small experiments to real-world use.

However, despite the challenges, several emerging practices make the uncertainty more manageable.

Managing the unknown

The most common advice is simple: roll out AI in stages. Start with a small, clearly defined use case, track costs closely, and expand only once you understand what it actually costs to run. For smaller businesses, it makes more sense to start with low-cost use cases where rough estimates suffice, rather than waiting for perfect financial clarity. The goal isn’t to avoid surprises, but to encounter them early when they are easier to manage.

One practical approach is to set fixed spending limits. Instead of treating AI as an open-ended expense, some companies assign a monthly budget for each team. If costs go over, it forces a review, whether that means improving how the system is used, switching to cheaper tools, or redesigning workflows. This mirrors how companies learned to manage cloud costs after realizing that unchecked usage could quickly spiral out of control.

Using open-source models is another option. Tools like LLaMA or Mistral can reduce usage-based costs, but they also require their own infrastructure and expertise. A mix of both approaches often works best, using simpler, cheaper models for basic tasks and reserving more advanced systems for complex work.

Most importantly, companies need to track costs from the start. That means monitoring usage, understanding which teams or tasks are driving spending, and reviewing it regularly. The businesses that struggle the most are not always the ones spending heavily, but the ones that only realize what they have spent when the bill arrives.

Uncertainty is the point

For companies that remain unsure about adopting AI, there is also the option of waiting it out.

AI pricing, regulation, and measurement tools are still evolving, and in a few years, the landscape may be clearer and more stable. But waiting comes with its own risk. Research from McKinsey & Company shows that the small group of companies already seeing meaningful returns invested early, before the economics were fully clear.

For many entrepreneurs, AI does not change the instinct to jump off a cliff; it changes the conditions under which the fall occurs. The ground is less visible, the materials are unfamiliar, and the cost of each adjustment is harder to predict. What separates the companies that navigate this shift from those that stall is not their ability to map the descent in advance, but their ability to build, test, and correct in motion, accepting that the plane will never be fully designed before it has to fly.

Quick Bits, No Fluff

• Meta's face problem: 75 civil liberties groups are demanding that Meta cancel facial recognition for Ray-Ban glasses, warning that it enables stalking, abuse, and ICE targeting without bystander consent.

• OpenAI buys a finance brain: OpenAI acquired personal finance startup Hiro, absorbing its team as it pushes deeper into AI-powered financial planning ahead of its IPO.

• AI influencers crash Coachella: AI-generated influencers flooded this year's festival with desert content and set reviews, never having left their servers.

Wednesday Poll

Meme Of The Day

The Toolkit

• Leonardo AI: AI image studio for concept art, product visuals, and styled assets at speed.

• Modal: Serverless Python platform to run AI and data workloads without touching infrastructure.

• QuillBot: AI paraphraser that rewrites, tightens, and clarifies text in multiple languages.

Rate This Edition