Before electric refrigerators, families needed an iceman to deliver ice blocks to keep their food cold. When the electric refrigerator was invented, it wiped out the need for ice delivery. Ice companies tried to save their businesses by buying faster trucks. But that wasn’t the issue. They optimized the wrong solution instead of adapting to the new technology.
Many companies chasing AI today are making the same mistake.
Survey after survey confirms the same thing: most organizations are struggling to scale AI beyond a handful of isolated use cases and aren’t seeing meaningful business impact.
Most companies are attacking AI the same way they’ve attacked every previous wave of technology, by copying what others are doing, layering new tools on top of old processes, and hoping the outcomes will be different this time. It’s called "reasoning by analogy." And it’s why so many AI initiatives stall, overspend, and underdeliver.
Elon Musk has built SpaceX, Tesla, and more by refusing to accept that the way things have always been done is the way things must be done. His problem-solving framework, anchored in "first principles thinking," offers a better blueprint for approaching AI. It’s not complicated. But it requires discipline, honesty, and a willingness to change.
Here is how it works and what it looks like when applied to AI challenges inside a business.
First principles thinking means stripping a problem down to its most basic, provable truths and building up from there. Rather than asking "how have others solved this?" you ask "what is actually true about this problem?"
When Musk wanted to reduce the cost of rocket manufacturing, he didn’t look at what other aerospace companies were charging. He asked what rockets were actually made of, calculated the raw material cost, and concluded that the industry was massively overpricing the end product. That insight was the foundation for SpaceX.
Companies implementing AI need to apply the same rigor. Start by asking what business outcome you’re actually trying to achieve. Not "we want to use AI," but "we need to reduce our invoice processing cost by 40%" or "we need to close the books 3 days faster." The technology choice follows the outcome, not the other way around.
Invoice Processing Example: Instead of automating the existing approval workflow, step back and ask why invoices require manual intervention at all. The root cause is often mismatched data between purchase orders, receipts, and invoices. An AI solution built on first principles would focus on eliminating those mismatches upstream, so invoices never require human review to begin with.
WIP Accruals Example: Rather than using AI to speed up the manual accrual process, ask what drives inaccuracy in WIP estimates. The answer is usually a lag in project progress data and inconsistent coding by project managers. A first-principles AI solution addresses the data problem directly, automating the collection of real-time progress inputs and flagging coding anomalies before they distort the accrual.
Most processes are built on assumptions that made sense at the time but have never been revisited. First principles thinking requires surfacing those assumptions and testing whether they still hold.
Musk questioned the assumption that electric vehicles required expensive, custom battery packs. By sourcing and combining standard lithium-ion cells used in consumer electronics, Tesla dramatically reduced battery costs while the rest of the industry lagged behind.
In AI implementations, the most dangerous assumptions are process-level ones. "We have always routed invoices to three approvers." "Accruals are always prepared by the project accountant." These constraints feel permanent but are often just habits.
Invoice Processing Example: Companies assume that every invoice needs human review for compliance reasons. When examined carefully, most compliance checks are rule-based and can be fully automated. AI can apply those rules consistently at scale, with exceptions flagged for human review rather than every invoice.
WIP Accruals Example: Companies assume the project manager is the only reliable source of percent-complete estimates. AI can cross-reference labor hours posted, materials consumed, and milestone completion data to produce an independent estimate, reducing dependence on a single subjective input and improving accuracy.
There is a difference between constraints that are real and constraints that are inherited. First principles thinking forces you to separate the two.
Musk was told that reusable rockets were technically impractical. He did not accept that as a hard constraint. He challenged it, invested in solving the engineering problems, and created a model that fundamentally changed the economics of space travel.
In AI, the inherited constraint most companies accept is their existing process. AI gets layered on top of a workflow that was designed for humans doing manual work. The result is incremental improvement at best.
Invoice Processing Example: The inherited constraint is a three-way match process that requires a human to reconcile purchase order, receipt, and invoice line by line. AI can perform that match in milliseconds. But if the process is designed to wait for human confirmation before payment is released, the speed advantage is lost. Challenging that constraint means redesigning the approval flow around AI-confirmed matches rather than human ones.
WIP Accruals Example: The inherited constraint is a monthly close cycle where project data is captured once a month. AI can pull and process project data continuously. But if the accrual schedule is built around a monthly cadence, that capability goes unused. Challenging the constraint means moving to a rolling accrual model that AI can support in real time.
Complex problems feel unsolvable until they are broken into their constituent parts. Each component can then be evaluated and addressed independently.
When designing the Tesla Gigafactory, Musk didn’t look at battery production as a single problem. He decomposed it into energy inputs, material flows, manufacturing steps, and logistics, then optimized each component separately before integrating them.
AI implementations that fail usually treat the business problem as a monolith. They try to automate an entire process in one pass and get stuck on the complexity. The better approach is decomposition.
Invoice Processing Example: Invoice processing breaks down into data capture, validation, matching, approval routing, exception handling, and payment release. AI can be applied to each component individually. OCR handles data capture. Matching algorithms handle three-way reconciliation. Rules engines handle routing. Exceptions go to humans. Each layer is solvable. Trying to solve them all at once usually means solving none of them well.
WIP Accruals Example: WIP accruals break down into data collection, percent-complete estimation, cost allocation, variance detection, and reporting. AI can be applied to data collection and estimation first, which are the highest-effort, lowest-accuracy components of the current process. That alone can reduce close time significantly before the full solution is complete.
Once you have challenged assumptions and decomposed the problem, the question is whether to improve the existing process or replace it. First principles thinking often leads to a fundamentally different design.
SpaceX didn’t improve the Space Shuttle. They built a different kind of rocket entirely, with reusability and cost reduction as the design parameters from the start.
Most companies automate the process they already have. A first-principles approach asks what the process would look like if it were designed today, for the first time, with AI as a native capability rather than an add-on.
Invoice Processing Example: A process rebuilt from scratch wouldn’t have an approval queue at all. Invoices that match validated purchase orders and receipts would be paid automatically. Only invoices that fall outside defined tolerances would ever reach a human. Straight-through processing would be the default, not the exception.
WIP Accruals Example: A process rebuilt from scratch wouldn’t require a project accountant to manually compile accrual data at month-end. AI would continuously monitor project activity and maintain a running accrual estimate throughout the month. The close process becomes a review and sign-off on a number that is already calculated, not a calculation that starts from scratch each period.
First principles thinking ultimately asks “what are the actual limits here?” Not the limits of current practice, but the limits of physics, economics, and logic.
Musk calculated that the raw materials for a rocket battery pack cost roughly 2% of what the industry was charging for the finished product. That gap between what was possible and what was being delivered became the target.
For AI, the equivalent question is “what would this process look like if cost and time were at their theoretical minimum?” That target defines the ambition. Current performance defines the gap. The AI implementation strategy should be designed to close that gap, not to make marginal improvements to the status quo.
Invoice Processing Example: The theoretical minimum for a matched invoice is milliseconds. Current average processing time in most companies is measured in days. AI-enabled straight-through processing can move the needle from days to hours for the majority of volume. That’s not a 10% improvement. It’s a structural change in how the process operates.
WIP Accruals Example: The theoretical minimum for WIP accrual accuracy is a real-time reflection of actual project activity. Current accruals are snapshots taken once a month, often based on estimates that are days old by the time they are recorded. AI can move companies from monthly snapshots to continuous estimates, reducing close time and improving the reliability of project financials.
AI will not deliver transformational results if it’s applied using the same thinking that designed the processes it is supposed to replace. Companies must be willing to go back to first principles to see return from AI. Big budgets and dozens of pilots don’t automatically grant an advantage. Challenge what you’ve always assumed, and design something better. Don’t focus on the technology. Focus on the thinking.
At Trenegy, we help organizations deploy AI solutions that enhance efficiency, improve decision-making, and create immediate business impact. To chat more about this, emailinfo@trenegy.com.
Trenegy is a non-traditional consulting firm based in Houston, Texas.
