What Entrepreneurs Can Learn from Weapons of Math Destruction
Mara had run her bakery-supply startup for two years before applying for a loan to fund a delivery van. The online application took twenty minutes, and within seconds an automated underwriting system returned a decision: denied, citing unspecified "risk factors." No loan officer had read her business plan or noticed that her one missed payment years earlier traced to a supplier's shipping delay, not mismanagement. The system had compressed signals loosely tied to her ability to repay into one score. Mara could not see which factor sank her application, and the support line only repeated that the decision was final.
The whole book, compressed
This edition works through 4 core lessons — take them one at a time, not all at once.
What's different for your conditions
Tap whichever ones sound like you. Nothing here is sent or saved — it's just for your own noticing.
0 of 5 selected
The core lessons, applied to you
Why the numbers can look neutral and still be wrong
A credit or risk model looks like pure arithmetic, but every input it weighs reflects choices made by the people who built it. When those choices draw on historical lending patterns, they can quietly reproduce old biases against certain neighborhoods or business types, without anyone intending harm. This does not mean every scoring system is broken; a transparent model with tested inputs and a real appeals process can beat an inconsistent human reviewer. The danger grows when a model's logic is hidden and its errors compound unchecked.
When a model treats a thin credit file or unusual revenue pattern as a red flag rather than a data gap.
Before signing with a lender or platform, ask in writing what categories of data factor into account decisions.
Where scoring already touches your business
These systems show up across tools entrepreneurs rely on daily: a lending platform's underwriting engine, an insurance carrier's pricing model, a marketplace's seller-rating algorithm, and dispatch software allocating gig work. Each is an example of a broader, well-documented pattern in which consequential decisions are automated at scale, with little visibility into how they work. A vendor-management system might quietly deprioritize your listings for patterns common to legitimate seasonal businesses.
When you assume a rejection reflects your business's quality rather than a mismatch between your data and a model's assumptions.
Track every automated decision affecting your business, including the date, stated reason, and reference number.
Advocating for your business inside the system
You have more leverage than it feels like in the moment. Ask directly which factors a decision relied on, and whether a human can review your case. Many lenders and platforms have an appeals process customers rarely use simply because they do not know to ask. Keep organized records of your performance so you can present a fuller picture than the system captured. This is not about gaming a score; it is about giving the humans behind it context the model lacked.
When entrepreneurs assume an automated denial is unappealable and move on without asking a single question.
Request a written explanation of any adverse decision, then ask specifically whether human review or a documented appeal exists.
The debate you can't resolve alone
No amount of individual diligence fully closes the gap between how these systems are built and how accountable they are to the people they affect. Meaningful change, such as required transparency about scoring factors or independent audits of lending algorithms, depends on regulators and public pressure operating beyond any business owner. This is an ongoing societal conversation, not a settled argument, and entrepreneurs are one of many stakeholders with reason to care.
Understanding why a model can misjudge you competes with the urgency of getting a loan approved today. Documenting every decision takes time you may not have between orders and payroll. Pushing for human review can feel riskier than quietly reapplying elsewhere. Balancing these means picking battles: document routinely, but escalate only when the stakes justify it.
Name one thing, honestly
- Which automated decisions currently shape access to capital, insurance, or customers for your business? - Do you know whether the platforms you depend on offer any human review process? - What would change in how you operate if you assumed every score could be appealed?
Good. Now look back at the "Try this" line under whichever lesson feels most connected to what you just wrote, and treat that as this week's actual assignment — not all 4 at once.
The practices from this edition
0 of 3 selected
Run this for seven days
Request a written explanation of any adverse decision, then ask specifically whether human review or a documented appeal exists.