One Must-Read Article

The Last Mile: How Bad Data Kills AI Implementation

Last week, I introduced three big ideas from Andreessen Horowitz’s Big Ideas 2026 report—predictions that matter now for growth-stage companies. This week, I’m diving into the first one: the last-mile data problem.

It’s the unglamorous issue quietly preventing AI from working inside many businesses.

Most companies want AI to solve real problems. But when you look closely at why AI pilots stall or disappoint, the root cause is rarely the technology. It’s this: Their data is a mess.

The constraint isn’t models. It’s that roughly 80% of operational knowledge lives in emails, PDFs, photos, and people’s heads—places AI can’t reliably access or trust.

For most growth-stage companies, the AI gap isn’t talent or tools. It’s operational readiness.

The Bottleneck Nobody Talks About

Jennifer Li at Andreessen Horowitz describes the issue as “data entropy”—”the steady decay of freshness, structure, and truth inside the unstructured universe where most corporate knowledge lives.”

Translation: your company’s real knowledge—the information that determines whether you win deals, deliver on time, and retain customers—is scattered everywhere:

• Email threads acting as your informal CRM

• PDFs buried in shared drives

• Excel files with conflicting versions

• Slack conversations that double as documentation

• Screenshots and photos from the field

• Handwritten notes and half-completed forms

This isn’t just messy. This is why AI doesn’t work for you yet.

Why This Matters More Than the Models

Here’s what most AI vendors won’t tell you: The models are already good enough.

GPT-4, Claude, Gemini—these tools are remarkably capable. What they can’t do is impose clarity on operational chaos.

When AI hallucinates—producing plausible but wrong answers—it’s usually not a model failure. It’s a data failure.

• Your sales AI can’t prioritize leads when the CRM is stale and the real intelligence lives in forwarded emails.

• Your support AI can’t give accurate answers when product docs, tickets, and implementation reality all contradict each other.

Andreessen Horowitz frames this as a venture-scale infrastructure problem. For operators, it shows up much closer to home.

The Last-Mile Problem

A16z (Andreessen Horowitz newsletter) calls this “the last-mile problem of data inside firms.” Not the exciting work of building new models—but the unglamorous work of getting operational data clean enough to use.

Think about package delivery. Amazon can predict demand, optimize routes, and automate warehouses. But someone still has to carry the package from the truck to your door. That last hundred feet is the hardest part.

Your data has the same issue. You have systems. You have integrations. But the last mile—the context that makes information meaningful—is trapped in formats AI can’t reliably use.

What This Looks Like in Practice

The Construction Company
An $8M contractor tried using AI for bid estimation. They had project histories, supplier pricing, and labor costs. But the real intelligence lived in field photos and site notes—handwritten observations about soil conditions, access constraints, and hazards. AI couldn’t reliably interpret that context. The estimates were no better than those produced by their experienced estimator.

The Professional Services Firm
A consulting firm wanted AI to scope projects based on past work. The project data existed in their Project Management system. But the real story—what caused overruns, what delighted clients, which teams performed best—lived in emails, buried documents, and senior consultants’ heads.

This is where most AI discussions stop. This is where operators actually win.

What You Can Actually Do

Good news: you don’t need a data science team or custom infrastructure.

Start with one high-value workflow
Pick the single process where bad data costs you the most—sales qualification, project estimation, customer on-boarding.

Map the real data sources
Don’t just list systems. Identify where actual intelligence lives, including informal sources: emails, photos, conversations, notes.

Create capture systems for unstructured data
Structured call templates. Photo documentation standards. Consistent naming and filing. Recorded walkthroughs with transcripts. Brief written summaries after key conversations.

None of this is AI. This is operational discipline—and it’s the prerequisite for AI to work.

Use AI to clean historical data—slowly
Once new data is captured cleanly, use AI to extract structure from historical data in small batches: one project, one client segment, one quarter at a time.

Measure data quality, not just AI output
Track whether AI has the information it needs, whether it’s current, and whether sources conflict.

Here’s My Take

Your AI problem can be a data problem. Your data problem is an operational discipline problem.

Operational discipline has always separated companies that scale profitably from companies that grow chaotically.

Companies that solve their data problem first build compounding advantages. Clean operational data improves every decision—and every decision generates better data.

For most companies, the window to fix this deliberately—before it becomes reactive and expensive—is probably 12–18 months.

Andreessen Horowitz is betting billions that enterprise data infrastructure will be one of the largest opportunities of the next decade. But the real opportunity isn’t just for startups building tools.

It’s for operators who realize that fixing their data isn’t a tax on AI. It’s the foundation for everything that comes next.

Next week: How the best companies use AI to strengthen business models—not just cut costs.