Jonathan Haaswritingnowusesabout
emailgithubx
Jonathan Haaswritingnowusesabout

Why Your AI Strategy is Actually a Spreadsheet Strategy

September 10, 2025·3 min read

Most enterprise AI transformations are solving problems that spreadsheets handle at 1/50th the cost. The misalignment is driven by career incentives,...

#ai#business-strategy#spreadsheets#overengineering#technical-debt#contrarian#enterprise

A Fortune 500 company spent $5M on an "AI-powered intelligent document processing platform." The system extracts numbers from PDFs and puts them in columns. An analyst with Excel could do this in 20 minutes. But "Excel" does not appear in transformation roadmaps. "AI" does.

Most enterprise AI initiatives are solving spreadsheet problems at 50x the cost.

The Recurring Pattern

The business problem "show me which products are selling" has been solved identically since 1985. Only the price changes.

1985: =VLOOKUP(). Cost: 100.Timetoproduction:onehour.2005:customwebapplicationwithaSQLbackend.Cost:100. Time to production: one hour. 2005: custom web application with a SQL backend. Cost: 100.Timetoproduction:onehour.2005:customwebapplicationwithaSQLbackend.Cost:200K. Time to production: six months. 2025: "AI-powered demand forecasting platform." Cost: $5M. Time to production: 18 months and counting.

Same answer. Same columns. Same numbers. The underlying computation is a filter and an aggregation. It was a spreadsheet operation in 1985, and it remains one in 2025.

The Incentive Misalignment

The technical explanation for this pattern is straightforward: these problems do not require machine learning. The actual explanation is organizational.

Nobody gets promoted for maintaining spreadsheets. Everyone gets promoted for "leading AI transformation." The executive incentive is to frame every data problem as an AI problem, regardless of whether AI adds value over simpler alternatives.

Vendors reinforce this. A spreadsheet costs 100andhaszeroswitchingcosts.AnAIplatformcosts100 and has zero switching costs. An AI platform costs 100andhaszeroswitchingcosts.AnAIplatformcosts500K/year with a three-year contract and migration costs that make departure prohibitive. The vendor is not selling a better solution. They are selling lock-in with an AI marketing wrapper.

Consultants complete the cycle. The engagement to "evaluate AI opportunities" will always find AI opportunities. The incentive structure guarantees it.

The Decision Framework

Three questions separate AI problems from spreadsheet problems.

Can a formula express the logic? If the transformation from input to output can be described as a deterministic rule -- lookups, aggregations, conditional logic, arithmetic -- it is a spreadsheet problem. AI adds no value to deterministic transformations.

Where is the complexity? Complex logic at simple scale: spreadsheet. Simple logic at massive scale: database with standard ETL. Both complex: potentially a machine learning problem. The overwhelming majority of enterprise data problems are simple logic at simple scale.

What is the cost of errors? A spreadsheet error is debuggable. Open the cell, read the formula, find the mistake. An AI error is opaque. The model produced the wrong output. Why? The debugging cost scales with the opacity of the system, and LLMs are maximally opaque.

Where AI Genuinely Applies

AI adds value when the task cannot be expressed as rules: semantic analysis of unstructured text at scale, pattern recognition in high-dimensional data (medical imaging, fraud detection in complex transaction networks), and generative tasks where the output space is too large for enumeration.

The pattern is clear. If the data fits in rows and columns and the logic is expressible as deterministic rules, it is a spreadsheet problem. If the input is unstructured, the patterns are latent, or the output requires generation -- then AI may justify its cost.

Most enterprise data problems are the former. The 5MAIplatformandthe5M AI platform and the 5MAIplatformandthe100 spreadsheet produce the same output. The difference is which one gets budget approval.

share

Continue reading

The 10-Minute AI POC That Becomes a 10-Month Nightmare

Five lines of Python and an API key produce a working demo. The gap between that demo and a production system contains failure modes the prototype...

The AI Agent Gold Rush: Why Everyone's Building Picks and Shovels

Most AI agent infrastructure is premature. The agents themselves barely work. The industry is selling Formula 1 equipment to people still learning to...

Two Minds in the Machine: Shared Context Is the Only Thing That Matters

I added Gemini to a codebase that already had Claude embedded. The useful discovery was about shared context files, not model capabilities.

emailgithubx