A full chapter from The 2026 AI Strategy, published open. The complete ebook is free; this chapter stands on its own.
The evidence. The average company now runs somewhere between 100 and 275 SaaS applications (BetterCloud / Productiv, 2024). Workers toggle between apps roughly 1,200 times a day, losing just under four hours a week simply reorienting, about 9% of work time (Harvard Business Review, 2022). Around 60% of the day goes to "work about work": chasing documents, status, and moving data between systems (Asana, Anatomy of Work). Stack those findings and a picture emerges: by industry estimates, a knowledge worker can spend roughly 200 hours a year doing nothing but moving between systems, before counting the work inside them.
The villain. Not your people, and not any one vendor. The villain is the operating model itself: running a business through a sprawling software stack is inherently labor-heavy. Somebody has to pull the data, key it into the next system, chase the approval, reconcile the mismatch. The human operating layer is the cost, not the license fee. I call it the software tax: the hours and headcount a business pays to operate its software to run the business.
The precedent. Every era pays an operating tax to its own technology. A century ago the telephone was run by rooms of switchboard operators: skilled people whose entire job was connecting the machine to the business. Automated switching did not end the phone call. It ended the job of operating the phone system, and no company that made the switch ever missed it. The software stack is this era's switchboard, and the human operating layer around it is just as removable.
The Fortune 500 plays. Large enterprises have attacked operating cost for decades with four disciplined plays: cost-to-serve analytics that find where the margin actually goes, process automation and process mining, shared services consolidation, and zero-based budgeting that makes every line of spend re-justify itself. Each play historically required a transformation office and a consulting budget. That was the barrier.
What changed. In bounded, instrumented workflows, AI agents can now perform defined operating tasks: pull the data, reconcile it, draft the output, route the approval. Connection standards let those agents work inside the systems you already own, and humans govern the seams where judgment, risk, and accountability sit. That is not a claim that agents run your business. It is a claim that the twenty-step manual chain can become a two-approval chain, and the stack stops requiring an army to run. In practice the plays become concrete quickly: an agent that reconstructs true cost-to-serve from your job logs and ranks accounts by real margin; a quote agent that turns a photo or description into a priced quote; an agent that runs invoice-to-job reconciliation; an AI front desk that handles scheduling and routine inquiries; an AI pass over the P&L that flags software spend returning no measurable value, which is the software tax made visible.
Where it stops. The software tax is a margin play: the same output, produced with less operating labor. Its twin, the capacity ceiling in the next chapter, runs the same mechanism in the opposite direction: more output from the same team. If your problem is margin, this door is yours. If your problem is demand you cannot serve, read on.
The metric. Cost reduction, measured as labor-hours per unit of work and the share of spend returning no value. Sized directionally for a $250M professional-services firm, the toggling floor alone runs to several million dollars a year, and the broader operating layer several times that, by the composite models we build from a company's own numbers.