techbullion.com

Dylan Pulver: Stop Asking Whether AI Works and Start Asking What an Hour of Your Team Costs

By Victor Miller·Published November 17, 2026·Business news, Technology

Dylan Pulver

Corporate AI spending has produced an enormous amount of nodding in conference rooms and negligible movement on the profit and loss statement (P&L). The pattern repeats across industries: a vendor runs a demo, executives agree it looks impressive, a pilot gets funded, and 12 months later nobody can point to a line item that changed. Dylan Pulver, who has spent years building automation for back-office operations, argues the failure starts with the question being asked. "Does AI work?" is unanswerable and therefore unfundable. What a bookkeeper's hour costs, fully loaded, is a number most finance teams can produce before lunch. That shift sounds small, but it is the difference between a technology conversation, which stalls, and a budget conversation, which gets decided.

Why The Capability Question Kills Projects

Pulver's objection to capability-first thinking is not philosophical. It is that the conversation has no terminus. "Someone shows a demo, everyone nods, nothing changes," he says. Every participant can agree the tool is impressive without anyone committing to anything, because impressiveness is not a budget category. Reframe the same discussion around hours and the numbers do the persuading. A team that spends a week every month reconciling data between two systems has a knowable annual cost: loaded hourly rate, times the hours, times 12. Against that sits a one-time build and a modest running cost. "Nobody has to believe in AI to read that math," Pulver says.

The framing cuts in both directions, which is what makes it useful rather than merely rhetorical. It gives serious automation a path through procurement, and it kills the vanity project. Plenty of AI spending happens because a board member asked what the company is doing about AI, not because anyone identified work worth removing. Under Pulver's test, that spending has nowhere to hide. If the hours do not add up, do not build it. Executives who adopt this discipline get a side benefit: they stop evaluating vendors on demo quality and start evaluating them on whether a specific recurring cost disappears.

Where The Margin Leaks

Finding the expensive work requires ignoring most of what a company documents about itself. Pulver does not open with the organizational chart or the software inventory. He asks people what they do on a Tuesday. The costly processes are rarely the ones anybody wrote down, and they are seldom the dashboards and reports executives look at. They are the weekly export pasted from one system into another, the invoice checked by hand because two totals never match, or the report someone rebuilds from scratch every month. "The margin leaks in the boring work," he says.

Three signals mark a candidate: it happens on a schedule, it follows the same steps every time, and the person doing it can describe it without thinking. Log each one with frequency, duration and owner, multiply by loaded cost, and a ranked list exists within a day, with the top three items usually accounting for most of the bleed. Pulver adds a fourth filter that operators tend to miss: count how many systems the process touches. Work that crosses system boundaries is where humans get used as glue, and glue is the first thing to automate. That single criterion explains why so much expensive manual labor survives inside companies that consider themselves well digitized. Each system is modern. The seams between them are staffed by people.

The Test That Separates Savings From Assistance

Most of what is sold as workplace AI is assistance, and assistance does not move unit economics. Pulver's test is blunt and worth adopting, exactly as he says it: after deployment, can this run while the person is on vacation? If not, the company bought a copilot. "The person is still there, still paid, still the bottleneck," he says. Real automation starts the process, runs it to completion, and pulls a human in only when something is off. But scope matters as much as autonomy. Saving clicks optimizes a step. Changing economics means owning the workflow from trigger to finished output in the system of record. The technically interesting part is never the obstacle. "The hard part is never the AI piece. It is the edges: the login that expires, the file format that changed last quarter, the exception nobody mentioned during scoping." One process Pulver rebuilt dropped from consuming a week of a person's month to under an hour of review.

The same logic explains why pilots stall. They are designed to prove AI works rather than to remove a cost, so they pick a demo-friendly problem, run it on clean data, and produce a result nobody owns. The alternative is deliberately straightforward. Pick one process with a known hourly cost before building anything, record the baseline, build the crude version in a week, then run it against real data alongside the human for a month and count. The output stops being a projection and becomes a measurement with a denominator. Then comes the step where most initiatives die: the hours have to leave the budget, through reassignment or not backfilling a role, and that is a management decision rather than a technical one. "The initiatives that reach the P&L have an owner who wanted the hours back before the pilot began, and the pilot is just the receipt."

What Operators Are Underestimating

As agentic systems take on multi-step work, Pulver expects readiness to hinge on three factors. The first is written process documentation, including the exceptions, since agents run on instructions and most firms have never written that file. The second is exception handling treated as a real job, because when agents run 95 percent of the routine, the human role becomes reviewing the flagged five percent. The third is clean system access, since businesses built on shared logins and spreadsheets will spend months on plumbing before any agent does useful work.

What he thinks is most underestimated is the compounding. A small team with ten processes, all automated, operates like a company three times its size, and that gap will show up in margins long before it shows up in headlines.

Follow Dylan Pulver on LinkedIn and at dylanpulver.com for more on back-office automation, agentic systems, and building AI projects that reach the P&L.