The Staircase Problem: Why Better AI Tools Make Bad Processes Worse
AI made every step of your business faster, yet the work still finishes on the same date it always did. The trap is simple: speeding up one task in a process you never redesigned just relocates the bottleneck, and quietly raises the odds something slips through.
A small law firm we know automated its first contract last spring. An engagement letter that used to eat a paralegal's afternoon now drafts itself in about ninety seconds. Real time saved, no asterisk. And yet contracts don't reach clients any faster than they did a year ago. The drafts pile up on one partner's desk, waiting for the same careful read they always got, except now there are five times as many of them.
This is the Staircase Problem, and once you see it you start seeing it everywhere.
Picture your business as a staircase. Each step is a stage of work: an inquiry becomes a proposal, the proposal becomes a job, the job gets reviewed, approved, delivered, and billed. Work climbs from one step to the next. AI lets you sprint up a single step, usually the one with drafting or research or a rough first pass in it. But if the staircase itself still spirals the way it did before, sprinting up one step gets nobody to the top sooner. It just hands a bigger, faster pile to whoever is standing on the next step.
The Cambrian Explosion of AI Agents
Over the past two years nearly every business has done the same thing: bought the tools. ChatGPT, Claude, Copilot, plus a swarm of niche assistants wrapped around the same handful of models. The market for AI agents is now measured in the billions, on a curve that looks less like ordinary adoption than a Cambrian explosion: new species of software appearing faster than anyone can catalog them.
The pitch is irresistible and, on its own terms, true: point a tool at a task your team already does and the task gets faster. Andrej Karpathy even named the new mode "vibe coding," letting software write itself from plain-language intent. Drafting, summarizing, first-pass research, data entry, the boring middle of knowledge work, all of it compresses.
So adoption is not the hard part. Adoption is the easy part. The hard part is that getting faster at a step is not the same thing as getting faster at the work.
What Happened to the People Who Went First
Software teams hit this wall first, which makes them a free case study for everyone else.
By 2025 a large share of new code was being written by AI, and developers reported genuine time savings on the writing itself. Then the measurements came in, and they were sobering. METR ran a careful controlled trial with experienced developers on real tasks and found they were about 19% slower with AI, even though the same people were convinced they were faster. The feeling of speed and the fact of speed had quietly come apart.
Where did the time go? Mostly to the next step. Bain found that actually writing and testing code is only about a quarter to a third of the journey from idea to shipped product. Speed up that slice and everything around it, review, coordination, approvals, becomes the jam. Review was the canary in the mine: teams pushed far more work for sign-off, and the sign-offs took longer, because there was more to check and it was harder to follow. Google's 2025 DORA research put numbers on it. AI pushed throughput up while pushing delivery stability down. More shipped, and more broke.
The tools, meanwhile, kept getting better. The 2026 generation can run for long stretches on its own and even catch and fix its own mistakes, which was supposed to settle the argument. It didn't. Better tools closed the quality gap on the fast step and left the rest of the process exactly where it was. When the tools were weak, you could blame the tools. Once they were strong, the only thing left to blame was the way the work was organized.
The Rigor Has to Go Somewhere
Here is the principle underneath all of it. Every process contains a fixed amount of judgment, checking, and care, call it rigor, and that rigor does not disappear when you automate the easy part. It relocates.
When a paralegal spent an afternoon drafting a contract, a lot of rigor was baked into the drafting itself: slow, deliberate, self-correcting work. Hand the drafting to AI and the draft lands in seconds, but the rigor it used to carry hasn't evaporated. It has moved downstream to the partner who now has to supply all of it at the review step, across five times the volume. The same holds for the clinic that auto-generates visit notes, the agency that spins up forty ad variations, the accountant whose software drafts the client memo. The first draft got cheap. The checking did not, and there is now far more of it to do.
That leaves three options, and only three. You can let the rigor pile up as a bottleneck, the partner's overflowing desk. You can quietly skip it and ship work nobody fully checked, which is the dangerous one. Or you can redesign the step where the rigor now lives so it can absorb the new flow. Most businesses, without ever deciding to, drift into one of the first two.
The Pattern Repeats in Every Office
The software story isn't really about software. It's a sequence, and it runs the same way in a dental practice, a real estate brokerage, a marketing shop, or a three-person operations team.
- A tool speeds up one concrete task — the proposal, the listing write-up, the intake summary, the invoice.
- The bottleneck jumps next door, to the step that was never built for the new volume: review, approval, the one person who signs off, the handoff to the client.
- The process itself has to change, not the tool. A weekly review cadence designed for ten items breaks at eighty.
- Roles shift. The person who used to produce the work now mostly directs and checks it, and the skill that matters becomes judgment, not output.
- The old scoreboard stops meaning anything. Counting drafts produced, calls logged, or listings posted measures the cheap step, while the expensive one, was this actually right and good, goes unmeasured.
The tell that a business is stuck on step one is volume without results: more proposals sent and no more deals closed, more content published and no more customers, more memos drafted and no more clarity. Speed without redesign produces motion, not progress.
Redesign the Staircase, Not the Step
The businesses that get real value from AI share one habit. They start with the bottleneck, not the tool.
First, find where work actually stalls. It's almost never the drafting; it's review, approvals, and handoffs, the steps where a human applies judgment. That's where AI has to genuinely help, with tools that check and flag rather than only generate, or where the process gets rebuilt to handle more flow.
Second, pick one or two processes and redesign them end to end, instead of sprinkling tools across everything at once. Client intake. Monthly reporting. Proposal to signed contract. Rebuild one staircase for the world you're actually in, then move to the next. Businesses that run twenty disconnected experiments tend to collect twenty disappointments. The ones that fully transform two processes get two wins they can build on.
Third, ask the uncomfortable questions before you buy anything. Which task here can AI genuinely speed up today? Where will the work pile up the moment it does? What would this process look like if we designed it this year, from scratch, knowing what these tools can do? What are we measuring that's about to become meaningless?
The hardest part is rarely the technology. It's being willing to admit that a process you built and refined over years is now the constraint, not the asset.
The sharpest line from the DORA research applies to a five-person firm as much as a five-thousand-person one: AI magnifies what's already there. Tight processes, clear ownership, and honest feedback get amplified. So do tangled handoffs, fuzzy responsibility, and the corners everyone quietly cuts, only faster.
So the question to sit with isn't which tool to buy. It's this: if you rebuilt this process from nothing today, knowing exactly what AI can and can't do, would it look anything like the one you have? When the honest answer is no, you've found your staircase. The work is to redesign it, not to keep polishing the one step that already moves fast.
Sources
Curious what this looks like for your business?
We help small and mid-sized teams turn AI from a headline into something that actually works. Let's talk about where to start.