The AI J-curve: why productivity drops before it soars

Somewhere between six and eighteen months into most AI programs, the sponsor asks why nothing shows up in the numbers. The licenses were bought and the copilots are deployed, and productivity is flat or slightly worse. At that point somebody concludes the technology was overhyped, or the vendor oversold it, or the team resisted.

Usually the right explanation is different. The pattern has a name and solid research behind it, and it is older than software.

The economics of the dip

Erik Brynjolfsson, Daniel Rock, and Chad Syverson called it the productivity J-curve, in a 2021 paper in the American Economic Journal: Macroeconomics. Their argument: a general purpose technology, and AI is one, delivers little value on its own. Getting value out of it requires complementary investments first. Processes have to be redesigned, people retrained, roles restructured, data cleaned up and connected. Those investments are intangible, so accounting expenses them and national statistics barely see them. The organization spends real money building invisible capital, and measured productivity drops while it happens. The climb comes after the complementary capital is in place. Their estimate for the computer era: by the end of 2017, accounting properly for intangibles would raise measured US total factor productivity by roughly sixteen percent.

The precedent is electricity. The economic historian Paul David showed that factories took decades to profit from electrification. Early adopters swapped the steam engine for one giant electric motor and got almost nothing, because the factory was still built around the old power source: overhead shafts, belts, floors stacked vertically to stay near the drive shaft. The payoff arrived when engineers put a small motor on each machine and redesigned the whole layout around the flow of work. Dropping a chatbot into an unchanged workflow is the giant-motor move.

Workslop, the visible symptom

In 2025 the dip acquired a name you can see in your inbox. Researchers at BetterUp Labs and Stanford’s Social Media Lab coined the term workslop for AI-generated output that looks like finished work but does not actually advance the task. Their survey of 1,150 US workers found that about 40 percent had received workslop in the previous month, and each instance took close to two hours to deal with. Receivers also downgraded the sender: roughly half saw colleagues who sent it as less capable than before.

Workslop happens because the tool speeds up drafting before the organization has rebuilt verification and standards around it. The cost of quality control quietly moves from the writer to the reader. Spread that across an org chart and you get the finding that made headlines the same year, an MIT Media Lab report that 95 percent of organizations had yet to see measurable return on generative AI spending. That number describes thousands of organizations sitting at the bottom of the curve with the complements unbuilt.

What determines how deep you sink

Four variables do most of the work. First, process dependence: AI dropped into a broken process automates the production of mess. Second, whether verification was designed or improvised. If nobody decided who checks AI output and to what standard, the checking lands on whoever trusted the sender. Third, whether usage was mandated ahead of redesign. Pushing adoption metrics before the workflow changes buys the dip without the climb. Fourth, measurement discipline. Organizations that track cycle time, rework rates, and verification load can actually see the complementary capital forming. Organizations that only look at quarterly output see noise.

Surviving it

The playbook follows from the mechanism. Fund the process redesign, the training, and the data work as real projects with budgets, since that spending is the thing the curve is made of. Sequence use cases by reversibility so early failures are cheap. Build verification into the workflow instead of leaving it as an unfunded courtesy, which is the single most direct control on workslop. And set expectations at the top before the dip arrives: this is a J-curve investment with a dated climb, and the board deck should say so.

The organizations that get hurt are the ones that panic in the trough and cut the complementary investment because the numbers look bad. That locks in the costs and forfeits the climb.

I am turning the diagnostic behind this piece into an interactive worksheet: where you are on the curve, your risk factors for a deep dip, and a realistic value timeline. Until it ships, the working version is available through the contact page.