The honest answer to the productivity question
Generative AI will not explode your team’s productivity overnight, and the reason is not the model, it is everything around it. That is the uncomfortable conclusion once you put the 2026 hype next to the data and the economic history. AI genuinely helps with research, boilerplate, first drafts, and review. The mistake is expecting an overnight 10x. The bottleneck is process, judgment, and an information-noise tax, not the capability of the tool.
I run a WordPress agency. I use these tools every day. This is not a contrarian pose for clicks. It is the view from inside delivery, where the gap between a convincing demo and a shipped, maintainable site is exactly where the promised gains tend to evaporate. The problem is not enthusiasm. The problem is the tempo we expect.
A 1987 quip that keeps coming back
Start with the line that named the whole phenomenon. In 1987 the Nobel laureate Robert Solow, writing in The New York Times, observed that “you can see the computer age everywhere but in the productivity statistics”. Computers were on every desk. The aggregate numbers did not move the way the brochures promised. That observation seeded what economists now call the productivity paradox, or the Solow paradox.
It would be convenient to dismiss this as ancient history from the era of beige towers. Erik Brynjolfsson of MIT did not. In a 1993 article he coined “The Productivity Paradox of Information Technology” and argued, carefully, that the absence of visible gains did not mean the technology was useless. It meant measurement, lag, and mismanagement were hiding the value. His more recent work argues today may rhyme with the late 1980s: the capability is real, the payoff is delayed, and a lot of the early disappointment is self-inflicted.
That is the first reason for skepticism about the overnight-10x story. We have run this experiment before, with a general-purpose technology at least as consequential as generative AI, and the productivity curve did not bend on the schedule the vendors wanted.
Electrification, or why 40 years is a real number
The deeper explanation comes from the economic historian Paul A. David. His argument is blunt: real gains arrive only once the world around a technology is rebuilt. His worked example is the electric motor. Factories did not get more productive the moment they swapped steam for electricity. They got more productive once they were redesigned around electricity, which took roughly 40 years.
The reason is physical and organisational. Steam-era factories were built around a single central engine and a forest of belts and shafts. Electricity’s real advantage, the unit drive motor on each machine, only paid off when factory owners abandoned the old multi-storey layout and built single-storey plants where machines could be arranged by workflow rather than by proximity to the driveshaft. The technology was available for decades before the buildings, the management, and the labour practices caught up. Only then did the statistics move.
Apply that lens to a WordPress agency in 2026. You can put an AI assistant inside the block editor today. But your intake process, your QA gate, your handoff to clients, your billing model, and your definition of “done” were all designed for a world where a human wrote every line. The motor is on the bench. The factory has not been rebuilt.
The number nobody quotes back: about 3 percent
Here is where the polemic gets specific. Early, MIT-adjacent claims floated efficiency improvements in the order of 40 percent for AI-assisted knowledge work. That number travelled fast because it justified budgets. Then a spring 2025 study from the National Bureau of Economic Research measured it more carefully and found the average measured efficiency gain from AI tools was about 3 percent. Not 40. Three.
Worse for the hype, the same body of research found that on more complex, judgment-heavy tasks, AI sometimes hurt motivation and quality. That matches what I see in practice. Asking a model to scaffold a custom block, draft a migration plan, or summarise a plugin changelog: net positive, sometimes large. Asking it to make an architectural call about whether a client needs a headless front end or a hardened classic stack: net negative if anyone treats the output as an answer rather than a starting position to argue with.
Three percent is not nothing. Compounded across a year, across a team, three percent is real money. But three percent is not a revolution you reorganise your hiring plan around in a quarter. It is the kind of gain that shows up only if you keep the discipline to capture it, and disappears the moment you let the tool generate work nobody needed.
Volume is not value
The more useful framing of where the gains actually come from is in an EY report on AI and productivity. Its central move is to separate productivity of volume from productivity of value. Generative AI extends automation beyond repetitive tasks into knowledge and creative work, which is genuinely new. But the gains that matter, EY argues, are in quality, creativity, and the speed of decisions, not in raw output.
In an optimistic model, the same report puts AI’s contribution on the order of 3 percent added to global GDP by 2033. Note the date. Not by next quarter. By 2033. That is, again, the timeline of process change, not model releases.
The volume-versus-value distinction is the most practically useful idea in this whole debate for an agency. If you measure your AI adoption by how many words, drafts, or tickets it produces, you will declare victory while quietly drowning. If you measure it by how fast you reach a good decision and how little you have to redo, you will find out whether it is working.
Three ways the gains leak away
Across every wave of disappointing technology adoption, the same three mechanisms recur. They are worth naming because each has a direct WordPress counterpart.
| Mechanism | The general pattern | The WordPress 2026 version |
|---|---|---|
| Tools without process change | Buying the technology but keeping the old workflow, so it becomes a costly layer on top of how you already worked | Bolting an AI panel onto an unchanged intake, build, and QA pipeline and expecting it to compound |
| The adaptation lag | Standards, regulation, and culture must change before the gains land, per Paul A. David | Teams, contracts, and client expectations still assume hand-written work end to end |
| Side-effects that eat the gains | In 1966 computers generated more reports than managers could read | A deluge of AI-generated drafts, briefs, and summaries of little real value that someone still has to triage |
That third row is the one I would underline. The 1966 problem was real: the machine could generate more management reports than any manager could absorb, so the bottleneck moved from producing information to reading it. The 2026 version is the flood of AI-generated everything. Pull requests no one fully reviewed. Documentation generated to satisfy a checklist. Client-facing copy drafted in seconds and then argued over for an hour. The model made the cheap part cheaper and shoved the cost into the expensive part: human attention and judgment. That is the information-noise tax, and it is invisible on any dashboard that counts output.
Two real cases from delivery
Abstractions are easy to nod along to, so here are two contrasting cases from the kind of work an agency actually does.
Case one. A WooCommerce store with more than 30 plugins and a time-to-first-byte sitting around 1.8 seconds. The owner had heard AI could “fix performance” and wanted us to point a tool at it. We did use AI heavily, but not the way they imagined. It was excellent at the bounded, legible parts: triaging which plugins fired the most queries, drafting the first version of a transient-caching layer, generating test cases for a checkout regression suite, explaining an unfamiliar third-party hook. What it could not do was decide that the real fix was removing eleven plugins and rebuilding two as a small custom plugin, because that decision required weighing the client’s roadmap, their in-house skills, and their tolerance for risk. The AI compressed the legible 60 percent of the job. The illegible 40 percent, the part that was actually the bottleneck, it could not touch. Net result: faster, yes. Ten times faster, no. The judgment was always going to be the rate limiter, and our WordPress maintenance and care work is mostly judgment.
Case two. An Elementor site that fell over under a campaign traffic spike. Here the temptation was to let AI generate a wall of “optimisation” recommendations, and it happily would have. That is exactly the information-noise trap. A model will produce twenty plausible suggestions, eighteen of which are generic and two of which matter, and now a senior person has to spend their scarcest resource, attention, separating signal from confident-sounding noise. The win came from a human deciding to cache aggressively at the edge and defer the page builder’s heaviest widgets, then using AI only to draft the implementation. The thinking was the slow part. The typing was the fast part. AI accelerated the typing.
The pattern across both: AI is a multiplier on the parts of the work that were never the constraint. The constraint is judgment, and judgment does not have a 10x button.
What this means for WordPress 7.0 and AI features
WordPress 7.0 ships with native AI capabilities deeper in the editing and workflow surface than anything before it. That is genuinely useful, and I am not arguing against it. I am arguing against the tempo people expect from it.
If you treat WordPress 7.0’s AI features as a layer to bolt onto your current process, you will get the 1966 outcome: more drafts, more suggestions, more “content,” and a team quietly spending its recovered time reviewing material that should never have been generated. That is mechanism one and mechanism three working together.
If instead you treat it as a reason to redesign the workflow, the way factory owners eventually redesigned around the electric motor, the gains are real and durable. Concretely, that means deciding in advance which tasks AI drafts, which a human must review before anything ships, and which you simply refuse to automate because the judgment cost is the point. It means measuring cycle time and rework rate, not word count or ticket throughput. It means being honest that a WordPress developer’s value is shifting toward the decisions AI cannot make, not away from them.
Pricing is the same story. Clients sometimes assume AI should make everything cheaper because the typing is cheaper. The typing was never the expensive part. The architecture, the security posture, the performance budget, the accessibility compliance, the decision about what not to build: that is the work, and it is why our pricing stays individual rather than a per-word commodity. AI lowers the cost of the legible tasks and raises the relative value of the illegible ones.
So what should you actually do
None of this is an argument to ignore AI. It is an argument to set the right tempo and the right measurements. If I had to compress the senior take into a handful of operating rules:
- Adopt AI for the bounded, legible tasks first: research, boilerplate, first drafts, test scaffolding, code and content review. That is where the measured gains genuinely live.
- Redesign the surrounding process before expecting compounding returns. The tool is the electric motor. You still have to rebuild the factory.
- Budget for the information-noise tax explicitly. Every piece of AI-generated work is a draft that costs human attention to review. If reviewing it costs more than writing it, you have lost.
- Measure value, not volume. Cycle time, rework rate, and decision speed tell you whether AI is helping. Output counts tell you nothing useful and flatter you into complacency.
- Keep humans on the judgment calls. On complex, judgment-heavy tasks the NBER data shows AI can hurt quality. Treat its output as an argument to test, never an answer to ship.
The point is the tempo, not the tool
The productivity paradox is not a prediction that AI fails. Solow’s computers eventually did show up in the statistics, decades later, once the world reorganised around them. Brynjolfsson’s lag resolved. David’s electric motor remade manufacturing, forty years on. The technology was real every time. The disappointment was always about expecting the gain on the calendar of the marketing department rather than the calendar of organisational change.
Generative AI in 2026 is the same shape of story. About 3 percent measured today, plausibly much more by the early 2030s, but only for the teams disciplined enough to rebuild their process around it and honest enough to count value instead of volume. The problem was never enthusiasm. The problem is the tempo we expect. Get the tempo right and the paradox is not a warning. It is a roadmap.
Last updated: 31 May 2026.
