AI Productivity Gains Outside Tech

On Sept. 25, 2026, Goldman Sachs strategists told clients that the five largest US hyperscalers will raise AI infrastructure spending by more than half in 2027, to $1.2 trillion, as Bloomberg reported. The returns aren’t keeping up.

In a survey of working scientists, respondents reported saving nearly seven hours a week with AI, and nearly half of those who saved time spent more than 25% of that gain verifying what the model produced, Campus Technology reported on Sept. 23, 2026. Gross hours fall. The booked saving is smaller, and in high-stakes technical work it can shrink toward noise.

That difference is the argument. AI pays off clearly where a wrong answer is cheap to catch, as in a support queue of short, policy-bound replies. It pays off far less where checking the output costs the same expertise as producing it, which describes research, legal drafting, and most financial analysis. The $1.2 trillion committed for 2027 is justified by the first category and will largely be deployed into the second.

Hyperscaler Capex: The 2027 Run Rate

Bloomberg’s Sept. 25, 2026 report put the five largest US hyperscalers on track to increase AI infrastructure spending by more than half next year, to $1.2 trillion, citing Goldman Sachs strategists. A separate Seeking Alpha summary of Goldman’s view described US hyperscaler capital deployment reaching $1.4 trillion in 2027 and framed the shift as one from experimentation to implementation, with consumer AI agents opening a new phase of growth. Different scopes, same direction.

Buyers of compute should not read that build-out as an imminent price cut. Capacity under construction does not move inference pricing until it is live and running at load, so through late 2026 and early 2027 the newest accelerators still price at scarcity. Depreciation from this wave lands on hyperscaler income statements over multi-year schedules, which puts the earnings drag well after the cash has left.

The Verification Tax on Time Savings

The step that vendor case studies skip is the one that matters: scientists save close to seven hours each week and mostly reinvest that time in research rather than cutting hours, while nearly half of the time-savers burn more than a quarter of that gain checking model output. In technical work, verification is not overhead to be optimized away; it is the control that makes the output usable at all.

A larger study of a generative AI system used by 5,000 customer support agents found productivity and retention both rose while stress fell, per TechTarget’s coverage. Short replies checked against written policy are cheap to evaluate, which is why the net gain there is cleaner than in research or legal drafting. The two results do not conflict. They bracket the range.

The rule that falls out of both: net savings track the cost of catching a wrong answer. Where checking is fast and cheap, most of the gross gain survives to the P&L. Where checking demands the same expertise as the original work, the tax eats the saving and the business case has to rest on something else, such as coverage during peak volume or retention. Forbes’ October 2026 piece on separating real gains from extra work makes the same point from the management side: adoption counts for little if the rework it creates stays invisible in the workflow.

Four Data Points Side by Side

These measurements answer different questions and should not be averaged into one verdict on AI. Read them as a portfolio of evidence with different scopes and different horizons.

Signal	What it measures	Reported figure	Source and date
Hyperscaler capital spending	Planned 2027 outlays by the five largest US hyperscalers	More than 50% growth to $1.2 trillion	Bloomberg citing Goldman Sachs, Sept. 25, 2026
Research time saved	Self-reported weekly hours saved by scientists, minus validation time	Nearly seven hours saved; nearly half spent over 25% of the gain verifying	Campus Technology, Sept. 23, 2026
Support agent performance	Productivity, retention and stress among 5,000 agents using a generative AI assistant	Productivity and retention rose, stress fell	TechTarget
Headcount at AI-intensive firms	Hiring patterns across 21,000 firms ranked by generative AI spending intensity	Higher headcount and more entry-level hiring	TechRepublic citing Ramp and Revelio data

Adoption by Industry and Firm Size

Tech and finance set the pace, and the skills that spread next will reach healthcare and education on a slower schedule, according to MIT Technology Review’s Oct. 20, 2026 review of TripleTen research. The concentration matters for benchmarking. A productivity gain measured at a software company tells you almost nothing about what a hospital billing department will see, because the tasks, review requirements, and data access rules differ.

The Census Bureau’s Business Trends and Outlook Survey gives the most granular public view, breaking AI use down by industry, firm size, state, and city from September 2026 responses, as Motley Fool’s research summary lays out. Firm size is the split worth watching. Small firms adopt through hosted APIs and shrink their fixed costs, while large firms build internal platforms and absorb integration costs that never appear in a subscription line item.

What workers actually do with the tools is narrower than product marketing suggests. ZDNet’s 2026 survey work found usage clustered around writing, summarizing, and search-style lookup, with far less uptake on the autonomous, multi-step task completion that dominates demos. That gap between deployed use and showed capability is where most failed pilots start.

Hiring Data and the Job-Loss Story

Data from 21,000 firms compiled by Ramp and Revelio associates intensive generative AI adoption with higher headcount and more entry-level hiring, per TechRepublic’s report. The finding is correlational. Firms spending heavily on AI are often growing for other reasons, and causality could run either direction. It does rule out the strongest version of the claim that AI spending mechanically eliminates entry-level roles.

The composition of those roles is shifting. PwC’s assessment is that entry-level AI workers now need skills closer to senior level, Computerworld reported, because the junior tasks that used to build judgment are the ones AI handles first. A firm can hire more people and still raise the bar for each hire. The result is a workforce that grows while the standard for every individual hire rises.

Costs also migrate outside the deploying firm. Politico’s October 2026 analysis traces early AI job losses and cost shifting onto workers, courts, and infrastructure, while The Economist notes that an AI-driven productivity lift expands the economy but can also worsen fiscal positions in heavily indebted rich countries if gains accrue to capital faster than to tax receipts. Neither effect shows up in a vendor ROI calculator.

Measuring Verification-Adjusted Gain

Gross time saved is the wrong number to take to a budget meeting, because it ignores the review step and the per-call cost. The snippet below instruments a support-reply workflow so you can compute net minutes saved alongside latency and token usage, using the OpenAI Python client.

import os, time
from openai import OpenAI

client = OpenAI() # reads OPENAI_API_KEY from the environment
MODEL = os.environ["MODEL_ID"] # pin the exact model id you benchmarked

def draft_reply(ticket_text: str) -> dict:
 start = time.perf_counter()
 resp = client.chat.completions.create(
 model=MODEL,
 messages=[
 {"role": "system",
 "content": "Draft a reply to the customer. Cite the policy id you relied on."},
 {"role": "user", "content": ticket_text},
 ],
 temperature=0,
 )
 latency_s = time.perf_counter() - start
 usage = resp.usage
 return {
 "draft": resp.choices[0].message.content,
 "latency_s": round(latency_s, 2),
 "prompt_tokens": usage.prompt_tokens,
 "completion_tokens": usage.completion_tokens,
 }

def net_minutes(baseline_min: float, edit_min: float, verify_min: float) -> dict:
 # gross saving is what a vendor demo shows; net saving is what finance can book
 gross = baseline_min - edit_min
 return {
 "gross_saved": round(gross, 1),
 "net_saved": round(gross - verify_min, 1),
 "verify_share_of_gross": round(verify_min / gross, 2) if gross > 0 else None,
 }

# Note: not production-ready. A real harness needs per-call cost from current
# pricing, timeout and retry handling, PII redaction before the API call, and
# verify_min logged from the agent UI instead of self-reported by the agent.

The third output matters most. If verification runs above roughly a quarter of the gross time saved, the business case stops resting on hours and has to rest on quality, retention, or coverage during peak volume. Those are legitimate returns, but they need their own baseline, measured before deployment rather than reconstructed afterward.

Two things to watch as 2027 capacity comes online. First, whether inference prices fall fast enough to change the net calculation, which depends on how full the new data centers run rather than on announcements. Second, whether verification tooling improves, since every hour of human checking removed from the loop converts directly into booked savings. Until both move, the honest number to report is the net one.

Key Takeaways:

Goldman Sachs strategists expect the five largest US hyperscalers to spend more than half again on AI infrastructure in 2027, reaching $1.2 trillion, per Bloomberg’s Sept. 25, 2026 report.

Time savings from AI are real but partly consumed by verification: scientists reported nearly seven hours saved per week, and nearly half of them spent over 25% of that gain checking outputs.

Adoption is concentrated in tech and finance, with Census Bureau survey data showing wide gaps by industry and firm size.

Data from 21,000 firms links intensive generative AI spending to higher headcount, which cuts against the simplest job-loss story.

Track verification-adjusted net savings, not token volume or gross hours, before you sign a multi-year compute commitment.

More in-depth coverage from this blog on closely related topics:

Sources and References

Sources cited while researching and writing this article: