Your AI Tools Are Creating Work, Not Saving It

A new ResumeTemplates.com survey of U.S. workers (summarized at metaintro as “The AI Productivity Trap”) found that 31% of workers say their workload increased after their company rolled out AI. Some of those workers report they are now expected to do two to four times as much work as before. Read that twice. The tool pitched as a workload reducer is, for roughly one in three workers, a workload multiplier.

That number is easy to dismiss in isolation. It stops being dismissible the moment you line it up against two other 2026 datasets that were collected independently and tell the same story. ActivTrak watched the actual clickstream. BCG ran a controlled study on cognitive load. All three arrive at the same place. More AI tools, past a point, produce more work, not less.

Most AI strategies are built on the opposite assumption. That is the problem.

Quick Verdict

Finding	What the Data Shows
Workload increase after AI	31% of U.S. workers report workload went up, not down (ResumeTemplates.com)
Multiplier effect	Some affected workers now expected to do two to four times as much work
Email time after AI	Up 104% (ActivTrak, 443M hours observed)
Chat and messaging	Up 145%
Daily focused work	Down 23 minutes per person
Productivity cliff	Self-reported productivity rises up to 3 AI tools, then plummets at 4+ (BCG via HBR)
The label researchers used	”AI brain fry”
What this means for you	Your AI rollout needs a tool ceiling, not another seat license

The 31% Number Is Not a Measurement Error

The ResumeTemplates.com survey asked a simple question. After your company adopted AI tools, did your workload go up, stay the same, or go down? Thirty-one percent said up. Some of those workers now report being expected to do two to four times as much work as before. The tool pitched as a workload reducer is functioning, for a meaningful share of the workforce, as a workload multiplier.

That result is counterintuitive enough that the instinct is to find a methodology flaw. The problem is that two other studies using completely different methods converge on the same conclusion.

ActivTrak’s 2026 State of the Workplace didn’t ask workers how they felt. It watched what they did. Across 443 million hours of digital activity from 163,638 employees at 1,111 organizations, time spent in work applications went up 27% to 346% across the board after AI adoption. Email climbed 104%. Chat and messaging climbed 145%. Meanwhile, daily focused work sessions dropped by 23 minutes per person.

This is the key point. The survey data and the observational data agree. Workers say they are doing more. The instrumentation confirms they are doing more. The AI was supposed to absorb the low-value tasks. Instead, the low-value tasks multiplied.

Why More AI Creates More Work

There are three mechanisms, and each one is visible in the data if you know where to look.

1. AI output triggers more communication, not less. Draft an email in ChatGPT, send it, get a reply, run the reply through Claude to summarize it, ask a follow-up. Each AI-assisted interaction is faster individually and produces more interactions in aggregate. The 145% jump in chat and messaging is not a bug. It is the predictable output of a workforce where every message costs less to produce. Volume expands to fill the new bandwidth.

2. Review load scales with generation speed. If an AI writes five drafts in the time it used to take you to write one, you now have five drafts to read, edit, fact-check, and approve. The generation step got fast. The review step didn’t. For most knowledge work, the review step was the actual bottleneck. I wrote about this pattern in the shadow AI piece. Speed on the wrong side of the workflow does not equal productivity.

3. Expectations reset upward the moment adoption is announced. The ResumeTemplates.com data caught this directly. Managers who know their team has AI raise the bar for what a week’s output looks like. That is rational from a competitive standpoint and punishing from a workload standpoint. The worker absorbs the raised expectations and the new tooling load at the same time.

None of these mechanisms are about AI being bad. They are about a rollout that added capability without subtracting demand. You cannot add a powerful tool to a workflow and expect the workflow to shrink. Workflows expand to consume new capacity unless someone explicitly shrinks them.

The BCG Finding That Nobody Wanted to Hear

Then there is the BCG study published in Harvard Business Review, which ran on 1,488 full-time U.S. workers in March 2026. This one is the most uncomfortable of the three, because it puts a specific number on the ceiling.

Productivity rose when workers went from one AI tool to two. It rose again at three. At four or more, it dropped. Not plateaued. Dropped. Self-reported productivity plummeted. Mental fatigue jumped 12%. Information overload jumped 19%. Thirty-four percent of affected workers reported actively planning to quit.

BCG labeled the effect “AI brain fry.” The term got mocked on LinkedIn for two weeks. The underlying finding survives the mockery.

Three is the ceiling. Four is the cliff. If your company has licensed seven AI tools for the average employee and is wondering why the expected productivity gains never showed up, that is your answer.

The researchers also quantified the downstream cost. Workers experiencing brain fry made 11% more minor errors and 39% more major errors. This is the part the cost-savings calculations never include. You save an hour on drafting, then lose three hours on a major error that a less-overloaded worker would have caught.

How Does the Three-Tool Ceiling Show Up in Real Org Charts?

Here is the shape of it in a typical knowledge work team. Count the tools.

1. A Copilot or ChatGPT Enterprise license for general writing and synthesis.
2. A purpose-built tool inside the team’s workflow (Gong for sales, Glean for search, Intercom Fin for support).
3. A code or data tool if the role touches either (Cursor, GitHub Copilot, Hex, Julius).

That is three. The BCG data says this is the sweet spot.

Now count what most enterprises have actually deployed. Copilot across the suite, plus a marketing AI, plus a support AI, plus a coding AI, plus a meetings AI, plus whatever three vendors landed in Q1 procurement cycles. Seven to ten is common. The average marketing employee in the BCG sample was using between four and six AI tools daily. Marketing was the function with the highest brain fry rate. Twenty-six percent reported symptoms.

If your tool count is five or more per worker and your productivity metrics are flat, that is not a mystery. That is the ceiling.

The 40-Second Definition

The AI productivity trap is the pattern where adding AI tools increases the total time workers spend on work, reverses self-reported productivity gains past three tools, and shifts the bottleneck from generation to review and communication. It is measurable at the worker level (ResumeTemplates.com, 31%), at the organizational level (ActivTrak, 443M hours), and at the cognitive level (BCG, 1,488 workers). The fix is a tool ceiling, not more training.

Why the Industry Can’t See This Clearly

Every AI vendor’s pitch deck contains a productivity gain number. “30% faster drafting.” “10 hours a week saved.” “50% reduction in cycle time.” Those numbers are real. They are also measured on the wrong axis.

Vendor metrics measure per-task speedups. They do not measure aggregate workload. They do not measure review time, coordination overhead, error rates from cognitive fatigue, or the increase in meetings required to discuss the outputs the AI generated. When you stack seven per-task speedups on a single worker, each one real, the combined effect is the ActivTrak pattern: more time across every category, less focused work, worse outcomes.

This is the same structural problem I laid out in why your AI tool stack is about to get smaller and more expensive. The per-tool ROI math does not aggregate. Three tools that each save an hour do not necessarily save three hours. Often they save less than one.

Sales teams selling these tools can show you the per-tool number. The CFO asking why the consolidated headcount plan didn’t change is asking the right question and getting shrugs.

What a Three-Tool Ceiling Actually Looks Like

The implication isn’t that AI doesn’t work. It’s that the deployment pattern most companies are running is actively producing the problem the studies keep documenting. Buy whatever’s on the market, give every worker access to everything, measure tool-level ROI, repeat. Here is what the fix looks like in practice.

Step 1: Audit Actual Tool Count Per Worker

Not purchased licenses. Actively used tools. Run an internal survey or pull usage data. The number you get is almost certainly higher than you assume. The CIOs I talk to guess three and find six.

For each worker, list every AI tool they touched in the last week. Include browser extensions. Include the ChatGPT tab they keep open. Include the free tools IT doesn’t know about. If the count is above three for more than 25% of your workforce, you are in brain-fry territory. That is your real productivity problem, and no training module fixes it.

Step 2: Consolidate to Three Roles

Pick the three functions AI will serve for each role, and pick one tool per function. A marketing copywriter doesn’t need six AI writing tools. They need one writing tool, one research tool, and one analysis tool. Total. If they currently have nine, you have six to retire.

This will be politically hard. The tool champions inside the company have opinions. The vendors will call. The procurement director who negotiated the contract will push back. Do it anyway. Every tool past three is a direct cost in self-reported productivity, per the BCG data, before you even count license spend.

I laid out the underlying consolidation logic in the agent sprawl prevention piece. Same framework applies to tool count.

Step 3: Redesign Workflows Around Review, Not Generation

The reason AI created more work is that it made the generation step cheap and left the review step alone. Reverse that. For every workflow, ask where the review bottleneck is, and fix it before adding any AI generation capability.

Example: if an AI writes five draft emails and a human now has to read five instead of writing one, the net time is often higher, not lower. The fix is an AI review layer, or a workflow change that commits to the first draft with a quick human check, not an elaborate review of every option.

Step 4: Install a Tool Budget at the Team Level

Not a dollar budget. A count budget. Every team gets three AI tool slots per role. Adding a fourth requires removing one of the existing three. This forces the tradeoff conversation that most orgs are avoiding. It also cuts vendor proliferation at the source.

The teams I’ve seen run this well treat the three slots like a roster. Switching a tool is a decision. Running five tools in parallel is not an option. The productivity numbers recover within a quarter.

Step 5: Measure Workload, Not Tool ROI

Stop asking “did this tool save time?” Start asking “is total worker output higher and is total worker input lower?” Those are different questions. The first gets a yes from every vendor. The second is where the actual productivity gain lives, and where the current deployment pattern is failing.

I’ve written the instrumentation layer for this in the AI productivity metrics piece. Revenue impact per worker, cycle time for complete workflows, and focused work hours per day are the metrics that catch the trap. Tool-level usage metrics don’t.

Three Mistakes That Keep the Workload Climbing

Stacking tools instead of sequencing them. The impulse to give every worker access to every AI tool feels democratic and it produces brain fry. Three tools picked deliberately beat seven tools picked by committee. Every time.

Measuring adoption instead of outcomes. “85% of our workforce uses Copilot weekly” is an adoption number. It tells you nothing about whether workload went up or down. The ResumeTemplates.com data suggests a high-adoption workforce is often a high-workload workforce. If adoption is your only metric, you will not see the trap until it is built into your culture.

Assuming the fix is training. Workers don’t need more training on their seventh AI tool. They need their seventh AI tool taken away. Training is the vendor’s preferred remedy because it keeps the license. It is rarely the correct one when the root cause is tool count, not tool mastery.

The Uncomfortable Implication

If the BCG ceiling is real, and three converging datasets suggest it is, the economics of enterprise AI adoption in 2026 are different than the pitch. The CFO who approved licenses for seven AI tools per employee was told each tool produced a productivity gain. In isolation, that was true. In combination, the gains reverse. The spend compounds. The productivity flatlines. The error rate climbs.

The 20% of firms that are capturing the vast majority of AI value, which I wrote about in 74% of AI gains go to 20% of firms, are not the ones with the most tools. They are the ones who picked the right three and aligned workflows around them. The other 80% are buying more tools and wondering where the productivity went.

It went into 104% more email, 145% more chat, and the 23 minutes of focused work that vanished from the average day. It went into the review queue, the error rate, and the 34% of affected workers BCG caught planning to quit.

Your AI tools are creating work. Not because AI is broken. Because the deployment pattern most companies are running is a tool-count problem dressed up as a strategy. The fix is a ceiling, a consolidation, and a workload metric. Three tools, picked on purpose. That is the version of AI productivity that actually works.

Your first move this week: count the AI tools your top ten workers actually use. If the average is above three, you have your answer.

---

Related Reading:

• Shadow AI Is Costing You More Than Productivity. Here’s How to Fix It.
• Why Your AI Tool Stack Is About to Get Smaller and More Expensive
• Agent Sprawl: An SMB Prevention Guide for 2026
• AI Productivity Metrics Are Dying. Revenue Impact Is Taking Over.
• 74% of AI Gains Go to 20% of Firms. Here’s Why.