AI tools are often framed as time-savers. In many situations, they are. A controlled study of professional writing tasks found that access to ChatGPT reduced average completion time by 40% while improving rated output quality.[1]
But speed gains are not universal. In a 2025 randomized trial involving experienced open-source developers working on familiar repositories, participants took 19% longer when AI tools were available.[2]
When using AI tools actually makes things slower, the problem is usually not that the tool takes too long to respond. The problem is that the whole workflow takes longer once prompting, reviewing, correcting, narrowing, comparing, and reshaping are included.
Fast output and fast completion are not always the same thing.
What “Slower” Really Means in This Context
This kind of slowdown is easy to miss because the first visible step looks fast.
A response appears in seconds, which creates an immediate sense of progress.
But generation time is only one part of a task.
The more useful measure is time to a finished result.
That may include:
- Explaining the task
- Providing context
- Waiting for output
- Reading what was produced
- Checking facts
- Correcting tone or structure
- Asking follow-up questions
- Comparing versions
- Reformatting or transferring the result
- Making the final decision yourself
If those additional steps take longer than the work AI saved, the overall workflow becomes slower even though the first draft arrived quickly.
That is the central idea of this article.
Why Fast Output Can Still Create More Work
Generative AI is very good at producing plausible material quickly.
That is part of what makes it useful.
But a polished-looking answer is not automatically accurate, complete, or ready to use. NIST identifies confabulation as a recurring generative-AI risk in which systems can confidently produce false or erroneous content.[4]
That matters for workflow speed because uncertainty creates review work.
A paragraph that looks finished may still need:
- Fact-checking
- Tone correction
- Missing context added
- Overconfident wording softened
- Incorrect assumptions removed
- Important exceptions restored
So the effort may shift rather than disappear.
Drafting becomes faster.
Review becomes heavier.
Whether the trade is worthwhile depends on the task.
How the Slowdown Usually Happens
There are several common ways AI adds friction.
Prompt Overhead
Some tasks need so much explanation that preparing the prompt becomes a substantial part of the job.
If the task depends on a particular audience, tone, history, exception, format, source, or constraint, all of that may need to be communicated before the tool can produce something useful.
For a large task, that setup may be worthwhile.
For a small task, it may be faster to do the work directly.
That is why using AI for a short everyday task does not automatically make the task more efficient.
Sometimes the explanation takes longer than the execution.
Review Burden
The more accuracy matters, the more expensive review becomes.
A rough brainstorming list may need only a quick glance.
A factual explanation, professional email, technical document, or public-facing article may need much closer checking.
NIST recommends verifying sources and citations in generated outputs where reliability matters because fluent output can still contain errors.[4]
This is where perceived speed can become misleading. The tool may save five minutes of drafting but create ten minutes of verification.
Iteration
One prompt does not always produce a usable answer.
A person may then ask for:
- A shorter version
- A different tone
- More detail
- Less detail
- Another structure
- A new example
- A combination of two earlier versions
Each revision is fast individually.
But the total time can accumulate. At some point, the person may discover that they are editing around the tool rather than finishing the original task.
Tool Friction
There is also ordinary workflow overhead.
Copying and pasting.
Uploading files.
Moving between tabs.
Re-entering context.
Fixing formatting.
Comparing an AI result with the original document.
None of these steps feels significant on its own. Together, they can erase part of the time saved during generation.
The Task Itself May Be the Real Problem
AI tends to help more when the task is already reasonably clear. If the person has not decided what they want, the tool can generate many possible directions without resolving that uncertainty.
For example:
“Give me another version.”
“Try a different angle.”
“Make it shorter.”
“Actually, combine the first and third versions.”
The system is producing material rapidly.
But the difficult work is still happening somewhere else: deciding what the final result is supposed to be.
In this situation, AI can make the process feel active while adding more material around an unresolved decision.
That is not necessarily a model failure. The task itself has not yet been defined clearly enough.
When Checking Becomes More Expensive Than Drafting
Some work is naturally review-heavy.
A factual article may need source verification.
A technical explanation may need details checked against documentation.
A sensitive message may require careful wording.
A recommendation may depend on conditions the AI does not know.
For these tasks, review is not an optional finishing touch. It is part of the actual job.
The faster first draft therefore matters less if the checking stage carries most of the responsibility.
This is also where this article differs from What AI Can Do Well — and What It Consistently Gets Wrong.
That article explains why generative AI can produce both useful language and recurring errors.
This one asks a narrower workflow question: What happens to total completion time once those errors and uncertainties have to be checked?
When the Hard Part Is Judgment, Not Wording
Some tasks look like writing problems but are really judgment problems.
The difficult part may be deciding:
- What matters most
- Which detail should be left out
- Which trade-off deserves emphasis
- What recommendation is appropriate
- What tone fits the situation
- Which source deserves trust
AI can generate language around those decisions. But generating language is not always the part consuming the most time.
If the person still has to make every important judgment afterward, the tool may add an extra layer rather than remove one.
A Useful Mental Model: The Food Processor
The original food-processor analogy works well here.
A food processor can save substantial time when preparing a large batch of repetitive ingredients.
For one small, precise cut, setup and cleanup may take longer than simply using a knife.
The machine is not failing.
The task does not justify the overhead.
AI tools can work similarly.
A long summarization task, multiple draft variations, or a large amount of repetitive transformation may justify the setup.
A two-sentence reply that needs careful personal judgment may not.
The useful question is therefore not:
“Is AI fast?”
It is:
“Is AI faster for this complete task?”
Too Much Output Can Become Its Own Burden
Another source of friction is surplus output.
Generative AI can create:
- Five alternatives instead of one
- Twenty ideas instead of three
- A long explanation where a short answer was enough
- Several possible structures
- Multiple competing recommendations
More options can be useful. They can also create selection work.
Someone now has to read more, compare more, reject more, and decide more.
The effort has moved from creating possibilities to sorting possibilities. That trade can be worthwhile when exploration is the goal. It is less useful when the person already knew roughly what they wanted.
Research Shows Why “AI Saves Time” Is Too Broad
Studies of AI productivity do not support one universal answer.
In a 2023 experiment involving 453 college-educated professionals completing mid-level writing tasks, people using ChatGPT finished the tasks about 40% faster on average, while independently rated quality improved by 18%.[1]
In a very different setting, METR studied 16 experienced developers completing 246 real tasks in mature open-source repositories during early 2025. Allowing AI tools increased completion time by 19% on average.[2]
The researchers explicitly cautioned against generalizing that result to software development as a whole.[2]
And the picture continues to change.
In February 2026, METR reported that later data involving newer AI tools could not provide a reliable estimate of current developer productivity effects because participant and task selection had introduced substantial bias. The researchers said newer tools may be producing larger speedups, but the available data were too weak to estimate the size confidently.[3]
The useful conclusion is therefore not:
AI makes work faster.
Nor is it:
AI makes work slower.
It is:
productivity depends heavily on the task, the user, the tool, the amount of review required, and the workflow around it.
When AI Usually Saves Time
AI is more likely to create a real time advantage when:
- The task is sufficiently large to justify setup
- The goal is already clear
- Rough output is acceptable
- Errors are inexpensive to identify
- The result can be checked quickly
- Repetition or scale matters
- Generating alternatives is genuinely useful
Examples may include first-pass drafting, rough summarization, reorganizing supplied text, generating variations, or transforming material that already exists.
These are not guarantees. They are signs that the time saved during generation has a better chance of surviving the rest of the workflow.
For the broader task-fit question, What AI Tools Are Good At (And What They’re Not) covers that distinction more directly.
When AI Is More Likely to Add Friction
AI can be less efficient when:
- The task is extremely short
- Explaining the task takes almost as long as doing it
- The required output is highly specific
- Important context is difficult to communicate
- Accuracy requires substantial verification
- The task depends mainly on personal or professional judgment
- Generated alternatives create more decisions rather than fewer
This does not mean AI should never be used in those situations. It means the time benefit should not be assumed.
Sometimes a person may still choose AI because they want another perspective or because generating alternatives is useful even if it takes longer. Efficiency is not the only reason to use a tool.
A Better Way to Judge Whether AI Is Saving Time
One simple distinction makes this easier:
Do not measure how quickly the AI responds. Measure how quickly the task finishes.
If the tool produces a draft in thirty seconds but requires fifteen minutes of repair, the thirty-second generation time is not the meaningful number.
Likewise, if a task that normally takes an hour becomes a thirty-minute process even after checking and editing, the tool has produced a real efficiency gain.
The measurement should include the whole workflow.
Prompting.
Generation.
Review.
Correction.
Integration.
Decision-making.
Completion.
That is where the actual time saving—or slowdown—appears.
What to Expect From AI Tools More Realistically
AI is not a universal shortcut.
It is a tool that can change where the effort happens.
Sometimes generation replaces enough manual work to make the overall task substantially faster.[1]
Sometimes prompting, review, correction, and integration outweigh the benefit, as at least one controlled real-world coding study has demonstrated in a particular setting.[2]
And because AI systems and workflows continue to change quickly, those effects should not be treated as permanent properties of the technology.[3]
The most useful question is therefore practical: does this tool reduce the total work required to reach the result I actually need?
If it does, the tool is saving time. If it only creates faster-looking activity while extending the path to completion, it is adding friction.
References
- Noy S, Zhang W. Experimental evidence on the productivity effects of generative artificial intelligence. Science, 2023.
- Becker J, Rush N, Barnes B, Rein D. Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity. METR, 2025.
- Becker J et al. We are Changing our Developer Productivity Experiment Design. METR, 2026.
- Autio C et al., National Institute of Standards and Technology. Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile. NIST AI 600-1, 2024.
