Evidence · strength 2 of 5 · searched to 2 Nov 2026
Does AI help professionals write faster and better?
AI for drafting work documents may help, but the evidence is not yet strong enough for a firmer verdict. 3 of 3 studies found a clear gain in quality ratings.
What experiments found when people drafted work documents with generative AI: quality as judged by blind graders, time taken, and how sure we can be.
VerdictMay help
No good evidence. Studies point to a gain, but no factor named in advance explains who gains.
ImpactGain
3 of 3 studies found a clear gain in quality ratings.
Evidence strength2 of 5
We have low confidence in this estimate.
Cost per user a year$$
General AI assistants have free tiers; business plans list about $20–$30 a user a month (checked 2 Nov 2026).
Pilot rating under our v0.3 rules. The rules are published on our method page but not yet independently reviewed, so ratings may change as the method is checked and new studies arrive. Our method.
What the studies found
Quality (headline)
Gain- Noy and Zhang (2023): +18 Blinded grades from experienced professionals in the same occupations.
Our results show that ChatGPT substantially raised productivity: The average time taken decreased by 40% and output quality rose by 18%.
- Cruces et al. (2026): +1.24 SD Overall task score (reply email to the boss diagnosing a business problem and proposing a solution), low-education participants (p < 0.01).
The overall score rises by 1.242 SD for low-education participants and 0.834 SD for high-education participants, with both estimates significant at the 1% level.
- Cruces et al. (2026): +0.83 SD Overall task score, high-education participants (p < 0.01).
The overall score rises by 1.242 SD for low-education participants and 0.834 SD for high-education participants, with both estimates significant at the 1% level.
- Ju and Aral (2026): +0.323 points on a 7-point rating scale Human-rated ad text quality (7-point scale), human-AI vs human-human teams (SE 0.024, significant).
Human-AI 1.954 0.323 -0.136 -0.014 -0.094 (0.197) (0.024) (0.021) (0.021) (0.006)
- Ju and Aral (2026): −0.136 points on a 7-point rating scale Human-rated ad image quality (7-point scale), human-AI vs human-human teams (SE 0.021, significant).
we found that human-AI teams produced higher quality ad text but lower quality ad images compared to human-human teams (Table 2)
Output
Large gain- Ju and Aral (2026): +50% Ads submitted per worker in a 40-minute task, human-AI vs human-human teams.
Participants in human-AI teams submitted 50% more ads per worker, per unit time than their counterparts in human-human teams
Skill
Small gain- Cruces et al. (2026): +0.17 SD Follow-up questions answered without AI, low-education participants (p < 0.05).
Prior AI access increases the follow-up score of low-education participants by 0.171 SD, significant at the 5% level.
Time
Gain- Noy and Zhang (2023): −40% Time taken on an occupation-specific writing task (press releases, short reports, analysis plans, delicate emails). Self-reported.
Our results show that ChatGPT substantially raised productivity: The average time taken decreased by 40% and output quality rose by 18%.
- Cruces et al. (2026): −1.514 minutes per task Task completion time, high-education participants (control mean 10.423 minutes) (p < 0.01).
AI assistance reduces completion time by 0.961 minutes for low-education participants, relative to a control mean of 10.697 minutes, and by 1.514 minutes for high-education participants, relative to a control mean of 10.423 minutes.
- Cruces et al. (2026): −0.961 minutes per task Task completion time, low-education participants (control mean 10.697 minutes) (p < 0.10). Not statistically clear.
AI assistance reduces completion time by 0.961 minutes for low-education participants, relative to a control mean of 10.697 minutes, and by 1.514 minutes for high-education participants, relative to a control mean of 10.423 minutes.
A minus sign on time means the task took less time. Bands are provisional; see how impact is banded. Effects are not pooled yet: each band is set from one result per study, and where studies use different units we show only the direction.
How sure are we?
- Starting level3 of 5, from 2 grade A or B studies.
- Down one: old models0 of 3 studies show they used a model released since May 2025 (the 18-month window); the rest used older models or do not say which.
- Flag (no deduction)Mostly working papers or preprints, not yet peer-reviewed (flagged, not deducted).
- Publication biasnot testable with fewer than 10 studies.
- Result2 of 5. We have low confidence in this estimate.
Behind the average
- Lower- vs higher-ability workers (baseline grade) (Noy and Zhang (2023)): Inequality between workers fell. In the working paper, low-graded workers gained quality and speed while high-graded workers kept quality and got faster (Science figures not opened).
- Education gap (Cruces et al. (2026)): Without AI, high-education participants outscored low-education ones by 0.548 SD; with AI the gap fell to 0.139 SD, closing about three-quarters of it.
- High-education participants, follow-up without AI (Cruces et al. (2026)): No clear carry-over gain once AI was removed (+0.071 SD, not significant), but no harm either.
- Output diversity (Ju and Aral (2026)): Human-AI teams produced more similar ads to each other (cosine distance -0.094), a homogenising effect.
How, where and why it works
Why it works
1 of 4. Studies suggest the AI takes over first drafts, with people editing; few measure this directly.
For whom and where
2 of 4. Weaker writers gained more in the best-known trial, narrowing the gap between people.
What it takes
0 of 4. Studies use short one-off tasks; they say little about rollout.
The full cost
1 of 4. Total task time includes reading and editing the draft, but costs are not counted.
Four separate scores from 0 to 4, never added up: how much the studies tell us about why it works, for whom, what it takes and what it really costs (the EMMIE method).
What to do next
Measure your pilot →If you are trying it yourself
- Use AI for a first draft, then rewrite in your own words.
- Check every fact, name and number.
- Compare the time with your usual way on a few documents.
If you are rolling it out
- Agree when AI drafting is allowed and how it is disclosed.
- Have experts rate samples blind, as the studies did.
What it takes
- Short tasks in experiments are not a full job: measure on your own documents.
- Keep a human editor responsible for every document.
What can go wrong
- Text can sound confident and still be wrong.
- Writing can become more alike when many people use the same AI.
The studies
| Study | Grade | Design | People | AI model (released) | Company role |
|---|---|---|---|---|---|
| Experimental evidence on the productivity effects of generative artificial intelligence Shakked Noy, Whitney Zhang, 2023. Science 381 (journal) | B | RCT, realistic task | 453 college-educated professionals (Science abstract); 444 in the March 2023 MIT working paper | ChatGPT (OpenAI) (Nov 2022) | none |
| Does Generative AI Narrow Education-Based Productivity Gaps? Evidence from a Randomized Experiment Guillermo Cruces, Diego Fernandez Meijide, Sebastian Galiani, Ramiro H. Galvez, Maria Lombardi, 2026. NBER Working Paper 34851 (working paper) | B | RCT, realistic task | 1,174 adults aged 25-45 in Argentina (low- and high-education groups); preregistered, AEA RCT Registry 0016607 | OpenAI GPT-4.1 (custom assistant embedded in the task interface) (Apr 2025) | none |
| Collaborating with AI Agents: A Field Experiment on Teamwork, Productivity, and Performance Harang Ju, Sinan Aral, 2026. arXiv:2503.18238 (preprint) | C | RCT, realistic task | 2,234 US Prolific participants (1,258 in human-AI teams, 976 in human-human teams) producing 11,024 ads; 1,168 separate raters; ads run on X for about 5M impressions | OpenAI gpt-4o (gpt-4o-2024-08-06), with DALL-E 3 for images (Aug 2024) | none |
Grades: A, a well-run randomised trial in real work; B, a strong natural experiment in real work or a well-run trial on realistic tasks; C, a trial with clear limits or a matched comparison; D, before-and-after, surveys or case studies. How studies are graded.
Does the rating hold up?
- Run without company-run studiesNo change: no such studies.
- Run without working papers and preprints1 studies: strength 0, verdict no good evidence, headline band gain.
How up to date is this evidence?
Method →Evidence searched to 2 Nov 2026. Studies cover models released Nov 2022 to Apr 2025. Rating computed 5 Nov 2026 under Module 15 v0.3 (pilot).
How this page was made: AI research agents searched for studies and read each paper; a second pass checked every number on this page against the paper itself, with the quote and where it appears. The rating is computed by fixed rules, not chosen. No person outside AI Adoption has reviewed it yet. Editorial owner: AI Adoption editors. Spot a mistake? Corrections.
Disclosure: AI Adoption uses AI tools, including Claude from Anthropic, in its research and editorial work. No company pays for or has any say in a rating.