AI time savings are not a fixed percentage that applies to every legal task. A recent international survey reports that 62% of participants save between 6% and 20% of their weekly working time. That is an important signal, but it does not prove the impact in a particular firm. The task, baseline, review and rework determine the real result.
Reading the latest figure correctly
The 2026 Future Ready Lawyer Report from Wolters Kluwer is based on an online survey of 810 legal professionals from law firms and legal departments in the United States, China and several European countries. Germany was included; Switzerland and Austria were not. The survey ran from 8 to 25 August 2025.
Sixty-two percent of respondents reported saving between 6% and 20% of their weekly working time through AI. According to the report page, the average was close to 10% of the working week. The unit matters: the finding concerns a percentage of weekly time, not six to twenty hours saved.
The measurement method matters too. This is self-reported evidence. Participants estimated their own time savings; their work was not timed under controlled conditions and compared with a manual control group. The result is evidence of perceived efficiency, not a guarantee for a particular matter.
Adoption, use and productivity are different measures
Current studies often use similar language for different things. The 2026 ILTA Technology Survey reports that 94% of 508 responding firms are engaged with generative AI. The figure includes use and exploration. It does not show how often individual professionals use a tool or how much time it saves.
A product-usage event answers another question again: it can show that a workflow was started, but not whether it was completed, useful or faster. A started Review may be completed, abandoned or followed by substantial rework. Usage therefore remains separate from productivity. The existing guides cover practical AI implementation in law firms and adoption in legal departments.
A reliable evaluation must keep three levels separate:
Level | Suitable measure | What it does not prove |
|---|---|---|
Adoption | approved access, active users | actual use |
Use | started and completed workflows | time savings or quality |
Impact | time, rework, material corrections, outcome | transferability to other tasks |
Mixing these levels can turn many logins into a claim about efficiency or one fast task into a general return on investment. Neither conclusion is justified.
Why broad claims of 40% to 70% are not reliable
The previous version of this CASUS article opened by claiming that Legal AI could save between 40% and 70% of the time spent on document analysis, research and contract review. There was no single study covering all three task types, markets and working methods. The range is therefore not used in this revision.
Spectacular examples involving millions of documents are not a useful average for day-to-day legal work either. Large-scale extraction, an open research question and negotiation of a bespoke clause have very different baselines. The share of automatable work and the required review effort differ accordingly.
The same caution applies to benchmarks of answer quality. A strong score on standardised questions does not measure the time required for a full matter or the quality of a different jurisdiction. Accuracy and time are separate outcomes.
The task determines the possible gain
High efficiency is most plausible where the workflow repeats, the instruction is clear and the output format is consistent. Examples include initial structuring of a document collection, extraction of defined fields or comparison of recurring contracts against a maintained playbook. A structured Review is a measurable working step, not evidence of a better result by itself.
The gain is less predictable for open legal questions, unusual definitions, incomplete records or strategic judgement. A fast first answer may create additional review time. A persuasive but incorrect finding does not save time when the responsible professional must detect and repair it.
Switching between tools also belongs in the calculation. If a result must be copied from a separate application, reformatted and inserted in the right part of a document, part of the analytical gain may disappear. Technical generation time alone is therefore not a useful productivity metric.
A pilot needs a baseline
A reliable pilot starts before the tool is used. Select five to ten recurring, completed tasks for which the outcome and actual effort are already known. The matters should be similar enough to support comparison, but not so artificial that they stop reflecting real work.
For each task, record at least:
active working time without AI;
active working time with AI;
review and rework time;
material corrections before approval;
completeness against the requested output;
the responsible professional's decision on usability.
Time must be measured consistently. Counting research pauses in the manual version while timing only technical generation in the AI version creates an artificial advantage. Learning, prompt adjustment and transfer into the working document also belong in the total.
Quality is not an optional condition
Faster work is valuable only when the output meets a predefined quality threshold. For contract review, suitable criteria may include relevant issues found, unsupported findings, missed exceptions and correct references to the agreement. For research, source support, currency, jurisdiction and verifiability may matter.
Define the assessment before the pilot begins. Otherwise, the threshold can move later to favour the desired result. A single easy or difficult matter can also distort a small sample. Report the median, range and outliers alongside any average.
Time saved is not automatically revenue gained
Even a methodologically sound time saving does not answer the commercial question. Under hourly billing, faster work may initially mean fewer billable hours. Under fixed fees or capped budgets, the same efficiency may improve margins. In either model, value arises only when the released capacity is used well or the benefit is made visible to clients.
Research published by the Thomson Reuters Institute in August 2026 illustrates the gap. It draws on 116 interviews with law firm leaders and 2,527 lawyers identified as stand-out performers by clients. Many see an AI strategy, but far fewer connect it confidently with workflows, client value and commercial implementation.
Internal reporting should therefore separate time, quality and commercial use. A pilot can succeed operationally before a revenue effect is measurable. Conversely, a fast task without adequate review may look inexpensive initially and create costs later.
An honest result is usually task-specific
A useful evaluation does not end with a statement such as "Legal AI saves 10%." It might conclude that the median working time for an initial review of recurring agreements fell across eight comparable matters while rework remained stable. For open legal questions, the sample may have been too small and the variation too wide for a conclusion.
This form is less dramatic, but much more useful for a purchasing and implementation decision. It shows where the tool belongs, which review step is required and where the evidence remains insufficient.
Teams that want to test the process can evaluate CASUS with their own work. Define the task, baseline, quality threshold and responsible human review before the first comparison.
Frequently asked questions
How much time does Legal AI save on average?
There is no reliable average covering every legal task. In the 2026 Future Ready Lawyer survey, 62% reported saving between 6% and 20% of weekly working time. It is self-reported evidence, not a guarantee for a specific workflow.
Does six to twenty percent mean six to twenty hours?
No. The survey refers to the percentage of weekly working time. It must not be restated as a number of hours.
Do international studies apply to Switzerland, Germany and Austria?
Only with caution. Tasks, organisations, jurisdictions and products differ. Research provides comparison points, while the effect of a specific workflow should be tested on the team's own matters with a documented method.
Which time belongs in the measurement?
Include preparation, use of the tool, human review, rework and transfer into the working document. Measuring generation time alone understates the real effort.
How many matters does a pilot need?
Five to ten recurring matters can provide a starting point for an internal comparison. Small samples remain uncertain, so report the median, range and outliers rather than turning them into a general market figure.
Why is a usage rate not proof of productivity?
A usage rate shows that a feature was opened or started. It does not show completion, required review or whether the result met the quality threshold.
