Designing against automation bias: why we refuse to build an accept-all button
Legal software should reduce mechanical work, not critical thought. Here is why we intentionally introduce friction into the contract review process.
Across our active accounts, we monitor the accept rate of inline redlines proposed by Review. By the time a firm has tuned its playbook for six months, it is common to see accept rates exceed 90%. Software vendors usually celebrate this metric as proof of accuracy. We view it with caution.
High accept rates look excellent on a performance dashboard, but they introduce a silent, structural risk. When a system is consistently correct, the human operator stops reading and starts clicking. Our objective is to accelerate legal workflows by removing mechanical drudgery. It is not to turn a junior associate into a rubber stamp.
The danger of a smooth interface
Automation bias is well documented in aviation and medicine. It is the tendency for a human to defer to a highly reliable machine, eventually stopping the independent verification of its outputs. In contract review, this manifests as a dangerous complacency. If ZAAN-7B Counsel correctly identifies and resolves a missing clause nineteen times in a row, the associate reviewing the twentieth contract is highly likely to click accept without reading the surrounding text.
Consider a limitation of liability clause in a complex master services agreement. Review flags a missing carve-out for gross negligence and wilful misconduct. It proposes a redline based on the firm’s playbook floor. The legal logic is entirely sound. However, the commercial reality of this specific transaction—perhaps a low-margin engagement where the counterparty holds all the leverage—might dictate conceding the point to close the deal.
If accepting the system’s redline is too frictionless, the lawyer stops evaluating the commercial context. They simply execute the mechanical task of clearing notifications from the sidebar. The nuance of the transaction is lost to the efficiency of the interface.
Structural friction in Review
This is why we intentionally design friction into our interfaces. The most frequent feature request we receive from new deployments is an ‘Accept All’ button for redlines. We will never build it.
Instead, Review 4.2 categorises deviations into severity tiers: High, Medium, and Note. We enforce different interaction patterns for each. For minor stylistic deviations flagged as a Note, the user can dismiss or accept the change quickly. But for a High-severity flag—such as a missing survival provision for a data breach indemnity, or a governing law mismatch—the interface deliberately slows the user down.
The ‘Insert Redline’ button in the Word add-in remains inactive until the reviewer expands the citation. They must look at the specific playbook rule or the precedent clause before they can apply the change. It is a delay of perhaps one second, but it forces a cognitive context switch. It breaks the rhythm of mindless clicking and forces the reviewer to acknowledge why the change is being made.
Forcing the context switch in Draft
We apply a similar philosophy to generation. In Draft 3.9, when a user asks the system to draft a bespoke termination right, the output is never inserted directly into the active document. It appears in a side panel.
The generated text is accompanied by citations drawn from the firm’s matter history via Recall. The drafter must explicitly drag the clause into the document or click to insert it at the cursor. Furthermore, hovering over the generated sentences highlights the precedent material used to model the phrasing. We want the lawyer to see immediately that a specific sentence was adapted from a 2024 software licensing agreement. If that precedent is misaligned with the current transaction’s risk profile, the user has the necessary information to discard the draft before it pollutes the current document.
Metrics that matter more than speed
We measure first-pass review time rigorously. Reducing the hours spent on routine vendor agreements is a primary reason firms deploy our platform. However, speed is a byproduct of reducing mechanical tasks, not the objective of the review itself.
True efficiency means the software automatically cross-references defined terms, calculates limitation periods, and verifies ethical walls. It means the lawyer does not have to scroll across sixty pages to check if a capitalised term was actually defined in the schedule. But the evaluation of risk allocation takes time. A standard non-disclosure agreement might legitimately be reviewed in four minutes. A 120-page share purchase agreement should not, regardless of how quickly the model processes the text.
If we observe in our telemetry that a user is clearing twenty High-severity flags in thirty seconds, the software has not made them faster