There is a category of work where the reading is not optional, not skimmable, and not deferrable, because a statute put a clock on it. My example is the one I know from operating a product in the space: the data subject access request. Someone asks an organization for every piece of personal data it holds about them, and in the UK the organization has one calendar month to comply. The requester's file might touch emails, HR records, complaints, call notes, tickets, years of accumulated paper. And before any of it goes out the door, someone must find every reference to the requester and, just as critically, protect everyone else: the law requires third parties' information to be redacted or their disclosure justified, name by name, page by page.
Understand what that job is, mechanically. It is exhaustive reading under deadline, where the cost of a miss runs in both directions. Miss the requester's data and you have failed the statutory duty. Miss a third party's name and you have breached someone who never asked to be in this story at all. The traditional method is paralegals, spreadsheets, and evenings: a person reads perhaps fifty or sixty pages an hour of this material with the attention it actually needs, which prices a ten-thousand-page request in the hundreds of hours before a single judgment is made. Organizations facing this either burn the staff, hire outside help at outside rates, or quietly sample instead of reading, and the third option is the industry's open secret, because nobody can staff exhaustive reading on a one-month clock.
This is the cleanest real-world case I know for the machine readers, because the machine does the one thing the job actually requires and humans structurally cannot deliver at volume: it reads every page. Not the likely pages. Every page, the attachments, the duplicate threads, page forty of the exhibit nobody opened, at the same attention on page nine thousand as on page one. Every name found, every identifier flagged, every candidate redaction marked with its location. What used to be the whole engagement, the finding, compresses from hundreds of hours to a background job.
But here is the part that makes this a field note rather than a product pitch, and it is the same lesson every domain in this series keeps teaching: the reading was never the hard part. It was just the expensive part. The hard part survives fully intact, because the law aims it directly at human judgment. UK guidance sets out a balancing test for third-party information: is consent feasible, is the person identifiable in context, is disclosure reasonable weighing everyone's interests. That test cannot be batch-processed, because it is not information retrieval. It is a decision about competing rights, made per name, per context, and someone accountable has to make it and be able to defend it later.
So the workflow that actually works splits exactly along that line. The machine reads everything and produces the complete inventory: every occurrence, every name, nothing missed. The human works from the inventory and spends their hours where the statute wants them spent, on the balancing calls, the judgment redactions, the defensible reasoning. The professional's job description quietly inverts: from finder who occasionally judges to judge with a perfect finder. And the quality moves in the direction people don't expect. The old method's real risk was never bad judgment; it was good judgment applied to an incomplete picture, because sampling was all the clock allowed. Exhaustive reading under the judgment of the same professional is simply a better process than either the machine or the person alone.
Two hard-won cautions from the trenches, for anyone applying this pattern to their own deadline reading, whether it is discovery, diligence, audit, or claims. First, demand provenance on everything. A flag that says "personal data, page 4,102, line 12" is usable; a summary that says "we found 340 names" is not, because the professional has to stand behind each call, and you cannot stand behind an aggregate. The trace back to the page is what keeps the human genuinely in the deciding rather than ceremonially attached to it. Second, never let the completeness of the machine's reading launder the confidence of the conclusions. Read everything, then decide like it is your name on the response, because legally speaking, it is.
The one-month clock is not getting longer, the files are not getting smaller, and the people doing this work were never the problem. They were rationing the only resource the job consumed, and the ration just ended. Ten thousand pages by Friday used to be a staffing crisis. Now it is a Tuesday-morning job queue and a Thursday of proper judgment, which is what the statute wanted all along: not heroic reading, but considered answers, on time.