Review Steps You Should Not Hand to a Bot
The four-page lease sits on the kitchen counter. A tenant reads it once, skims the pet addendum, initials where the landlord marked, and signs. What the tenant has just done, without quite deciding to, is hand the most consequential document of the next twelve months to a skim. The signature is real. The reading was not.
This is the quiet failure mode of automated review. Software can now summarize contracts, flag clauses, check listings, scan receipts, and grade resumes. The tools are competent enough to be trusted and cheap enough to be used everywhere. What gets lost is the class of review that depends on a human being actually looking, holding the object in question in mind, and applying judgment that no training set contains. Some reviews should never be delegated, and it is worth naming which ones and why.
The lease clause that governs the deposit
A summarizing tool reads the deposit paragraph and reports: security deposit equal to one month's rent, refundable within thirty days of move-out. Accurate. Complete, as far as the summary goes. What the summary misses is the sentence buried two sections later: refund subject to a cleaning fee assessed at the landlord's discretion. The two clauses sit far apart. Each is unremarkable alone. Together they mean the deposit is not really one month's rent; it is one month's rent minus whatever number the landlord chooses to write down.
No summarizer is built to hold the whole document in tension at once. It compresses each section independently, and compression is exactly the operation that dissolves cross-references. The tenant who reads the whole thing, slowly, on paper, catches the pairing because the pairing is a feeling as much as a fact: a small cold click of recognition. Software does not get the click.
Reading the room tone in a recorded meeting
Transcript tools do a decent job on the words. They do nothing with the pause before the finance lead answers the question about the hiring freeze. That pause, three seconds of silence in a room where nobody usually hesitates, carries more information than the answer that follows it. Automated minutes render the exchange as a clean Q and A. A person in the room hears the hesitation and remembers it three weeks later when the freeze is announced.
The same asymmetry shows up in written review. A manager scanning an AI-generated performance summary sees that an employee "consistently meets deadlines" and "collaborates well." The summary is not wrong. It is the residue after every specific friction has been boiled off: the missed handoff in March, the client call that went sideways, the project that finished late but finished. Those are the facts a promotion decision actually turns on, and they survive only in the memory of a person who was paying attention.
The apartment listing with the wrong photograph
Listing aggregators and AI rental scouts will compare a unit against a renter's stated criteria and return a confident shortlist. What the scout cannot do is notice that the kitchen window in the third photograph looks directly onto a brick wall two feet away, or that the "sunny corner unit" has a fire escape bolted through the middle of its only window. A machine matches attributes. A person standing in the room registers dread.
Some of this can be caught from photographs, but only by someone who has trained an eye on the specific tells: water stains on a ceiling photographed at a flattering angle, a baseboard heater with a scorch mark, a door that does not meet its frame. The review that matters here is a walk-through, and a walk-through is a form of review that cannot be automated because the reviewer has to be physically present to be unsettled.
Grading the essay that sounds fluent
Automated essay scoring has been around for decades, and it has always favored prose that sounds like it knows what it is doing. Fluency is easy to score. Whether the student actually understood the novel is harder. A tool trained on thousands of past essays learns the surface features of good writing, and a student who has learned the same surface features can pass without engaging the text at all.
A teacher reading the same essay catches the tell immediately: a paragraph that summarizes the plot with no argument, a thesis that announces a theme without committing to a reading, a closing sentence that gestures at significance and says nothing. The teacher's judgment is not more accurate because teachers are smarter than software. It is more accurate because the teacher has read the book and can tell whether the student has.
The term sheet with two friendly sentences
Founders reviewing a term sheet with a summarizer get a bulleted list: valuation, option pool, liquidation preference, board composition. Each line is correct. What the summary cannot convey is the tone of the liquidation preference clause, which is written to sound standard and is not. A 1x participating preferred with a full ratchet is a different animal from a 1x non-participating preferred, and the difference lives in the adjectives.
Lawyers who review these documents for a living read for the sentences that do not belong: the extra clause, the unusual carve-out, the definition that quietly widens. A summarizer trained to extract key terms will extract the key terms and move on. The clause that does not belong does not appear on any list of key terms. It appears because a human being reads every word and stumbles on the one that is out of place.
Scanning the receipt before the return window closes
Expense tools now photograph receipts, extract the vendor and amount, and file them. Useful. The review that should not be automated is the one a person does at the counter, holding the receipt and the item together, checking the line items against what was actually bought. A receipt scanner will file a duplicate charge without complaint. A person notices the second coffee charged twice because the total feels wrong.
This is the smallest example and maybe the most instructive. The review is not really about the receipt. It is about the moment of contact between the document and the thing it describes. Every automated system severs that contact. The document becomes data, the thing becomes inventory, and the gap between them stops being anyone's job.
The code review that catches the missing test
Static analysis and AI code reviewers catch style violations, obvious bugs, and security patterns. They are genuinely good at this, better than tired humans at two in the morning. What they miss is the absence of a test for the case nobody thought to handle. A missing test is not a line of code. It is a hole, and tools review lines.
A senior engineer reading the same diff notices the hole because the diff does not mention the edge case that bit the team last quarter. That is memory, not analysis. The reviewer holds the team's history in mind and checks the new code against it. No tool has access to the specific institutional memory of one team's prior failures, which is precisely the knowledge that makes the review valuable.
What these reviews share
Every example above turns on one of three things: a cross-reference across a long document, a physical or emotional detail that only a present person registers, or a piece of institutional memory that no training set contains. Automated tools compress, extract, and match. The reviews that matter here require holding the whole thing in mind at once, noticing what is absent, and knowing what happened last time.
The temptation to delegate is real, and it is not irrational. Reviewing a lease line by line takes an hour. Summarizing it takes four seconds. The hour is the cost of catching the clause that governs the deposit. The four seconds is the cost of not catching it, paid later, with interest.
The signature that requires a reading
There is a version of this that is not about distrust of software at all. It is about what a signature means. A person signs a lease, a term sheet, an offer letter, a return authorization. The signature is a claim that the signer read the document. Delegating the reading to a summarizer does not transfer the claim; it hollows it out. The signature stays. The reading behind it disappears.
So the practical rule is narrower than a general suspicion of automation. Use the summarizer for the documents nobody will ever hold a person to. Use it for the email chain, the meeting notes, the article. Do the reading yourself for the documents where the signature is the point, and do it slowly, on paper, with a pen, because the hand moves at the speed of attention and the eye does not.
The tenant in the kitchen, the one with the four-page lease, eventually signs. The question is whether the pen was moving across a document that had been read or across a summary that had been trusted. The difference does not show up in the signature. It shows up eleven months later, on the day the deposit does not come back.
