85% Complete and Materially Absent

You cannot review your way out of AI slop, and slop was never the failure that mattered.

By four in the afternoon I was reading the eighth document of the day, and I stopped and called a peer about it, because I could no longer tell whether the problem was the document or me.

Twelve pages. Nothing in it was wrong. It covered the right ground, named the right constraints, and reached a recommendation nobody could call incorrect. It also went nowhere. It accumulated, section by section, and never committed to a position hard enough for me to push against.

He’d read three like it that week. Then he said the part I’ve been chewing on since. There’s no way in. You can’t redirect a document. You can only comment on it.

I should say plainly that I have produced plenty of these. I have asked for them, written them, reviewed them closely, and skimmed them while telling myself I’d come back. Nothing that follows is an accusation I’m standing outside of.

Multiply that by every decision your organization made this week. Six people approve something significant, each assuming one of the other five checked the premise. None did. There was no premise to check. There was a draft, and drafts are not arguments. They are positions that have already won.

Aaron Horwath described the resulting mood in Noema last week, and he’s right about it. Reading it, I felt like he was describing me. He’s writing about why the work stopped feeling like anything. This article is narrower. It’s about the specific design that produces the feeling, and about the fix your organization has already chosen, which will not work.

Because the slop problem is now consensus. Everyone has read the same discourse, reached the same diagnosis, and converged on the same answer: better prompts, tighter review standards, a human in the loop. That answer addresses the wrong failure. The one that matters has been sitting in the research literature since 1991.

One disclosure before I use it. 85% complete is my number, not a measurement. Not the fraction of work finished — the fraction at which a document becomes unopposable. Below it, a draft invites help. Above it, it’s finished. At 85%, objecting looks like obstruction, and the missing fifteen contains everything that mattered.

And one boundary. This is about documents where a judgment is the deliverable: strategy, architecture, design, policy, the memo that decides something. Not code. Code has tests, and a test doesn’t care who wrote the thing or what it displaced. The failure below only happens where the artifact is the reasoning.

The fix everyone is reaching for

The workflow is standard: draft with AI, circulate, collect comments, revise, ship. Every step has names on it. It photographs beautifully in a process review.

Look at where the artifact enters. It arrives first, before the problem has been talked through and before anyone has committed to a position worth arguing with. Everything downstream is reaction, and reaction has a ceiling. You can improve margins. You cannot restructure a premise without appearing to attack whoever produced it, so nobody does.

So organizations scale up the review, and the review is now enormous. Adaptavist’s 2026 survey of 2,500 knowledge workers across five countries found 42% spending more time verifying AI output than they save by using it, and 55% saying it is reducing overall team efficiency. ISC2 found 63% of cyber professionals saying validation work is growing their workload. Every hour of it is booked as diligence.

Now separate two things your organization currently treats as one.

The first is slop: output that is wrong. It has errors in it. It is visible, it is expensive, and review does eventually catch it — badly and at enormous cost, but it catches it. This is the failure everyone is fighting, and the fight is at least aimed at something real.

The second failure produces no errors whatsoever. The twelve-page document I called my colleague about was accurate. The reasoning held. No reviewer would ever have found a problem in it, because there was no problem in it. The damage was done to the thinking that would have happened if the document hadn’t arrived first. You cannot inspect for an idea nobody had.

Even on the first failure, review performs worse than you think. The International AI Safety Report 2026 states the established finding: automation bias undermines competence by discouraging active reasoning and verification, and reviewers both miss what the system failed to flag and act on advice they should have rejected. A 2025 systematic review in AI & Society covering 35 peer-reviewed studies across human factors, cognitive psychology and HCI lands in the same place. Humans reviewing automated output under-detect its errors. That’s not a training problem; it’s the documented behavior of the review step.

On the second failure, review isn’t weak. It’s blind, and it has been measured that way since before most of your engineers were born. Jansson and Smith, Design Studies, 1991: show people an example solution alongside a design problem and their subsequent designs reproduce its features — even when the example’s flaws are pointed out to them explicitly.They called it design fixation. Read that with your review process in mind. Your quality gate is staffed by people who have already been captured by the thing they are inspecting, and naming what’s wrong with it does not release them. Thirty-five years later, we made the fixating artifact free, instant and mandatory, and the entire remediation strategy consists of inspecting it harder.

The phase that disappeared was never on anyone’s process diagram. Two people talking badly about an unfinished problem, arriving somewhere neither started, produced no artifact and no timestamp. A survey of 32,352 workers across 47 countries found half of employees now use AI instead of collaborating with colleagues. That phase was invisible, so when a tool made it skippable it was skipped, and no dashboard registered the loss.

Three things to take back

1. Ban the document until the disagreement has happened. This is a peer step, not a governance step. Before a decision goes anywhere near a leadership forum, the people who will build the thing get thirty minutes on the problem with no deck, no draft, no AI in the room. The output is a shared understanding of what you’re actually solving, not a straw man to react to. Fixation is a finding about exposure: what constrains the thinking is having seen a solution first. The only intervention that follows is thinking before exposure, which is why this has to happen at the peer level and early — the moment it becomes a leadership ritual it acquires an agenda, and the candor it depends on is gone.

2. Require the record of what was rejected, and make it attributable. Every document above the threshold carries a field: options considered, why each was rejected, and who argued for it. Block review on an empty one. Writing that section is the only step in the workflow that cannot be completed unless a disagreement happened, so the requirement doesn’t detect the missing argument. It forces one. Attribution is what stops the field being generated alongside the document it appears in — three plausible alternatives, dismissed in a sentence each, laundering the absence of argument into evidence of one. This one doesn’t work without the first. The names have to come from somewhere real.

Make the name a credit. The person recorded against a rejected option isn’t the person who was wrong; they’re the reason the chosen path had to beat something. Then settle the bets quarterly: pull the rejected options from the last two quarters and check which turned out to be right. Say once, in writing, that this record is never an input to performance review, and mean it — the moment it reaches a calibration session, people stop filing real alternatives and you have bought yourself an expensive new form of paperwork.

Premortems work for this reason. Gary Klein’s method, built on Mitchell, Russo and Pennington’s 1989 finding that imagining an outcome has already happened improves the ability to identify its causes by around 30%, does one structural thing: it forces the paths not taken into the room before commitment hardens. It works because it legitimizes doubt. In a planning meeting, the person raising the ugly objection looks disloyal. In a premortem, naming it is the assignment.

3. Build in latency where it matters. Judgment needs a gap between question and answer. Name the small set of decisions that get a mandatory overnight, written but not circulated until someone has slept on it. Then defend the delay in public the week it costs you something. Teresa Amabile’s decades of research at Harvard Business School found that people produce their most creative work when driven by the interest and challenge of the work itself, and that the environments which kill creativity are political, risk-averse and relentlessly outcome-focused. Speed is not free. It is purchased from the same account.

The trade

The first recommendation gives back the argument and the third gives back the thinking time, and both are parts of the work your best people have been quietly missing. The second asks for something harder: a record kept honestly and read back four times a year, which is the kind of discipline that gets adopted in January and abandoned by April. It survives here because it produces a number. Every quarter you learn how often the rejected paths turned out to be the right ones. Judgment stops being the invisible thing you’re asked to protect on faith and becomes something you can put next to velocity on a slide. You are not being asked to be brave about a worse number. You are being handed the only evidence that would tell you whether your organization’s decisions are any good, which is a thing almost nobody knows about their own company.

You will not hear about any of this from below. The people who miss the argument are still there and have gone quiet, because raising it now reads as resistance to AI, and that is a career position nobody wants on their record. Ask someone to name the last decision they disagreed with and didn’t say so.

My colleague was right, and he understated it. There’s no way into the document. There was also no way into the decision, and by the time either of us read anything, there hadn’t been for days.

Your teams didn’t stop arguing because they stopped caring. They stopped because you made arguing optional, and optional is how a thing disappears with nobody choosing to end it.