Disability Management Runs on Disclosure. Your AI Runs on Data.

There is a moment I have watched more than once, and it never gets less uncomfortable.

An organisation puts automation into its absence or disability process. Six months later the numbers look better. Absence records are cleaner. Accommodation requests are down. Reported durations are shorter. Responses are more consistent. Somebody builds a slide.

And the program is in worse shape than it was before.

Not because the system made a bad decision. Often it has not made a single bad decision anyone can point to. The damage happened somewhere the system was never pointed, and nothing in the reporting can see it.

I want to explain that mechanism, because I do not think it is widely understood, and because it is the first thing I would look at in any workplace that has put automation near a disability file.

The input is not data. It is a person deciding to speak.

Every part of this process begins with disclosure.

Nothing happens in disability management until somebody tells you something they did not have to tell you. That they are struggling. What they actually cannot do, as opposed to what they have been managing to fake. What the medication does to their afternoons. That the accommodation on paper is not the accommodation they are getting. That they are not going to make it through another quarter like this one.

Every accommodation is built on top of that act. Every return-to-work plan. Every early intervention that actually intervenes early. The file does not exist until a person with good reason to be afraid decides to say something anyway.

That is what psychological safety buys you in this specific domain. Not comfort, and not morale. It is the supply line for the only information the process runs on. Take it away and you do not get a slower program. You get a program running on a fiction.

Anyone who has worked a caseload already knows the shape of this. The file that arrives at month nine when it could have been resolved at month one — and the reason it arrived at month nine is that the person spent eight of them deciding whether it was safe to say anything.

Automation changes the disclosure calculation before it changes any decision

Here is the part that gets missed.

The governance conversation about AI at work is almost entirely about outputs. Is the model biased. Is the decision lawful. Is there a human in the loop. Those are real questions and I spend a good deal of my time on them.

But every one of them describes what happens after information arrives.

An employee’s decision to disclose happens before. It is not a legal calculation or a technical one. It is a risk assessment, made quickly, mostly unconsciously, and it turns on a single question: what happens to me if I say this?

Now change one thing. The employee learns — or assumes, or hears in the lunchroom — that there is a system reading absence patterns. Scoring duration. Flagging files. Predicting who is likely to be off a long time.

They do not stop having the condition.

They stop telling you about it.

They take the day as a personal day instead of a sick day. They spread absences out so the pattern looks less like a pattern. They under-report on the form. They do not ask for the accommodation they need, because asking creates a record and the record goes into something they cannot see. They work out what a good answer looks like, and they give it.

That behaviour has a name in my field. Masking is the effort of appearing fine at work when you are not. It is exhausting, it is expensive, and people do it because it is very often the rational choice. (I am developing a measure for it. It is in development and not yet validated, and I will not pretend otherwise.)

What automation does is raise the price of honesty and lower the price of masking — without anyone deciding to do that, and usually without anyone noticing it has happened.

The trap: your data improves while the truth degrades

This is the part I would put on the slide, if I were making the slide.

When people work out what a system rewards, they produce what it rewards. So every metric the automation was bought to improve will improve.

Cleaner absence data, because the messy absences got recorded as something else. Fewer accommodation requests, because people stopped asking. Shorter reported durations, because people came back before they were ready and did not say so. More consistent responses, because everyone found the answer that does not raise a flag.

Your instrument looks like it is working at precisely the moment it has stopped measuring anything real.

And you cannot see it from inside the system, because the one signal that would tell you — a person volunteering an inconvenient truth — is the exact thing the deployment suppressed. You have built a mirror and pointed it at itself.

Economists have a version of this: a measure that becomes a target stops being a good measure. What I am describing is the disability-management case, and it is worse than the general one for a specific reason. In most domains the gamed measure is a proxy for the thing you care about. Here the suppressed disclosure is the thing you care about. It is not a proxy for early intervention. It is early intervention.

So “is the model biased?” is the wrong first question

It is not an unimportant question. It is just not the first one — and treating it as the first one is how a program passes every audit it commissions and still fails the people it exists for.

The audits currently on offer split in two, and this problem falls through the gap between them. The hiring audits stop at the point of hire; nothing follows a person into absence, accommodation or return to work. The enterprise governance reviews ask whether you have a management system in place. Neither asks whether the automation changed what your employees are willing to say to you.

Nobody asks it, in my experience, because it does not look like an AI question. It looks like a culture question. It is both — and the technical people and the accommodation people are in different rooms, so it lands on nobody’s desk.

This is the first link in a longer chain — what a system asks for, what it refuses to decide, and whether a person can check what it did. I set the whole chain out in The Truth Supply Chain.

What I would actually look at

Six things. None of them require you to understand a model.

One. Did disclosure move? Compare accommodation requests, early self-reports and voluntary disclosures for the twelve months before deployment against the twelve after. If they fell, that is not a success story and it should not be reported as one. It is the first thing I would want explained.

Two. What have you told people, in plain words, before they had to ask? Non-disclosure is a rational response to uncertainty, and specificity is the antidote. An employee who knows exactly what the system reads, what it does with it, who sees the output and how to challenge it, is deciding under known conditions. An employee who knows only that “there’s some AI thing now” will assume the worst — and will usually be assuming something worse than the truth.

Three. Which outputs decide, and which only inform — and is that written down anywhere? Not what the vendor says. What is written, where, and whether the people using it every day would give you the same answer.

Four. What does it do when information is missing? Say so, or score anyway? A system that produces a confident number from an incomplete file is not helping you. It is manufacturing a problem you will hear about eighteen months from now.

Five. Can the employee actually use it? Open the employee-facing screen — not the dashboard, the screen your worker sees. The clearest example is the smallest one: a leave request form that times out, with no way to save and come back, in front of somebody with a fatigue or cognitive condition. Nobody designed that. Nobody tested it. And it is doing real work against exactly the people the process exists for.

Six. If somebody asks how a decision about them was made, can anyone reconstruct it? Not the outcome — the reasoning. Eighteen months later. In front of somebody who was not in the room.

Whether any of this is legally required where you operate is a question for your counsel, and I put those questions to counsel rather than guessing at them. What I am describing is not a compliance floor. It is whether the program still works.

The line that matters most

I have built one of these systems, so let me be concrete about where I put the boundary in my own.

It runs on 49,178 Canadian tribunal decisions in a Canadian-region database. It checks its own citations — format, then existence against the source table, then whether the passage actually says what the answer claims — and a change that drops that accuracy is blocked before it ships. It states its own coverage gaps and declines to answer rather than guessing.

And it never decides anything. The deterministic part produces the findings. The language model explains them and drafts. That separation is written down, and it is the first thing I look for in anybody else’s.

I also built the business case that put AI into a disability operation. It was funded, it worked, and the pilot measured roughly a twenty percent reduction in administrative time. I am not against this technology. A great deal of the administrative weight in disability management is worth automating, and the people carrying it deserve to be freed from it.

What I am against is automating the paperwork and quietly automating away the reason a person was willing to open the file in the first place.

The verdict

Automate the file. Do not automate the conditions that made someone willing to tell you the truth.

Because the system cannot tell you when that has happened. The metrics will say you are fine. They will say you are better than fine.

And the first sign you get that they were wrong will be a file that arrives at month nine, or a complaint you cannot reconstruct a decision for, or a grievance, or somebody’s leave that did not have to happen.

By then the question is not whether the model was fair.

It is why nobody told you.