Digital Appeals & Records is intended as a research and drafting aid for hearing officers. The proposed workflow organizes supporting records while leaving decisions with the authorized person.
This was not the easy product decision. The hearings backlog is real, the political pressure to clear it is real, and the temptation — explicitly voiced by some of our early prospects — was to push AI further into the decision. We didn't. This post explains why.
What the temptation looks like
Administrative hearings — unemployment appeals, Medicaid eligibility appeals, child support determinations, license revocation hearings — produce backlogs measured in months. Hearing officers are scarce. The work is bounded by precedent, regulation, and a defined record. From a distance, it looks like exactly the kind of work AI should automate.
Vendors get asked, repeatedly, whether they can:
- Predict how a case will likely rule based on similar past cases
- Recommend a ruling for the hearing officer to "approve"
- Draft a decision the officer can edit and sign
- Auto-rule cases that are "clear" and route only ambiguous cases to humans
Research assistance and automated decision-making have different risks and governance requirements. This article makes no delivery-date commitment for either capability and does not propose automated rulings.
None of them are products we will build.
Why we drew the line where we did
The intended boundary keeps case decisions with the hearing officer. Research, citation checking, record organization and drafting assistance should be evaluated against the agency's procedures.
Three reasons for the line, in increasing order of importance.
The legal reason
Administrative hearings exist because due process requires a human decision-maker to weigh evidence and apply law to facts. The structure of administrative law assumes a decision-maker who can be cross-examined on appeal, who has discretion within statutory bounds, and whose reasoning can be reviewed. An AI system that recommends an outcome — even one the officer could override — substitutes a different kind of decision-making. The reviewing court can't cross-examine the model. The model's reasoning isn't visible. The discretion shifts somewhere else.
Legal scholarship on this is consistent: AI as decision-support is acceptable; AI as decision-maker raises constitutional and statutory concerns that have not been resolved. We chose to build inside the consensus.
The institutional reason
Hearing officers are professionals. They have training, accountability, and a body of decisions they own. Their job is not a bottleneck to be removed — their job is the substance of due process. A product that treats them as a bottleneck is a product that misunderstands the work.
The design goal is to support hearing officers with research and drafting tools while preserving their decision authority. Adoption and usability require evaluation with the people who would use the tools.
The behavioral reason
Anchoring is the most-studied bias in decision-making research, and it's the one we worry about most in AI deployment. If the system shows a hearing officer a "predicted outcome" or "AI-suggested ruling," that prediction will anchor the officer's reasoning even if the officer believes they are independently evaluating the case. The behavioral literature is unambiguous on this. It is also unambiguous that the anchoring effect persists when the user is told to ignore the suggestion, when the user is an expert, and when the user is a hearing officer specifically.
Once you put a predicted outcome on the screen, the officer's independent judgment is no longer fully independent. We will not put a predicted outcome on the screen.
The right question is not "can AI rule on a hearing." The right question is: "should we build a system that anchors the hearing officer toward a specific ruling?" We answered no.
What "mentor" actually means in practice
The line is not abstract. It shows up in specific product decisions:
- Digital Appeals & Records retrieves relevant precedent — but does not score precedent by "favorability." The officer sees the precedent organized by topic, jurisdiction, and date. The officer evaluates relevance.
- Digital Appeals & Records drafts language — for findings of fact, conclusions of law, procedural sections. The drafts are templated against the officer's prior writing style and the case record. The drafts do not include a recommended outcome paragraph; the outcome paragraph is the officer's to write.
- Digital Appeals & Records organizes the record — exhibits, testimony, prior filings — into a navigable structure. It surfaces inconsistencies (a date in one exhibit conflicting with testimony) but does not weigh them.
- Digital Appeals & Records checks citations — verifies that cited statute and regulation are current, that quoted language matches source, that case citations are accurately formatted. This is mechanical work, and the AI is good at it.
- Digital Appeals & Records does not predict outcomes — there is no "this case is 73% likely to be reversed" feature, because there cannot be, and we will not build one.
The audit dimension
A deployed workflow should record AI suggestions, human edits and decisions with appropriate access controls. The completeness and usability of that audit record must be verified for the implementation.
A reviewable workflow should capture suggestions, acceptance or rejection, edits and final language with timestamps. Confirm logging coverage, retention and legal requirements for the implementation.
What we tell prospects who push for more
Some prospects ask us to add the predicted-outcome feature anyway. They argue the officer can ignore it, that hearings are routine, that the backlog is the real problem. We listen, and then we say no. We say it for the legal, institutional, and behavioral reasons above. We also say it because we believe the line we drew will be the right line in five years, when the case law on AI in administrative decisions is more settled, and we'd rather hold it now than retrench from it later.
If your agency is sizing AI for hearings work, we're happy to walk through what Digital Appeals & Records does and — equally importantly — what it doesn't. Schedule a conversation →.