Your legal team blocks an AI feature. When should you accept that and when should you push back?

Corvidale Screening runs background checks for employers who hire fast and in bulk: retail chains, staffing agencies, home care providers. ScreenPulse is the model that reads every report and scores how much attention it needs. Delwyn Bircher owns it. Corvidale's general counsel, Callixtus Dalziel, blocked two different features inside one year. One block came with a rule attached. The other came with nothing. Delwyn treated them exactly the same way, and only one of those calls was right.

How to answer this, stage by stage

Nobody in the room is grading whether you sound respectful toward legal. They're grading whether you can name the one question that tells you which kind of no you're actually holding.

Let's learn

What does a lawyer's no actually tell you?

Corvidale Screening runs background checks for employers who hire fast and in bulk: retail chains, staffing agencies, home care providers. ScreenPulse reads every report that comes back and gives it a score, so a recruiter knows which ones need a close look and which ones don't. Recruiters using that score clear a clean report in about six minutes instead of twenty-two.

Both blocks stopped a feature cold. Only one of them was actually standing on something.

Delwyn Bircher owns ScreenPulse. In one year, Corvidale's general counsel, Callixtus Dalziel, blocked two different features built on top of it.

The first was Auto-Clear: let ScreenPulse clear a report on its own, no recruiter needed, once the score came in low enough. Part of that score leaned on arrest-only flags, an arrest with no conviction attached. Callixtus blocked it and named his reason: the EEOC's own 2012 guidance on arrest and conviction records, because arrest data tracks who gets stopped, not who did something wrong, and using it that way can treat one group differently without anyone deciding to on purpose. Four straight quarters of Corvidale's own data backed him up.

The second was Verify Assist, months later: a feature that just turns four separate verification documents into one plain paragraph for a recruiter to read. No score, no decision, nothing automated. Callixtus blocked that too. His reason: "I'm not comfortable with AI touching hiring records." No rule, no case, no scenario. Just that.

That's the actual mistake, not which lawyer was right which time. Auto-Clear's block deserved full respect the first time it landed, and it took a client's audit to teach Delwyn that the hard way. Verify Assist's block deserved a direct question, and it took a new hire noticing the gap to get one asked at all.

What I would leave alone: Corvidale's legal team also reviews the wording on the candidate consent form every quarter, standard language, no model behind it, nothing about how a score gets made. Delwyn never questions that review. It's routine, it's cheap, and getting it wrong costs almost nothing next to a scoring decision.

The lesson: a block that names a real, specific risk is telling you something true about the model, not being difficult. A block that never gets more specific than a feeling is legal doing what any team does under pressure, reaching for the safest-sounding no. They look identical from the outside. The only way to tell them apart is to ask, every single time, not just on the weeks you've got the energy for the argument.

Now here is the same thing as a story

Use this version when you've got room to let it land. A spreadsheet believes a percentage. A person believes a Tuesday.

Delwyn Bircher has spent four years building screening products at Corvidale, and she can read a flagged background report faster than most of the recruiters ScreenPulse was built for. Give her a report with six line items and she'll tell you which one actually matters before you finish your coffee.

For most of that time, whenever Callixtus blocked something, Delwyn had one habit. She asked him, directly, what rule or case the block was actually standing on. It worked. Twice that first year, the answer changed her mind completely, and both times the feature shipped better for it.

Then came Auto-Clear.

Corvidale had pitched Larkhaven Eldercare, a home care staffing agency that screens something like a hundred and fifty caregivers a week, on a version of ScreenPulse that could clear a clean report with nobody looking at it. Callixtus blocked it before it reached a single client, and this time his answer had a number in it: the EEOC's 2012 guidance on arrest and conviction records, and four quarters of Corvidale's own data behind it. Black applicants were showing up flagged on arrest alone, no conviction, at more than three times the rate of White applicants in the same zip codes.

Delwyn had heard Callixtus reach for "I want to be careful here" twice that same quarter, on features that turned out fine after one conversation. She read this one the same way. She pushed, her VP backed her, and a limited version of Auto-Clear went live for Larkhaven anyway, six weeks ahead of the redesign Callixtus had actually asked for.

It ran clean for six weeks. Then a state labor auditor pulled Larkhaven's hiring file for an unrelated reason, and noticed the same pattern Callixtus had already named: Black applicants cleared out at a lower rate, arrest-only, no conviction, than everyone else. Larkhaven's operations director called Corvidale directly, furious, and gave them ten days to explain it in writing or lose the contract.

The habit that had worked for a year broke on the same day the one block that mattered finally got tested for real.

Three people spent the better part of three weeks on the response: pulling every flagged report, rebuilding Auto-Clear around convictions only, writing the audit letter twice because the first draft undersold how serious it was. A hundred and twenty hours, and Larkhaven stayed, barely.

After that, Delwyn stopped asking. Any time Callixtus said no, she accepted it, full stop, and moved the feature off the roadmap without another word. It felt like the responsible thing. For a while, nobody noticed the difference, because nothing else got blocked.

Then, four months later, Callixtus blocked Verify Assist: a feature that reformats four separate verification documents, a state registry check, an elder-abuse registry check, an education check, a license check, into one clean paragraph. It doesn't score anything. It doesn't decide anything. It just saves a recruiter from reading four PDFs by hand. His reason, in the block memo: "I'm not comfortable with AI touching hiring records." Delwyn read it, nodded, and moved on.

Thirteen weeks went by. At Larkhaven alone, recruiters kept hand-compiling those four documents themselves, about twelve extra minutes a candidate, a hundred and fifty candidates a week. Nobody was tracking it as a cost. It just looked like the job.

Then a new PM joined Delwyn's team and sat in on a roadmap review. She pointed at Verify Assist, still parked in the blocked column, and asked the only real question in the room: "Did legal ever say why, on this one?"

Delwyn didn't have an answer. She had never asked.

Here's the replay. Same block, same memo. This time Delwyn goes back to Callixtus and asks him directly: what's the specific rule or case behind this. He tells her, honestly, that there isn't one. It's discomfort, the general kind, not the EEOC kind. So they build a lightweight version together, with a one-line disclaimer Callixtus writes himself, and it ships within three weeks. Larkhaven's recruiters get their twelve minutes back on every one of a hundred and fifty reports a week, starting the same month.

Same lawyer, same team, two blocks four months apart. One had a citation, and Delwyn should have trusted it the first time instead of needing an audit to believe it. The other had nothing, and she should have asked twice instead of needing a new hire to notice.

What I'd tell myself, back in the meeting where Auto-Clear's block first landed: the block that has a number in it and the block that doesn't sound exactly the same read out loud. You have to ask which one you're holding. I stopped asking right when it mattered most, and it cost thirteen weeks before anyone caught it.

PICK, and what fifteen minutes buys you

Not permission to fight every no, and not a reason to accept every one either. PICK only earns its place here if it turns a block into a question instead of a coin flip.

One path costs a conversation. The other costs weeks, no matter which direction the guess goes.

Four branches, one root question: did anyone actually ask, or did the deadline decide on its own.

The trade worth saying out loud: asking costs a short, visible pause, about fifteen minutes, right when a deadline makes any pause feel like the wrong move. That pause is worth paying every time, because the alternative, guessing which kind of no you're facing, only shows its real cost after the hours are already spent: a client nearly lost on one side, thirteen weeks of quiet waste on the other.

And if you want to be sure it really works, try it somewhere else

Same four letters, a claims adjuster's queue instead of a hiring pipeline, and the AI-specific risk moves from arrest records to a fraud score.

One feature genuinely needed the citation it got. The other needed someone to notice it never got one.

Windward Mutual uses ClaimGuard to score how likely a new insurance claim is to be fraudulent, so adjusters know which files need real scrutiny. Someone proposes Auto-Deny: skip adjuster review entirely and deny any claim ClaimGuard scores above a fraud-risk cutoff. Legal blocks it, and names the reason: the state's Unfair Claims Settlement Practices Act, which requires a documented, individual review before any denial, a real statute with real penalties attached. Position: accept it, and redesign so every denial, however high the score, gets a human sign-off with the score attached as one input, not the whole decision. Impact: skip that redesign, and Windward risks real regulatory penalties and bad-faith lawsuits, the same shape of risk Corvidale faced with arrest-only flags, a model quietly making a call the law says a person has to make. Cost asymmetry: building the sign-off step in from day one costs a few extra seconds per claim for the adjuster. Shipping Auto-Deny blind first and retrofitting the sign-off after a regulator notices costs a full compliance review and a real fine. Kill criteria: does legal name the actual statute, section number included? Here, yes, so the feature gets built, just with a person still holding the pen. Separately, someone proposes Claim Summarizer, condensing a claim's photos, repair estimate, and police report into one paragraph, no scoring, no decision. Legal blocks that one too, on "we're not ready for AI anywhere near claims," no statute, no case. Same test, same answer: push back, ask which specific rule this breaks, and when none arrives, ship the summarizer with a plain disclaimer instead of leaving it parked for a quarter.

Swap the trigger and it still runs.
Speed: an interviewer caps you at ninety seconds. Skip straight to the test: does legal name something specific, or not, before anything else gets said.
Cost: no time for even the fifteen-minute question before a decision is due. Fine, but say the test out loud and make whoever owns the block answer it, instead of letting the deadline decide by default.
The model got better, for real: say ClaimGuard's fraud score later gets independently audited and clears the four-fifths line with real margin. That's the moment a stricter feature like Auto-Deny earns another look, not before it.

Where people run it wrong.
They let a block that sounds serious stand unquestioned, because arguing with a lawyer feels riskier than losing the feature.
They treat "no citation yet" as "no citation ever," and never go back to check once real evidence would exist.
They accept or fight a block based on which lawyer sent it, not on what the block actually says.

How to use it live. If you're ever asked whether a legal block should stand, buy yourself a second with one plain question, said out loud: "did they name the actual rule, or are we just being careful?" That question is the whole method, asked instead of stated.

Three steps, and the whole method lives in whether step two ever actually gets asked.

Flashcards (tap any card to flip it)

Check yourself Score: 0 / 0