The short answer: AI can own posting and formatting in a search, and assist on sourcing lists and scheduling. It should never decide who advances, or write why a named candidate didn’t. Those two steps are opinions about people. With 33.9% of finance and insurance firms already using AI somewhere, that line belongs on paper.
Written for the HR lead or chief of staff who runs hiring at a financial services firm of 50 to 500 people. The line works the same way at a family office, a non-profit or a professional services firm.
A third of US finance and insurance businesses, 33.9%, reported using AI in a business function in spring 2026, per the Census Bureau’s Business Trends and Outlook Survey. About 39% expected to be using it within six months. Nationally, the current rate across all US businesses was 19.8%. Hiring is where that adoption meets a firm’s most sensitive material: the names of people who are quietly in a search. Names are where the risk starts. A stored opinion is where it lands.
In the searches we run for firms this size, one person usually owns the whole hiring process, and much of what they hold is quiet. A role nobody has announced. A shortlist of people still employed somewhere else. North Bridge anonymizes by default for the same reason: an example stops being safe the moment it attaches to a real person.
This guide covers why an opinion about a named candidate changes the rules, how to sort each step of a search, and where a tool should own, assist or stay out. It then covers when to redraw that line and three signals that show it’s holding. The method draws on the Precision Screen, the separate skills and fit screen North Bridge runs on every placement.
Why does an opinion about a named candidate change the rules?
An opinion about a named candidate changes the rules because a stored opinion stops being private. Once a tool holds it, it’s a record someone else keeps. A tool that holds that opinion makes a record the firm no longer controls. SHRM’s 2026 research lists recruiting first among HR uses of AI, at 27%.
A search produces two kinds of output. The first is artifacts, such as the posting, the calendar invites, the deduplicated list and the tidy summary of a resume, and none of those carries a judgment about anyone, though some carry names. The second is opinions about named people, like the note that says this candidate isn’t right and explains why. That note is the most sensitive thing a search creates.
The problem is that tools blur the two. A meeting summary looks like an artifact. If it records that the hiring manager doubted a candidate’s judgment, it’s an opinion with a name on it, stored wherever the tool stores things. In SHRM’s 2026 State of AI in HR research, 62% of organizations were using AI somewhere in the business, but only 39% had implemented it in HR. SHRM reads that gap as HR lagging the rest of the business. Hiring is where rules are least settled.
For a firm whose clients value discretion, the cost of getting this wrong isn’t a bad hire but a stored opinion that outlives the search. We covered the wider shifts in our earlier piece on AI in modern hiring; this guide is about where the line sits.
How do you sort a search into what a tool can own?
Sort a search by what each step produces, not how long it takes. A tool can own a step with no name on it and assist one that carries names. An opinion about a named person stays with a person. Three habits do the sorting: label each output, keep two steps human, check what tools keep.
1. Label each step by its output, including the summaries
The detail most firms miss is the summary. An interview summary, a call recap or a ranked shortlist with comments reads like admin, so it gets automated along with the scheduling and nobody reads it back. If it contains a view about a named person, it belongs in the opinion column, whoever or whatever wrote it.
2. Keep two steps human: who advances, and why
North Bridge runs the Precision Screen on every placement, a formal two-step screen that evaluates skills fit and cultural fit separately. The split is a useful model here: skills fit rests on evidence, while cultural fit is a judgment about a named person, as is the decision to advance or decline. Keep those judgments, and the written reason, with a person. Structured reference checks sit on the same side of the line.
3. Check what each tool keeps after the search closes
Ask every vendor three questions in writing: what the tool stores, where, and for how long. A quiet search ends when the hire starts, but the records it created can sit in a vendor’s system for years. For firms FINRA supervises, its 2026 oversight report asks whether cybersecurity programs cover vendors’ use of generative AI.
List every tool that saw a candidate’s name in your last search. A clean list is short, and you can say what each one kept. If you can’t, the confidential part of that search was never yours to promise. Stuck on one? Compare notes with us
The line on one page: own, assist or stay out
The governing rule is simple. A tool owns a step when its output carries no name and no opinion. It assists when the output carries a name but no opinion, and stays out when it carries both. SHRM found 49% of organizations have policies regulating employees’ AI use. That figure doesn’t tell you what a hiring tool may store.
The table below sorts a typical search into the three columns.
| Step in a search | What it produces | The tool’s role |
| Writing and posting the role | An artifact, no name attached | Own |
| Scheduling interviews and sending invites | An artifact that carries names | Assist, with a person sending the final invite |
| Building and deduplicating a sourcing list | Names, no opinion yet | Assist, with a person choosing who is contacted |
| Summarising an interview or call | Often an opinion about a named person | Stay out, or a person edits before anything is stored |
| Deciding who advances | An opinion about a named person | Stay out |
| Writing why a candidate was declined | An opinion about a named person | Stay out |
The columns work because they follow the record, not the effort. Scheduling takes the most time and carries little risk, while a decline note takes two minutes and carries the most.
The middle rows matter most, because a sourcing list holds the names of people who may not have told their employer they’re looking. A tool can help build it, but a person decides who gets the call.
The offensive play is to write this table before the next search opens and send it to every vendor you use, so their answers tell you which tools fit a quiet search. If speed is the worry, why vetting outranks speed makes the case that a careful search rarely costs what people fear.
If a tool ranked or screened candidates for a New York City role in your last search, find its bias audit. It should be under a year old. Without one, you carry exposure forward. Want a second view of your last search? Let’s talk it through
Three moments to redraw the line
Redraw the line when the search, the market or the tool changes, not on a calendar. Three moments justify it: a search that has to stay quiet, a hire in a new jurisdiction, and a vendor adding a feature. SHRM found 57% of HR professionals in states with AI employment rules weren’t aware of them.
- A search that has to stay quiet. An unannounced role, or a replacement while the incumbent is still in the seat, moves every step one column to the right. Keep it off tools that store candidate names unless you can say where those names go.
- A hire in a new market. Chicago, New York and London each come with their own expectations. New York City’s Local Law 144 requires a bias audit within one year of use and 10 business days’ notice to candidates. Illinois’ Department of Human Rights published proposed rules on May 15, 2026 that would require employers to tell employees when AI is used in employment decisions. In London, the Information Commissioner’s Office expects employers to tell candidates when automated decisions are used and how to request a human review.
- A vendor adds a feature. A scheduling tool that starts writing interview summaries has moved into the opinion column. Recheck it on the day the feature ships, not at renewal.
Automate the timing, never the human moment. A reminder to recheck tools at the start of each new search costs nothing, while the decision about a named person stays a conversation between people.
Weak enforcement isn’t a reason to relax, because the rules around it are still moving. New York State’s Comptroller audited enforcement of Local Law 144 and found the city surveyed 32 companies and identified one issue of non-compliance. Auditors reviewing the same companies found at least 17 instances of potential non-compliance, and the city received only two complaints in two years. Illinois and the UK both moved in 2026.
Three signals that the line is holding
Three signals show whether the line holds. Track the tools that saw a candidate’s name in each search, the share of decisions with a reason a person wrote, and the days from notice to first tool use. Aim for a list you can name in full, a reason on every decision, and notice that meets each market’s rule.
Read them together, because each failure points somewhere different. If the tool count stays short but written reasons slip, the risk is a habit forming, so restate who writes the decision. If reasons hold but the tool count climbs, features are drifting into the opinion column, so recheck vendors. If both slip, the line was never written down.
Policy alone won’t carry this, because knowing a rule and acting on it are different things. Among HR professionals who knew their state’s AI rules, SHRM found only 12% had taken steps to comply. The same research counted 19 of the most populous states with AI laws or regulations relating to employment. Awareness is not practice.
Set the baseline before the next search, not after a complaint, by writing down which tools touched your last three searches and what each one kept. Firms that keep finding the same gaps can read our process warning signs.
Run the summary test: open the last three interview summaries a tool wrote and find where each lives. All three should sit in systems your firm controls. If one doesn’t, a declined finalist’s file has left the building. Planning a quiet search? Ask how we’d search
Frequently asked questions
A hiring tool belongs wherever no name meets an opinion
Draw the line by what a step produces. With 33.9% of finance and insurance firms using AI and 57% of HR professionals in regulated states unaware of the rules, the risk sits in the stored opinion. Firms that get this right keep tools on the artifacts. Decisions stay with a person.
The tools, the vendors and the rules for your hiring desk are yours. Running confidential searches with the Precision Screen on every placement is where North Bridge fits in. We do that through direct-hire and contract searches built for quiet roles. If your next search can’t be announced, talk to North Bridge before any tool sees a name.
References
- “Large Firms With at Least 20 Employees Biggest AI Users.” US Census Bureau, Business Trends and Outlook Survey. Published May 26, 2026. Verified live October 13, 2026.
- “The State of AI in HR in 2026: 5 Critical Insights for CHROs.” SHRM (reporting SHRM’s State of AI in HR 2026 research). Published April 3, 2026. Verified live October 13, 2026.
- “GenAI: 2026 FINRA Annual Regulatory Oversight Report.” FINRA. Published December 10, 2025. Verified live October 13, 2026.
- “Automated Employment Decision Tools (AEDT).” NYC Department of Consumer and Worker Protection. Published undated. Verified live October 13, 2026.
- “Artificial Intelligence in Employment.” Illinois Department of Human Rights. Published May 19, 2026. Verified live October 13, 2026.
- “Automated decisions can streamline the hiring process with the right safeguards in place.” Information Commissioner’s Office. Published March 31, 2026. Verified live October 13, 2026.
- “Enforcement of Local Law 144: Automated Employment Decision Tools.” Office of the New York State Comptroller. Published December 2, 2025. Verified live October 13, 2026.
