AI Guard Scoring · Compliance
AI Scoring of Security Guards: What US State AI Laws Already Require From Your Patrol Data
The Argument in Washington Is Not Your Problem
For most of 2026, the national conversation about artificial intelligence has been about frontier models — what they might be able to do, who should be allowed to build them, and whether anyone can slow them down. It is a real debate. It is also almost entirely irrelevant to a security contractor running four hundred officers across six states.
There is still no federal AI act. What exists instead is a patchwork of state law, and three of those states have already written rules that reach something very ordinary: how you evaluate the people who work for you.
If you calculate a completion percentage from patrol data and that number influences who gets the good post, who gets written up, or who gets moved to supervisor, you are inside the scope of at least one of these regimes today. Most security companies have not noticed, because they do not think of themselves as AI users. That belief is the exposure.
What Actually Counts as an "Automated Decision System"
California's employment regulations on automated decision systems took effect on October 1, 2025. They define an ADS as a computational process that makes a decision, or facilitates human decision-making, regarding an employment benefit. The definition is deliberately broad: the process may be derived from artificial intelligence, machine learning, algorithms, statistics, or other data processing techniques.
Read that list again. Statistics. A spreadsheet that ranks officers by completion rate is a computational process built on statistics, and if a manager uses that ranking to decide who gets disciplined, it facilitates a decision about an employment benefit.
The regulations do carve out ordinary software — word processors, spreadsheets, IT tools, navigation, spellcheck — but only on the condition that the tool is not making a decision regarding an employment benefit. The carve-out is about use, not about product category. The same spreadsheet is outside the rules when it tracks fuel spend and inside them when it ranks guards for a promotion.
This is the single most misunderstood point in the field. Vendors advertise "AI-powered" to sell, and buyers assume that the absence of that badge means the absence of obligation. The regulatory test runs the other way around: it starts with the decision, then looks back at what produced it.
The Three Regimes That Already Reach Guard Operations
| State | Status | Who is covered | Key duties | Records | Enforcement |
|---|---|---|---|---|---|
| California | In force since October 1, 2025 | Employers with five or more employees | No discriminatory outcomes from an ADS; liability extends to an employer's agents, including vendors | Four years | Private right of action under FEHA |
| Illinois | In force since January 1, 2026 | Employers with one or more employees in Illinois | Notice when AI is used in covered employment decisions, including discipline, discharge, and promotion | Not specified in the statute | Complaint under the Illinois Human Rights Act |
| Colorado | Effective January 1, 2027 | Employees and job applicants | Pre-use notice, post-adverse-outcome disclosure, meaningful human review by a person who can override | At least three years | Attorney general only |
California — in force since October 1, 2025. Applies to employers with five or more employees. Discriminatory outcomes from an ADS are unlawful whether or not anyone intended them. Records connected to the system — selection criteria, relevant outputs, audit findings — must be preserved for four years. Two features make this the sharpest of the three: liability extends to an employer's "agent," which reaches third-party vendors acting in a traditionally employer function, and the Fair Employment and Housing Act carries a private right of action. An aggrieved officer does not need a regulator to act.
Illinois — in force since January 1, 2026. House Bill 3773, enacted as Public Act 103-0804, amended the Illinois Human Rights Act at 775 ILCS 5/2-102. It applies to employers with one or more employees in Illinois. Covered decisions are not limited to hiring: recruitment, promotion, renewal of employment, selection for training or apprenticeship, discharge, discipline, tenure, and the terms and conditions of employment are all in scope. Employers must give notice when AI is used for those purposes, and using ZIP code as a proxy for a protected class is itself a civil rights violation. Liability attaches to effect, not intent.
Colorado — effective January 1, 2027. Senate Bill 26-189 replaced the state's earlier AI act with a narrower but sharper framework built around "consequential decisions," defined as decisions relating to an individual's access to, eligibility for, or compensation related to employment, among six other domains. Employees and job applicants are explicitly covered. Deployers owe a pre-use notice and a post-adverse-outcome disclosure. Individuals can request meaningful human review and reconsideration, and the reviewer must be a person with authority to approve, modify, or override the decision, trained to conduct the review, and must not simply defer to the system's output. Records demonstrating compliance must be kept for at least three years. Developers owe deployers technical documentation covering intended uses, categories of training data, known limitations, and instructions for human review. Enforcement sits with the attorney general alone; there is no private right of action. The attorney general filed proposed implementing rules on August 11, 2026 and released a revised interim draft on October 6, 2026, with a public hearing scheduled for October 26, 2026.
Where This Touches an Actual Guard Operation
Five situations, all routine:
- Shift and post allocation by score. The officer with the best completion rate gets the daytime corporate post; the officer with the worst gets the 11pm industrial route. That is a term and condition of employment.
- Discipline after a missed round. A write-up generated because the patrol record shows a gap. Illinois names discipline explicitly.
- Pay banding. Any performance tier that moves an hourly rate touches compensation, which is the core of Colorado's consequential-decision test.
- Promotion to supervisor. Named in all three regimes.
- Client-requested removal. A client reviews a monthly report, sees one officer's numbers, and asks for a replacement. The decision looks like a client decision. The employment consequence is yours.
None of these require an AI feature. They require a number and a manager who acts on it.
The Notice Problem Nobody Has Solved Yet
Illinois requires notice. What a compliant notice looks like is another matter. The Illinois Department of Human Rights published proposed implementing rules on May 15, 2026, then temporarily withdrew them in June and canceled the scheduled public hearing. The statutory duty did not go away with the rules. Employers are obliged to give notice under a standard that has not been finalized.
Waiting for certainty is the wrong move here, because the obligation is already live and the enforcement mechanism — a complaint under the Human Rights Act — does not depend on rulemaking. A defensible interim notice describes, in plain language, which tools are used, which employment decisions they inform, what data they draw on, and how an officer can ask for human review. That is not a legal opinion; it is the shape every one of the three regimes converges on.
Why an Explainable Score Is Easier to Defend Than a Smart One
Every one of these frameworks asks for essentially the same three things: tell people the tool is in use, keep the records that show what the decision rested on, and make sure a human with real authority made the call.
A black-box model fails all three quietly. You cannot notice what you cannot describe, you cannot retain an explanation you never had, and a supervisor who cannot interrogate an output tends to defer to it — which is precisely what Colorado's rules say meaningful review is not.
A rule-based score fails none of them. If the number is "checkpoints verified divided by checkpoints assigned," a supervisor can open the underlying record, see which rounds were completed and when, see the incident that explains a deviation, and form an independent judgment. The number becomes evidence in a decision rather than a substitute for one.
This is worth stating plainly to anyone selling you the opposite: in an employment context, sophistication is a liability and legibility is an asset.
How Trinity Guard® Is Built, and Where the AI Actually Sits
We separate the two functions deliberately, and it is worth being precise about which is which.
The performance score is rule-based. Trinity Guard® scores each officer's performance as a percentage at the end of every shift, calculated from completed and missed patrol tasks and subtasks against the assigned patrol structure. There is no model inferring anything about the officer. A supervisor can reconstruct route history per shift and per site and see exactly what produced the figure. Incidents, photos, and vehicle logs attach to the same timeline, so a documented incident supplies the context for a deviation — and the supervisor, not the system, decides whether that deviation was operationally justified.
The AI validates scans, not people. Checkpoint QR scans are validated server-side in real time to detect faked or duplicated scans. That model makes a judgment about an image, not about an employee. When it flags something irregular, it surfaces the scan for inspection rather than issuing a verdict. The technology flags; the human judges.
Trinity Agent is a query tool, and we treat it as in scope. A manager can ask Trinity Agent which officers performed best over a period. Under California's definition, a tool that surfaces that ranking facilitates human decision-making about an employment benefit — so we do not claim it sits outside the rules. The honest position is not exemption, it is documentation. The Agent answers from the same verified patrol records the supervisor can open directly, its output is reproducible, and nothing it returns is an employment decision until a person makes one.
We designed it this way before any of these laws existed, for a different reason. A scoring system that punishes blindly is surveillance with a number attached, and officers see through it in a week. Fairness was an operational necessity first. It turned out to be the compliance posture as well.
If you want the operational side of this — which patrol metrics are worth tracking and which ones quietly make operations worse — that is covered in guard patrol analytics and route history reports, and the case for using the same data to recognize good work is in Beyond Surveillance.
A Seven-Point Checklist
- Inventory. List every tool that produces a number, ranking, or flag about an individual officer — including spreadsheets.
- Classify by use. For each one, write down which employment decisions it informs. Tools that inform none are outside scope; say so in writing.
- Notice. Issue a plain-language notice in every state where you employ, not only where the law is clearest.
- Human review. Name the reviewer for each decision type. Confirm they have authority to override, and that they are expected to look at the underlying record.
- Retention. Four years in California, three in Colorado. Set the longer one as your standard. Your patrol audit trail is where most of this already lives.
- Vendor documentation. Ask each vendor how the number is computed, what its known limitations are, and what they will provide for your disclosure obligations. Colorado will require developers to hand this over from 2027; ask now.
- Re-check annually. Other states are considering similar rules, Colorado's implementing rules are not final, and Illinois has not finalized its notice standard.
If you are setting up patrol structures and reporting from scratch, the mechanics are in the setup guides, and the platform overview is at guard tour patrol system.
The Short Version
The AI story that gets the headlines is about models that do not exist yet. The AI story that will reach your business is about a percentage in a monthly report and what a supervisor did with it. The companies that come out of this well will not be the ones that avoided technology. They will be the ones who can show their work.
Key Facts
- Under California's employment regulations, an automated decision system does not have to involve artificial intelligence — a computational process built on statistics or an algorithm is covered if it facilitates a decision about an employment benefit.
- Illinois has required employers to notify workers of AI use in covered employment decisions since January 1, 2026, and those covered decisions include discipline, discharge, and promotion.
- Colorado's Automated Decision-Making Technology Act takes effect January 1, 2027 and requires a named human reviewer with authority to approve, modify, or override the output.
- A completion percentage calculated from patrol records is an evaluation input, and its regulatory status depends on what the employer does with it, not on how it was computed.
- Trinity Guard® scores patrol completion by rule, uses server-side AI to validate checkpoint scans rather than to judge officers, and leaves the employment decision with the supervisor.
FAQ
Not necessarily, and it does not have to. California's employment regulations define an automated decision system to include processes derived from algorithms or statistics, so a rule-based percentage can be covered if it facilitates a decision about an employment benefit. The label on the tool is not the test.
It depends which one. California's rules have been in force since October 1, 2025 and apply at five or more employees. Illinois has applied since January 1, 2026 and reaches employers with a single employee in the state. Colorado applies from January 1, 2027. Other states are considering similar rules.
Under Colorado's framework, the reviewer must be a person with authority to approve, modify, or override the decision, must be trained to conduct the review, must consider the relevant available evidence, and must not simply default to the system's output. A supervisor who signs off on a ranking without opening the underlying record is not conducting meaningful review.
Not instead of. California's regulations extend liability to an employer's agent, which can reach a vendor, but that adds a defendant rather than removing one. Colorado goes further in the other direction: contract terms indemnifying a developer against liability for its own antidiscrimination violations are void as against public policy.
Four years for automated-decision-system records in California, and at least three years for compliance records in Colorado. Setting four years as a single company-wide standard is simpler than tracking two clocks.
No. The platform calculates completion against the assigned patrol structure and surfaces irregular scans for inspection. Every employment consequence — discipline, allocation, promotion — is a supervisor decision made on the record the platform preserves.
Run a 14-Day Guard Tour Pilot
See exactly what your patrol records would show if someone asked. Verified checkpoints, reconstructable route history, and a completion score a supervisor can open and explain.
Start your pilot →About the author: Gyula Györfi is the founder of Trinity Guard® and a former police commander with 26 years of police service, 18 of them in command. He builds patrol software that has to work where security work actually happens: on the phone in an officer's hand, on a night shift.
