AI Franchisee Selection and Territory Awards 2026: Scoring a Contract, Not a Hire

Five Ways the Legal Map Differs From Hiring

The employment statutes mostly do not apply

A franchisee is a business buying a licence, not an employee. Title VII, the ADEA and the state fair employment acts are built around the employment relationship, and a franchise award normally sits outside it. Teams that learned AI compliance from the hiring literature therefore reach for controls that are not the operative ones here.

The contract statute does apply, and it is older and broader

The right to make and enforce contracts free of race discrimination reaches commercial relationships directly. No employment relationship is required, no agency charge has to be filed first, the limitation period is long, and the remedies are not capped the way some employment remedies are. A franchise award is exactly the kind of contract formation it was written for.

Financing drags a second regime in

Where the franchisor, an affiliate or a preferred lender extends or arranges credit, the decision becomes a credit decision. Credit law prohibits discrimination on protected bases and requires a statement of the specific principal reasons for an adverse action — a requirement that a scoring model with no reason codes cannot satisfy.

Franchise relationship law adds a state layer

Several states restrict termination, non-renewal and transfer refusal without good cause and with notice. A model-driven renewal score that no one can explain does not produce good cause, and it will be read as pretext where the outcome pattern is unfavourable.

Disclosure rules reach the criteria themselves

The disclosure document describes the relationship a prospect is buying. Where selection, territory and renewal are materially driven by an internal scoring system, saying nothing about it is a disclosure decision, and saying something inaccurate about it is a worse one.

Six Places a Model Touches the Award

Exposure is a property of the position in the pipeline rather than of the model, and the earliest position is the one with the least documentation.

Lead scoring

Inbound prospects are ranked for follow-up by a model trained on which past leads closed.

Where it turns: The quietest failure in the pipeline. Nobody is rejected, so no adverse action is recorded — candidates simply never get called back. The pattern only becomes visible in aggregate, which is precisely the form a disparate impact analysis takes.

Candidate qualification

Liquidity, net worth, credit, industry experience and prior ownership are scored into a go or no-go.

Where it turns: Facially neutral financial criteria carry well-documented demographic correlation. They may be entirely defensible as business necessity, but only if someone wrote down the necessity and tested whether a less restrictive threshold would have worked.

Interview and assessment scoring

Structured interviews, personality inventories or video assessments scored by a model.

Where it turns: This is the stage most likely to have been bought from a vendor selling into HR, which means it arrives with employment-framed assurances that address the wrong statute and leave the contract claim untouched.

Territory allocation

Site and territory availability assigned by a model optimising forecast revenue.

Where it turns: Demographic data is an input to almost every trade-area model. Where the optimiser systematically steers candidates toward or away from areas that correlate with protected characteristics, the resulting map is the evidence — and the housing and lending analogues make the theory familiar to plaintiffs.

Financing and terms

Fee structure, royalty ramp, development schedule or lender routing set by score.

Where it turns: Differential terms on a protected basis is the oldest claim in the book, and where credit is involved the adverse action notice must state specific principal reasons. A score band is not a reason.

Renewal and non-renewal

Performance models flag which franchisees to renew, re-price or exit.

Where it turns: Performance data is downstream of the territory you assigned and the support you provided. A model that ranks outcomes without accounting for those inputs will penalise the franchisees your earlier decisions disadvantaged, which is how one decision becomes a pattern.

What Becomes Discovery

None of the following is exotic. It is the ordinary exhaust of a scoring system, and it exists whether or not anyone has looked at it.

The training set, and which historical awards it learned from, including the years before anyone reviewed those decisions.
Every input feature, plus the proxies — zip code, school, prior brand, referral source — that stand in for something you never intended to use.
Score distributions by outcome, which is where a four-fifths style comparison gets computed by someone other than you.
The threshold history, because a cut-off that moved after a quarter of poor conversions is a decision with a date on it.
Human override logs, which either show independent judgment or show a rubber stamp at a rate that proves the model decided.
Vendor documentation, marketing claims and validation reports, which will be read against what the system actually did.

Six Controls That Actually Fit This Regime

Decide, in writing, that the model recommends and a human decides

Then instrument it. An override rate near zero converts your recommendation engine into the decision-maker as a matter of evidence, regardless of the policy document.

Run the impact analysis before the model is deployed, not after a complaint

Compute outcome rates across groups at each stage. Privilege considerations are real and worth planning for, but the analysis you never ran is not protected by anything, and the data to run it later will exist anyway.

Write the business necessity down next to each threshold

A liquidity requirement is defensible when it is tied to a documented failure-rate analysis, and fragile when its origin is that someone picked a round number in 2019.

Generate real reason codes at every declination

Required where credit is involved and useful everywhere else. Reason codes force the model to be explainable at the level of the individual decision, which is the level at which it will be challenged.

Keep territory optimisation away from demographic targets

Use traffic, competition, tenancy and economics. Where a demographic variable improves the forecast, note that the improvement is exactly the correlation a plaintiff will point at.

Put diligence into the vendor contract

Validation evidence, notice before behaviour changes, audit export, and an answer on whether the vendor may be named directly. Vendors supplying scoring tools have already been sued alongside their customers in the hiring context.

Questions Franchise Development Teams Ask

Franchisees are not employees. Does discrimination law reach how we select them?

Yes, through a different door than the one most AI governance programmes were built around. The employment statutes are keyed to the employment relationship and mostly do not reach the award of a franchise. The federal right to make and enforce contracts without race discrimination is not so limited: it applies to commercial contracting directly, requires no employment relationship, does not require an administrative charge before suit, carries a long limitation period and is not subject to the damages caps that apply to some employment claims. Credit law adds a second route wherever financing is extended or arranged, and several states regulate termination, non-renewal and transfer refusal in the franchise relationship itself. The practical consequence is that a team which imported its controls from hiring has controls aimed at a statute that does not apply, and gaps under the ones that do.

Our model only ranks leads for follow-up. Nobody is rejected.

That is the stage with the least visibility and, for that reason, the most exposure. A ranking that decides who gets a call back allocates the scarce resource in the pipeline, and a candidate who is never contacted receives no decision to appeal, no reason and no record. From the outside it is indistinguishable from not being interested — until someone computes contact rates by group across a year of leads and the gap appears in aggregate. Two properties make it worse than an explicit rejection. The lead model is usually trained on which past leads closed, so it reproduces whatever pattern the previous decade of awards contained. And because nobody classified it as a decision system, it typically has no documentation, no validation and no retention policy, so the only surviving record is the raw log. Score the leads if you must, but measure the contact rate.

Can a territory allocation algorithm create a discrimination claim?

It can, because trade-area models are built on demographic data by design. A model that recommends who should be offered which territory is making an allocation decision about the value of the contract on offer, and where the resulting assignments correlate with the protected characteristics of the candidates, the map itself is the evidence. Plaintiffs have a well-developed analogue to borrow from in housing and lending, where steering claims turn on outcome patterns rather than intent. There is also a second-order effect that shows up later: territory quality drives performance, and performance drives renewal, so a steering pattern at award time produces an apparently neutral performance gap years afterwards. Two disciplines help. Keep demographic variables out of the candidate-facing allocation step even where they improve the forecast, and test assignment outcomes by group as a standing report rather than as an investigation.

We use a third-party scoring vendor. Does that shift the risk?

It spreads the risk without removing yours, and the trend runs the other way. In the hiring context, vendors supplying algorithmic screening have been pulled into litigation directly rather than sitting behind their customers, on theories that a tool making the operative decision across many clients is acting in the relevant role itself. Nothing about that theory depends on the relationship being employment. Meanwhile the franchisor remains the party that chose the tool, set the thresholds, and awarded or declined the contract. Two practical implications. Your diligence file is part of your defence, so validation evidence, documented threshold decisions and a record that you asked about impact testing are worth more than an indemnity from a small vendor. And the contract should require notice before model behaviour changes, because a silent update to someone else's model is a change to your selection criteria.

Do we have to disclose the scoring system in the disclosure document?

Treat it as a live question rather than an obvious no. The disclosure document exists to describe the relationship a prospect is buying, and where selection, territory assignment, financing terms or renewal are materially driven by an internal scoring system, that is a fact about how the relationship works. The greater danger is not silence but inconsistency: describing selection as a relationship-based review conducted by the development team, while the operative decision is a model threshold, creates a mismatch between the document and the practice that is uncomfortable to explain later. Where the system also produces performance projections or territory forecasts shown to prospects, the earnings-claim rules engage on their own terms, with their own substantiation requirement. The safe posture is to describe the process accurately at a level that does not give away the model, and to make sure the description matches what the pipeline actually does.

What should we keep, and for how long?

Keep more than feels comfortable, because the limitation periods here are long and the artefacts cannot be reconstructed. The record that matters is the decision record: inputs used, score produced, threshold in force on that date, model version, human reviewer and what the reviewer changed. Add the periodic impact analyses and the business necessity rationale for each criterion. Two failure patterns are worth designing against specifically. The first is a vendor holding the only copy of the score history, which becomes unavailable when the contract ends — export it continuously. The second is a threshold that is edited in a configuration screen with no history, so the value in force on the date of a challenged decision is unknowable. Version thresholds like code. A defensible file is not the one that shows no disparity; it is the one that shows you measured, understood, and can say what the rule was that day.

The Contact-Rate Test

Pull last year's inbound franchise leads and compute one number: the share that received a human contact, broken out by the geography of the enquiry.

If that number varies sharply and nobody in the organisation can say which feature produced the variance, you do not have a lead-scoring tool. You have an unexamined selection criterion operating before anyone has been told they were considered.

Awarding a Franchise Is Not Hiring. The Scoring Model Does Not Know the Difference.