What's the Smallest Hiring Process a Two-Person Company Can Defend?

What Are the Four Artifacts?

Four documents, and an afternoon is enough to make them. The criteria for the job, written before you look at a single application. One question set every candidate answers. Dated notes on each candidate rated against those criteria. And a short paragraph naming why the person you hired won it. Nothing else has to exist for the file to hold up.

The order is the whole trick. Criteria written after you meet three people describe those three people, and a reason written a month after the decision is a reconstruction. Each artifact is cheap while it is being made and impossible to make later, which is why the two-founder version of this is not a smaller version of a corporate process. It is the same four things with nobody to delegate them to.

What each one is doing, in the order you write them:

1. The criteria. Five to eight requirements drawn from what the job actually does in its first six months, each marked required or preferred, each with a sentence describing a weak, adequate and strong answer. This is the artifact everything else refers back to. 2. The question set. The same questions, in the same order, for every candidate, rated on a common scale, with both of you agreed in advance on what an acceptable answer contains ⁶. Follow-ups can go anywhere; the base set does not move. 3. The notes. Written during or immediately after each conversation, dated, tied to a criterion rather than to an impression. "Walked through a pricing decision she reversed and why" is a note. "Sharp" is not. 4. The reason. Three or four sentences: which criteria decided it, where the finalists actually differed, and what the runner-up was missing. If you cannot write it without naming a criterion, the decision was made on something you did not measure.

A fifth habit costs nothing and matters more than it looks: put the criteria and the question set somewhere both founders can see them before the first call, not after. Two people improvising the same interview is not a structured interview, and the file will show it.

Which Laws Reach a Two-Person Company?

Fewer than most founders assume, and never zero. At one employee you are covered by the requirement to give equal pay for equal work. At fifteen, the federal bans on discrimination by race, color, religion, sex, national origin, disability and genetic information attach. At twenty, age discrimination joins them. Below fifteen, those federal statutes do not reach you, and state or local law often does ¹.

California is the sharpest example of why the federal number is the wrong number to plan against. Its Fair Employment and Housing Act defines an employer as any person regularly employing five or more persons ⁷, so a two-founder company is three hires from being covered by a statute whose thresholds are set by a legislature, not by you. Other states set their own thresholds. Check the one you actually hire in before you decide the question is theoretical.

The Uniform Guidelines on Employee Selection Procedures (the federal standard for anything used to make a hiring decision) are applied in the enforcement of Title VII and Executive Order 11246 ⁹, so at two people they are not law for you yet. They are still the cheapest template available, because they are the standard your process will be read against the moment you cross a threshold, and building to them now costs an afternoon rather than a retrofit.

There is also the part no threshold governs. A hiring file is what answers an unemployment claim, a contract dispute, an investor's diligence questionnaire and your own memory in eighteen months, when the second person you hired asks why the first got the higher band. Two founders who can produce dated criteria and dated notes are in a different conversation from two founders who can produce a Slack thread. Asking candidates directly about how they work with AI has its own limits worth knowing before you write the question set: what you can legally ask about AI use covers them.

Write the Criteria Before You Look at Anyone

Start from the duties, not from the ideal person. List what the role actually does in its first six months, then name the five to eight requirements each duty implies. Mark each one required or preferred. Beside each, write what a weak, an adequate and a strong answer looks like. That last column is the one founders skip, and it is the one that makes two people rate the same answer the same way.

The federal sequence is the same sequence, which is convenient: analyze the job, decide which competencies the interview will assess, write the questions, then build the rating scale, and keep a record of how you did it ⁶. At two people the job analysis is an hour with a document rather than a panel of subject-matter experts, and it is still a job analysis.

One trap is worth naming, because it is where most homemade criteria go wrong. Content validity, the argument that a selection procedure is defensible because it samples the work, does not support measuring traits and constructs directly. The Guidelines name intelligence, aptitude, personality, commonsense, judgment and leadership among the traits and constructs a content strategy cannot demonstrate the validity of ³. So "good judgment" is not a criterion. "Chooses which of two customer escalations to answer first and says why" is a criterion, and it is testable.

The same rule kills the other common shortcut: criteria describing a knowledge or skill the person will be taught on the job are outside what a content strategy supports ³. If your onboarding teaches the tool, the tool is not a requirement.

Write the bar as a behavior and it survives the whole process: it becomes the interview question, the note field, and the sentence in the decision reason. Setting that bar for AI work specifically is its own exercise, worked through in what a defensible AI proficiency bar looks like.

Why a Sales Record and a Data Record Aren't the Same Record

Because defensibility rests on job-relatedness, and job-relatedness is a property of one job. A selection procedure is supported by content validity only to the extent that the behaviors it makes a candidate demonstrate are a representative sample of the behaviors of the job in question ³. The sales job's behaviors and the data job's behaviors are different behaviors, so the file that defends each is a different file.

In practice that means the job analysis, not the interview, is the part that changes per role. The Guidelines ask that it focus on the work behaviors and the tasks attached to them, that the behaviors chosen for measurement be the critical or important ones making up most of the job, and, where a behavior produces something, that the work product be analyzed too ³. A sales hire's record has a discovery conversation and a handled objection in it. A data hire's record has a query, a result, and what the candidate did when the result disagreed with the brief.

The consequence founders underestimate is that one reusable interview across every opening is the least defensible artifact in the set, precisely because it is the most convenient one. Reuse the format. Rewrite the content per occupation.

Two further rules follow from the same principle. A selection procedure that has adverse impact is treated as discriminatory unless it has been validated, and where two procedures serve your purpose about equally well, the one with the lesser adverse impact is the one to use, which is why the Guidelines fold a search for suitable alternatives into the validity work itself rather than leaving it as an afterthought ⁴. And a vendor's paperwork does not transfer the duty: the employer remains responsible for ensuring its selection procedures are valid, whatever the vendor's documentation says ⁵.

If the role is one where the resume has stopped telling you anything, the cheapest job-related artifact is usually a short piece of the actual work, and the trade-off is laid out in replacing the resume screen with a work sample.

Keep Three Numbers and Four Files per Opening

The Uniform Guidelines let anyone employing a hundred people or fewer keep a deliberately short record: for each year, the number of people hired, promoted and terminated for each job by sex and where appropriate by race and national origin; the number of applicants for hire and promotion on the same breakdown; and which selection procedures were used, standardized or not ². Three counts a year, plus the four artifacts per opening.

That provision exists to keep the burden proportionate, and it is the single most useful thing in the Guidelines for a company this size. It is not a pass on evidence, though. Where you have reason to believe a procedure is screening one group out, the same section expects you to hold whatever validity evidence for it you have ².

Collecting the demographic counts feels wrong to founders who have never done it. You are counting the process, not the people: the numbers are gathered separately from the hiring decision, never seen by whoever is rating candidates, and their only purpose is to let you notice a pattern you would otherwise learn about from a lawyer. A voluntary self-identification question, kept out of the file the interviewers read, is the standard way to do this.

Retention is where the growth curve bites. Once the federal statutes cover you, a personnel or employment record is preserved for one year from the date the record was made or the personnel action was taken, whichever is later. Once a charge has been filed, all relevant personnel records are preserved until final disposition ⁸. The practical version at two people: keep everything for a year minimum, keep it somewhere that is not one founder's laptop, and delete nothing while anyone is disputing anything.

What Changes at Five, Fifteen, and Twenty People?

Thresholds arrive, and each one attaches obligations to a process that is already running or is not. In California, five employees brings the state statute ⁷. Fifteen brings the federal bans on race, color, religion, sex, national origin, disability and genetic information discrimination; twenty adds age; a hundred brings the EEO-1 workforce report ¹. None of them announce themselves.

So the honest reason to build the four artifacts at two people is not that the law demands it today. It is that every threshold above assumes a documented process existed before the threshold was crossed, and the artifact you cannot manufacture retroactively is a dated one. A founder who started the habit at hire two has a file; a founder who starts it at hire fifteen has a file that begins at hire fifteen.

Tools are the other thing that changes on the way up, usually before the headcount does. The obligations that arrive with an assessment, a screen or a scoring tool attach to you rather than to the vendor. You are the one responsible for ensuring what you use is valid for the position and the purpose you use it for, however good the vendor's documentation is ⁵. Ask for the job analysis behind it and the occupations it covers, and treat a tool with neither as undocumented for your roles. See how Olive measures this.

When the first written rule does become worth having (usually the first time a candidate asks whether they may use AI on your take-home and the two of you answer differently), the shortest useful version is a page, not a handbook. What actually belongs in an AI hiring policy is the list, and most of it is decisions you have already made informally by then.

Read the evidence >