Contractor or Employee: What AI Can Tell You Before the IRS Does

Why Nonprofits Get This Wrong So Often

Start with the grant budget, the single most common origin of a nonprofit misclassification. A funder's template offers lines for personnel, fringe, and consultants. A program officer suggests the evaluation work belongs under consultants. The budget is approved, and the organization now has a document from a respected institution that appears to classify the worker. It does not. A grant budget governs how money may be spent and has no authority over employment law. If the person filling that line works under your direction, on your schedule, doing work central to your programs, they may well be your employee whichever column the money came from. The National Council of Nonprofits makes the same point, and it is worth repeating to your finance committee in those terms.

Then there is the hours assumption, so widespread it barely registers as a belief: this person only works ten hours a week, so they cannot be an employee. Hours worked is not a factor in any of the tests. A part time employee is an employee. Its cousin is the temporary assumption, that a role lasting only through the pilot must be contract work. Duration is weakly relevant to the permanency factor and decisive nowhere. Short term employees are a normal category, and payroll systems handle them fine.

The consultant who never left is the pattern that catches organizations which were originally in the right. A genuine consultant is engaged for a defined scope, brings their own methodology, serves other clients, and decides how the work gets done. Then the engagement renews, the scope broadens to whatever is needed that quarter, the person starts attending staff meetings, gets an organizational email address, begins supervising interns, and eventually holds a role everyone recognizes except the accounting system. Nobody decided to misclassify. The relationship changed shape while the paperwork stayed still. Our article on the consultant dilemma covers the strategic side of that dependency, and classification exposure is its legal shadow.

Sector-specific roles supply a steady stream. Teaching artists, workshop facilitators, tutors, and trainers are frequently paid on 1099s, sometimes correctly and often not. The distinction usually turns on control. An artist who brings an established curriculum, sets their own sessions, and teaches for several organizations looks very different from someone teaching your curriculum, in your building, on your schedule, under a program manager who observes and corrects them. Contract case managers raise the same question with higher stakes, because direct service delivery is almost by definition your usual course of business, which is precisely the factor that decides the answer in ABC test states.

Fiscal sponsorship confuses even careful organizations. In a comprehensive sponsorship, the sponsored project is a program of the sponsor, and the people doing the work are generally the sponsor's employees. In a grantor-grantee style arrangement the relationship differs again. What happens in practice is that a project leader treats their team as contractors because the sponsor handles the money, while the sponsor assumes the project leader handles the people. Both parties point at each other, and an auditor will not find that persuasive. If you sponsor projects or are sponsored, answer the classification question in the written agreement before anyone is paid.

What the Written Agreement Does and Does Not Settle

A great deal of false comfort rests on a signed independent contractor agreement sitting in a folder, so it is worth being precise about what that document accomplishes. In the IRS analysis the contract is evidence under the relationship-of-the-parties category, carrying real but limited weight, most useful where the other factors are genuinely balanced. In the economic reality analysis it matters far less, because the point of the test is how the relationship actually functions rather than how the parties describe it. Under a strict ABC test, prong A asks about control both under the contract and in fact, so the contract cannot save you if practice diverges from it, and prong B is untouched by contract language altogether.

The same limit applies to the worker's own preference, which comes up constantly. People ask to be paid on a 1099 because they like the simplicity, want to deduct expenses, or have other income and prefer to manage withholding themselves. None of it matters. Classification is not a term the parties can negotiate any more than minimum wage is, and a consent letter in the file is not a defense. If anything it can hurt, by showing the organization understood there was a question and settled it by agreement rather than by analysis.

What a contract genuinely can do is document the facts supporting the classification you believe is correct, and then constrain your own behavior so those facts stay true. A well-drafted contractor agreement describes deliverables rather than hours and supervision, leaves method to the contractor, assigns no work schedule or named supervisor, does not require attendance at unrelated staff meetings, does not prohibit other clients, does not grant employee benefits, and does not run indefinitely on automatic renewal. Those provisions are not decoration. Each one answers a factor a reviewer will ask about.

Which makes reading your existing agreements against those factors the more useful exercise, and one most organizations have never done because it is tedious and requires knowing what to look for. Templates picked up from a previous job or a web search are full of language that reads like employment: working hours, reporting to a supervisor, compliance with the employee handbook, at-will termination, performance reviews, exclusivity. An agreement requiring a contractor to follow your personnel policies asserts behavioral control on its face. Worse, even clean agreements are often undercut by practice, when contractors are put on the all-staff calendar, given a supervisor and a badge, or asked to request time off. Evidence of practice beats contract language in every test discussed here. Our guide to AI-assisted contract review describes the general technique, and classification language is one of its higher-value applications.

Running the Self-Audit, and Where AI Earns Its Place

A classification self-audit is not conceptually difficult. It is an inventory of everyone you pay who is not on payroll, a structured factor analysis of each one, and a written record of what you concluded and why. Few nonprofits have done it because the work is laborious and requires knowing which questions to ask. Thirty 1099 recipients means thirty analyses across three tests, each needing specific evidence about how the work actually happens. That is weeks of work for a finance director who does not have weeks, and it is exactly the shape of problem where AI moves something from theoretically-should-do to actually-done.

Building the factor-by-factor worksheet. Ask a model to produce a structured worksheet covering the IRS common law categories, the economic reality factors, and the three ABC prongs, with each factor expressed as a question about observable facts rather than a legal conclusion. Not "does the organization control the worker" but "who decides what days and hours this person works, and what happens if they decline an assignment." The output is a fact-gathering instrument, applied identically to every contractor, because consistency is what makes the resulting record credible. One worksheet completed the same way for all thirty people tells a coherent story. Thirty different memos do not.

Reading your actual agreements. Feed in your contractor agreements, with identifying details removed where you can, and ask the model to flag language asserting control over how the work is done, imposing schedule or location requirements, referencing employee policies or benefits, creating exclusivity, or establishing an indefinite relationship without a deliverable. Ask it to quote the clause and name the factor it touches. This is tedious pattern-matching review that models do well and humans do inconsistently at volume. The output is not a verdict on any agreement. It is a list of clauses to look at.

Comparing the description against the reality. Put a contractor's scope of work beside a description of how the engagement has actually run, written independently by the person who manages them, and ask the model to identify divergences. Managers describing their own working relationship honestly will say things like "she attends our Monday team meeting" and "I approve his schedule each month" without registering those as classification facts. The model surfaces the mismatch. This pairs with the discipline of keeping accurate HR records and role documentation, since a job description that no longer describes the job is a problem in more contexts than this one.

Summarizing your state's test, and drafting the interviews. Ask for a plain-language summary of how your state classifies workers for unemployment insurance and for wage and hour purposes, which version of the test applies, and whether occupational exemptions might cover your roles, then ask for the questions to take to counsel. Treat every word as a research brief requiring verification, because jurisdiction-specific legal detail is where models are most confidently wrong and classification law ages badly in training data. Separately, ask for a manager interview guide with neutral, non-leading questions about scheduling, instruction, equipment, supervision, and evaluation. A manager asked "is this person a contractor" will say yes. A manager asked "walk me through how you assign her work each week" will tell you what you need to know.

Producing the written record. The most durable product of the exercise is a dated, organized document showing that the organization asked the question in good faith, gathered the facts, applied the tests, and reached a conclusion with reasons. That record is valuable whether it finds everything fine or finds three roles that need to change. Models are good at turning messy worksheet output into a clean memorandum per role with the evidence attached. Store it with the care it deserves, and build the retention period into your records retention schedule rather than letting it accumulate in someone's drive.

If the Review Finds a Problem

Discovering a likely misclassification is not the disaster it feels like, but the next steps genuinely require counsel, and what follows is a map of what exists rather than a recommendation. Several of these paths involve filings or admissions that change your position, and choosing among them depends on facts about your filing history, your state, and your exposure that no article can assess.

Prospective reclassification. The simplest response is to move the worker onto payroll going forward, which resolves ongoing exposure and stops the clock. It does not by itself address past periods, and it raises the practical question of what you tell the worker, since a change in treatment invites the reasonable question of why. Handle that conversation deliberately and with advice. Reclassification also has budget consequences that surprise people, since employer payroll taxes, benefits eligibility, and workers compensation premiums add meaningfully to the cost of the same person doing the same work. Model that before committing, which is a natural extension of the work in our guide to nonprofit payroll.

The Voluntary Classification Settlement Program. The IRS operates a program allowing eligible taxpayers not under an employment tax examination to voluntarily reclassify workers as employees prospectively with limited liability for past federal employment taxes. Participants apply on Form 8952, and the terms involve paying a reduced amount based on compensation for a recent tax year, with relief from interest and penalties on that payment and from employment tax audit for prior years as to those workers. The IRS frequently asked questions on the VCSP set out eligibility, including consistent treatment and filing of required information returns. The limits matter: this is federal employment tax relief only, it does not resolve state tax, unemployment insurance, workers compensation, or FLSA claims, and entering it is an affirmative filing with the IRS.

Section 530 relief. A separate and frequently misunderstood provision can protect an organization from federal employment tax liability for past periods even where workers were in fact misclassified, if it meets three requirements: it consistently treated the workers, and all workers in substantially similar positions, as non-employees; it filed all required information returns, meaning timely Forms 1099, consistent with that treatment; and it had a reasonable basis, which may rest on judicial precedent, a prior IRS audit that did not challenge the classification, longstanding recognized industry practice, or another reasonable basis. This is a defense that arises in examination rather than something you apply for, and the reporting consistency requirement is the one organizations fail by accident. A single late or missing 1099 can break it, which is a good argument for taking information return discipline seriously and one of the quieter reasons clean records pay off in an audit preparation context.

None of the federal mechanisms reach state unemployment insurance, state withholding, workers compensation, or state wage claims, and some states run their own voluntary disclosure programs while others do not. Since state agencies are where classification disputes most often begin, a plan addressing only the federal side is a partial plan. Whatever path you take, document the decision and its reasoning the same way you documented the review. An organization that found a problem, sought advice, and acted is in a fundamentally different posture from one that did nothing, and the difference is visible only if it was written down at the time.

Making the Review a Habit Rather Than an Event

A one-time audit fixes the backlog and then decays, because the conditions that created the problem are still in place. New grants arrive with consultant lines, engagements renew and broaden, a contractor picks up a supervisory task because someone left. The fix is to attach classification checkpoints to the moments when the facts change, which is less work than periodic review and catches problems while they are still cheap. Three do most of the work. Before anyone new is paid outside payroll, the worksheet gets completed and someone with authority signs off, ideally built into the procurement or onboarding workflow so it is the default rather than an extra step. At renewal, a brief re-examination asks whether the engagement still matches what the agreement describes. And any material scope change, particularly supervision or system access, triggers a fresh review rather than a paperwork amendment.

Grant budget development deserves its own checkpoint because it sits upstream of everything. When a program budget is drafted, the question of whether a role is employment or contract work should be asked and answered then, with the budget built accordingly, including employer payroll taxes and benefits if the answer is employment. Organizations that skip this end up with an approved budget that cannot accommodate the correct treatment, which is the precise pressure that produces the wrong decision. Raising it while the funder can still be asked about moving a line is enormously easier than discovering it after the award.

Finally, name an owner and give the board a line of sight. Classification belongs with whoever owns HR and finance compliance, and in the same annual governance review as your other personnel policies. Boards do not need the detail, but they should know the organization has looked at this, when, and what it found, because the exposure is material enough to be a board-level risk item. Organizations that keep a current employee handbook and a documented set of personnel practices already have the habit. Adding classification to that rhythm is a small extension of something you are doing anyway.