About Severance.help

Why this exists

Severance is one of the largest financial decisions most people will make with the least information. You are handed a number and a deadline, usually on the worst workday of your career, and asked to evaluate an offer against a benchmark nobody has ever shown you. The people who know what is normal work in HR, or are employment lawyers, and neither is available to you on the afternoon it happens.

Meanwhile the information that does exist is scattered across law firm marketing pages written to generate consultations, forum threads where a single anecdote stands in for a benchmark, and generic AI tools that will confidently write you a letter without ever asking whether the offer was any good.

Severance.help is built to answer the first question people actually have, which is not “can something write me a letter” but “am I getting a reasonable deal?”

How the assessment works

We would rather explain the method than ask you to trust a black box with your employment history. The assessment is deliberately not a single prompt to a chatbot.

1. Structured inputs. You answer defined questions rather than writing free text, so every part of the result traces back to something specific you told us.
2. Deterministic calculations. Weekly pay, severance weeks, total package value, and weeks per year of service are computed with fixed arithmetic. No model is asked to do maths.
3. A documented factor set. Tenure, level, compensation, employer size, the reason employment ended, your performance record, and the terms of the agreement are evaluated against a defined rule set, so that similar situations produce similar answers rather than varying with the phrasing of a prompt.
4. AI reasoning and explanation. Claude, made by Anthropic, works from that grounded picture to produce the recommendation and to explain it in language you can act on. It interprets a structured factual context rather than improvising from raw document text.

Every assessment records which model and which prompt version produced it, so results can be reviewed and improved rather than merely trusted.

What we will not do

• We will not hide that the recommendation is AI-generated. It says so on the result.
• We will not tell you your employer broke the law. That is a legal conclusion, and we are not lawyers. Where your answers describe circumstances that deserve a lawyer’s eye, the assessment says the circumstance may warrant review.
• We will not invent precision. No made-up medians, no “87 percent chance of getting more money”. When we publish benchmarks, they will come from our own data and we will say what they are based on and how large the sample is.
• We will not promise an outcome. Nobody can tell you what your employer will do.
• We will not sell your information. See the privacy policy.

How this is paid for, and what happens next

Right now, nothing here costs anything and there is nothing to buy. That is deliberate. The assessment, the result, the explanation, and the negotiation letter are free, and the first useful thing cannot sit behind a paywall if the goal is to be the place people go when an offer arrives.

Over time, two things follow from doing that well. Aggregated, anonymized data from thousands of assessments will support genuine severance benchmarks, which is what will eventually let the tool tell you how your offer compares to real reported offers rather than to general practice. And some people, after seeing their assessment, will want an employment attorney rather than a website. Connecting those people with independent attorneys is the likely business model. It is not built yet, and when it is, it will be disclosed plainly on the page where it appears.

Who runs it

Corrections

If something on this site is wrong, we want to know. Corrections and the basis for what we publish are covered in the editorial policy.