Every product meeting in 2026 has the same gravitational pull. Someone says "users have questions," and forty seconds later the whiteboard says chatbot.

We build software for families of kids with disabilities. Parents arrive with questions like "the school says my son doesn't qualify for an IEP, is that allowed," or "what does ABA actually cost if insurance denies it." Those are exactly the questions a chatbot is good at answering fluently.

We built a forum instead. Here is the reasoning, because I think it generalizes to more domains than people expect.

The failure mode is not being wrong. It is being wrong confidently.

An LLM that does not know the answer does not act like it does not know. It produces something in the shape of an answer, at the same confidence as a correct one.

For most products that is an annoyance. You ask for a regex, you get a broken regex, you notice in ten seconds because the test fails. The feedback loop is tight and the cost of a wrong answer is one wasted minute.

Now change the domain. A parent asks whether the district can refuse an evaluation. The model says something plausible. There is no test suite. The parent does not have the expertise to catch the error, which is the entire reason they were asking. And the feedback loop is not ten seconds, it is the length of a school year, by which point a legal deadline has quietly passed.

The variable that matters is not accuracy. It is cost of a wrong answer multiplied by time to detect it. When that product is large, generated confidence is the wrong interface, no matter how good the model is.

Three things a forum has that a chatbot structurally cannot

Answers carry provenance. On the forum, the reply says who is talking. A parent three years into due process, a speech therapist, a special education attorney. A reader weighs "here is what happened when I filed" differently from "here is what the statute says," and they should. A chatbot flattens every source into one anonymous voice with no way to weigh anything.

Disagreement stays visible. Ask about a therapy approach and you get four replies that do not fully agree. That is not a bug in the content, it is the actual state of the field, and a parent making a decision deserves to see it. Chat interfaces are built to resolve into one answer. When the honest answer is "practitioners disagree and here is the shape of the disagreement," a single synthesized response destroys the most useful information in the room.

It gets more correct over time instead of less. A wrong forum answer gets corrected by the next person, in public, attached to the original. A wrong chatbot answer is generated fresh in a private session, seen by one person, and leaves no trace. One of those systems accumulates knowledge. The other regenerates from scratch every time and never learns it was wrong.

The part nobody warns you about

The technical build is not the hard part. Threads, categories, auth, moderation queue. It is a known shape and you can stand it up quickly.

The hard part is that a forum with no answers is worse than no forum. A chatbot works on day one for user one. A forum needs a critical mass of people who answer before it is useful to people who ask, and the first stretch is genuinely unglamorous: you seed real questions, you find practitioners willing to answer for nothing, you answer things yourself, and you accept that the ratio looks bad for a while.

That cold start is the real reason teams pick the chatbot. Not because it is better. Because it is instant, and a community is not.

A few things that made it survivable:

  • Seed from questions people actually asked, not questions you imagine. Support inboxes and search logs are full of real phrasing, and real phrasing is what other parents search for later.
  • Categories should match how users think, not how you store data. Parents do not think in terms of your schema. They think "school stuff" and "therapy stuff."
  • Answer latency beats answer quality early on. A decent reply in an hour builds a community. A perfect reply in four days does not.
  • Moderate hard on tone from day one. In a vulnerable niche, one dismissive reply costs more than ten helpful ones gain. The norms you allow in week one are the norms you have forever.

When the chatbot is the right call

I am not making a general anti-LLM argument, and the honest version of this post has to include the other side.

Use the model when the answer is deterministic and checkable. Where is this in the docs. What does this error mean. Summarize this thread. We use LLMs internally for exactly that.

The line I would draw: if a wrong answer costs the user more than the right answer saves them, do not generate the answer. Route to a human, cite a source, or say nothing. In special education, where a missed deadline can cost a kid a year, that line sits much closer than it does in most software.

The result is Special Needs Forum, where families ask about IEPs, ABA therapy, and school placement, and get answers from people who have been through it. It is free and it does not require an account to read.

If you are into the less philosophical side of this work, I also wrote about the schema mistake that cost us money on the directory side, which is a more conventional Postgres war story.

I work on Special Needs Care Network, helping US families find special needs schools and ABA therapy providers.