1,723 people showed up for Life After Loss on January 22 and 23, 2026. Not a product launch, not a conference keynote — a two-day gathering about how to hold grief. I keep coming back to that number because it tells me something about scale that most of us building conversational systems have not accounted for.
In 2026, a lot of people in the United States are grieving something they cannot name on a form. Eroding institutions. Rising violence. The sense that democratic ideals are more fragile than they were taught. Researchers call this ambiguous loss: grief with no clear ending and no closure, arriving in waves that get triggered by events nobody scheduled.
I build bots for a living. And ambiguous loss breaks almost every assumption baked into the way we design them.
We Design for Resolution. Grief Does Not Resolve
Open any conversational design doc and you will find the same skeleton: intent, slot filling, fulfillment, confirmation, close. The happy path ends. There is a terminal node. We measure ourselves on containment rates and task completion, which both assume the user arrived with a problem that can be finished.
Grief has no terminal node. Someone talking to your support bot about transferring a deceased parent’s account is not completing a task — they are in the middle of something that will keep returning for years. The flow ends. Their experience does not.
The practical consequence: stop modeling these conversations as funnels. Model them as long-lived state that can be re-entered at any time, from any point, with no penalty for coming back to the same thing six times.
What Trauma-Informed Actually Means in Code
Bereavement care practitioners have been clear that the field needs trauma-informed and culturally responsive approaches. That phrase sounds like policy language until you translate it into implementation details.
- No required fields. If someone cannot type the word, let them continue anyway. Forced disclosure is a design choice, and it is a bad one.
- Always-available exit. Every turn needs a path out that does not feel like failure. No “are you sure?” friction on leaving.
- Short turns. Grief affects the brain and body — attention, memory, processing. A 400-word empathetic wall of text is not kindness, it is cognitive load.
- Fragmentary input handling. Expect half-sentences, typos, repetition, silence. Your NLU should not ask someone to rephrase because the confidence score dipped.
- No hardcoded grief model. Do not ship the five stages as a state machine. Mourning practices, timelines, and language vary across cultures and families. Make that configurable content, not control flow.
Sustained Support Is a Scheduling Problem
One of the clearest gaps named in bereavement care is sustained support beyond the early weeks following a loss. The attention arrives fast and leaves fast. Week one brings casseroles and condolence emails. Month four brings nothing.
That gap is something software is genuinely good at closing, and it is not complicated. It is a job queue. Check-ins at week ten, month six, the anniversary. Opt-in, cancelable, low-pressure, and not framed as a reminder of death — framed as an open door. The hard part is not the cron expression. The hard part is that nobody prioritizes a feature whose engagement metrics look terrible by design.
If you build this, you also need to be honest about what the bot is. It is not a therapist. It is not a grief counselor. It is a tool that can notice timing, hold context, and route to humans. The most useful thing a system can do here is be a competent front door to the kind of community 1,723 people drove or logged in to find.
The Detail We Keep Losing
There is a second meaning to grieving the loss of details, and it sits closer to home for those of us shipping these systems.
Our whole stack is built on compression. We summarize transcripts. We embed sentences into vectors. We roll conversations into topic clusters and sentiment scores. Every layer throws away specifics in exchange for scale, and most of the time that trade is correct.
Grief is where it stops being correct. The specifics are the entire point — the particular phrase someone used, the exact thing they are afraid of losing next, the name they mentioned once in turn three. A psychologist describing grief attacks talks about carrying forward what was good in a relationship so it becomes part of you. That is a memory system. It is lossy in some places and stubbornly precise in others.
So when you design memory for a bot that will meet people on their worst days, ask what your summarizer is discarding. Keep verbatim where it matters. Store the name. Store the date they told you about. Let the system be specific, because the person on the other side is grieving something very specific, and a reply built on averages will tell them immediately that nobody was really listening.
🕒 Published: