Chat (Alt-Y): the vision behind the decision

Before committing to a chat direction and the front-end architecture around it, we need a sharper answer to one soft question — what is chat ultimately for? This breaks it into concrete questions, each tied to the build decision it actually drives.

The one question that splits the whole tree

Is chat a destination or a feature?

• Destination — over time Alt-Y becomes a primary way people use online: you type or talk, it summons the doc, edits it, answers. The grid and forms become surfaces that chat drives.
• Feature — chat stays an assistive panel next to a UI that remains primary. It helps, explains, fills a field; nobody runs their day from it.

Almost everything else — how much to invest, whether the top bar is chat-centric, whether a framework migration is even worth it — falls out of this one answer.

The vision has four dials

Naming which corner we are aiming at is the decision.

Table stakes regardless of corner: the AI is grounded in company knowledge. That is a requirement, not a strategic choice — it should not absorb the debate.

It is a good north-star — the question is which slice to fund first.

The concrete questions

★ = answer these first; they move the decision the most. Each question is tagged with the architecture decision it unblocks.

A. Purpose & users

• ★ What job is chat hired for? Pick the dominant spine: (a) help / onboarding, (b) data-entry copilot that mutates OneSheet docs, (c) query / analytics ("show me all overdue invoices"), (d) automation / agent runner, (e) command / navigation.Unblocks: how deeply chat must read and write online state — the single biggest driver of front-end effort.
• ★ Who specifically uses it? Name and rank the 2–3 personas that matter — e.g. accountant at an invoicing customer, field tech in CMMS, admin power-user — not "customers."Unblocks: a power-user daily driver justifies a real app; an occasional helper does not.
• Daily driver or occasional tool? Always-open sidebar used dozens of times a day, or invoked a few times a week?Unblocks: justifies or kills the top-bar redesign and the framework investment.

B. Relationship to online / OneSheet

• ★ Does it read the current doc's context automatically?
• ★ Does it write / mutate the doc, or only suggest?
• Is it scoped to the current doc, or the whole account?Unblocks: the entire integration architecture — and exactly what the prototype already exercises, so there is real signal here.

C. Autonomy — the "virtual workers / agents-as-a-service" question

• Does chat ever act without a human in the loop — on a schedule, watching for a condition, finishing a task and reporting? Or always turn-by-turn?
• If autonomous: what is the unit of work a virtual worker owns? Name one concrete job (a recurring report? chasing overdue invoices? CMMS triage?), not "tasks generally."Unblocks: autonomy needs a job runner, state, scheduling and an audit trail — a real backend plus app shell.

D. Social / multiplayer

• Does multi-person chat even exist in online today to put AI into? If not, "AI in company chats" means first building company chat — a much bigger swing.
• When AI is in a group thread: summoned (@mention) or always listening? Acts, or only answers?Unblocks: multiplayer real-time presence is the single strongest argument for a serious front-end framework.

E. Knowledge flywheel — "all chats as company context"

• Are chats stored and indexed as durable, retrievable company knowledge, or ephemeral? Which chats — human-AI only, or human-human too?Unblocks: a permissioned retrieval layer plus governance — a real backend commitment.

Its value compounds with use — but only if customers consent to ingestion, which is the next question.

F. Consent & governance — the bound

• ★ What is the consent / visibility model? Opt-in per user or company-mandated? Can people speak off-record? Who can the AI surface whose chats to — and does retrieval respect existing access rights?

Why this is a priority and not a footnote: always-on, plus AI inside human conversations, plus all-chats-as-context, deployed to EU SMBs, is one wording away from "we record and surveil every employee conversation" — a GDPR / works-council exposure, a trust killer, and a sales objection, not a feature. Decide it on purpose.

G. Success & horizon

• What does "working" look like in 12 months? One north-star number or behavior — e.g. "X% of edits originate in chat," "every new customer uses it in week 1," "support tickets down Y%."Unblocks: whether to optimize for adoption breadth or power-user depth.

The part that is genuinely ours to win

Everyone — Notion, ClickUp, Workday, Salesforce — is selling a generic agent bolted onto a chat box. None of them can correctly edit a specific customer's invoice or CMMS record, because they do not own the structured data model and the business rules. We do.

The save-correctness already built into OneSheet — rights-on-save, value-link forward-sync, money / VAT recompute — is exactly what makes an agent that acts through OneSheet trustworthy where a generic agent cannot be. And the same access-rights discipline gives us permission-aware retrieval for the flywheel: "the AI only surfaces what you are allowed to see" becomes a correctness guarantee, not a promise.

That turns both the agent and the flywheel from me-too into a moat — and turns the governance landmine into a differentiator, provided it is built in from the start, not bolted on after.

What to decide

1. Pick the corner. Settle the four dials — at minimum Destination-vs-Feature and the consent model.
2. Separate MVP from north-star. The MVP likely sits at the minimal end of the four dials, the north-star at the maximal end (see the table above). Size the architecture decision for that near-term slice, not the whole north-star.
3. Ground the soft questions in evidence, not a meeting:
   • Frequency and edit-vs-query ratio → pull from actual Alt-Y usage. That answers the daily-driver and read/write questions with numbers.
   • Personas, autonomy appetite, consent tolerance → 3–4 short customer conversations beat a strategy offsite.