Skip to content

Proposal: AI-assisted tech-support responder for Slack #440

@BigGillyStyle

Description

@BigGillyStyle

TL;DR — recommendation for discussion

Our knowledge lives only in Slack messages — we have no docs or FAQs to load
into an off-the-shelf tool. That single fact points us toward building a small
bot (Option B)
rather than buying one:

  1. Build small, on GCP (Option B). A lightweight bot reads our channels
    directly (Slack's API works on the free plan, within the 90-day window),
    captures resolved Q&A — e.g. a lead reacts ✅ to a good answer — into a
    permanent store, and surfaces the closest past answer when a question repeats.
    Pennies a month, and it's the only approach that actually uses the messages
    we already have before the 90-day clock erases them.
  2. Buying first won't help us yet. The off-the-shelf tools (AnswerHub, etc.)
    answer only from a knowledge base we hand-curate; none can read our existing
    channel messages.
    With no docs to seed them, they'd start empty.
  3. The unavoidable cost is human effort, not software — someone has to own
    keeping the knowledge current. No tool removes that.

Everything below is the detail behind this.


Problem / why now

Volunteers in our Slack tech-support channels answer a lot of the same
questions
and have to constantly monitor threads. That's an ongoing,
unrewarded burden.

There's also a clock running: our free Slack plan only keeps messages for 90
days
. Every good answer our volunteers write quietly disappears after three
months. So even before we talk about AI, there's value in simply capturing
those answers somewhere permanent.

This issue is a proposal to discuss at standup — not a finished design. Goal: pick
a direction.

Goal & non-goals

  • Goal: reduce volunteer load by automatically suggesting (or drafting)
    answers to common tech-support questions, using answers we've already given.
  • Non-goal (for v1): fully autonomous posting with no human in the loop, or
    replacing volunteers. We want to assist, not auto-pilot.

The common backbone (true for almost any version we build)

New question in Slack
        │
        ▼
1. Look up similar past answers   ← cheap, no AI needed
        │
        ▼
2. (optional) Ask a small AI model to phrase a reply from those answers
        │
        ▼
3. Bot drafts into a private "leads" channel → [Approve] / [Reject]

Two ideas do most of the work:

  1. Capture knowledge before Slack deletes it. A small listener saves resolved
    Q&A pairs — e.g. when a lead reacts ✅ to a good answer — into permanent
    storage (a simple database). This solves the 90-day problem regardless of
    which option below we pick.
  2. Human-in-the-loop by default. The bot proposes a draft to leads for one
    click to approve, rather than posting directly to members. This caps cost and
    prevents the bot from confidently posting something wrong.

A couple of terms, since they show up below:

  • Retrieval / search = finding the closest past answer. Cheap.
  • LLM = the AI model that writes a fresh reply. The optional, paid part.
  • Embeddings = turning text into numbers so "similar questions" can be found.
    Very cheap, but not literally free.

Options (roughly even trade-offs)

Option Build effort Deploy & maintain Monthly cost (ballpark)
A. Buy / no-code — an existing Slack AI app, or Slack's own built-in AI search Low Low — vendor hosts it $0 on free tiers, rising to roughly $10–30+/mo if it scales; some need a paid Slack plan
B. Light build — search only, no AI writing — capture knowledge + a bot that surfaces the closest past answer Medium Low–medium — runs on GCP, scales to zero when idle roughly $0–5/mo (GCP free tier; no AI tokens)
C. Full build — search + AI-drafted reply + human approval (the backbone above) Medium–high Medium roughly $0–10/mo (GCP free tier + a few cents of cheap-model tokens at our volume)

Option A (buy). Fastest path, least to maintain. We trade money and some
control for not writing/operating code. Good for a cheap pilot to see if this is
even worth it.

What's actually on the market (June 2026 search):

  • Slack's own AI — paid Slack plans only. Out for us (we're staying free).
  • AnswerHub — best free-tier fit on paper. Has a real $0 plan that's
    per-workspace, not per-user
    (unlimited FAQs, 3 knowledge-base files, 1
    website), with flat paid tiers above ($62 / $110 a month — still not per-seat).
    On the Slack Marketplace. But its knowledge base is hand-curated only
    confirmed from its own site, it answers from uploaded FAQs/files/websites and
    cannot import or learn from our existing Slack messages. With no docs to
    seed it, it starts empty. Its only "growing" mechanism is forwarding a new
    unanswered question to an expert and saving that reply — useful going forward,
    but it never mines the history we already have.
  • Question Base — older write-ups mention a free tier, but its current pricing
    starts at $1,000 for a 2-month trial, then about $6,000/yr per seat. Effectively
    enterprise pricing. Out.
  • Perfect Wiki$9 per editor/month, 5-editor minimum, 3-year billing,
    and Microsoft-Teams-first. Poor fit.

Why this matters for us specifically: every bought tool answers from a
knowledge base we curate (uploaded FAQs/docs) — none can read our existing
channel messages.
Since our knowledge lives only in those messages, a bought
tool would start empty and need us to hand-write FAQs to seed it. A custom
bot can read the channels directly
via Slack's API (works on the free plan,
limited to the 90-day window), which is exactly why building fits our situation
better than buying. Also note the free Slack plan caps us at 10 app
integrations
total.

Option B (light build). No AI writing at all — it just finds and links the
most relevant answer we've already given. Surprisingly effective for FAQ-style
support, and essentially free to run on GCP (Cloud Run scales to zero, so an idle
bot costs nothing). Lowest risk of "the AI made something up."

Option C (full build). Adds a small AI model that phrases a real reply from
the past answers, with a lead approving before it posts. B is a strict subset of
C
— we can ship B first and layer the AI on later only if plain search isn't
good enough. Cost stays in the pennies because we'd use a cheapest-tier model and
only call it when needed.

On model choice: the key is to use a cheap, fast model, not the flagship. The
top-end models (e.g. GPT-5.5 at about $5 in / $30 out per million tokens)
are overkill and pricey for this. The right tier is the fast/cheap class —
Gemini 3.5 Flash ("less than half the cost of other frontier models") or
GPT-5.4 mini/nano (about $0.20–0.75 input per million). At our volume that's
pennies a month. Exact pick is a later detail, not something to lock in now.

Trade-offs worth knowing before we commit

  • The hard part isn't the AI — it's getting clean Q&A out of messy threads.
    Real support threads wander ("nvm, fixed it"). The ✅-react-to-save approach
    works but depends on volunteers tagging good answers.
  • "Free" search still has a tiny cost (embeddings). Pennies, but not zero.
  • Secrets & permissions: we'd manage Slack tokens and decide who can
    approve drafts.
  • Ownership: for a volunteer team, "who maintains this" matters more than the
    bill. The upside: scale-to-zero means an unused bot costs next to nothing.

Suggested direction for discussion (not a mandate)

Because our knowledge lives only in messages (no docs to load into a bought tool):

  1. Build Option B small — a bot that reads the channels, captures resolved
    Q&A before the 90-day deletion, and surfaces the closest past answer. This both
    stops the knowledge bleed and starts deflecting repeat questions, for pennies a
    month.
  2. Add Option C later (AI-phrased replies with lead approval) only if plain
    search proves not good enough — it's a strict add-on to B, not a redo.
  3. Buying (Option A) stays a fallback, worth it only if we'd rather hand-write
    FAQs than write code — but it can't leverage the messages we already have.

Open questions for standup

  • Is there any budget, or must this stay essentially $0?
  • Which channels? Just tech-support, or others?
  • Auto-post vs. approve-first — comfort level?
  • Who owns/maintains it?
  • How fast does a response need to be (minutes? same-day?)?

Metadata

Metadata

Assignees

No one assigned

    Labels

    enhancementNew feature or requeststandupSomething people want to discuss at the next standup

    Type

    No type
    No fields configured for issues without a type.

    Projects

    Status
    Backlog

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions