CivicAI for staff

CivicAI is the AI intelligence layer of CivicOS. This page covers the four AI surfaces available to staff — representatives, government admins, NGO staff, moderators, and platform admins.

Every CivicAI output is a suggestion or draft. Nothing publishes, assigns, or decides on its own. Every AI-generated string in the UI carries an AI-generated · review badge so you always know what came from CivicAI and what came from a human.

The four surfaces

Surface	Where	For
Summarize discussion	Petition, issue, and consultation detail pages	Reading a long thread quickly.
Draft with AI	New announcement page (/org/:id/announcements/new)	Turning a brief into a publishable announcement.
Community Intelligence	Top of the organization dashboard (/org/:id)	Understanding what's happening across your whole community.
Analytics Narrator	Top of the admin Overview page (/admin, admins only)	Reading the platform's health without staring at charts.

Summarize discussion

When a petition, issue, or consultation has real conversation attached, you'll see a Summarize discussion button. Click it and CivicAI reads the entire thread (up to 200 comments, oldest first) and produces:

• A one-paragraph TL;DR you can read in ten seconds.
• Themes — the recurring topics people are raising.
• A sentiment mix — positive · neutral · negative across the whole thread.
• Top asks — what the community is specifically requesting, in their own words where possible.
• Recommended actions — 2–4 imperative-mood suggestions ordered by impact.
• Official responders — who from your side has already engaged.

The summary is cached for 30 minutes. Clicking again within that window returns the same summary instantly — you'll see a Cached badge instead of Fresh. This keeps costs down and lets you revisit the summary while you draft a response.

Consultations are supported too. When you summarize a consultation, CivicAI reads the structured responses (Q&A) rather than free-form comments, but the output is the same shape.

Who sees the button. Representatives, government admins, NGO staff, moderators, and platform admins. Citizens don't see it — this is decision-support for people who act on the discussion.

Draft with AI

On the New announcement page, above the title/body form, you'll find a Draft with CivicAI panel. Type a rough brief (at least 20 characters — the more concrete, the better), pick a tone and audience, and click Draft with AI.

CivicAI returns:

• A short title (max 80 characters, no emoji).
• A body — 2 to 5 short paragraphs in the tone you asked for.
• A few key points — scannable one-liners covering what changed, what you're doing, and what the reader should do or expect next.

Click Apply to form to fill the title + body fields on the actual announcement form. Both fields get an AI · edit to remove chip that disappears the moment you edit — so you always know at a glance what's still AI-generated and what you've made your own.

Tones.

Tone	Use when
Formal	Policy changes, official notices, regulatory items.
Friendly	Community events, milestones, weekly updates.
Urgent	Time-sensitive service disruptions, safety notices.
Empathetic	Announcements about outages, apologies, disruptions.

Audiences.

• General public — assume the reader may not know your org.
• Members only — assume they already do.

Ground rules the model follows. CivicAI is instructed never to invent facts, dates, quotes, phone numbers, or statistics you didn't put in the brief. If you don't mention a repair date, the draft won't promise one. The output is only as specific as your brief.

Drafts are not cached — a second click on the same brief gives you a fresh variation to compare.

Community Intelligence

On your organization dashboard (/org/:id), the Community Intelligence tile lets you generate an aggregate digest of what's happening across your whole community — not one thread, but the story the numbers + discussions tell together.

Click Generate insights and CivicAI pulls recent issues, petitions, and their comments from your community (capped at 25 issues, 15 petitions, 12 comments per thread), analyzes them, and returns:

• Activity counts — issues, petitions, comments in scope.
• A one-paragraph TL;DR — the top story right now.
• Recurring themes across all items.
• An overall sentiment mix.
• Top asks across the whole corpus.
• Recommended actions — 3–5 next steps ordered by impact.

The tile picks the community from your active community (or your primary community, if you haven't switched). If you don't have a community set, the tile prompts you to pick one first.

Cached for 1 hour — because community-wide themes don't flip in an afternoon.

Analytics Narrator (platform admins)

At the top of the admin Overview page (/admin on the admin console), platform admins have a CivicAI narrator tile. Click Narrate the numbers and CivicAI reads the same metrics shown below (users, communities, issues, petitions, moderation queue) and returns:

• A one-sentence headline — the top story in the numbers.
• A narrative — 2–3 short paragraphs weaving specific numbers into sentences.
• Highlights — 3–5 standout figures worth surfacing.
• Trends — where things are improving, plateauing, or slipping.
• Recommendations — 2–4 next steps grounded in what the numbers say.

Cached for 15 minutes. This is the surface that turns "here are 40 counters" into "here's what's happening on the platform right now."

Campaign surfaces (Community Funding)

Surfaces for organizations running a fundraising campaign. All of them are drafts you edit — nothing is written into a form, or published, until you click Use this.

Surface	Where
Draft campaign	New campaign page (/org/:id/campaigns/new)
Draft donor update	The update composer on your campaign page
Draft final report	The "File final report" form

Two more exist as API endpoints without a button yet: classify campaign (suggesting a category) and campaign impact (a plain summary of where a campaign stands).

Three things about these are worth knowing, because money makes them different from the announcement drafting above.

CivicAI knows what you reported, not what you spent. Donations settle straight into your bank account — CivicOS never holds the money and cannot check any of it. So every draft describes your spending as what your organization reports, never as verified fact. It will not write that something has been confirmed, because nothing has been.

It will tell you what is missing. Campaign drafts come with warnings, shown above the draft itself: costings a reviewer will ask for, claims your published record does not support, a goal that does not match the described work. Hearing that from a draft is cheaper than hearing it from a rejection two days later. The impact summary similarly reports gaps — money raised but not accounted for, milestones with no progress, long silences. That is deliberate. A summary that flattered a campaign which had gone quiet would be worth nothing to a donor deciding whether to give again.

It never does the arithmetic on unexplained money. In a final report draft, the amount still unaccounted for is calculated from the ledger, not written by the AI. That figure is frozen onto your public campaign page the moment you file, and it does not change afterwards even if you publish more spending later — so it is not a number anyone should be generating.

What CivicAI does not do

• It doesn't auto-publish. Every draft, summary, insight, and narration requires a human click before anything ships.
• It doesn't route or assign. Assignment still runs through the normal community + organization flow.
• It doesn't send messages. No notifications, no emails, nothing on behalf of the org.
• It doesn't remember previous chats. Each call is independent — no conversation history, no personalization, no user profile.
• It doesn't decide anything about money. No campaign is approved, rejected, paused, or prioritised by CivicAI. Platform admins have an advisory tool that flags campaigns worth opening first in a review queue; it cites the evidence behind every observation, offers the innocent explanation alongside it, and changes nothing. A person makes every decision.

Trust and audit

Every AI response includes the model name and the timestamp it was generated at. Every UI surface that renders AI output shows an AI-generated · review badge until a human has explicitly edited or approved the content.

Full audit persistence (ai_generations table with prompt hash, tokens, latency per call) is on the roadmap.

Rate limits

CivicAI endpoints share the same "Standard authoring" rate-limit budget as the org authoring endpoints. If you hit the limit while iterating on a draft, you'll see the standard rate-limit toast — wait a minute and try again.

Costs and provider

CivicAI runs on Google's gemini-flash-latest by default (an evergreen alias so the platform always uses the most current fast tier). Model selection is a deploy-time config, not a per-request choice — see the developer guide for details.

Coming next

Roadmap features you'll see later:

• AI-drafted responses on issues — same pattern as announcement drafting, but on every discussion thread.
• Multi-channel announcements — SMS / social / email variants alongside the main body.
• Multilingual — Yoruba, Igbo, Hausa, Pidgin.
• Duplicate detection at report time — reduce noise before it reaches your inbox.