Feature
AI social assistant
SignalDeck's AI assistant drafts posts, platform variants, inbox replies, report summaries, and campaign plans for a person to review. Drafts arrive labeled "AI draft", generation runs with tool use disabled so the model can't approve or publish anything, and every content version passes automated checks that block approval when a factual claim has no source or the copy breaks the workspace's brand rules.
What the assistant does
Drafts for people to review, with checks before anything is approved.
Six drafting jobs
Generate a post, generate platform variants without overwriting your edits, suggest an inbox reply, summarize a report from its metrics, analyze performance from synced provider metrics, and propose a campaign plan from a brief. Each runs as a queued job, and the composer or campaign page shows the output when it's ready.
Labeled until a person accepts
Suggestions show in the composer as "AI draft — not reviewed". Nothing lands in a post until someone chooses Accept into draft, and the post still goes through submission and approval like any human draft. Suggested replies are never sent automatically.
Source-grounded generation
Attach source snapshots to a request and the draft is written from them. Source text is passed to the model as fenced, untrusted data, and a snapshot from another tenant fails the job instead of being used.
Unsupported claims block approval
Dates, numbers, quotations, availability, attribution, and impact statements are detected in every content version. Each must match a source snapshot, be confirmed by a reviewer as an owner statement, or be removed; until then the version is review-blocked and can't be approved.
Brand rules enforced at review
The workspace brand profile lists prohibited claims, required disclosures, and words to avoid. Prohibited claims and missing disclosures fail the check; avoided vocabulary raises a warning. Unsupported superlatives and risky claims (guarantees, "risk-free", medical or financial promises) fail too.
Agent plans need approval
Describe a goal and the agent proposes a plan of steps drawn from a fixed tool registry, each with required scopes and a risk level. Plans wait in pending approval; a requester can't approve their own plan without the reviewer permission, and only approved plans execute.
Budgets and policy at execution
Agent execution checks the workspace policy and cost budget first. When a hard-stop budget is used up, execution stops with a clear error. Steps that act on a social provider, such as scheduling or sending a reply, are blocked from agent execution.
Provenance on every output
Each generation records the provider, model, prompt template, brand profile version, source snapshots, and a SHA-256 hash of the output, plus token usage and credits for metering.
How it works
From prompt to approved post, step by step.
1. Request a draft
In the composer, describe what the post should say and pick destinations. The request needs the ai:generate permission and an idempotency key, and it is written to the audit log as a sensitive action.
2. Generation runs as a job
A worker calls the model through SignalDeck's model gateway with no tools attached. If no AI provider is configured, the job fails closed with a clear error rather than returning placeholder text.
3. Accept, then edit
Suggestions appear labeled as AI drafts. Accept one into the draft, edit it, and save. Generated text is a starting point, not an approved post.
4. Checks run on the version
Saving a content version runs factual support, brand voice, near-duplicate (against the workspace's recent versions), superlative, risky-claim, link, character-limit, accessibility, and rights checks per destination.
5. Review and approve
A blocking failure stops approval with REVIEW_BLOCKED. AI review can add findings and suggestions, but it can never clear a failed check or grant approval.
6. Policy changes trigger re-checks
When the brand profile changes, unpublished versions are marked for re-evaluation and can't be approved until the checks run again under the new rules.
Controls
Who can do what.
Permissions per route
Every AI route declares its permission (ai:generate), API scope (posts:write, inbox:write, reports:write, analytics:read), and the separate ai rate-limit bucket, so heavy AI use doesn't throttle publishing.
Kill switch
An organization-level AI generation kill switch stops queued generation jobs without touching publishing.
Owner-approved voice examples
Only an owner or admin can approve the voice examples stored on a brand profile version.
Plans and limits
What each plan includes.
AI credits are included on Agency, Scale, Command, and Enterprise; Starter has no AI credits. Agent workflows are included on Scale, Command, and Enterprise. Image and video generation (the AI creative studio) is included on Agency, Scale, Command, and Enterprise. Credits are metered from the token usage recorded on each job.
Where it fits
Drafting help that stays behind your review.
Agencies
Draft first passes and per-network variants for each client workspace, then route them through internal and client approval. The brand profile of each workspace drives its checks.
Brand teams
Generate launch copy from attached source material, so every date, figure, and quote is tied to a snapshot before legal or comms review.
Creators
Turn a prompt into a draft and platform variants in the composer, then edit in your own words before scheduling.
Nonprofits
Impact figures and other factual claims in an appeal must be tied to a source snapshot or confirmed by a reviewer before approval.
Plan limits
AI credits per month.
- Agency
- 1,000
- AI credits per month
- Scale
- 5,000
- AI credits per month
- Command
- 15,000
- AI credits per month
- Enterprise
- Custom
- AI credits per month
FAQ
Frequently asked
Can the AI publish or reply on its own?
No. Generation runs with tool use disabled and any approval or publish fields in model output are dropped. Drafts and suggested replies need a person to accept them, and posts still need approval before they're scheduled.
What happens if a draft contains a claim we can't back up?
The content version is review-blocked. Each detected claim (a date, number, quote, availability, attribution, or impact statement) has to match an attached source snapshot, be confirmed by a reviewer as an owner statement, or be removed before anyone can approve it.
Does it learn our brand voice automatically?
Not on its own. The brand profile you write (prohibited claims, required disclosures, vocabulary, and owner-approved voice examples) is what the review checks enforce.
Which plans include AI?
AI credits are included on Agency, Scale, Command, and Enterprise; Agency is the first plan with credits and Starter has none. Agent workflows start on Scale and are included on Scale, Command, and Enterprise.
What can the agent do?
It plans with a fixed set of tools, such as listing workspaces, creating post drafts, reading metrics and pending approvals, drafting replies, and generating report drafts. A plan with an unsupported action is blocked, plans must be approved before they run, and steps that act on a social provider are blocked from agent execution.
Can another client's sources leak into a draft?
No. Source snapshots and recent posts used for checks are filtered to your tenant, and a generation request that references another tenant's snapshot fails instead of running.
What isn't supported?
Fully autonomous posting, auto-sent replies, and agent steps that act on a social provider aren't supported. Rejected drafts aren't fed back into memory automatically, and Starter has no AI credits.
Drafts with guardrails
Let AI write the first draft, and keep review in charge.
AI credits start on Agency. Book a walkthrough to see the checks on your own content.