Getting Started

Core Concepts

Understand the building blocks of Datashift.

Queues

A queue defines a review workflow. Think of it as an inbox for a specific type of decision.

Review Type— What kind of decision reviewers make
Assignment Strategy— How tasks are distributed to reviewers
Reviewers— Human reviewers, AI reviewers, or both

Example: "Refund Approvals", "Content Moderation", "Contract Reviews" — each with different reviewers and decision types.

Tasks

A task is a unit of work submitted for review. When your agent needs human input, it creates a task.

Data— The content to be reviewed
Context— Background info for reviewers
Summary— Short description shown in the queue

Tasks progress through states:

PendingQueuedReviewed

Reviewers

Reviewers make decisions on tasks. Datashift supports two types:

Human Reviewers

Team members who review tasks via the Console. Can also receive notifications via Slack, Discord, or Teams.

AI Reviewers

LLM-based reviewers that process tasks automatically. Configure with a model, prompt, and parameters.

Review Types

The review type determines what kind of decision reviewers make.

TypeDescriptionExample
ApprovalYes/no decisionApprove or reject a refund
ClassificationSingle category selectionBug, Feature, Support, Spam
ScoringNumeric ratingLead quality 1-5
LabelingMultiple labelsurgent, customer-facing, security
AugmentationEdit the contentFix AI-drafted copy before sending

Assignment Strategies

Control how tasks are distributed to reviewers.

StrategyDescriptionBest For
ManualTasks sit in pool; reviewers self-assignSmall teams, variable workloads
Round RobinAuto-assigns to next reviewerEven workload distribution
AI FirstAI reviews first, then humanHigh volume, AI pre-screening
AI LastHuman reviews first, AI audits afterwardQuality assurance, compliance checks

Reviews

A review is a decision made on a task. Each review contains:

Result— The reviewer's decision
Reviewer— Who made the decision
Feedback— Optional explanation or notes

Tasks can have multiple reviews (e.g., in AI First mode, the AI reviews first, then a human reviews).

Reviewer Channels

Reviewers can make decisions from wherever they already work. Datashift supports multiple review channels so reviewers don't need to learn a new tool.

Console

Full-featured web dashboard with dedicated review UIs for each review type. Keyboard shortcuts for fast reviewing. Best for high-volume or complex reviews.

Slack

Review tasks directly from Slack using interactive modals. Receive notifications when tasks are assigned and submit decisions without leaving the conversation.

Discord

Button-based review interactions in Discord channels or DMs. Reviewers click to approve, reject, or open the full review in the Console.

Microsoft Teams

Adaptive Card-based reviews in Teams channels. Supports tenant-level security verification and threaded conversations.

Each reviewer can configure their preferred notification channels in their profile settings. Reviews from any channel are recorded identically — the task doesn't know or care where the decision came from.

AI Reviewers

AI reviewers are LLM-based agents that process tasks automatically using structured output. They can work alongside human reviewers or handle tasks independently.

Provider— OpenAI (GPT-4o) or Anthropic (Claude)
Prompt— Instructions for how the AI should review tasks
Parameters— Temperature, max tokens, and model selection

AI reviewers are most powerful when combined with human reviewers using assignment strategies:

AI First

The AI reviews every task first. If it's confident, the result is used directly. If not, the task is escalated to a human reviewer. Reduces human workload by handling straightforward cases automatically.

AI Last

A human reviews first, then the AI audits the human's decision. Useful for compliance checks and quality assurance — the AI flags disagreements for further review.