1. Ground
Give AI source material, retrieval, and context that match the task.
The workload library
Twenty-one AI workloads to assess against your own process.
Every sample below is classified four ways so you can choose with discipline: is the work deterministic or non-deterministic, when does a human stay in the loop, what can AI safely do today, and how ready is it. The spine underneath all of them is the same: the determinism bridge.
Non-deterministic (the AI half)
Useful for language and judgment support. Define the quality gate and catch errors before they affect customers or records; an unexplained accuracy percentage is not a deployment standard.
- Draft, summarize, classify, qualify
- Answer from approved sources (RAG)
- Recommend next best action
- Detect sentiment, risk, anomalies
The determinism bridge
Deterministic (the systems half)
Rules validate and execute approved steps. Reliability still depends on correct inputs, permissions, error handling, and reconciliation.
- Validate, approve, calculate
- Write the record, schedule the task
- Enforce permissions and policy
- Log evidence; escalate exceptions
2. Constrain
Allowed actions, forbidden zones, confidence thresholds, required citations.
3. Review
Human approval for money, public posting, deletion, legal, security, irreversible actions.
4. Execute
Deterministic workflows handle API calls, writes, calculations, records.
5. Log
Prompt version, input, output, user, model, cost, evidence, corrections.
6. Improve
Turn misses into better prompts, cleaner data, sharper escalation.
How to read each workload
Read the proposed boundaries before choosing a pilot.
Type
Non-det probabilistic AI
Det rules-based system
Bridged AI + systems together
Human in the loop
Before action review before the reply, send, or record write
After action audit a sample only within approved low-risk limits
Decision owner a person makes the consequential decision
What AI can do
Execute within limits perform only the actions explicitly approved for this workflow
Recommend propose, human chooses
Draft human saves/sends
Gated never alone
Readiness
Candidate common starting point, subject to your readiness checks
Pilot needs additional proof before live use
Hold not yet / high risk
Readiness is specific to your business.
Before a live pilot, confirm: a mapped process and named operator; permitted, current data; tool access limited to the job; tested errors and exceptions; a human review or escalation path; a baseline and acceptance criteria; and a way to pause or roll back. The badges describe proposed patterns, not certification that a workload is safe to deploy in every setting.
For example, “Execute within limits” with “Before action” means AI prepares the work but cannot execute until the operator approves it. “After action” applies only to explicitly approved low-risk actions.
Featured walkthroughs
Six workloads, in detail
These six are illustrative workflow designs; account recovery requires additional security review. Each needs a named operator, current sources, tested permissions, and a measurable baseline before live use. The full sample walkthroughs are here.
The catalog
Fifteen more, ready to map to your business
Same four tags. Read the badges before you buy a tool; they tell you where the human stays, what AI may touch, and whether to deploy or pilot.
Readiness
AI can
Productivity
Meeting notes & actions
Auto minutes with owners and due dates from a transcript.
Non-detBefore actionDraftCandidate
Productivity
Email triage & drafts
Sort, prioritize, and draft replies; you send.
Non-detBefore actionDraftCandidate
Data
CRM enrichment & hygiene
Fill and tidy low-risk fields; flag conflicts.
BridgedAfter actionExecute within limitsPilot
Legal & ops
Contract / doc summary
Summarize terms and risks; never the final word.
Non-detDecision ownerRecommendPilot
Sales
Proposal / RFP drafting
First-pass proposal from your library and the brief.
Non-detBefore actionDraftCandidate
Enablement
Knowledge-base Q&A (RAG)
Prepare sourced answers; approve before replies and escalate gaps.
Non-detBefore actionExecute within limitsPilot
Support
Ticket classify & route
Propose a route within approved limits; audit corrections.
Non-detAfter actionExecute within limitsCandidate
Customer
Sentiment & churn risk
Flag accounts for human review and follow-up.
Non-detDecision ownerRecommendPilot
Finance
Invoice / receipt extraction
Pull line items; approve before posting.
BridgedBefore actionDraftPilot
Executive
Weekly operating report
Data extract + variance commentary; leader approves.
BridgedBefore actionDraftCandidate
People
Onboarding / training content
Role-specific guides; an SME reviews.
Non-detBefore actionDraftCandidate
Marketing
Content from voice / notes
Turn a founder's voice into repeatable drafts.
Non-detBefore actionDraftCandidate
People
Resume / candidate screening
Surface a shortlist; never reject alone, bias-checked.
Non-detDecision ownerRecommendHold
Finance & CX
Refund / credit decisions
AI prepares the case; a human approves; rules execute the approved amount.
BridgedDecision ownerGatedHold
Finance & ops
Forecast / anomaly flagging
Flag outliers for an analyst to confirm.
Non-detBefore actionRecommendPilot
The outcomes map
Every workload runs the same path
Trigger to outcome, AI and systems each doing what they're good at, with the human gate exactly where risk lives.
Use it like this: pick one Candidate workload with a clear action boundary, define the human gate before you build, measure the value, then graduate the next one. Keep everything marked Gated/Hold behind a human until trust and data are earned.
Go deeper
Want more examples?
This is a starter set of twenty-one. Want the full map? The Skynet Wars AI Capability Index charts the AI capability domains by category, with maturity signals and the leading tools in each.