The domain is a distraction
Scroll any job board and the postings blur together: “build agentic workflows” for support, for finance, for security, for healthcare. It reads like fifty different jobs. It isn't. Underneath, nearly every agent runs the same loop — perceive → plan → act → observe → repeat — wired to different tools. What changes the engineering is not the industry. It's two questions:
- How much autonomy does it get? Does a human approve each action, or does it act on its own?
- What does a wrong action cost? Can you undo it with a click, or does it move money, deny care, or halt a production line?
Plot those two axes and the fifty use cases sort themselves into four zones — each with a different build strategy. Here is the map.
The map: autonomy × cost of error
Each numbered dot is a use case from the table below; color marks its zone. Positions reflect a typical production deployment — the same task can move as you dial autonomy up or down.
Ship first — assistive · reversible
Low stakes, human stays in control. The place to earn trust and learn your eval. Draft, don't do.
Let it run — autonomous · reversible
Cheap to be wrong, so let it act. ROI comes from volume; add a circuit-breaker and monitor drift.
Human in the loop — assistive · costly
The human is the guardrail. Optimize recall + explanation; make review fast and auditable.
Danger zone — autonomous · irreversible
Autonomy over costly actions. Spend the most on verifiers, rollback, and tight write-scoping — or don't automate it yet.
The 50 use cases
Twelve domains, one loop each. A is autonomy and C is cost of a wrong action, both on a 0–10 scale for a typical deployment. Click any row (or the icon) for a one-line explanation, the retrieval it needs, and a hover tooltip.
| # | Use case | Domain | A | C | Zone | RAG | Main failure mode | Info |
|---|---|---|---|---|---|---|---|---|
| 1 | Support ticket resolution | Customer | 5 | 4 | Ship first | RAG-core | Wrong action on an account | |
| 2 | Live-chat deflection | Customer | 6 | 3 | Let it run | RAG-core | Bad answer at scale | |
| 3 | Voice IVR agent | Customer | 6 | 4 | Let it run | RAG-support | Mishears intent | |
| 4 | Returns & refunds | Customer | 4 | 6 | Human in loop | RAG-support | Wrongful refund | |
| 5 | Community moderation | Customer | 7 | 4 | Let it run | minimal | Over- / under-blocking | |
| 6 | Personal shopping assistant | Customer | 4 | 5 | Ship first | RAG-core | Buys the wrong item | |
| 7 | Bug-fix / PR agent | Software | 4 | 5 | Ship first | RAG-core | Plausible-but-wrong diff | |
| 8 | Code review | Software | 3 | 4 | Ship first | RAG-support | Misses a real bug | |
| 9 | Test generation | Software | 5 | 3 | Ship first | RAG-support | False-green tests | |
| 10 | Documentation generation | Software | 5 | 2 | Ship first | RAG-support | Stale / wrong docs | |
| 11 | SRE incident remediation | IT / Ops | 6 | 9 | Danger zone | RAG-core | Worsens the outage | |
| 12 | Access provisioning | IT / Ops | 5 | 6 | Human in loop | RAG-support | Over-grants access | |
| 13 | Cloud-cost optimization | IT / Ops | 6 | 6 | Danger zone | minimal | Breaks prod to save $ | |
| 14 | Log / observability triage | IT / Ops | 7 | 3 | Let it run | RAG-support | Buries a real signal | |
| 15 | SOC alert triage | Security | 6 | 7 | Danger zone | RAG-core | Missed breach | |
| 16 | Vuln scan → patch PR | Security | 4 | 6 | Human in loop | RAG-support | Bad patch | |
| 17 | Phishing detection & response | Security | 7 | 6 | Danger zone | RAG-support | Blocks legit mail | |
| 18 | Threat-intel enrichment | Security | 8 | 2 | Let it run | RAG-core | Noise | |
| 19 | SDR research & outreach | Sales / Mktg | 5 | 4 | Ship first | RAG-support | Spammy outreach | |
| 20 | Lead scoring & routing | Sales / Mktg | 7 | 3 | Let it run | minimal | Misroutes a hot lead | |
| 21 | Marketing content | Sales / Mktg | 4 | 4 | Ship first | RAG-support | Off-brand copy | |
| 22 | SEO optimization | Sales / Mktg | 6 | 3 | Let it run | RAG-support | Ranking hit | |
| 23 | Ad creative & A/B | Sales / Mktg | 5 | 5 | Ship first | minimal | Wastes ad spend | |
| 24 | Competitive-intel monitoring | Sales / Mktg | 8 | 2 | Let it run | RAG-core | Stale intel | |
| 25 | Social-media management | Sales / Mktg | 5 | 5 | Ship first | RAG-support | Brand blowup | |
| 26 | Accounting close / reconciliation | Finance | 4 | 8 | Human in loop | minimal | Misstated books | |
| 27 | AP / invoice processing | Finance | 5 | 7 | Human in loop | minimal | Duplicate / wrong payment | |
| 28 | Expense-report audit | Finance | 6 | 4 | Let it run | RAG-support | Missed fraud | |
| 29 | Financial-report analysis | Finance | 3 | 6 | Human in loop | RAG-core | Wrong number cited | |
| 30 | Fraud / transaction review | Finance | 6 | 8 | Danger zone | minimal | Blocks a good customer | |
| 31 | Trading research / signals | Finance | 3 | 7 | Human in loop | RAG-core | Bad signal → loss | |
| 32 | Contract review & redline | Legal | 3 | 7 | Human in loop | RAG-core | Missed liability clause | |
| 33 | KYC / AML onboarding | Legal | 4 | 8 | Human in loop | RAG-core | False clear → fine | |
| 34 | Regulatory-change monitoring | Legal | 7 | 4 | Let it run | RAG-core | Missed rule change | |
| 35 | eDiscovery / doc review | Legal | 4 | 6 | Human in loop | RAG-core | Missed privileged doc | |
| 36 | Patent / prior-art search | Legal | 4 | 5 | Ship first | RAG-core | Missed prior art | |
| 37 | Recruiting sourcing & screening | HR | 4 | 6 | Human in loop | RAG-core | Bias / bad rejection | |
| 38 | Employee onboarding automation | HR | 6 | 5 | Let it run | RAG-support | Broken access day 1 | |
| 39 | Internal HR helpdesk | HR | 6 | 3 | Let it run | RAG-core | Wrong policy answer | |
| 40 | Performance-review drafting | HR | 3 | 5 | Ship first | RAG-support | Unfair review | |
| 41 | Clinical documentation / scribe | Healthcare | 4 | 7 | Human in loop | RAG-support | Wrong chart note | |
| 42 | Prior authorization | Healthcare | 4 | 7 | Human in loop | RAG-core | Denied care | |
| 43 | Medical coding & billing | Healthcare | 5 | 7 | Human in loop | RAG-core | Miscoded claim | |
| 44 | Pharma literature research | Healthcare | 4 | 5 | Ship first | RAG-core | Missed study | |
| 45 | Insurance claims processing | Industry | 5 | 8 | Human in loop | RAG-support | Wrong payout / fraud | |
| 46 | Supply chain / procurement | Industry | 5 | 7 | Human in loop | RAG-support | Overcommit spend | |
| 47 | Manufacturing QA / control | Industry | 9 | 9 | Danger zone | minimal | Safety / scrap | |
| 48 | Real-estate due diligence | Industry | 3 | 6 | Human in loop | RAG-core | Missed title issue | |
| 49 | Personalized tutoring | Education | 5 | 4 | Ship first | RAG-core | Reinforces an error | |
| 50 | Travel planning & booking | Travel | 5 | 6 | Human in loop | minimal | Non-refundable mistake |
Which ones actually need RAG?
The RAG column above answers a question we get constantly: does this agent need retrieval-augmented generation, or not? The honest answer is that it splits three ways — and the split doesn't follow the domain. What matters is where the ground truth lives.
RAG-core — retrieval is the task (22)
The answer lives in a large, changing, unstructured corpus that won't fit in context — a clause library, a code base, filings, résumés, the litigation document set. Retrieval quality is the ceiling on quality. Contract review, financial-report analysis, eDiscovery, prior-art search, and recruiting screening all live here.
RAG-support — grounds tone, policy, or context (19)
The task is driven by structured tools or generation, and retrieval fetches a policy, a style guide, or nearby context to keep the output grounded — code review, marketing content, clinical scribe, expense audit.
Minimal — no corpus to retrieve (9)
The truth is structured (query the DB — accounting, fraud, lead scoring), streaming (sensor, log, or market data — manufacturing QA, travel search), or generated fresh (ad creative). Reaching for a vector store here just adds latency and a failure mode.
The rule: you need RAG when the grounding knowledge won't fit in context and can't be reached by a precise API call. Two caveats — many core cases are hybrid (support and claims retrieve policy docs and hit structured order systems), and code retrieval is RAG in mechanism even though the corpus is code, not prose.
What every one of them shares
Strip away the domain and the skeleton is identical. If you've built one well, you've built the shape of all fifty.
The value is tool use, not chat
The model is the reasoning glue; the leverage is in the API, DB, and file calls it makes. None of these are a chatbot.
State, memory, and a stopping rule
Every one needs working memory of the task, often long-term memory of history, a clear definition of “done,” and an escalation path when confidence drops.
Evaluation is the bottleneck, not generation
Generating a plausible action is easy. Verifying it was the right one — with tests, guardrails, ground-truth checks, or a human — is the hard engineering everywhere.
The same failure family
Confident-but-wrong actions, errors that compound over a long horizon, and brittle tool integrations. The mitigations differ; the failures don't.
What actually makes them different
Autonomy and cost of error are the headline axes, but a handful of secondary dimensions decide the rest of the architecture:
| Axis | Low end | High end | Shows up as |
|---|---|---|---|
| Reversibility | Undo with a click | Irreversible / physical / regulated | How much rollback you build |
| Latency budget | Minutes–hours OK | Real-time / sub-second | Model size, caching, streaming |
| Horizon length | 1–3 steps | Long multi-stage plans | Planning, memory, checkpointing |
| Read vs. write | Read-only synthesis | Mutates money / records / hardware | Write-scoping, approvals |
| Determinism | “Good enough” answer | Auditable, reproducible | Logging, eval, temperature |
| Topology | Single loop | Orchestrated specialists | Multi-agent + a verifier |
| Regulatory weight | None | HIPAA / SOX / GDPR / AML | Audit trails, PII handling |
A useful rule of thumb: engineering effort scales with (cost of a wrong action) × (degree of autonomy). That product — the diagonal toward the top-right of the map — tells you how much you'll spend on verifiers, guardrails, and human-in-the-loop design. The domain barely enters into it.
Why the map is really a budget
The top-right of this map is where autonomy meets irreversibility — and it's exactly where a workflow earns its keep. A single monolithic model can generate an action, but it can't check its own work; a workflow can spend a real verifier — execute the tests, re-derive the number, get a second model to refute the first — before anything touches the world. We've argued that a workflow is the only thing that can use a verifier, and that's the whole reason the danger zone is buildable at all.
But verifiers, extra agents, and review loops cost tokens and latency — so the answer isn't “always build the biggest workflow.” It's to find the cheapest topology that still clears your quality bar at the autonomy and cost-of-error your use case demands. That is precisely what AutoAW searches for: it co-evolves topology, prompts, models, and tools to land the right point on the frontier for your task — a cheaper, careful agent where that suffices, a verifier-heavy compound system where the stakes require it.
Have a use case on this map?
Tell us where it lands — autonomy and cost of error — and we'll help you find the workflow that fits it.