Can LLMs generate workflows without touching proprietary data?
Explores whether LLMs can orchestrate task automation by composing API calls rather than directly accessing confidential information, and whether this approach preserves security while handling unpredictable tasks.
Robotic Process Automation automates repetitive processes but fails on spontaneous, unpredictable tasks. FlowMind uses LLMs to generate workflows on the fly, with a design built around two constraints critical in finance. First, a generic "lecture" prompt recipe grounds LLM reasoning in reliable APIs — which both mitigates hallucination and, crucially, eliminates direct interaction between the LLM and proprietary data or code, so confidential information never enters the model. Second, it presents high-level descriptions of auto-generated workflows so users can inspect and give feedback. (It also ships NCEN-QA, a finance QA benchmark.)
The keeper is the architectural separation for high-stakes/confidential settings: let the LLM compose calls to vetted APIs rather than touch the data — the model orchestrates, the APIs hold the data and do the work, and the human inspects the plan. This is a security-and-trust pattern, not just a capability one.
This sits in the vault's agentic-tool/workflow thread. The API-grounding-for-security move complements Can codified expertise let non-experts match specialist output? (codify domain knowledge into scaffolding) and the runtime tool patterns in Can agents discover tools dynamically instead of pre-selecting them? — FlowMind generates the workflow up-front and keeps data out of the model.
Inquiring lines that read this note 25
This note is a source for these research framings, grouped by the broader line of inquiry each explores. Scan the bold lines of inquiry; follow any specific question forward.
How much do stated privacy concerns predict actual privacy-protective behavior?- Can differential privacy during generation eliminate leakage at scale?
- Can systems bound what gets disclosed per query without user consent?
- How does grounding LLM reasoning in APIs reduce hallucination in workflow generation?
- Can System 2 oversight prevent information overload in LLM-assisted work?
- Can code generation substitute for dynamic function calling in production workflows?
- How does data residency constraint change the economics of self hosted LLMs?
- Why does pre-computed workflow generation work better than runtime tool discovery for data security?
- How does workflow-level validation reduce false positives from over-tainting sensitive data?
- How does delegated workflow adoption differ from conversational chatbot usage patterns?
- When do employees shift from one-off AI queries to regular workflow integration?
- What task characteristics determine whether delegation is safe for users?
- What mechanisms let users efficiently override or terminate automated services?
- Can natural language permissions be reliably translated into enforceable access controls?
- What privacy-preserving alternatives exist for delegated agent authentication?
- Can deontic specifications replace runtime negotiation of task delegation?
Related concepts in this collection 2
This note in its neighbourhood — explore the map, then jump to a related concept in the list below.
Click a node to walk · click center to open · click Open in graph to see this note in the full knowledge graph
-
Can codified expertise let non-experts match specialist output?
When domain knowledge is captured as explicit rules and principles in an AI agent's scaffolding, can non-experts produce work at expert quality levels without consuming scarce specialist time? This explores whether structured knowledge codification dissolves organizational bottlenecks.
both put capability in the scaffolding/APIs around the model rather than in the model's direct data access
-
Can agents discover tools dynamically instead of pre-selecting them?
Explore whether agents can find needed tools during execution rather than choosing from a fixed set upfront. This matters for long-horizon tasks where relevant tools cannot be known in advance.
contrast: FlowMind generates a workflow up-front with data kept out of the model
Related papers in this collection 8
Papers most semantically related to this note, ranked by cosine similarity in the embedding space.
- FlowMind: Automatic Workflow Generation with LLMs
- Plan-Then-Execute: An Empirical Study of User Trust and Team Performance When Using LLM Agents As A Daily Assistant
- FLOWSTEER: Prompt-Only Workflow Steering Exposes Planning-Time Vulnerabilities in Multi-Agent LLM Systems
- A Practical Guide for Designing, Developing, and Deploying Production-Grade Agentic AI Workflows
- Assessing and Mitigating Data Memorization Risks in Fine-Tuned Large Language Models
- LLMs Corrupt Your Documents When You Delegate
- Autonomy Reshapes How Personalization Affects Privacy Concerns and Trust in LLM Agents
- Design Patterns for Securing LLM Agents against Prompt Injections
Original note title
grounding LLM workflow generation in reliable APIs avoids direct model contact with proprietary data while handling spontaneous tasks