AI can now write quick workflows on the fly — but what makes a tool trustworthy enough for a whole company to depend on?

This explores the gap between tools people improvise for themselves and software an organization formally adopts and depends on. A note on scope: the collection has no studies of spreadsheets specifically. What it does have is a cluster of recent work on the same divide as it reappears with AI, where models can now generate workflows on the fly. That work sharpens the old question. The difference isn't mainly about how hard the tool is to build. It's about who can see the work, who can check it, and who decided to use it.

Start with the trade-off at the center of improvised tools: they bend easily, but they don't spell out what they do. Studies of AI-generated analysis interfaces found that turning a chat into structured widgets made results clearer but made them harder to change mid-task. The flexible version, meanwhile, forced non-programmers to think like engineers Do generated analysis UIs really work better than chat?. The authors argue this tension can't be avoided, and it maps neatly onto spreadsheets versus enterprise systems. One bends to whatever you need today. The other locks down a specification so many people can rely on it. A related benchmark shows how much can be lost in the move from improvised to specified: generative UI tools quietly failed to build about a quarter of the design rationale they were given Do generative UI tools actually implement their stated design rationales?.

The second difference is reliability you can check. Teams putting AI workflows into production found that flexible, protocol-based tool access (MCP, a standard way for AI agents to connect to outside tools) produced unpredictable failures. They returned to explicit, hard-wired function calls. A survey of 306 practitioners found that 85% of production teams build custom agents rather than using frameworks Why do protocol-based tool integrations fail in production workflows?. Institutionalized software tends to give up flexibility in exchange for repeatable behavior. Auditability works the same way: wrapping a coding agent in a layer that keeps persistent records and an audit trail made its work recoverable without changing the agent itself Can orchestration layers make coding agents more auditable?. A more surprising finding explains why this matters more as tools get better. Weaker models damage documents visibly by deleting content, while frontier models corrupt them silently in ways that look fine on the surface Does model capability change how documents degrade?. Improvised workflows rarely have anyone checking for that kind of damage.

The third difference is organizational rather than technical, and it may be the most important. Evans argues that making tools easier to build doesn't solve the real bottlenecks. Most workers don't see their own tasks as automatable, and enterprise adoption needs decisions that cross departments and budget cycles Does easier tool-building actually solve enterprise adoption problems?. Adoption itself spreads socially. At Microsoft, engineers' peer networks predicted who first tried new AI coding tools better than seniority or tenure did Do social networks drive adoption of new coding tools?. An improvised tool spreads person to person. An institutional one gets there by a decision.

So is there a path between the two? Some work suggests improvised work can be turned into institutional assets without a top-down rebuild. Agent Workflow Memory pulls reusable sub-routines out of past runs and stacks them into larger ones, with gains of 24 to 51% Can agents learn reusable sub-task routines from past experience?. FlowMind generates one-off workflows that only call pre-approved APIs, so the improvisation happens inside a fenced-off area of trusted components Can LLMs generate workflows without touching proprietary data?. That points to an answer you might not have expected. The useful future may not be choosing between the spreadsheet and the enterprise system. It may be an approved set of building blocks that people can improvise with safely. That arrangement has its own risk, though: when workflows are generated on the fly, a crafted prompt can steer how they're put together before any security checks run Can prompts alone reshape multi-agent workflows without system access?.