Originally published at mauriciojuba.com
ARTICLE · GENAI · ENTERPRISE
A practical GenAI ideation pipeline for enterprise teams.
GenAI quickly became the industry’s favorite hammer — and suddenly everything looks like a nail. In enterprise environments that turns into a predictable mess: bloated backlogs, vague “AI everywhere” mandates, and solutions that exist mostly to justify the hype.
When I joined a GenAI initiative inside a large corporation, the initial direction wasn’t a product strategy. It was a sentence: “Find use cases across the company where GenAI could be applied.” That sounds proactive. In practice it creates ambiguity, fear and noise.
So we built a pipeline that does something far more valuable than brainstorming: it forces clarity, feasibility and accountability before you invest. That pipeline became the MAGIC framework.
MAP · ANALYZE · GRADE · INPUT · COMMIT
Why GenAI ideation fails in enterprise
Before the framework, stakeholders usually split into two extremes:
FEAR
RESISTANCE
“AI in my area sounds like job risk.”
·Information hoarding
·Defensive positioning
·Political friction
·Delayed collaboration
HYPE
OVERCONFIDENCE
“GenAI can solve everything.”
·Solution before problem
·Ignoring workflow reality
·No cost awareness
·No data constraints
On top of that, many proposed “GenAI use cases” don’t need GenAI at all. Automation, process redesign or better information architecture would solve 80% of them — faster, cheaper, safer. And there was a constraint we couldn’t ignore: token cost made naive experimentation expensive at scale.
Traditional ideation wasn’t enough. We needed a pipeline designed for uncertainty, risk and data reality.
Map
identify where GenAI could matter
Structured discovery — not brainstorming. Where does the work actually break down? Which team owns the workflow? Who are the POCs and decision-makers? What’s the real job-to-be-done behind the request?
OUTPUT
→One problem statement, with context
→One primary user group
→One workflow moment where the pain happens
→A list of stakeholders + POCs
If you can’t map that, you don’t have a use case — you have a suggestion.
Support agents handle 300+ tickets/day across 5 systems.
·What is happening?
·Why does it matter?
·Where in the org does it live?
Case rejection due to missing structured information.
·What event creates friction?
·When does the breakdown happen?
Tier 1 Support Agent.
·Who experiences the pain?
·What are they trying to accomplish?
High cognitive load + repeated manual summarization.
·Which exact step breaks?
·What happens today?
Copy-paste between tools + manual tagging.
- ·What is the workaround?
Support Ops Lead · Salesforce Admin · IT Security.
·Workflow owner
·POC
·Decision maker
·Data owner
Analyze
understand the workflow, not the fantasy
Where GenAI hype meets reality: interviews and shadowing, step-by-step workflow mapping, and finding the friction, waste and decision points. Separate “I don’t like this tool” from “this step is structurally broken.”
OUTPUT
→Workflow map (as-is)
→Pain points tagged by step
→A clear definition of what success looks like
Most “GenAI requests” hide a simpler truth: the process is unclear, not unintelligent.
INTAKEIncoming request created via internal tool.DECISION NODEMultiple intake paths, inconsistent fields.
CASE CREATIONAgent manually structures request.FRICTIONMissing structured information.WASTECopy-paste between systems.
CONTEXT GATHERINGAgent checks 3+ systems for supporting data.FRICTIONHigh cognitive load.WASTERepeated manual summarization.
VALIDATIONSupervisor reviews request.FRICTIONCase rejection due to incomplete data.DECISION NODESubjective approval criteria.
ESCALATIONRequest forwarded to specialized team.WASTERe-entry of previously known information.
DELIVERYFinal output delivered to end-user.FRICTIONTurnaround time unpredictable.
Grade
data readiness + risk & guardrails
The step most teams skip — and why most GenAI initiatives crash later. Can we even do this (do we have the data, is it accessible, structured, reliable, who owns it, how fresh)? And should we (what’s sensitive, what must never leave the boundary, when do we escalate to a human, what are the fail states)?
OUTPUT
→A data-readiness score (even if qualitative)
→A risk classification
→The first real handle on token cost
You can’t estimate cost without understanding inputs, retrieval and output constraints.
·Is the data programmatically accessible?
·Are APIs available?
·Is it locked in PDFs / emails?
·Who owns the data?
·Who can authorize usage?
·Who is accountable for misuse?
·Is the data complete?
·Is it consistent?
·Is there historical noise?
·PII?
·Regulated data?
·Cross-border restrictions?
·Estimated token volume?
·Frequency of inference?
·Long-context requirements?
Input
simulate reality before building
Not testing the UI — testing the clarity of the request. Imagine a powerful but literal executor that does exactly what you ask, no more, no less. If the instruction is vague, the result is vague. Using everything uncovered so far, construct the request with only action verbs.
OUTPUT
- →Not features. Not interfaces. Actions.
If the request hides ambiguity, the output amplifies it.
Commit
build or walk away
Creation is not automatic. Does the data layer support this? Are missing APIs a blocker? Is model capacity enough? Does token cost scale? Is the risk acceptable? Is the outcome measurable?
OUTPUT
→If no → we don’t build. We document the reason and move on.
→If yes → we define the MVP scope deliberately.
A decision, made on evidence — not enthusiasm.
·Data structured?
·Required APIs exist?
·Model capacity sufficient?
·Estimated monthly token cost
·Infra cost
·Opportunity cost
·Compliance exposure
·Escalation coverage
·Failure tolerance
What changed after MAGIC
→Bad use cases died early — and cheaply.
→Good use cases moved faster, because the missing info surfaced quickly.
→GenAI stopped being “a technology initiative” and became a product-decision discipline.
That shift reduces cost, risk and organizational fatigue — the real enemy in enterprise GenAI. If you want to apply MAGIC: start with three candidate use cases and run it end-to-end. If you didn’t kill at least one, your filter is too soft.
TL;DR
Most GenAI ideation fails because it skips data reality and risk. MAGIC forces clarity before investment — turning “we should use AI” into “this problem deserves AI, under these constraints.”
Top comments (0)