AI Use Case Identification

Key Facts

• Analysis of 600+ enterprise use cases shows most reduce to 6 fundamental primitives
• Impact/Effort scoring should be reassessed quarterly — high-effort use cases today may become low-effort as models improve
• AI works best for repetitive high-volume tasks and skill-bottleneck tasks where AI bridges knowledge gaps
• 92% of enterprises plan to increase AI spending; only 1% feel they've achieved AI maturity
• Governance, security, and cost efficiency are the decisive factors for scaling beyond pilots — not model capability
• Piloting phase (39% → 13% of organisations) is over: enterprises are now failing at the scaling problem

The 6 Use Case Primitives

Analysis of 600+ enterprise AI deployments shows most use cases reduce to one of six patterns. Knowing the primitive lets you reuse solutions across business units and estimate effort more accurately.

[Source: OpenAI Identifying and Scaling AI Use Cases, 2025]

Each primitive applies across every department. The same summarisation primitive works for legal, finance, HR, and engineering. Build the primitive once; deploy it across business units.

Agentic execution is the hardest primitive to deploy (highest governance requirement, most failure modes) and the highest potential value. Start with the other five first.

The Impact/Effort Framework

Score each candidate use case on two axes, plot on a 2x2 matrix, and prioritise.

Impact factors (value delivered)

• Revenue impact or cost reduction magnitude
• Number of people or transactions affected
• Frequency — daily tasks beat quarterly ones
• Criticality to core business process

Effort factors (implementation cost)

• Data availability and quality
• Integration complexity with existing systems
• Change management and adoption risk
• Compliance and governance overhead

Prioritisation matrix

Reassess quarterly. A use case rated high-effort in 2024 may become low-effort in 2026 as foundation models improve. The framework is not static.

Where AI Works Best

Repetitive high-volume tasks

Tasks performed many times per day, currently done by humans following a fixed process, with well-defined success criteria. AI delivers consistent performance without fatigue.

Signals this is the right fit:

• The task is documented (there's a runbook or SOP)
• Success is measurable without human judgment
• Volume is high enough that small efficiency gains compound

Skill-bottleneck tasks

Tasks where specialist knowledge is scarce and waiting for review creates a bottleneck. AI can provide 80% of the value at 1% of the specialist's time.

Signals this is the right fit:

• A specific expertise type (legal, medical, technical) is the constraint
• Generalists currently wait on specialists for knowledge they need frequently
• The specialist's time is better spent on edge cases, not routine applications

Tasks that don't fit well

• Creative work requiring genuine novelty (AI assists; does not replace)
• Tasks requiring legal accountability (human must remain in the loop)
• Extreme precision requirements with no acceptable error rate
• Tasks where explainability of the decision is required by regulation

The Scaling Bottleneck

Finding use cases is easy. Scaling them is where most enterprises fail.

Common failure modes when scaling:

1. Data access — the use case needs data that IT or legal won't expose to the AI system
2. Governance — no approval process for AI-generated outputs in regulated workflows
3. Integration — AI output can't feed back into existing systems of record
4. Change management — employees won't adopt the new workflow without investment in training and incentives

Governance, security, and cost efficiency have become the decisive factors. Not model capability. See landscape/enterprise-ai-adoption for the full governance picture.

Running an AI Discovery Sprint

A practical framework for identifying use cases in a business unit in one week:

1. Interview users — what tasks are repetitive, frustrating, or require knowledge you don't have?
2. Map to primitives — categorise each candidate as one of the 6 primitives
3. Score impact/effort — apply the 2x2; identify quick wins and future investments
4. Pick one quick win — build one high-impact/low-effort use case to demonstrate value
5. Document the pattern — what data, integration, and approval process was needed
6. Replicate — apply the same primitive across other business units before moving to harder primitives

The quick win serves as internal proof of concept and builds the trust needed to tackle higher-effort use cases.

For the AI Engineer

When advising a business on what to build, this framework makes you more credible than model recommendations:

• Leads with business value, not technology capability
• Produces a prioritised backlog, not an open-ended wish list
• Sets realistic expectations about what's easy vs what requires real investment
• Identifies governance requirements early, before they become blockers

The skill set to run this: stakeholder interviews, use case pattern recognition, and the ability to estimate technical effort. This is as much a software engineering skill as an AI skill.

Common Failure Cases

Selecting an agentic execution use case as the first deployment, before any simpler primitives have been validated Why: agentic execution has the highest governance requirements, the most failure modes, and the steepest integration complexity; teams attracted to the highest-value primitive underestimate the change management and trust-building required before an organisation will let an agent act autonomously. Detect: the pilot is late, over budget, or blocked on legal/IT approval; the team is building governance processes from scratch in parallel with the product; stakeholders are nervous about autonomous actions. Fix: use the 6-primitive ordering deliberately — pilot summarisation, classification, or extraction first; use the quick win to build organisational trust and the governance infrastructure that agentic execution will require.

Impact/Effort scores are set once and never reassessed, causing high-effort use cases to remain deprioritised after model improvements make them viable Why: the framework depends on effort estimates that are a function of current technology; a use case rated high-effort in 2024 because of poor extraction accuracy may be low-effort in 2026 after model quality improvements, but the backlog is never revisited. Detect: the prioritisation matrix still shows use cases from 18+ months ago in the "high effort, reassess later" quadrant; the team has not updated effort scores since the initial assessment. Fix: schedule quarterly backlog reviews where effort scores are re-estimated; specifically re-evaluate any use case that was blocked on model capability, data quality, or integration complexity — all three evolve rapidly.

Discovery sprint interviews only the loudest internal advocates, missing the use cases with the highest actual volume Why: interview-based discovery surfaces the use cases that motivated, articulate employees want to talk about; it systematically misses the high-volume repetitive tasks that workers have stopped noticing because they are so routine. Detect: the use cases identified are all creative or knowledge-intensive tasks; no high-frequency, low-complexity tasks appear in the backlog despite the organisation processing millions of routine transactions. Fix: supplement interviews with process mining — analyse system logs, ticket queues, and workflow data to find the highest-frequency human tasks; the best quick wins are often tasks that nobody thought to mention because they seem too simple.

Connections

• landscape/enterprise-ai-adoption — the broader adoption context and governance gap
• agents/practical-agent-design — agentic execution primitive in depth
• evals/methodology — how to measure whether a use case is actually delivering value
• synthesis/rag-vs-finetuning — which AI approach suits which use case type
• synthesis/llm-decision-guide — model and infrastructure selection per use case
• synthesis/cost-optimisation — cost structure when scaling use cases
• synthesis/software-engineer-to-ai-engineer — SE skills that transfer to use case analysis

Open Questions

• Do the 6 primitives hold across all industries, or do vertical-specific patterns emerge?
• How do you pick the right quick win — what criteria matter most for demonstrating value to leadership?
• At what point does an agentic execution use case require the full governance stack from day one?