The biggest mistake I see is starting with:
"Let's build an AI agent."
High-ROI teams start with:
"What outcome do we need — and what's the lightest-weight agentic approach that achieves it?"
A simple ROI model (that business owners can actually use)
Value usually comes from one (or more) of these:
- Deflection: fewer tickets/calls reaching humans
- Cycle time reduction: faster quotes, faster approvals, faster onboarding
- Throughput: same team handles more volume
- Quality: fewer errors, better compliance, better consistency
- Revenue lift: better conversion, faster follow-up, less leakage
Costs/risks to account for:
- Human oversight time (approvals, exception handling)
- Tool/API costs
- Data/security/compliance work
- Change management (training, adoption, trust)
If you can't name the KPI, you don't have a use case yet.
The "Use-case Selection Matrix" (impact vs feasibility)
Think in a 2×2:
- High impact / Low risk → start here (best pilots)
- High impact / High risk → needs approvals + careful rollout
- Low impact / Low risk → don't overbuild (Level 1 or basic automation)
- Low impact / High risk → avoid
Winning patterns (and which agent level they usually map to)
Pattern 1: Triage → route → summarize
Best when: you have lots of inbound requests (support, ops, HR) and humans waste time figuring out "where does this go?"
Level: often Level 2 first; Level 3 when you need deep integrations.
Pattern 2: Research + synthesis (decision support)
Best when: teams spend hours assembling briefs, comparisons, vendor research, market scans.
Level: usually Level 1. Tools like Deep Research are designed for multi-step research and deliver documented reports with citations. (help.openai.com)
Pattern 3: Draft + human approval (the "80% agent")
Best when: you want speed but must keep accountability.
Examples: customer responses, internal comms, proposals, meeting notes.
Pattern 4: Structured extraction (turn messy inputs into clean records)
Best when: you have PDFs/emails/forms and your CRM/ERP needs clean fields.
Level: Level 2 to start; Level 3 at scale.
Pattern 5: "Knowledge assistance" that is grounded in your content
Best when: humans repeatedly ask the same policy/process questions.
Level: Level 2 for many teams (e.g., Custom GPT with knowledge attachments). Note the knowledge feature has defined file limits (useful for playbooks, not infinite corpora). (help.openai.com)
For larger internal corpora, you typically shift toward retrieval-backed approaches (e.g., hosted file search on vector stores) so answers come from your source material. (platform.openai.com)
Losing patterns (red flags that kill ROI)
Red flag 1: "Replace the whole team"
If the first version is meant to fully automate a complex role, you'll spend months on edge cases and trust issues.
Red flag 2: Unclear process + unclear owner
Agents don't fix broken workflows. They amplify them.
Red flag 3: High-stakes actions without approvals
If an error causes financial loss, compliance exposure, or brand damage, start with approvals and auditability. Risks like prompt injection and data leakage become real as soon as tools and untrusted inputs are involved. (platform.openai.com)
Red flag 4: No measurement plan
If you can't measure before/after (deflection, handle time, cycle time, conversion), you can't prove value — which kills adoption.
A "mini PRD" template for scoping an agent use case (non-technical)
Before you build anything, write one page:
- User: who uses it (and when)?
- Job: what decision/action does it support?
- Inputs: what data does it read?
- Outputs: what does it produce (and in what format)?
- Guardrails: what it must never do
- Human-in-loop: where approval is required
- KPIs: 1–2 measurable outcomes for 30 days
If you can do this in 30 minutes, you likely have a real use case.
Next step
If you want help prioritizing, I run a short use-case workshop: we pick the best 1–2 workflows, choose the right level (1/2/3), and define success metrics for a pilot.