From ChatGPT Chat to Custom Agents That Work: Five Capability Layers
Many teams treat ChatGPT Enterprise or a strong model API as “digital transformation done.” Everyone can chat—yet order lookups stay manual, policies stay in folders, and price changes stay verbal. ChatGPT solves safe access to a strong model; a custom AI Agent solves finishing one business task inside permissions. Between them you usually need five capability layers. Positioning: /blog/enterprise-custom-ai-agent-vs-chatgpt. GeonAI delivers layer by layer; /agents are capability references only (not a public trial of 363+ presets). Email [email protected].
Five layers at a glance
| Layer | What it covers | If missing |
|---|---|---|
| 1 Entry & identity | SSO/login, visitor routing, session ownership | No audit of who asked |
| 2 Intents & flows | Intent table, slots, state machine | Every chat starts from zero |
| 3 Knowledge | RAG + ACL + citations/versions | Stale answers, cross-store leaks |
| 4 Tools | Allowlisted APIs, receipts, confirms | Can talk, cannot act |
| 5 Governance | Handoff, QA, release/rollback, weekly review | No accountability, cannot scale safely |
Layers are acceptables, not slogans. Artifact checklist: /blog/custom-agent-delivery-artifact-checklist. More than a prompt: /blog/custom-agent-more-than-a-prompt.
Layers 1–2 — entry, identity, intents
- Without enterprise identity there is no audit trail or ACL boundary
- Without an intent table the model only chats—it does not walk a business path
- Split external entries from internal assistants; do not share one unbounded prompt
Layers 3–4 — knowledge and tools
- Knowledge: updatable store + citations + ACL (/blog/enterprise-rag-knowledge-base-agent)
- Tools: allowlisted gateway, receipt codes, writes default to confirm (/blog/agent-function-call-mcp-integration)
- Layer 3 without 4 = answers without action; layer 4 without 3 = API calls on stale facts
Layer 5 — governance
- Boundary tables and human confirm: /blog/custom-agent-scope-boundary-in-contract
- Pilot harm metrics and scale gates: /blog/custom-agent-30-day-pilot-metrics
- “Fully automatic” without governance is unfit for external promise channels
Common fallacy
“Enterprise seats for everyone = digitized.” Seats buy access, not process or accountability. Likewise, a stronger model does not replace ACL, tools, or gates. Renaming a role “digital employee” also does not mean RPA or a custom Agent is live—lane split: /blog/what-is-digital-employee-vs-custom-agent.
Suggested path
- Start with layers 2+3: intent table + knowledge assistant with citations
- Add layer 4 read-only tools (order/stock lookups)
- Add writes, public entries, and full layer-5 governance last
- Accept each step with evals and gates—not “feels chatty”
How to brief GeonAI
Say which layers you already have and where you are stuck (e.g. “ChatGPT Enterprise, no tool gateway”). Email [email protected], /pricing, or Live chat. We fill layers—we do not require all five on day one.
Frequently asked questions
Must we launch all five layers at once?
No. Internal assistants can start with 2+3; external promise channels should have a minimum version of all five.
If we have ChatGPT Enterprise, do we still need custom?
For general office help, seats may be enough. To look up/change business systems under ACL with auditable acceptance, you need the custom layers.
How do the five layers relate to “more than a prompt”?
That post stresses KB/tools/handoff; this one also surfaces entry/identity, intents/flows, and governance for kickoff checklists.
Tools first or knowledge first?
Most scenarios: knowledge with citations first, then read-only tools, then writes—writes without governance fail loudest.
Does a no-code layer count as custom?
Fine for validation; external entries and core transaction writes usually still need a controllable gateway and contract boundaries. See /blog/custom-agent-vs-no-code-workflow-builders.
Which layers do the 363+ presets cover?
They illustrate capability types—not your identity, ACL, tool allowlist, or governance process.