L2 grounds the model in your data. Still human-initiated, but the system retrieves from internal documents, databases, tickets, transcripts. Output quality jumps because the context contains permissioned company data, not the open internet. Where most "AI ROI" stories actually live in 2026.
Technically: a vector store, a retrieval pipeline, a reranker, generation, and a packaging layer (Custom GPTs, Claude Skills, Gemini Gems, or a turnkey like Glean or Hebbia). L2 is where evals start showing up, because hallucination becomes measurable when the source-of-truth is internal documents instead of the open web.
The orchestrator at L2 is still the human. Lightweight glue (Zapier AI, n8n, Dify, Flowise, Langflow, Superagent) connects retrieval to a downstream step, but the person initiates the turn.
| Function | In practice | Signal |
|---|---|---|
| Support | A grounded help-center bot answers the FAQ band; humans take the rest | 25–40% deflection |
| Finance | Copilot pulls from the close package; the CFO writes the narrative | −60% CFO draft time |
| Sales | Grounded, sourced objection answers cut by segment | Cited in seconds |
| Legal | Cited answers with section references, Hebbia for diligence, Claude Skills for everyday | Cited Q&A |
| Internal | Glean stitches Slack, Drive, Salesforce, Notion, permissions intact | Search that works |
"Good" L2 evals are a golden dataset per copilot, a scored baseline, a target, and a regression run on every prompt or tool change, with adoption measured per-cohort rather than by license count. Teams that ship L2 copilots without an eval habit cannot reach L3 without rebuilding, so we start scoring at L2, not L3.
| Category | BOD Pick | Best fit | Pricing signal |
|---|---|---|---|
| Enterprise search / RAG | GLGlean | 1,000+ FTE orgs with SaaS sprawl | $15-25/seat/mo |
| Analyst-grade docs | HBHebbia | PE diligence, M&A, legal review | Mid-to-high six figures |
| Authored capabilities | CLClaude Skills | Portable, markdown, runs on Bedrock/Vertex | Token + seat |
| Authored capabilities (alt) | GPCustom GPTs / Workspace Agents | Sales enablement, internal Q&A | Bundled with ChatGPT Biz/Ent |
| BYO RAG (vector) | PCPinecone (managed) / pgvector | Embed in product / Postgres-heavy | $50-700/mo+ |
| RAG framework | LILlamaIndex | BOD L2-L5 default | OSS + cloud |
| Vertical copilots | HVHarvey · Clay · Writer · Intercom Fin | Legal / sales / marketing / support | Per-vertical SaaS |
| Hyperscaler-native |  AWS Q · Azure AI Search · Vertex | When the cloud is the constraint | Bundled |
| Evals | BTBraintrust primary; Langfuse OSS | Quality-as-engineering vs. cost-sensitive | Tiered |
| OSS alternative | OXOnyx | Sovereignty-first mid-market | Free (self-host) |
L2 is the substrate the agent layer above it depends on. Skipping it doesn't save time. It pushes hallucinations into production at L3. Buy Glean if you're 500+ FTE and your SaaS estate is sprawling. Build with Claude Skills + Pinecone + LlamaIndex if you're going to embed AI in your own product.
Grounded retrieval is a real step, but a human still drives every task, so unit cost barely moves. The value inflects one tier up, at L3, where the agent runs the workflow.
L2 is necessary substrate, not the destination. The curves cross before L3, which is why we don't stop here.
A wave of tools market themselves as agent platforms with low-code or no-code builders. For the right workflow they're the fastest way onto the ladder: a cheap rung that proves a workflow before it earns a code-first build. We treat them as a validation substrate with a named graduation test, not a permanent architecture for anything that becomes high-stakes.
| Category | Representative tools | The shared ceiling |
|---|---|---|
| Workflow-automation incumbents (LLM steps added) | Zapier · Make · n8n | Rule-based at heart; they hit a wall on judgment, exceptions, and multi-step reasoning. Eval and trace depth are shallow. |
| Horizontal no-code agent builders | Relevance AI · Lindy · Gumloop · Stack AI · Dify · Flowise/Langflow | You compose within the tool's harness with limited control of the prompt wrapper and routing; eval depth and step-level replay vary; logic lives in their canvas. |
| Retrieval / search layers (feed agents, don't orchestrate) | Onyx (OSS) · Glean (managed) | These are retrieval, not orchestration. A search layer is not an agent platform; action and orchestration live above it. |
| Vertical / industry-specific agents | Harvey · Hebbia · Sierra · Decagon · Writer · Clay · Intercom Fin | The harness is locked: you compose within the vendor's opinions and can't control the wrapper and routing where output quality is decided. |
When the workflow is bounded and low-stakes, a business operator should own it, the data already lives in connected SaaS, there's no hard audit or eval bar, and speed-to-validate beats control. The newer enterprise entrants (Stack AI, Sierra, Decagon) have largely closed the old governance gap, with SOC 2, RBAC, and SSO in the box, so the 2026 ceiling is harness control, eval depth, and step-level inspectability, not security.
If the tool owns the governed foundation, wrap the locked layer rather than rebuild it.
Grounded help-center bot reads the knowledge base + recent tickets. Humans inherit the long tail. Sets up the L3 graduation to ticket-resolving agents and the L4 graduation to multi-agent tier-1/2 with edge escalation.
CFO draft time -60%. Copilot pulls from close package; CFO ships the narrative. Sets up the L3 graduation to AI-assisted FP&A cycle compression (10d → 2d).
Objections, segment cuts, competitive intel: sourced, cited, fast. Sets up the L3 graduation to AI-enriched lead routing with measured conversion uplift.
Harvey for legal-heavy; Hebbia for diligence; Claude Skills for everyday. Sets up the L3 graduation to contract extraction at scale, feeding deal-desk risk scoring.
Glean stitches the SaaS estate with permissions intact. The "wow" demo of L2. The graduation sign is users asking for the action, not just the answer.
Help-center search, in-product Q&A, summarization. Feature-flagged, per-cohort usage. The L3 graduation makes the feature the core workflow, not the side panel.
A function-specific copilot wired to your L1 warehouse, packaged with an eval harness, governed through your hyperscaler's native AI gateway. Finance variance, sales enablement, support deflection, or marketing content: pick one, ship it, measure it.
Talk to BOD