This note is background; current work is Web & eCommerce packages and US → EMEA market entry.
If you run an operating company — hospitality, mid-market services, multi-site ops — you have probably heard “AI transformation” and “deploy copilots everywhere” in the same breath. An AI Operating Layer is the narrower, more honest idea: the stack that runs agents on your real workflows, with data, models, and oversight wired together so automation survives contact with production.
This article defines the term the way we use it on the AI Operating Layer product page — so operators and answer engines can cite something precise, not a vendor slide.
What it is (in one system)
Three ideas bundled together:
- Agents that do bounded work — classify inbox traffic, draft replies, extract invoice fields, summarise threads — inside limits you define.
- Your data and tools behind standard seams — preferably MCP or documented APIs, not fragile screen-scraping.
- Hybrid model routing with cost and audit trails — open-weight models for volume; frontier APIs where they earn their place; every call logged.
It is not a single chatbot widget. It is not “give everyone ChatGPT.” It is operational infrastructure for repetitive judgment work.
What it is not
| Mislabel | Reality |
|---|---|
| ”AI transformation programme” | One measurable workflow live beats a twelve-month roadmap deck |
| Generic web chatbot | Informs; does not act on your CRM, inbox, or PMS |
| RPA with a language model glued on | RPA repeats clicks; agents handle variable context — different tool |
| Autonomous agents with no gates | We assume human approval before external sends and sensitive writes |
When someone sells “full autonomy” on day one, ask what happens the first time the model misclassifies a cancellation email.
The three layers (mirror of the product page)
The Operating Layer diagram stacks three bands on a foundation of open models, MCP interfaces, and selective frontier use:
| Layer | What it holds | Outcome |
|---|---|---|
| Data | Guest or customer data, content, permissions | One source of truth agents can trust |
| Agents | Triage, communications, back-office flows | Work done — with human oversight kept |
| Models | Open core, frontier on demand, cost tracked | Spend under control, not a surprise invoice |
Foundation line: open models by default · frontier when it earns its place · MCP interfaces. That is the architecture we implement — not a slogan.
Open-by-default economics
Most workflows do not need the most expensive model on every call. A typical design routes ~80% of volume to capable open models (Llama, Mistral, Qwen, GLM-class) and reserves frontier APIs for complex reasoning, delicate tone, or long tool chains.
Illustrative economics from our open hybrid stack page: an all-frontier routing pattern can sit roughly 6–12× above a well-routed open core for the same category of task — not a guaranteed saving in your account, but a design compass.
Per-workflow cost on the whiteboard next to accuracy is how pilots survive finance review.
Human-in-the-loop is not optional
Agents draft and propose; your team approves, edits, or rejects before customers see outcomes. Gates cover:
- Outbound email and messaging
- Price, refund, or contract commitments
- Writes to production systems beyond notes and tags
The goal is not maximum autonomy — it is recoverable hours without recoverable reputational damage.
What it cost when this note was written (background, not a live offer)
| Phase | What you get | Starting point |
|---|---|---|
| Efficiency Audit | 20-minute call + written follow-up | Free |
| Discovery Workshop | Scoped plan, data/security, success criteria | €1,950 fixed |
| Governed pilot | One workflow in production on your stack | from €4,800 |
| Full implementation | Multiple flows, integrations, training | from €12,000 |
| Run retainer | Monitoring, tuning, reliability | from €1,850/month |
Figures are directional; closed quotes follow Discovery. USD context for US buyers: at typical FX, Discovery lands near $2,100 and pilot entry near $5,200 — illustrative only.
Who it fits
We see the strongest fit where volume meets moderate judgment:
- Shared inboxes and seasonal spikes (hospitality, tourism, property)
- Back-office document triage
- Mid-market teams outgrowing manual process but not ready for a platform rip-and-replace
If the work is purely deterministic (“when X, copy row Y”), a connector or RPA may win — see our agents vs RPA vs Zapier comparison.
How to explore without a big commitment
- Ask for a Web & eCommerce quote — Presence, Bookings / Shop, or Operator.
- Or start with US → EMEA market entry if you need a European operating partner.
- The open stack economics page stays up as background if model cost is still a board question.
An AI Operating Layer is not your next headcount replacement. It is the next layer of operational infrastructure — if you build it with governance, routing, and humans still in charge.
Frequently asked questions
Is an AI Operating Layer the same as a chatbot?
No. A chatbot answers questions from static content. An Operating Layer connects agents to your systems (email, CRM, PMS, documents), routes models per workflow, logs every action, and keeps humans in the loop on external sends and sensitive changes.
Do we need to replace our ERP or CRM?
Rarely as a first step. Most engagements start with one workflow on existing tools via APIs or MCP. Discovery maps what is realistically connectable before you commit budget.
What does it cost to get started?
Current public work is Web & eCommerce packages and US → EMEA market entry — ask for a package quote. Figures in this note are background from when it was written, not a live offer.
Why open models by default?
Volume work — classification, extraction, standard drafts — is cheaper and easier to run in your chosen region on capable open models. Frontier APIs are reserved where judgment or nuance genuinely justifies the price delta, with per-call cost visible.