Skip to main content

AI deployment · enterprise customer operations

I take AI agents from demo to production where failure is expensive.

Ten years turning enterprise strategy into delivery, the last stretch inside a tier-1 enterprise running GenAI on tens of thousands of customer conversations a day. This is the log: what I shipped, what broke, and the tools I built along the way.

See the workRun the business case
70K+
AI-handled conversations a day in production
10+ yrs
enterprise transformation: Big 4 advisory, global SIs, in-house
70
European companies benchmarked on GenAI customer-service maturity
§ 01Selected work

Deployments, tools and experiments

Each entry follows the same format: the problem, the architecture, three numbers, what broke. Production work is anonymised.

All entries
  1. LOG-012025–26

    Migrating a 70K-conversations-a-day voicebot from NLU to hybrid GenAI

    Moving a regulated, high-volume customer-service assistant off intent classification without breaking what customers and the business rely on.

    Shows → End-to-end ownership of a production migration: architecture split, evals, KPI redesign, vendor SLAs, change management.

    In productionTier-1 Italian enterprise · regulated · voice + chat
  2. LOG-022026

    A vendor-neutral business case for AI agents

    An open calculator that ties an AI agent to the P&L: containment ramp, repeat contacts, per-minute vs per-resolution pricing, payback.

    Shows → Translating agent performance into money a CFO will sign off on, and knowing where the numbers lie.

    LivePublic tool · runs in the browser
  3. LOG-032026

    GenAI Evaluation Kit: measuring whether a conversational agent actually works

    The evaluation method from production, packaged: three quality pillars, 18 LLM-as-judge prompts, a golden-dataset template and a scoring dashboard.

    Shows → Eval mindset: golden datasets, judge calibration, quality gates before real traffic.

    ShippedDigital product · sold to teams running GenAI assistants
  4. LOG-042026

    Deployment in a box: a voice agent for a fictional utility, end to end

    A full customer deployment, compressed: discovery, a voice agent with three tools over MCP on simulated APIs, evals, a KPI dashboard and an honest post-mortem.

    Shows → What a deployment strategist does in the first 90 days, in a form you can run.

    In buildOpen-source reference deployment · fictional customer
  5. LOG-052026

    AI-to-AI call lab: what happens when a customer's agent calls your voicebot

    A test bench where a customer-side AI agent calls an enterprise voicebot. Measuring loops, mutual detection, fraud surface and cost.

    Shows → Thinking one step ahead of the customer: the next failure mode of voice deployments.

    In buildResearch prototype · follows up my paper on AI-to-AI calls
§ 02Tools

Tie the agent to the P&L

Most AI-agent business cases are a vendor slide. This one shows its assumptions.

Tool · business case calculator

Live

Containment ramp, repeat contacts, per-minute vs per-resolution pricing, payback.

Vendor-neutral. Plug in your volumes and prices, see where the two pricing models cross, and print a two-page summary for finance with every assumption on it.

Open the calculator

// the line most business cases skip

per_minute = p_min × (c·d_contained + (1−c)·d_escalated)

per_resolve = p_res × c

// they cost the same at

c* = p_min·d_esc / (p_res + p_min·(d_esc − d_cont))

// below c*, pay per resolution. above, per minute.

§ 03Method

How I run a deployment

Five steps, each one written up or built into a tool. The order matters more than the tooling.

  1. 01

    Discover

    Map the contact reasons, volumes and the money behind them before touching a model.

    Business case tool

  2. 02

    Decompose

    Split the journey into deterministic flows (auth, payments) and generative ones (triage, FAQ, troubleshooting).

    Hybrid architecture

  3. 03

    Evaluate

    Golden datasets from real transcripts, LLM-as-judge calibrated against humans, regression tests on every change.

    Evaluation method

  4. 04

    Ship in stages

    Lab → shadow → pilot → rollout. Real traffic only once the system fails cheaply in the lab.

    Staged rollout

  5. 05

    Make it matter

    Report task completion, cost per resolution and effort, not intent accuracy. Tie every number to the P&L.

    Cost per resolution
§ 04Writing

Notes from the field

All writing
§ 05About

Consultant by training, builder by choice

Luca De Angelis

I spent a decade in enterprise consulting, from Big 4 advisory to global system integrators, and then moved in-house to build. Today I work on GenAI voice and chat assistants for a tier-1 enterprise: architecture, evaluation, KPIs and the slow work of getting operations, legal and vendors to trust a probabilistic system.

The pattern I keep seeing: deployments rarely fail on the model. They fail on integrations nobody scoped, metrics that stopped meaning anything, and stakeholders who were sold a demo. That gap between demo and production is the job I want to do.

Working toolkit

  • Voice & chat agents
  • LLM-as-judge evals
  • MCP / tool integrations
  • Conversation design
  • Human-AI routing
  • KPI & business cases
  • Next.js · Python
  • Claude Code
Education
MBA, LUISS Business School
Before
Big 4 advisory · global SIs
Languages
Italian · English
§ 06Contact

Deploying agents in Europe? Let's compare notes.

lucadeang@hotmail.itLinkedIn