The production AI stack: 5 layers between a demo and a system

Quoted tools are neutral ecosystem references, not a closed list.

In this article

The 5 layers of a production AI system

1. 1DataOne source, one owner
2. 2Automated loadsUpdates itself, alerts on failure
3. 3AISearches, respects permissions, is tested
4. 4Go-liveSame in test and real use
5. 5MeasurementAccuracy, cost and failures

Example: from demo to system

Example: a family-owned distributor tests an assistant that answers customer balances. The demo uses an exported spreadsheet; production reads the ERP every night.

Example: customer balance question

Demo · example 5 min per question

1. 1Exported sheetWeekly copy: 20 min
2. 2Someone runs itOn their laptop: 5 min per question
3. 3AnswerNobody reviews it: errors surface days later

Production · example 10 s per question

1. 1ERPRead every night: 0 staff minutes
2. 2Automated loadRetries and alerts: 5 min if it fails
3. 3AssistantAnswers in 10 s, same in test and use
4. 4Weekly dashboardAccuracy and cost: 15 min per week

Demo vs production, layer by layer

Layer	Demo	RecommendedProduction
Data	Spreadsheet exported by hand	A single source, with an owner per field
Automated loads	Someone runs a file when they remember	Runs daily on its own, retries and alerts on failure
AI	An instruction that worked once	Searches your data, respects permissions and is tested before each change
Go-live	Works on a laptop	The same system in test and in real use
Measurement	Nobody measures	Accuracy, cost per answer and failures, every week

RecommendedProduction

Data

Demo

Spreadsheet exported by hand

Production

A single source, with an owner per field

Automated loads

Demo

Someone runs a file when they remember

Production

Runs daily on its own, retries and alerts on failure

AI

Demo

An instruction that worked once

Production

Searches your data, respects permissions and is tested before each change

Go-live

Demo

Works on a laptop

Production

The same system in test and in real use

Measurement

Demo

Nobody measures

Production

Accuracy, cost per answer and failures, every week

For your technical team: names and tools per layer

Data = data layer (PostgreSQL, Snowflake). Automated loads = pipeline orchestration (Airflow). AI = retrieval, permissions and evals (LangChain). Go-live = container deploy (Docker, Kubernetes). Measurement = monitoring and observability (PostHog). Neutral references, not a closed list.

The model is the part that changes most

6 months

is OpenAI's minimum notice before retiring a generally available model. Real case: gpt-5.1 was deprecated on 2026-10-01 and leaves the API on 2027-04-01.

Criterion: switching models must not touch the other four layers; your quality tests tell you whether the new one answers as well. The AI layer in detail: what is RAG and what is MCP.

What it costs per month and when it pays back

What a production system costs per month

Economy: gpt-6-lunaUSD 0.10 input / USD 0.50 output per million tokensAdvanced: gpt-6.1-solUSD 2.00 input / USD 10.00 output per million tokens

Monthly costs

Real queries per month

5,000

Quality test runs (evals) per month

4

200 questions per runServer, database and monitoring (typical range, example)

USD 150

Team maintenance hours per month (example)

16 h

Technical team hourly cost (example)

USD 25

What it saves (example)

Minutes saved per query

3 min

Hourly cost of the person who does the work today

USD 6.95

PY minimumPYG 12,683 per hour (MTESS) ≈ USD 2.23EC minimumUSD 482 ÷ 173.33 h (2026 SBU) ≈ USD 2.78PY real cost (example)minimum × 2.5 ≈ USD 5.58EC real cost (example)minimum × 2.5 ≈ USD 6.95

Adjust by hand

USD 6.95

Estimated floor: the minimum-wage hour. Paraguay converted at PYG 5,694.47 per dollar (BCP, 9 Oct 2026). Real cost: ×2.5, an example of two minimum wages plus contributions. Sources: MTESS, Ministry of Labor (EC), BCP.

One-off

Setup (example range: USD 5,000–15,000)

USD 8,000

Total cost per month: USD 553Pays back in: 6.8 months

Monthly cost breakdownestimated

• Model (queries)the cheapest partUSD 2.75 · 0.5 %
• Model (quality tests)USD 0.44 · 0.1 %
• Server and monitoringUSD 150 · 27 %
• Team hoursUSD 400 · 72 %

Model tokens are 0.6% of the monthly cost.

Tokens: the unit AI providers bill for the text the model reads (input) and writes (output). OpenAI puts 100 tokens at about 75 English words.

Total cost per monthUSD 553

Gross savings per monthUSD 1,738

Net savings per monthUSD 1,184

Pays back in6.8 months

Model prices: OpenAI API pricing, standard tier, checked 2026-10-11. Economy gpt-6-luna: USD 0.10 input and USD 0.50 output per million tokens. Advanced gpt-6.1-sol: USD 2.00 and USD 10.00.

Assumptions: 3,000 input and 500 output tokens per query or test question; 200 questions per test run. Hourly cost of the person who looks up the answer today: anchored to the minimum-wage hour (Paraguay: PYG 12,683, MTESS, ≈ USD 2.23 at the BCP rate of 9 Oct 2026; Ecuador: USD 482 SBU ÷ 173.33 h ≈ USD 2.78, Ministry of Labor). The starting value is Ecuador's example real cost: minimum × 2.5 = USD 6.95. Server, team hours, minutes saved and setup are examples: replace them with your figures. With the starting values: model USD 3.19 (including USD 0.44 of tests), total USD 553 per month, gross savings USD 1,738 (250 h × 6.95), net savings USD 1,184 and payback in 6.8 months.

Go-live checklist

Production-ready: 10 points by layer

Copy checklist

• Data: every field the AI uses has a single source and an owner.
• Data: it is defined which information is personal or confidential and who can see it.
• Automated loads: data updates on its own, retries and alerts on failure.
• AI: there is a set of real questions with their correct answer, approved by the business area.
• AI: when it has no answer, the system says so and hands off to a person.
• AI: switching models does not force rebuilding the other layers.
• Go-live: the same system runs in test and in real use.
• Go-live: compliance approved where the data is processed (cloud, country, provider) under Law 7593/2025 (PY) or the LOPDP (EC); for banks, also against the BCP's SIB (PY) or the Superintendencia de Bancos (EC).
• Measurement: every week you see cost per answer and accuracy.
• Measurement: one owner reviews that dashboard and decides.

Where the data is processed, in Paraguay and Ecuador:

• Paraguay: Law 7593/2025, published 2025-11-27; takes effect on 2027-11-27 (art. 57). Data may go to another country only with adequate protection or contractual safeguards (art. 19).
• Ecuador: LOPDP, Registro Oficial 459 of 2021-05-26.
• Banks: also check what your supervisor requires: the BCP's SIB (PY) or the Superintendencia de Bancos (EC).
• By country: AI for banks and fintech in Paraguay · AI for banking and cooperatives in Ecuador.

Example pilot by layer, with deliveries every 2 weeks

1. Week 1-2Data and use caseOne frequent question, one data source and defined permissions.KPI: single source with an owner
2. Week 3-4Automated load and first versionThe assistant runs in a test environment with data that updates itself.KPI: % correct on the test questions
3. Week 5-6Real use with one teamWeekly dashboard and a go / no-go decision.KPI: time and cost per query vs. today