Diagnosis and first AI pilot for an SME

What you will achieve now

A decision or test applied to "Diagnosis and first AI pilot for an SME".

• 1 Understand the criterion
• 2 Do a small practice exercise
• 3 Save evidence

A brief note with what you did, what went well, what failed, and what you would review later.

• Turn an everyday problem into a small, measurable pilot.
• Choose carefully between a local tool, cloud, or a hybrid design.
• Prevent the first test from accessing, sending, or modifying data without human control.

The entry rule: process before tool

Start with a task that happens often, has a clear input, and can be reviewed before it causes an effect. Strong first candidates are usually classifying emails, summarizing meetings, extracting fields from documents, or drafting reply drafts.

Do not promise hours saved or immediate ROI up front. State a hypothesis and measure it: savings, quality, and real cost depend on the data, the team, review, and maintenance.

FICHA DE DESCUBRIMIENTO (15 minutos)

1. Tarea repetitiva:
2. Quién la hace y cuántas veces por semana:
3. Tiempo actual por caso y errores frecuentes:
4. Entrada mínima necesaria (no "todos los datos"):
5. Salida que una persona pueda comprobar:
6. Acción que el sistema NO puede hacer todavía:
7. Datos personales, financieros, de salud o secretos implicados:
8. Herramientas ya utilizadas (correo, Excel, CRM, Drive...):
9. Persona responsable de revisar cada salida:
10. Métrica de éxito y fecha de revisión:

Prioritize with a table, not with enthusiasm

Pick three candidates and score them from 1 to 5. Prefer high impact and frequency, but low data exposure and low harm if the output is wrong. If a task affects payments, contracts, people selection, health, or rights, do not use it as a first pilot.

CASO                         IMPACTO  FRECUENCIA  REVISABLE  RIESGO  DATOS
Clasificar emails               4          5           5        2       2
Extraer campos de facturas      4          4           4        3       4
Enviar respuestas a clientes    4          5           2        4       3

Regla inicial: empieza por la opción más revisable y de menor riesgo.
No sumes puntos de forma ciega: un riesgo alto descarta el piloto aunque el ahorro parezca atractivo.

Choose the architecture for the case

• Cloud tool: good for a draft without sensitive data, with quick setup and a known usage-based budget.
• Local or self-hosted: useful when you need more control over where data is processed and you can manage the hardware, backups, updates, and access.
• Hybrid: separate data or tasks—for example, anonymize and classify locally, and use an external service only for non-sensitive text.

Local does not mean automatic compliance, zero cost, or total security. Cloud is not always unacceptable either. Document purpose, data, provider, access, retention, cost, and owner before you decide.

30-day plan: one case, four decisions

• Week 1 · Baseline. Collect 10 to 30 authorized or synthetic examples; measure time, errors, and current cost. Define what a correct response looks like.
• Week 2 · Safe prototype. The system only creates a preview or a draft. It does not send emails, post accounting entries, delete, or charge.
• Week 3 · Evaluation. Compare output with human review. Record wrong fields, abstentions, review time, incidents, and cost.
• Week 4 · Decision. Keep, fix, expand, or drop the pilot. Dropping it is valid: avoid paying for automation that does not help.

REGISTRO MÍNIMO DEL PILOTO

caso: "extracción de facturas a borrador CSV"
modo: "vista previa; sin importación contable"
muestra: "20 documentos autorizados o anonimizados"
campos: [emisor, fecha, número, base, IVA, total]
aceptación: "campos críticos correctos o marcados como dudosos"
revisor: "administración"
incidencia: "guardar documento ID, campo, motivo y corrección"
coste_a_medir: "suscripción/API + equipo + tiempo de revisión + mantenimiento"
fecha_decisión: "YYYY-MM-DD"
acción_prohibida: "pago, asiento contable o envío automático"

Four pilots that are usually safe when tightly scoped

• Email: classify and propose a draft; a person decides whether to send it.
• Meetings: generate summary, decisions, and tasks; the host corrects before sharing.
• Invoices: extract JSON/CSV with a fixed schema, alerts, and a confidence threshold; administration validates before export.
• Internal RAG: answer only over a small set of authorized documents, cite the source, and abstain when it cannot find one.

Exit checklist before you expand

• The responsible person can explain what the flow is for and stop it.
• The team knows what data goes in, where it is hosted, who can access it, and how long it is kept.
• There is a test with normal, ambiguous, and wrong cases; failures are logged.
• Irreversible actions require explicit human approval.
• Cost includes subscriptions or API, infrastructure, energy, support, backups, and review time.
• If personal data or a high-risk case is involved, the relevant data protection officer/advisor has been brought in.

If you have saved the evidence for this lesson, continue with "Safe local pilot with Ollama". If not, repeat the check before moving on.

Learn from a verifiable reference

Aulafy separates stable concepts from data that changes—versions, prices, models, and commands. Check the course sheet for the review date, verified scope, and primary sources.

"Editorial review" means structure, claims, and sources have been reviewed. It does not mean every command has been executed: when a technical test exists, it will be labeled as such.