Practical guide

Practical guide: how to design automation and AI that keeps working after the demo

Automation does not mean adding AI everywhere. It means designing a workflow with inputs, states, rules, data, actions, exceptions and a way to observe and improve it once it is in production.

Principles

Six ideas for keeping an automation working

Process before tool

First understand what happens today, who is involved, what data enters and what the expected outcome should be.

AI where judgement helps

AI can interpret language, intent or context; rules, permissions and limits remain part of the design.

APIs and webhooks with purpose

Tools connect when moving data or triggering actions reduces work and avoids information silos.

Model routing

Not every task needs the same model: cost, speed, privacy and complexity can define the route.

Observability and traceability

States, errors and events make it possible to know what happened and decide whether to retry, fix or intervene.

Human-in-the-loop

Good automation knows when to stop, ask for more information or hand the case to a person.

1. What to automate first

Strong candidates are usually repetitive processes with sufficiently clear rules and a concrete outcome: answering common questions, collecting the same information, creating a request, alerting a person, recording an opportunity or preparing a booking.

2. What should not be automated yet

Processes that change every week, sensitive decisions, tasks with too many exceptions or flows that are not clearly defined usually need structure and human judgement first. Automating a confusing process only executes confusion faster.

3. What a simple workflow should look like

Event → context → validation → decision → action → state → follow-up. The channel may be web, form, WhatsApp or chatbot; the result may be a booking, task, alert, CRM update or human handover.

4. How systems connect

APIs, webhooks and connectors let one tool deliver information or react when an event occurs. Integration should include enough permissions, validation, idempotency and error routes to avoid duplicated actions or lost data.

5. How to choose which AI model to use

An AI Gateway can separate the application from a specific provider and route tasks by complexity, cost, latency or privacy. Premium models can be reserved for difficult reasoning while lighter or open-weight models handle repetitive work where they fit.

6. What to measure and maintain after launch

It is not enough to know that it “works”. Review volume, errors, retries, handovers, conversions, operating cost and drop-off points. Rules, prompts, knowledge, models and integrations become part of evolutionary maintenance guided by real usage.

7. What a professional implementation should cover

Implementation turns the design into production: configuration, integrations, credentials and permissions, rules, validation, testing, error routes, minimum documentation and controlled go-live. After that, a new phase begins: support, monitoring and evolution according to the agreed service.

Control

Production automation needs control, not blind autonomy

The objective is not to remove people from every process. It is to automate repeatable work while keeping limits, permissions, observability and human intervention for judgement, exceptions and sensitive decisions.

Clear scope, permissions and expected states.

Human handover and exception routes when judgement is required.

Errors, retries and critical journeys tested before production.

Want to know which part of your process is worth automating first?

We review the process first and recommend only the improvements that can make the work clearer, faster or easier to follow.

See all ways to get in touch

Put the criteria in context

From architecture principles to concrete customer-facing layers

These examples show where the same design criteria appear in WhatsApp, the website, web chatbots and our own internal operation.

Talk to Nova on WhatsApp