From the outside, an automation can look simple.
A customer writes. The AI replies. A record is created. An appointment is booked. A notification goes out. It all seems to happen in seconds.
What you don't see is the work done beforehand, so that the sequence runs reliably and keeps running when something comes up that the demo never showed.
Because automating does not mean connecting two tools and hoping the workflow holds up on its own. In a real business there is incomplete data, permissions, opening hours, duplicates, status changes, exceptions, APIs that fail, customers who write in unpredictable ways, and decisions that should not be left to a machine.
That invisible work is precisely what separates an impressive demo from a system that is actually useful.
Key idea: a professional automation is not built to work when everything goes well. It is built to know what to do when something goes wrong.
1. Before automating, you have to understand the real process
The first mistake often happens before anyone writes a line of code: trying to automate a process that is not yet well understood.
In a company, almost no process is as linear as it seems.
For example, “handling a quote request” can involve:
- receiving an inquiry via WhatsApp, the website, email or phone;
- checking whether they are a new customer or an existing one;
- understanding what they need;
- checking whether the service is a fit;
- asking for missing information;
- deciding who should handle it;
- preparing the quote;
- recording that it has been sent;
- remembering when to follow up;
- updating the status if they accept, decline or stop replying.
If we only automate “sending a quote”, we will probably speed up one small part of the problem and leave everything else untouched.
That is why the work starts with much simpler questions:
What happens today? Who is involved? What information is needed? What decisions get made? Where is time lost? Where do errors appear? Which situations need human judgment?
Automation starts by understanding the business, not by choosing a tool.
2. Not everything that can be automated should be
Technical feasibility should not be the deciding factor.
Some tasks are repetitive, predictable and easy to validate. They are usually good candidates.
Others depend on context, negotiation, sensitivity, professional accountability or serious consequences. In those cases, automation can help, but it should not make the call.
A practical way to see this is to split tasks into three levels:
| Type of task | Example | Sensible approach |
|---|---|---|
| Repetitive and predictable | Logging a form submission in a CRM | Direct automation |
| Variable but bounded | Classifying an inquiry and preparing a draft | AI + rules + review when needed |
| Critical or ambiguous | Making a decision with legal, financial or personal impact | AI support, mandatory human oversight |
A good implementation also knows when to say “we would not automate this”.
That is not a limitation of the project. It is part of the design.

3. Designing the flow means deciding what happens in each state
When a process becomes a system, it needs clearly defined states.
A request is not just “a lead”. It can be newly received, waiting for information, qualified, assigned, scheduled for a meeting, at the proposal stage, awaiting follow-up, won, lost, discarded or archived.
Each state can have different rules.
If a person has already decided the next action, an automation should not overwrite it. If a customer has already rejected a proposal, they should not receive an automated sequence written for someone who is still interested. If mandatory data is missing, the system must know whether it can continue or should stop.
This is where much of the real work lies: modeling the process before automating it.
Without clear states, the automation ends up deciding on ambiguous data.
4. Data is part of the automation
An automation can only work with the information it receives.
And in a real business, that information rarely arrives clean.
A phone number may come with or without a country prefix. A company name may be written in several different ways. The same customer may use two email addresses. A form may arrive without an optional field that later turns out to be important. Two different channels may create two records for the same person.
That is why many workflows depend on work around normalization, validation, deduplication, required fields, formats, internal identifiers and the links between contact, company, opportunity, project or invoice.
In addition, if personal data is involved, the design must apply a basic GDPR principle: process only data that is adequate, relevant and limited to what the stated purpose requires. The AEPD stresses that data minimization must be built into the default design, not bolted on at the end.
A workflow can be technically correct and still leave you with a CRM nobody can use if the data is poorly stored.
In practice: many automations do not fail because the AI gives a bad answer. They fail because the system does not know whether it is working with the right customer, the right status or the right version of the data.
5. Integrations are more than “connecting apps”
When two systems connect, you have to decide what information moves, in which direction and under what conditions.
An integration between WhatsApp, a calendar and a CRM may need to answer questions like these:
- What is the source of truth for availability?
- What happens if an appointment is changed in the calendar?
- Who is allowed to cancel?
- What happens if the customer writes while the calendar is temporarily unavailable?
- When is the contact created in the CRM?
- How do we avoid booking two appointments from the same message?
- What CRM information can the assistant look up?
APIs and webhooks move the data. Designing the behavior is still our job.
Connecting two apps is where an integration starts, not where it ends.

6. AI should be used where it adds value
Not every step needs AI.
If a clear rule can solve a task, it will usually be more predictable and easier to audit.
AI starts to pay off when we have to handle less structured information: interpreting what a person means, classifying an inquiry, summarizing a conversation, extracting data from a message or drafting a reply suited to the context.
A robust architecture usually combines:
rules + structured data + integrations + AI + human intervention.
This idea aligns with management frameworks such as NIST's AI Risk Management Framework, which organizes the work around governing, mapping, measuring and managing risk across the lifecycle, and with ISO/IEC 42001, which treats AI as a management system that must be maintained and continuously improved.
7. Exceptions are where serious work begins
Demos usually show the ideal path.
A customer asks something expected. The data is there. The API responds. The calendar has an open slot. The system gives the right answer.
But a real business throws up other situations: information is missing, an external service goes down, duplicates appear, no slots are open, the conversation shifts topic, a person steps in manually or a credential has expired.
The system needs to decide whether to ask, retry, wait, log, hand over or stop.
Designing these exceptions usually takes more work than building the main path.
And it is time well spent.
8. Security, privacy and permissions are also part of the process
An automation may have access to sensitive information about the business and its customers.
That is why you need to decide:
- what data each component actually needs;
- which user or service can read it;
- who can modify it;
- what information should never reach an AI model;
- how consent is recorded when it is required;
- how secrets, tokens and credentials are protected;
- which actions need human confirmation.
The rule should be simple: each system and each person should only access what they need to do their job.
Adding AI does not remove security obligations. In some workflows, it makes them even more important.

9. Testing is more than clicking through once
An automation is not validated just because one path ran correctly.
You need to test the normal case, incomplete data, duplicates, external errors, retries, status changes, permissions and the real channels and devices people actually use.
A retry, for example, should not create two contacts, two invoices or two bookings.
Testing is not a cosmetic phase at the end of the project. It is part of development.
10. Automation needs traceability
When something happens automatically, we need to be able to answer later:
What happened? When? With what data? Which system acted? What result did it produce?
That is why activity logs, sync statuses, the date of the last run, request IDs, change history and alerts for failed integrations matter.
NIST treats AI risk management as a continuous process across the lifecycle. For a business, that means something very concrete: if we cannot see what the system is doing, it is hard to govern and fix.
Knowing that “the workflow is running” is not enough.
11. Measure before attributing results
An automation should improve a specific part of the process: reduce manual re-entry, speed up responses, stop follow-ups from slipping through the cracks, improve record quality or make follow-up easier.
But we should not automatically credit it with every improvement that follows.
If a campaign drives traffic, we need to tell apart impressions, clicks, visits, sessions, leads, opportunities and sales. Each system may be measuring a different stage of the journey.
Professional measurement also means recognizing when the numbers do not add up, rather than bending them to tell a nice story.
12. Maintenance is part of the automation
A workflow can run perfectly today and fail a few months later, even though nobody seems to have touched anything.
An API changes. A permission expires. The business changes its hours or services. A tool changes its limits. The team starts using a field differently. A new exception appears.
ISO/IEC 42001 does not treat AI management as a project that ends at go-live, but as a system that must be maintained and continuously improved. For an SME, that idea is simple: anything that enters daily operations needs an owner, follow-up and review.
Our rule of thumb sums it up:
A good automation is not the one that looks most advanced. It is the one that keeps working when nobody is watching.
13. The difference between a demo and a real system
A demo answers one question:
Can we make this happen?
A business system has to answer many more:
- Does it work with real data?
- Does it happen consistently?
- Does it avoid duplicates?
- Does it respect permissions?
- Does it know when to stop?
- Does it leave a record?
- Does it recover from a failure?
- Can a person intervene?
- Can we measure what it has done?
- Can we keep it running when the business changes?
The first question shows technical capability. The others decide whether the system can do real work.
A simple example: a request that arrives via WhatsApp
Imagine a business that wants a WhatsApp agent to collect requests and help book an appointment.
The demo version could be:
Message → AI → calendar → appointment.
In a real business, it usually needs something closer to this:
- Identify the intent of the message.
- Check whether the contact already exists.
- Collect only the information that is needed.
- Validate that the requested service exists.
- Check real availability.
- Propose valid options.
- Confirm before creating the appointment.
- Record the source and the context.
- Avoid duplicates if the customer repeats the message.
- Hand over to a person if the request is out of scope.
- Log an error if the calendar does not respond.
- Let the team change the appointment afterwards without the bot recreating it.
It is still “automating an appointment.”
But now we are talking about automation that can live alongside a real business.
What you should ask before automating a process
Before choosing a platform, agent or AI model, these questions are usually more useful:
- What problem do we want to solve, exactly?
- How does the process work today?
- Where does repetitive work pile up?
- What data does each step need?
- What is the source of truth for that data?
- Which decisions can be handled with rules?
- Where does AI add value?
- Which decisions need a person?
- Which exceptions do we already know about?
- What happens if an integration does not respond?
- How will we avoid duplicates?
- How will we know the system has done the right thing?
- What are we going to measure?
- Who will maintain the process when it changes?
If these questions have no answers yet, choosing a tool is premature.
Frequently asked questions
In practice: automating well does not mean adding more automatic steps. It means cutting work and errors without losing control of the process.
Which process should an SME automate first?
One that is repetitive, stable enough, has clear data and a well-defined operational problem. If the process changes every time or nobody can explain how it works, it is better to sort it out first.
When do you need AI, and when are rules enough?
Rules work well when conditions are clear and predictable. AI adds value when there is natural language, classification, information extraction or variability that isn't worth turning into hundreds of rules.
How do I know if an automation is working well?
Checking that it runs is not enough. You also need to watch for errors, exceptions, duplicates, timing, the human intervention required and its effect on the process's real goal.
If you want to go further on operations and maintenance, read how to move from using tools to governing a system or request an initial assessment.
Conclusion
The visible part of an automation is usually the smallest part.
Behind it sit process analysis, data structure, rules, integrations, AI, security, permissions, testing, error handling, traceability, measurement and maintenance.
It is not about adding as much technology as possible.
It is about making the process clearer, more reliable and easier to run.
And before asking which tool we need, there is a far more useful question:
What should happen, from start to finish, for this process to work better?
Reference sources
This article does not reproduce the text of these sources; it uses them as technical references to explain the problem from a business point of view.