Contents8
- 01What is automation without AI?
- 02Why does putting AI everywhere cost so much and make a project drift?
- 03At which steps is AI justified?
- 04How do you know whether your task needs AI?
- 05Rule, AI and human: how do you combine them in one flow?
- 06Does an automation without AI need monitoring?
- 07Where to start
- 08Frequently asked questions
Not every automation needs AI, whatever the trend of putting it everywhere might suggest. An automation without AI is a series of written rules that run on their own: when an invoice arrives, record it and tell the accounts team. It impresses less than an agent in a demo, but it gives the same result on every run. This guide, from our business process automation and bots section, helps you choose, step by step, what comes down to a rule, an AI or a human.
In short
A rule you can write down can be automated without AI, for less money and with a consistent result.
An AI step adds a cost on every run and a result that can vary: it has to earn more than it costs to check.
AI is justified at a specific step, when the input or the output is language, not for the whole process.
With or without AI, an automation in production needs monitoring: without AI, it does not get things wrong, but it can fall silent.
What is automation without AI?
A customer fills in the quote form on your website:
A quote request, handled by rules
- The record
The workflow creates it in the CRM.
- The assignment
The record goes to the sales rep for the area, according to the postcode, and they are notified.
- The alert
If the stated budget exceeds a threshold, management is notified too.
No language model is involved: the form fields are already structured, and the same case always follows the same path. The same request, received as a freely written email, would need a reading step: that is where, and only where, a model could be useful.
The tools that run these workflows charge for their own use: Zapier counts tasks, Make counts credits, n8n counts executions, as our comparison n8n, Make or Zapier for an SME sets out. As long as no step calls a model, there is no token bill on top. That is the starting point of our automation offer: less re-keying, with a counter of the cost per execution and per month.
Why does putting AI everywhere cost so much and make a project drift?
A cost on every run
A rule costs whatever its execution costs in the tool, whatever the length of the text processed. An AI step adds a call to the model, billed according to the tokens it reads and produces: the price grows with the size of the documents and with the context you send it. A step that costs almost nothing during testing can weigh heavily at a year’s real volume. Our guide to the cost of AI in business details this calculation.
A result that can change
A rule always gives the same result for the same input; a language model does not. Anthropic’s glossary, from the maker of the Claude models, says so about temperature, the setting that controls how much randomness goes into the answers:
Even with temperature set to 0, the results will not be fully deterministic and identical inputs may produce different outputs across API calls.
That does not matter for the first draft of an email. For an amount carried into your accounts, it is a risk that calls for checks, a log of decisions and a way of putting errors right.
Drift that starts small
In June 2025, the research firm Gartner predicted that more than 40% of agentic AI projects would be abandoned by the end of 2027, because of soaring costs and unclear business value, according to a Reuters report carried by Yahoo. It is a forecast, not a record, but the mechanism is easy to recognise: an AI step added where a rule would have done, then a second, and soon nobody can say what each one earns, or why the bill is rising.
AI projects multiply their costs when they put a model where one is not needed: every unnecessary call is paid for on every run, and every variable result calls for one more check. It is the surest way to waste a budget and send a project adrift. Our analysis of the return on investment of AI in business shows why the return is slow to come when the tool is bought before the task has been chosen and measured.
At which steps is AI justified?
A language model pays off where a rule cannot cope, that is, when the input or the output is language:
- reading free text: the email “the box arrived open, two glasses broken” has to become a complaint, with the right reason and the right order;
- settling an ambiguous case: a request that could fall under customer service or sales, and that keywords classify badly;
- writing a first draft: a reply or a summary, read by someone before it goes out.
In each of these cases, AI takes up one step, not the whole process: fetching the email, finding the order and opening the ticket remain rules. An AI step in a workflow is not an agent, incidentally, and the difference is explained in AI agent: what it is, and what it is not. Anthropic goes the same way in its guide Building effective agents: it recommends “finding the simplest solution possible”, and points out that workflows “offer predictability and consistency for well-defined tasks”.
How do you know whether your task needs AI?
Ask five questions at each step, not of the whole process:
- Can you write the rule down? If a new colleague can apply it with a one-page sheet, a workflow can too.
- Is the input structured? A form or an API response can be read without a model.
- Must the result be identical every time? An amount or a status calls for a rule, which does not vary.
- What does a mistake cost? If the action commits the company, such as a payment or an email sent to a customer, plan for human validation, with or without AI.
- What is the volume? Multiply the cost of a call to the model by the real number of runs over a year before choosing.
If the first three answers are yes, the step does not need AI. If one is no, AI may be justified at that step, and at that step alone.
Rule, AI and human: how do you combine them in one flow?
The most robust flows mix all three. Let us go back to the supplier invoice, received this time as a PDF by email, with a layout that changes from one supplier to the next:
Who does what, from the email received to the accounting entry
- A rule
Fetch the email and the attachment. The path is always the same.
- The model
Read an invoice with a variable layout. The document is not structured.
- A rule
Match the amount to the order. The calculation must be exact every time.
- A human
Approve a discrepancy or a new supplier. The decision commits the company.
- A rule
Create the accounting entry and file the document. The output format is fixed.
Our bookkeeping data-entry agent follows this split of roles: the model reads the photo of a receipt sent on Telegram, automatic checks look for duplicates, test that the amounts are consistent and apply a confidence threshold before any entry is made, and a human approves the record with a single button press. It is one of the tools we run ourselves. Customer support follows the same split: our guide to automating customer support, where AI sorts and humans review, shows where to place the rule, the model and the person.
A rule can also decide when the AI is called.
From the field
A saving that owes nothing to the model
On the translation pipeline we operate for Palmora Property, in six languages, each text segment gets a fingerprint, and only new or modified segments go for translation: after the first pass, 95% of segments come out of the cache without calling the AI again.
Does an automation without AI need monitoring?
Yes, for a different reason: without AI, it does not go wrong unpredictably, but it can fall silent. A source that changes format or an incomplete response, and the workflow stops, or worse, carries on with bad data.
We run our own n8n workflows in production, and two safeguards come up on every one of them. Unexpected data, such as a value missing from a lookup table, is handled by a deterministic rule, with a narrow scope, checked and reversible: with no AI in the loop, there is nothing to hallucinate. And an empty response from the source must never go into production: without a rule that refuses to publish an empty catalogue, a workflow showing green can empty a site overnight.
Our guide to monitoring your n8n automations describes the alerts and recovery that stop an overnight failure being discovered on Monday. And when the software you need to connect has no API, the way you get into it matters more than the choice of tool: we compare six ways in to automate software without an API. And if you plan to resell a service built on n8n, its licence, which is not open source, sets limits that we detail in what the n8n licence allows.
Where to start
The starting point comes down to three moves:
- The task.
Pick one that comes back every week.
- The sequence.
Describe it step by step, with the person who does it today.
- The five questions.
Ask them at each step: what comes down to a rule, an AI or a human soon becomes clear.
When describing the task, also note what it taught the person who did it, and who will take over that role once the task is automated: our analysis Will AI take young people’s first jobs? explains why.
That is how our automation offer works: we observe the task with the person who does it, then show you its cost per execution and per month, with a check-in at 30 days. If you do not yet know which task to pick, our digital and AI maturity audit is there to find out what to automate before you spend, and it rules out in writing the projects that do not hold up. And if the analysis concludes that an agent is justified, our guide to the AI agent in production, from demo to product shows what to plan for so that a team adopts it.
Frequently asked questions
How do I know whether my process needs AI?
Break it into steps and ask the question of each one, not of the whole process. A step whose rule can be written down, whose input is structured and whose result must be identical every time does not need AI. A model is justified where you have to read free text, classify an ambiguous case or write a first draft, and only at those steps.
Is an automation with AI less reliable?
It is less predictable: a model can respond differently to the same input. That does not make it unusable, but it calls for safeguards that a rule does not require: automatic checks on what the model produces, human validation for whatever commits the company, a log to understand an error afterwards. Well framed, an AI step saves time without losing control.
Does an automation without AI really cost less?
At run time, yes: it pays for no model call on any run. It does have a cost, though: it has to be built, monitored and fixed when a source changes format. Self-hosted, it costs only the server that runs it, plus maintenance time. The comparison is made over a year, at real volume, also counting the human time it frees up.
Can you put an AI agent in n8n without losing control?
Yes, provided you limit what it can do. n8n lets you require human approval before an agent uses a specific tool: the workflow pauses and sends an approval request, for example on Slack, on Telegram or in n8n’s chat. If it is refused, the action is cancelled and the agent is told. Keep this check for actions that are hard to undo.
Sources
- Anthropic’s glossary (new tab)platform.claude.com
- Reuters report carried by Yahoo (new tab)tech.yahoo.com
- Building effective agents (new tab)anthropic.com
- Human approval before an agent uses a specific tool (new tab)docs.n8n.io
Jim · founder of Brain Plus AI
Also in automation and bots
Choose and price8 min
n8n, Make or Zapier: which to choose for an SME?
All three tools connect your applications without heavy development, but they do not charge for the same thing and do not give you the same freedom. This comparison for decision-makers looks at the billing unit, hosting, licence and who maintains it the day it breaks.
Practical guide9 min
Monitoring your n8n workflows: no silent failures
An automation that fails without warning is worse than a manual task: nobody knows the work has not been done. Here is the minimal set-up that turns a silent failure into an alert, then into a rerun, using the tools n8n already provides.
Practical guide7 min
Automating software without an API: six ways in
Your business software has no API, or keeps it for a more expensive plan? There is usually still a door for automating it: here are the six, from the most solid to the most fragile, with what each one costs in maintenance.