# Marketing automation: what to automate once you have a system
Source: https://unrivals.com/blog/marketing-automation/
Site: UNRIVALS · Language: en · Updated: 2026-11-08
> Marketing automation repeats what you already wrote, thousands of times. What to automate, what stays human, how to test a step before you switch it on, and what 799 pages read on a real company show.
---
*Marketing automation* means software carries out the repetitive steps of marketing on its own, following rules you wrote. It sends an email after a quote request, routes the request to the right salesperson, puts the same description on hundreds of pages and pulls the numbers into a report.
**TL;DR.** Automate a marketing step only after you have done it by hand, written down the rule and named the person accountable for it, because the software repeats exactly what it receives. We analysed the filed accounts of 17,940 Romanian companies with sales above €3 million in 2024 that also filed accounts for 2025.
Of the 17,067 that reported staff in both years, 3,826 sold more than 10% more per employee without adding staff. And 1,428 of those same companies still ended the year with a lower net result than in 2024. Efficiency helps, and the decision about what deserves repeating stays with you.
We talk almost daily with founders running companies of a few million euros. Many can't find people, many decisions still pass through them, and orders arrive in waves. A natural question follows from that. What can we automate in marketing, so that everything stops depending on us?
Below you will find a map of the processes worth automating, a four-question test and an example run on a real company, with 799 pages read on 4 November 2026. [Our guide to AI as a company processor](/blog/ai-as-processor/) covers the role of AI models overall. Here we look at what your company repeats while you are not watching.
## What is marketing automation?
Marketing automation is the use of software that carries out repetitive steps based on a rule of the form "if X happens, do Y". After a quote request, for example, the software confirms receipt, creates the customer record in the CRM, the system where you keep customers and requests, and alerts the salesperson for that region.
The rule can be simple, like a description template for Google, or it can span several channels, like a series of emails sent according to what someone asked for. **What these have in common is no human at each repetition.** That is the line between automating and delegating.
When you ask ChatGPT for a text, you still read it before it goes out, and [our delegation sheet for ChatGPT](/blog/chatgpt-for-business/) covers that case. With automation, the rule runs at night and during your holiday, with nobody reading each result. That is why its mistakes cost differently.
Three different things now sit under the same name. The first is the fixed rule, written by a person, which always takes the same step. The second is a language model (an LLM, meaning an AI such as ChatGPT or Claude) asked to write the wording on the spot.
The third is the *AI agent*, software that receives a goal and picks its own steps. The more the software decides alone, the more the written rule and the checking matter.
The pressure to use them is real. In a [Gartner survey published in May 2026](https://www.gartner.com/en/newsroom/press-releases/2026-05-11-gartner-survey-reveals-marketing-leaders-expect-ai-automation-of-marketing-work-to-double-to-36-percent-by-2028), 402 chief marketing officers expect AI to automate 16% of marketing work in 2026 and 36% by 2028. **The share of companies that keep customer records in a CRM hints at how ready they are to automate.**
[Eurostat data for 2025](https://ec.europa.eu/eurostat/databrowser/view/isoc_eb_iip/default/table?lang=en) show that 28.5% of EU companies with at least 10 employees use CRM software, and 13.9% of Romanian ones. The figure only counts CRM use, and records can exist in other forms. Still, in roughly 7 of 10 such companies across the EU, customer records are not kept in such a system.
Company size explains the temptation. Of the 17,255 Romanian companies above €3 million in 2024 sales that declared employees in their 2025 accounts, half have 41 people or fewer, and 4,744 have fewer than 20. At that size, every departure hurts, and software that never resigns looks like the right answer.
## Why does automation repeat mistakes just as fast?
Automation repeats mistakes because the software executes the rule without judging it. It doesn't know whether the sentence is good, whether the data are correct or whether the step still makes sense after an offer changes. It receives an instruction and applies it a thousand times with equal confidence, whatever that instruction is worth.
The first principle from our methodology that applies here is *infrastructure before marketing*. It states that "the client rarely has a marketing problem, the client lacks infrastructure (systems, channels, positioning, a brand foundation), and marketing on a weak base burns money". **Automation on a weak base burns money faster.**
The cost of a weak base shows up in international surveys too. According to [Gartner's 2025 marketing technology survey](https://www.gartner.com/en/marketing/topics/marketing-technology), only 49% of companies' marketing tools are actively used. The gap is capacity that is paid for and left idle.
Data are the second problem. In [Validity's July 2025 report](https://www.prnewswire.com/news-releases/validity-releases-state-of-crm-data-management-in-2025-report-revealing-disconnect-between-data-quality-and-ai-implementation-302499899.html), based on 602 CRM users in the US, UK and Australia, 76% say less than half of their data is accurate and complete. Validity sells solutions for marketers, so read the figure with that in mind.
Think about what a rule does with data like these. If the customer record holds a wrong name, the automated email repeats it. If the industry field is empty, the customer receives another industry's offer. **A correct rule on wrong data produces tidy mistakes, sent on time.**
AI agents don't avoid the problem. In [another Gartner survey, from October 2025](https://www.gartner.com/en/newsroom/press-releases/2025-10-29-gartner-survey-finds-45-percent-of-martech-leaders-say-existing-vendor-offered-ai-agents-fail-to-meet-their-expectations-of-promised-business-performance), covering 413 marketing technology leaders, 45% of those testing or using agents say vendor-offered agents do not meet the promised performance. Half say their organisation lacks the technical and data readiness.
Aparajita Mazumdar, Senior Principal, Research at Gartner, draws the conclusion in the same release. "To fully realize the benefits of AI, CMOs must also look inward and address gaps in their technical and data infrastructure."
The second principle treats AI as a processor. Our statement says AI is a cognitive processor, an extension of how the company thinks, and that the result "depends on how precisely you tell it where to think". For automation, precision means the rule written before launch, the data it starts from and the point where a person checks the result.
## Does marketing automation increase profit?
Efficiency improves the odds of ending the year ahead, yet it does not guarantee profit, and the accounts do not show who automated anything. Romanian company filings show this fairly clearly. Sell more with the same people and you end up ahead more often, while more than a third of the companies that managed it still closed the year with a lower net result.
Romania publishes company accounts, so this can be checked on real numbers. We analysed the accounts filed with the Finance Ministry for 2025 and 2024, for the 17,067 companies with sales above €3 million in 2024 that declared employees in both 2025 and 2024.
Sales per employee means sales divided by headcount, in lei (the Romanian currency), without adjusting for inflation. Net result is the final profit or loss.
Share of companies with a lower net result in 2025 than in 2024
Sales per employee flat or lower, 7,459 companies68.8%
Sales per employee higher, 9,608 companies39.2%
Of those, over 10% more per employee with no added staff, 3,826 companies37.3%
Source · accounts filed with the Romanian Finance Ministry for 2025 and 2024, UNRIVALS calculation on 17,067 companies with sales above €3 million in 2024 and employees declared in 2025 and 2024
The first bar shows what happens to companies that do not become more efficient, with more than two in three ending the year with a lower result. The second bar shows the gain associated with efficiency. **Efficiency is worth pursuing, and the numbers say so plainly.** The red bar shows its limit.
Those 3,826 companies come closest to the promise of automation, more output with the same people or fewer. Of those same companies, 1,428 ended the year with a lower net result. The drop from 39.2% to 37.3% is very small, even though the jump in efficiency is large.
The accounts don't tell us who automated anything, or why profit fell. We have a hypothesis and we mark it as one. Efficiency lowers the cost at which you sell, and it leaves the reason customers choose you untouched, so the gain can leak into price or other costs.
The hypothesis ties to the third principle in our methodology. It says the marketing budget is for development, meaning strategy and architecture, "not for optimising the same channel forever", and that the right strategy lowers real cost.
**Automation makes the same channel less expensive to run.** If margin is your problem, [our guide to profit margin](/blog/profit-margin/) shows how to check it first.
## What does automation look like on a real company?
It is easiest to see on a template, one of the simplest forms of automation, a sentence written once and repeated across hundreds of pages. We chose TAMOS, the brand under which Conart S.A., a Romanian furniture factory in Zimnicea with 24 years of activity and 48 employees, sells online.
We published [our audit of TAMOS](https://unrivals.ro/deck/tamos) (in Romanian) in September 2026, from public data. According to [Conart's 2025 accounts, published by the Romanian tax authority](https://webservicesp.anaf.ro/bilant?an=2025&cui=14495295), the company had sales of 14.8 million lei and a net profit of 621,287 lei. In 2024, on sales of 16.9 million lei (about €3.4 million, above the threshold of the analysis above), net profit was 1.93 million lei. So TAMOS sits among the 7,459 companies in the first bar, with lower sales per employee in 2025 and a lower net result.
On 4 November 2026, we read all 799 product pages in the site's sitemap, the list of pages it gives search engines, and noted each page's title and the description the company gives search engines. Anyone can repeat the measurement, because the data are public.
TAMOS, 799 product pages, the description for Google
323contain the same Romanian wording, "advantageous prices direct from the factory"
40.4% of pages
505talk about price
63.2% of pages
0mention the 24 years of factory work or the 200,000-plus customers the company declares
in title or description
Source · direct reading of the pages on tamos.ro, 4 November 2026, UNRIVALS calculation
The red figure is what matters. **The company has two pieces of proof a competitor can't buy, and the template leaves them out.** The audit names exactly those 24 years and those customers as the only assets that are hard to copy. In product descriptions, price takes their place.
The identical wording leads us to read this as a template, although we can't see from outside how the site is set up. The Romanian word for low-priced, "ieftin", doesn't appear on product pages. It remains in the homepage description, the same sentence we flagged in the audit.
The site also has pages that look generated in series. Its general sitemap has 3,537 addresses, and 3,076 of them begin with "mobila-", meaning furniture, category pages included. In a random sample of 250 addresses, 220 are town pages such as "Mobila Stubeiu", and on 195 of those the description simply repeats the title.
Ahrefs estimates 31,927 monthly visits from unpaid Google search for the site's top 60 pages, in November 2026. At most six of those are town pages, bringing roughly 3.5% of the visits between them. We did not measure what the pages outside the top 60 bring.
Now we apply the four questions from the section below to the product descriptions. *Did it work by hand?* From outside we can't know. *Does the rule fit in one sentence?* Yes, you can read it on 323 pages. *Are the data correct?* All 799 pages have a title and a description, so the mechanics work.
*Who is responsible when it breaks?* We can't see the answer from outside, and we can see the effect. Both proofs are missing from all 799 descriptions.
We are not saying the template lowered Conart's profit. The accounts don't show that, and we can't see the pages' internal results. We are saying only that **the software carried the price criterion onto hundreds of pages without asking anyone.** This is automation in its simplest form, and the mechanism works flawlessly.
The good news comes from the same mechanism. **If the descriptions come from a template, one changed sentence reaches hundreds of pages.** Our hypothesis is "24 years in our own factory in Zimnicea. Over 200,000 customers".
You test it on one category for 90 days. You measure CTR, the share of people who click after seeing the result in Google, and the average price you sell at. Only then do you decide whether the sentence goes across the whole site.
## What should you automate in marketing, and what stays human?
Automate the steps with a clear rule, clean data and inexpensive mistakes, and leave the steps where a mistake costs a customer or a position to people. The line depends less on technology and more on what a wrong repetition costs you. The table below puts the usual marketing processes on one page.
| Process | What the software does | Switch it on when | What stays human | What breaks if you start early |
|---|---|---|---|---|
| Monthly report | pulls numbers from ads, site and CRM | the numbers mean the same thing every month | interpretation and the decision | a polished report on conflicting definitions |
| Quote requests | confirms, creates the record, alerts the salesperson | you have settled who answers, for which region and how fast | the conversation with the customer | the request reaches someone who has left the company |
| CRM data entry | fills the record from the form and from email | fields are few and mandatory | periodic cleanup | duplicate records and wrong names in emails |
| Page templates | put the title and description on hundreds of pages | you have decided what the company says about itself | the sentence in the template | the wrong criterion, repeated everywhere |
| Email series | sends messages based on what someone asked for | the series was tested by hand on real customers | the copy, stopping the series | messages that continue after the customer has bought |
| Lead scoring | ranks requests by signs of interest | sales can tell you which requests become contracts | the threshold and its review | salespeople chase high scores while the contracts come from elsewhere |
| AI-written copy | proposes variants from a brief | the position, the proofs and a list of bans exist | reading it and being accountable for it | invented numbers and promises you can't keep |
| Position and offer | nothing, perhaps the research | it does not get automated | the entire decision | a company that says what everyone says |
Read the table top to bottom, from the inexpensive mistake to the costly one. **The first rows start first, and the last row never starts.** A wrong report gets corrected next month. A wrong email sent to the whole list reaches your customers, and the effects can't be undone.
There is also a technical cost, paid on every future email. [Google's rules for email senders](https://support.google.com/mail/answer/81126?hl=en) require the reported spam rate to stay below 0.30%, and anyone sending more than 5,000 messages a day to Gmail accounts must offer one-click unsubscribe.
An automated series sent to people who don't expect it can damage your domain's reputation, so offers sent by your salespeople from the same address may struggle to arrive as well. That is why the stop rule matters as much as the start rule. **Every automation needs a written stop condition.**
The AI row deserves separate attention, because it is the most tempting. [Our guide to marketing with AI](/blog/marketing-ai/) explains how AI works from an existing strategy. For automation, one thing counts, namely that text written by a model doesn't go to customers until a person has read it.
## How do you know a step is ready to automate?
A step is ready to automate when you can answer yes to four questions. You have done it by hand and know it works, the rule fits in one sentence, the data are correct, and someone is responsible when it breaks. If one yes is missing, the work to do is that missing answer, and the software can wait.
The four questions before you switch automation on
Did it work by hand?
The step delivered when a person did it, enough times
Does the rule fit in one sentence?
A new hire can apply it without asking
Are the data correct?
You checked 20 records picked at random
Who is responsible when it breaks?
A name, an alert and the stop condition
The red link is the easiest to miss, because no vendor can sell it. A platform can give you the rule and the data. **Nobody can buy you the person accountable for the result.** Without that person, the automation runs until a customer tells you it broke.
The order comes from our sequencing principle, which asks us to educate the market and the team gradually. It states that "each stage has its own objective", and that skipping one means burning stages and losing. In automation, the stages are done by hand, written, checked, automated. Each produces the material the next one needs.
How many repetitions count as "enough" is not a scientific number. We recommend, as a rough guide, seeing the step work on at least 20 real cases, enough for exceptions to appear. An exception that shows up twice should enter the rule or go to a person.
The exercise below takes about an hour, and you do it yourself.
1. Search Google for `site:yourcompany.com` and read the descriptions of the first 30 results. Google can rewrite them, so check a few pages directly too. Note the sentence that repeats most.
2. Next to it, write what you have that a competitor can't say, with the number or proof behind it. Compare the two lines.
3. Pick from the table above one process that hurts and answer the four questions in writing. Put a name next to the fourth.
4. Pull 20 records from your CRM or from whatever records you keep. Count how many are complete and correct.
5. Give the rule a bad case, such as a record with no name or a duplicate request, and watch what it does.
Step 5 is the easiest to skip, because its result doesn't look good in a presentation. **A test on a bad case shows what the system will do on an ordinary day.** If the rule walks past the bad case without a word, it isn't ready.
## What have we automated, and what did we take back out?
We automate the gathering of public data, the record of each project's stage and mechanical checks, while decisions and sending stay with people. We got here through mistakes we measured. All the cases below are ours and are described at length in the linked articles.
The first case comes from email sending. Platforms give no error when a template variable, the slot filled with a recipient's name or company, is missing, and they quietly replace it with a blank. So in [our cold email method](/blog/cold-email-b2b/) we check every variable by code, read the messages for several real recipients and receive a copy in our own inbox before the others go out.
The second case is an automation we removed. On 23 September 2026, we tested whether an AI model could judge a text's clarity on its own.
It caught all 7 bad sentences planted on purpose, and it also flagged 11 of the 18 sentences in a text a human had already approved. **We dropped the automatic checker, and the final read stayed human.**
The third case concerns how we check our checks. We gave a quality gate, an automatic check, a deliverable we had broken on purpose, and the gate approved it. We told that story in [our article on the AI Brain, the memory system a company keeps for its AI](/blog/ai-brain-for-founders/), along with the fourth case.
A platform confirmed it had saved data that looked different when we read them again. Since then, any change on a system customers use gets read back separately from the confirmation.
The fifth case shows what remains after the repairs. On 3 October 2026, the quality gate rejected on first pass a batch of three AI-written articles, because our estimates appeared written as facts. That gate handles what can be checked mechanically, and a person goes on to read everything that leaves.
**The pattern in all five is the same.** Automation stayed where another rule or a person can check it, and it came out where it only gave a feeling of safety. What the company knows stays written in its documents, and programs can change.
## Which company layer does automation sit on, between L1 and L4?
Automation works on L1 and L2, where traffic, requests and sales live. It is coordinated from L3, and L4 gives it its content. Our methodology divides the work into four layers, from L1, performance marketing, meaning ads and traffic, to L4, positioning. The mistake appears when a company buys tools for L1 and L2 before making its L4 decisions.
What automation does on each layer
L4 · Positioning and category
Does not get automated, and decides which sentence gets repeated
L3 · AI Brain
Company memory, rules and checks
L2 · Revenue and sales process
Requests in the CRM, routing, email series, lead scoring
The red layer is the only one with no automation and the only one that decides the content of the others. **Every rule below carries forward a decision made above.** At TAMOS, the template works on L1, and the sentence inside comes from an L4 decision, made or not made. The software carries it without knowing.
L2 is the layer where time is easiest to win, because that is where requests get lost between marketing and sales. It is also where you see whether sales are under control, whether you know what happens between offer and order. A rule that alerts a salesperson and reminds them to follow up helps only if clear stages exist.
L3 is where the rules live, what we call the AI Brain. **There you keep the template sentence and the reason it was chosen,** along with the name of who can change it. Without this layer, an automation started today becomes, two years on, a rule nobody in the company can explain.
That is why [marketing architecture](/blog/marketing-architecture/) places automation after strategy and infrastructure. Across the layers, it all comes down to one sentence you can check in your company. *If you can't say which L4 decision a rule carries, you have no reason to run it.*
## What are examples of marketing automation tools?
Examples of marketing automation tools include HubSpot, Adobe Marketo Engage, [Salesforce Marketing Cloud](https://www.salesforce.com/marketing/b2b-automation/) and [Zoho Marketing Automation](https://www.zoho.com/marketinghub/). They handle email series, lead routing, scoring and campaign reporting. A CRM with built-in rules, or a site template, is automation as well, even if nobody calls it that.
The list says little on its own. **The difference comes from what you put into the tool.** The same platform can run a clear rule set for one company and a pile of forgotten campaigns for another. So read the question that follows as a question about your company as much as about the software.
## Which automation tool is the best?
No tool is best in general. Choose the one that connects to your CRM, that your team can run, and whose rules you can read and change without the vendor. For companies with no CRM, the first tool is the CRM itself, with a single automation switched on.
[Adobe](https://business.adobe.com/blog/2026-gartner-magic-quadrant-b2b-marketing-automation-platforms) and [HubSpot](https://www.hubspot.com/company-news/gartner-magic-quadrant-2026) announce they are Leaders in Gartner's 2026 ranking of B2B platforms. We read their announcements and not the full report, so treat that list as a starting point.
Four criteria matter more than the ranking. The first is the connection to where your data lives. The second is who runs the tool after the person who installed it leaves. The third is how easily you can read the active rules. Finally, consider how easily you can leave with your data.
**A large platform risks staying half used at a company with no written rules,** as in the Gartner survey. **Choose the tool after answering the four questions** for your first process. Tools change more easily than the habit of working with rules.
## What obligations come with automating messages and AI copy?
When you automate with AI, the company that decides to use it is accountable for the result, and in the EU the AI Act's transparency obligations have applied since 2 August 2026. In short, photorealistic AI images get a label, and so does public-interest text with no human review.
The practical consequence ties directly to automation. A flow that publishes public-interest text on its own loses the exemption from labelling, because the exemption requires a person who reads the text and is accountable for publishing it.
**The person reading the text is also your legal protection.** The details, paragraph by paragraph, are in [our guide to what the EU AI Act requires](/blog/eu-ai-act/).
Email has its own rules. Automated commercial messages fall under each country's electronic marketing law and under data protection rules, whether a person or a program sends them. **Sending on its own doesn't change who is accountable for it.** Before any series, ask whoever handles compliance in your company.
## What founders tell us about automation
Three worries come back in our conversations with founders, and each has a different answer. We put them here as we have understood them, because each can change the order in which you automate. Start with the one costing you now.
The first concerns people. Many tell us they can't find staff and that almost everything passes through them. Automation can take the report and the request tracking off their shoulders. But someone still has to think through a written rule, and that first writing takes your own time. **Automation frees up time only after it has cost you time.**
The second concerns sales, which arrive in waves with nobody able to say why. Here automation helps through record keeping, because every request lands in one place, with a name and a date next to it. **Only after a few months of correct records do you have something to improve.**
The third concerns strategy, which exists on paper and doesn't happen. The temptation is to put it into a program and hope the program runs it. *A program runs only what has already become rules.* Translating strategy into rules stays human work, and everything after it depends on that.
## Frequently asked questions
### Which automation tool is in demand in 2026?
We did not find a neutral ranking of demand, so we won't invent one. Two signals exist. Adobe and HubSpot announce Leader status for 2026, and [Chief Marketer reports on Gartner's 2026 spend survey](https://www.chiefmarketer.com/gartner-cmo-spend-survey-budgets-reflect-increase-in-consumption-based-martech-paid-media-spend/) that 62% of 401 chief marketing officers (CMOs) plan to invest more in marketing technology, while martech's share of the marketing budget fell to 19.4%, a five-year low (26.6% in 2021).
### Can you use ChatGPT to automate tasks?
Yes, within limits. According to [OpenAI's documentation](https://help.openai.com/en/articles/10291617-scheduled-tasks-in-chatgpt), ChatGPT can run tasks once or on a schedule, depending on your account. It suits monitoring and summaries you read yourself. We wouldn't let it send or publish on its own in the company's name.
### Is Zoho a marketing automation tool?
Yes. Zoho Marketing Automation is sold as a multichannel product for running campaigns and generating sales-ready leads, according to Zoho's own page. Whether it fits you depends on the criteria above, and especially on whether your CRM and your written rules are in place first.
### Is SFMC a marketing automation platform?
Yes, SFMC usually stands for Salesforce Marketing Cloud, and Salesforce presents Marketing Cloud on its B2B marketing automation page. The same page describes a shift from rules and triggers toward AI agents, so apply the four questions to whatever runs on its own.
### Do I need a CRM before automating marketing?
For automations involving customers, yes, or at least a single set of records kept correctly. Without one place where the customer record lives, the rules have nowhere to take data from. You can start with a simple CRM and a few mandatory fields the team really fills in.
## What does your company repeat while you are away?
**Every active rule speaks in the company's name each time it runs.** You wrote some of them, others were set up by someone who has left, and a few came with the platform. The first job is to find them and read what they say.
Then you choose. **Put on repeat only what you have decided to own in the customer's mind.** Everything else can wait until it passes the four questions.
If you want an outside eye on what your site and communication repeat today, [ask for a diagnostic](/#diagnostic). We show you, on public data, which criterion your rules carry forward and which proof is missing from them.