By 
Shreyas M V · Published · 13 min read
When Not to Use AI in Marketing: Tasks Where It Costs More Than It Saves
Skip AI for a marketing task when briefing it, reviewing it and redoing the rejected part takes longer than doing the task yourself. This page gives a break-even formula and seven tasks to hold back, each tied to a published source. It contains no tests of our own, and the hour figures in its examples are hypothetical.
When should you not use AI for a marketing task?
Do not use AI for a marketing task when the time to brief it, review it and redo the rejected part exceeds the time to do the task yourself. The tool's speed is not the issue. The checking is.
- Review time decides the cost, because you pay it on every output and it grows when a wrong answer is hard to spot.
- Hold back positioning, emotional customer messages, analysis on thin data, personalization built on personal data, replies in communities that ban generated text, and anything you cannot afford to review.
- Keep AI for output you can check at a glance, the case Guillaume Decugis described in the Marketing AI Institute's 2021 piece.
- The evidence is three published pieces, and we ran no tests. The hour figures here are hypothetical and the calls are our judgment.
The three top-ranking pages we read on this topic list drawbacks such as cost, limited creativity and privacy. None of them prices a drawback in hours. This page does, using a formula you can run with your own numbers.
How do you tell when AI costs more than it saves?
AI costs more than it saves when the hours spent briefing, reviewing and redoing exceed the hours the task takes by hand. Four numbers settle it: H, the hours to do the task by hand; G, the hours to brief and generate; R, the hours to review and fix one draft; and p, the share of drafts you scrap and redo by hand.
AI saves time when: G + R + (p × H) < H
The formula is ours, not a source's. It says the tool wins only when the brief, the review and the expected redo together stay below the original task.
The table runs the arithmetic for two tasks. The inputs are hypothetical, chosen to show how the result flips; they are not measurements, and Kanso did not time anyone.
| Input | Task A | Task B |
|---|---|---|
| Hours to do it by hand (H) | 1.0 | 2.0 |
| Hours to brief and generate (G) | 0.1 | 0.25 |
| Hours to review and fix one draft (R) | 0.2 | 1.5 |
| Share of drafts scrapped and redone by hand (p) | 0.1 | 0.5 |
| Redo time, p × H | 0.1 × 1.0 = 0.1 | 0.5 × 2.0 = 1.0 |
| Total with AI, G + R + p × H | 0.1 + 0.2 + 0.1 = 0.4 | 0.25 + 1.5 + 1.0 = 2.75 |
| Result against doing it by hand | Saves 0.6 hours (1.0 - 0.4) | Costs 0.75 hours more (2.75 - 2.0) |
Generation time barely moved between the tasks. Review time and the share of scrapped drafts did, and they turned a saving of 0.6 hours into a loss of 0.75. Most arguments for AI count G and skip R.
The formula also leaves out subscription fees and setup. Dentsu and the Medium post both list implementation cost as a barrier, and Dentsu says small and medium-sized businesses can find it a major one, so add those to the right-hand side if you pay for them.
Why does review cost decide the answer?
Review cost decides the answer because it is paid on every output and it rises when a mistake is hard to see. The Marketing AI Institute's 2021 article quotes Guillaume Decugis, CEO of Linkfluence, saying that recommendations which are right only 80 or 90 percent of the time still provide value, because humans can quickly tell the correct actions from the incorrect ones.
That is the piece's clearest case for using AI, and it carries a condition. The value holds only while the check is quick. A list of ten keyword ideas can be scanned in a minute, while a 1,500-word draft full of claims needs every name and number verified, which is our reading rather than a measured result.
Decugis also said that building systems where AI replaces humans entirely is very hard. So the review is never zero, and a plan that assumes it is zero is the one that fails.
Two other limits make review slower. The Medium post says many algorithms are black boxes, which makes errors harder to find and responsibility harder to assign. And Kanso's weekly growth checklist reports Google's guidance that AI-generated content should be manually fact-checked and reviewed before publishing, so skipping review is not only a quality risk.
Which seven marketing tasks should founders hold back from AI?
Seven kinds of work tend to cost more than they save: positioning and voice, emotional customer moments, analysis on thin data, work across disconnected tools, personalization on personal data, community replies and anything published without an affordable review. The table pairs each with the limit a source names and our call.
| Task | Limit the sources name | Where it comes from | Our call |
|---|---|---|---|
| Positioning, point of view and brand voice | Limited creativity and originality; repetitive, mechanical output | Dentsu; the Medium post | Do not delegate. Use AI only to critique your draft. |
| Emotional customer moments: complaints, apologies, cancellations | Emotion, feelings and subjective thinking; empathy and human touch | Or Shani in the Marketing AI Institute piece; Dentsu | Do not delegate. Write it, then have a colleague read it. |
| Analysis and forecasts on thin data | Capability depends on data richness; small evidence is hard; prediction can be uncertain | Adam Long and Pawan Deshpande in the Marketing AI Institute piece; Dentsu | Do not trust unreviewed. Collect the data first. |
| Work spread across tools that do not share data | Disconnected stack causes data loss and delays | Tom Gerace in the Marketing AI Institute piece | Connect the data before adding AI on top. |
| Personalization built on customer personal data | Privacy exposure; bias; unfair targeting; hard-to-explain decisions | Dentsu; the Medium post | Hold until someone qualified has reviewed the data handling. |
| Replies in communities that ban generated text | Hacker News does not allow generated or AI-edited text | Hacker News Guidelines, as summarised in Kanso's Hacker News guide | Do not use. Write every word yourself. |
| Anything published without a review you can afford | Replacing humans entirely is very hard; AI pays only while a human can sort right from wrong quickly | Guillaume Decugis in the Marketing AI Institute piece | Do not publish unreviewed. Budget the review first. |
How the table was put together: we took the limitation lists in the Marketing AI Institute article, Dentsu's page and the Medium post, kept the limits that map to something a founder actually does, and applied the review-cost test above to each. The sources describe limits. The last column is our decision.
Why should positioning and voice stay with the founder?
Positioning and voice should stay with the founder because review can fix a draft but cannot add an opinion the draft never had. Dentsu says AI can assist the creative process but cannot fully replace the human touch in producing original ideas, and the Medium post warns that over-reliance can produce repetitive, mechanical content that fails to resonate.
Here the rewrite rate p is high, because a draft with no point of view needs the point of view written in. When you rewrite more than you keep, you have paid for the draft and the writing.
A fair worry is that when competitors use the same tools, their output converges. None of the three pieces measures that, so treat it as reasoning, not a finding.
There is a cheap use that stays on the safe side of the test. Ask AI to attack your draft, list the objections a skeptical buyer would raise or point out vague sentences. You can judge each suggestion in seconds, and the voice stays yours.
Why not hand emotional customer moments to AI?
Complaints, apologies and cancellations are poor candidates because Or Shani, CEO of Albert, named emotions, feelings and subjective thinking as the obvious limits of AI in the Marketing AI Institute piece. Dentsu makes the same point, saying AI cannot replace the empathy and understanding that people build with customers.
Review hours are not the only cost here. A wrong tone reaches a person who is already unhappy, and an edit after sending is not possible. That is our analysis, not a sourced figure.
Routine questions are different. Dentsu lists responsive chatbots that answer customer questions among AI's advantages, so the line runs between a question with a checkable answer and a moment where the customer needs to feel heard.
When does AI analysis cost more than it saves?
AI analysis costs more than it saves when the data behind it is thin, scattered or unreliable. Adam Long, then VP of Product Management at Automated Insights, said AI is limited by the availability of data, and Dentsu says incomplete, inaccurate or biased data degrades the results.
Pawan Deshpande described two further limits in the same piece. AI struggles with very small amounts of evidence, and optimizing two dependent conversion rates at once, such as a landing page test and an ad creative test, becomes too complex, so the two are run independently.
For a founder, the small-data case is common. A company with few customers or little traffic has little for any model to learn from, and you cannot check a forecast against a history you do not have.
Tom Gerace, then CMO of EquiLife and founder of Skyword, added the tooling problem. When content, analytics, sales and channel data sit in separate platforms, data is lost, you see behavior only in aggregate and you cannot act in near real time. Fixing that costs less than any model you put on top of it.
Why is AI personalization risky for a small team?
AI personalization is risky for a small team because it runs on large volumes of customer data, and both Dentsu and the Medium post flag privacy as a core exposure. The Medium post names the General Data Protection Regulation and India's Digital Personal Data Protection Act, and Dentsu says breaches can damage a company's reputation and bring serious legal penalties.
The Medium post adds three more risks: bias in historical data that can reinforce stereotypes or exclude groups, discriminatory pricing or ad placement, and behavioral predictions used to push impulsive buying. Each is hard to catch by reading one output, which is where the review-cost test bites.
This is a marketing view, not legal advice. Kanso does not check regulatory compliance, so have a qualified person review how any tool stores and uses customer data before you switch personalization on.
Should you use AI to write replies in online communities?
Not on Hacker News, where, as of October 2026, Kanso's Hacker News guide reports that the site's guidelines do not allow generated or AI-edited text. Elsewhere, a reply is judged as one person's voice, and the other members decide whether it reads that way.
The arithmetic works against it too. A ten-minute reply leaves little room: five minutes to brief and read the draft and five to review it use up the whole task before any rewrite, with hypothetical numbers again.
Kanso's Reddit and Hacker News reply framework sets out what a good reply contains. It is written to be done by a person.
Where does AI still earn its place in marketing?
AI earns its place where output is quick to check and an error is cheap, which is the case Decugis described. Dentsu's list of advantages fits that test: scheduling social posts, sending emails, choosing send times, adjusting ad targeting and answering routine customer questions.
The Marketing AI Institute lists advertising, analytics, content marketing, communications and PR, customer service, ecommerce, email marketing, social media and sales as areas AI is beginning to affect. It also claims AI will bring more than $2 trillion in value to marketing and sales. The text we read gives no method for that figure, and the same page promotes an AI class, so we do not rely on it.
Nothing here argues for staying away from AI. Andre Konig, cofounder of Opentopic, urged customers and vendors to stay with applications that work and deliver results, and the formula tells you which applications those are for your business.
How do you test a task before you give it to AI?
Test it by timing one honest attempt from brief to publishable output and comparing the total with the time to do it by hand. This is a process we suggest, not one a source prescribes.
- Time the manual version. Record H for a normal instance of the task, or estimate it from the last time you did it.
- Run one attempt and time the brief. Record G, including the time spent supplying context.
- Review to your publishing standard. Check every name, number and claim, and fix the tone. Record R, which is the number people skip.
- Judge whether you would keep it. Repeat a few times, since one attempt can only be kept or scrapped, and estimate p from the results.
- Compare. If G + R + p × H is below H, plus any fees, keep using it. If it is above H on several attempts, stop.
- Test again later. The Marketing AI Institute piece dates from 2021, and tools change, so repeat the test every few months.
Put the person who knows the subject in the reviewer seat. A reviewer who cannot spot a wrong claim turns R into a guess.
What is documented and what is our analysis?
The documented points are what the three pieces and two of Kanso's own articles say; the formula, the examples and the calls are our analysis. Keeping them apart shows how much of this page you can verify.
Documented
- As of the Marketing AI Institute article of 8 November 2021, six experts named these limits: dependence on rich data, uncertain prediction and analysis, the difficulty of replacing humans, a disconnected marketing stack, weak handling of emotion, and unrealistic expectations.
- Dentsu's page (no date visible) lists high implementation costs, limited creativity, data privacy, dependence on data and lack of human touch as the disadvantages of AI in digital marketing.
- A Medium post (no date visible) lists privacy and security, bias, lack of creativity and emotional intelligence, opaque algorithms, job loss, cost and manipulative marketing.
- Guillaume Decugis said recommendations that are right only 80 or 90 percent of the time still provide value because humans can quickly sort correct from incorrect actions.
- As of October 2026, Kanso's Hacker News guide reports that the site's guidelines do not allow generated or AI-edited text.
- As of 13 October 2026, Kanso's weekly growth checklist reports Google's guidance that AI-generated content should be manually fact-checked and reviewed before publishing.
Our analysis
- The break-even formula, G + R + p × H against H, is our own arithmetic. No source in the research gives it.
- Every hour figure in the worked example is hypothetical, chosen to show the arithmetic. They are not measurements, and Kanso ran no test to produce them.
- The calls in the task table, such as do not delegate, are our judgment. The sources describe limits; they do not tell a founder what to do.
- The claim that review is the expensive term is our reading of Decugis's point, not something a source measured.
- The expert quotes on sentiment analysis and machine vision date from 2021. Some limits may have changed since, and we did not test current tools.
- The worry that identical tools produce identical output is our reasoning. None of the three pieces measures it.
The evidence is narrow. Two of the three pieces carry no visible date, one is a personal post, and none reports measured results. The sources are enough to name the limits and not enough to price them.
Where does Kanso fit, and where does it not?
Kanso fits the finding and drafting work, and it does not remove the review step this page is about. Its Writer Agent turns Search Console opportunities into long-form drafts in your brand voice for you to review and publish, and its Reddit, X and LinkedIn agents draft replies and posts for you to edit and post yourself.
Kanso does not auto-post and does not schedule X or LinkedIn posts, and its Hacker News Agent finds conversations and recommends what to do but never posts. The review hours stay with you, so run the break-even test before you rely on any draft. Plans are on the pricing page.
Limitations of AI in marketing: frequently asked questions
What are the disadvantages of AI in marketing?
The disadvantages named in the three pieces we read are cost, limited creativity, data privacy, dependence on data, lack of human touch, bias, opaque decisions and unrealistic expectations. Dentsu lists the first five of those for digital marketing, a Medium post adds bias, transparency and job loss, and the Marketing AI Institute's experts add disconnected tools and expectations. This page adds a cost view: each disadvantage matters in proportion to the review time it creates.
What are the limitations of AI in marketing?
The main limitations are its need for rich data, its uncertain predictions, its weak handling of emotion and the difficulty of replacing people entirely. In the Marketing AI Institute's 2021 piece, Adam Long said AI is limited by the availability of data, Pawan Deshpande said small amounts of evidence are hard to work with, and Guillaume Decugis said fully replacing humans is very hard. Because the piece is from 2021, test current tools before relying on a limit.
What are the risks of AI in marketing?
The documented risks are privacy and security failures, biased targeting, decisions nobody can explain and manipulative use of behavioral predictions. The Medium post names the GDPR and India's Digital Personal Data Protection Act in connection with the privacy risk, and Dentsu says breaches can damage reputation and bring legal penalties. For legal questions, ask a qualified lawyer; this page takes a marketing view only.
How is AI affecting marketing?
The Marketing AI Institute says AI is beginning to affect advertising, analytics, content marketing, communications and PR, customer service, ecommerce, email marketing, social media and sales. The same article claims AI will bring more than $2 trillion in value to marketing and sales by raising revenue or cutting costs, but the text we read gives no method for that figure, so we treat it as a claim. Dentsu lists personalization, automation, data analysis, optimization and customer experience as the gains.
Is it possible that marketing will be completely replaced by AI?
The sources we read do not support a full replacement. Guillaume Decugis said building systems where AI replaces humans one hundred percent is very hard, and Dentsu says AI cannot replace the empathy and relationships people build. They do not forecast jobs, though, and the Medium post raises job displacement as a concern, so the safe reading is that the work changes while review and judgment stay with people.
What are the advantages of AI in marketing?
Dentsu names personalization, automation of routine work, data analysis, optimization and a better customer experience. Examples it gives include scheduling social posts, sending emails, choosing the best time to send and answering customer questions with chatbots. Those are jobs where output is quick to check, which in our reading puts them on the safe side of the review-cost test on this page.
What are the pros and cons of AI in marketing?
The pros Dentsu names are personalization, automation, data analysis, optimization and customer experience, and its cons are cost, limited creativity, privacy exposure, data dependence and loss of human touch. The Medium post and the Marketing AI Institute's experts add bias, opaque decisions and unrealistic expectations on the con side. None of them prices the cons, which is what the break-even formula on this page is for.
What are the drawbacks of AI in digital marketing?
Dentsu lists five: high implementation costs, limited creativity, data privacy, dependence on data and lack of human touch. It adds that small and medium-sized businesses can find the costs a major barrier. A founder should add a sixth to that list, the hours spent reviewing output, because it recurs on every task.
What problems does AI cause in marketing?
The problems named in the sources are weak output from incomplete data, data lost between disconnected tools and repetitive, mechanical content. Dentsu ties output quality to data quality, Tom Gerace describes data loss across a disconnected stack, and the Medium post warns of repetitive content. The rework cost of all three is our estimate, not a measurement.
Which marketing tasks should a founder not give to AI?
A founder should not hand over positioning and voice, emotional customer messages, analysis on thin data, personalization built on personal data, replies in communities that ban generated text, or anything that will be published without a review. Each is a case where the check is expensive or the error is costly. The table on this page gives the source behind each call.
Sources
- 6 Limitations of Marketing AI, According to Industry Experts (Paul Roetzer, 8 November 2021), Marketing AI Institute
- Advantages and Disadvantages of Artificial Intelligence in Digital Marketing, Dentsu Indonesia (no date visible in the text read)
- The Dark Side of AI in Marketing: Risks and Limitations, Medium (a personal post, no date visible in the text read)
- The Weekly Growth Checklist: SEO, AI Search, Content, Competitors and Distribution, Kanso, 13 October 2026
- Hacker News for B2B Founders: When to Post, When to Comment, When to Stay Out, Kanso
- Writing Comments That Earn Trust: A Reply Framework for Reddit and HN, Kanso, 16 October 2026
Note: Facts are attributed to the Marketing AI Institute, Dentsu, a Medium post and Kanso's own articles, read on 17 October 2026. Only the Marketing AI Institute article shows a date (8 November 2021), so some limits it lists may have changed. The formula, the example figures and the task calls are Kanso's analysis, not research findings. This is a marketing view, not legal advice.
AI workflows