What you lose when AI starts handling your conversations and decisions

By now, we all know that AI can be extremely useful. I use it in my own work to organize notes, summarize documents, test ideas, and improve a rough structure. It can save time on the administrative jobs that otherwise sit between thinking and doing.

The problem starts when saving time becomes handing over responsibility. I’m seeing two habits that should concern marketers:

• People use AI to write messages that other people then use AI to summarize and answer.
• Marketers upload reports or exports and ask a language model to decide what the business should do without explaining the business behind the numbers.

Both habits produce something that looks polished, which is precisely why they’re easy to trust. A long email can hide a simple point. A detailed analysis can hide missing context. In each case, the output may look more substantial than the thinking behind it.

Speed is helpful, but it doesn’t remove the need for context, judgment, or a person who’s prepared to own the decision. AI can make marketing work faster, but it can also put another machine between two people and turn incomplete data into confident advice.

Below are some drawbacks to relying on AI to handle your communications, data, and contextual interpretation, as well as recommendations on how to use tools such as ChatGPT and Claude without losing context, trust, or control of the decision.

What happens when two AI assistants start emailing each other

Picture a simple exchange. One person needs to say that Thursday works and asks ChatGPT to make the message sound more professional. The recipient opens the resulting 400 words, asks another AI tool to summarize them, then asks it to draft a suitable reply.

The two tools are now conducting a polite conversation while the humans review generated text. At that point, everyone might as well make a cup of tea and let the two ChatGPTs get on with it.

It sounds silly because it is silly, but the commercial cost is real. Communication is supposed to help people understand each other. When a straightforward point is expanded, summarized, and expanded again, tone and intent can get weaker at each stage.

Hesitation may disappear. Urgency can be overstated. A cautious suggestion can return as a commitment nobody meant to make.

Longer messages also create more room for ambiguity. The recipient must work out which details matter, which were added for polish, and what the sender actually wants. If the message is simply that Thursday works, sending 400 words doesn’t make the answer more professional. It makes Thursday harder to find.

Dig deeper: AI works best when it removes work, not judgment

Why automated communication can be tricky in agency relationships

Automated communication matters particularly in agency work. Clients don’t only pay for tasks and reports. They pay for interpretation, judgment, and confidence in the people handling their marketing.

SEO updates often contain uncertainty. A ranking movement may be commercially important, or it may be noise. Traffic may rise for searches that will never produce a customer.

Leads may fall because visibility has weakened, or because something changed after the visitor reached the website. A generated report can describe the movement, but description isn’t the same as understanding.

Clients also reveal useful information in ordinary conversations. They mention that an important product has been unavailable, that the sales team is struggling to return calls, that prices have changed, or that inquiries feel different. Those comments can completely change the meaning of an analytics report.

A generic email can’t ask the unexpected follow-up question. It can’t notice that the client sounds uncertain about a priority. It can’t recognize when a neat recommendation conflicts with how the business actually works. Sometimes the most useful response is a short conversation, not a polished wall of text.

Repeated AI phrasing causes another problem. Different clients begin to receive the same voice, even though their situations are different. The message may be grammatically sound yet still feel impersonal.

Trust is built when a client knows that somebody has understood their business, not when every update arrives wearing the same shiny vocabulary.

Uploading data doesn’t upload the business

This habit is more serious because it can influence decisions. Marketers can upload Search Console exports, analytics reports, ranking data, sales figures, and campaign results. AI can sort the information, calculate changes, and find patterns quickly. Those are useful capabilities.

What it can’t do is infer every event that produced the numbers. Uploading a spreadsheet doesn’t automatically promote ChatGPT to managing director.

A report can be accurate and still provide an incomplete basis for action. The data may show what happened inside the measurement system, but not everything that happened inside the company, the market, or the customer journey. That creates a dangerous gap: The data can be correct while the recommendation is wrong.

Dig deeper: Why AI search is your new reputation risk and what to do about it

The variables missing from a marketing export

A standard marketing export rarely explains that a major promotion ended, prices changed, or an important product went out of stock. It won’t usually show that a sales team changed how it qualifies leads or how quickly it follows them up.

Measurement can change, too. A consent banner, broken form, or altered analytics configuration can make performance look different even when customer behavior hasn’t changed in the same way.

Seasonal demand, a new competitor, and changes to Google’s search results can all affect the figures. A shift in margins can mean that more revenue produces less profit.

AI doesn’t know which of these events matter unless someone supplies the context. Even then, the person supplying it must know what to mention.

The model can’t ask the warehouse manager about stock, the sales team about lead quality, or the customer service team about a new complaint unless the marketer recognizes that those conversations are necessary.

SEO examples that expose the context problem

Consider impressions rising while clicks remain flat. An AI summary may praise increased visibility. That’s a reasonable first observation, but it may be commercially useless.

If most of the new impressions come from broad informational searches while clicks for important services decline, the business isn’t necessarily moving in the right direction. The top line improved because two very different types of visibility were bundled together.

Now consider a fall in organic sessions after a consent or tracking change. A model that only sees the before and after numbers may diagnose an SEO decline and recommend technical or content work. The timing could make that answer look convincing.

Yet the website may be attracting a similar number of visitors while measuring fewer of them. Changing pages before checking the measurement would treat the symptom shown in the report rather than the cause.

A third example is a keyword gap analysis. A tool identifies hundreds of terms competitors rank for, and the AI recommends a large batch of articles. Again, the conclusion sounds logical.

But the real weakness may be that key service pages fail to answer the questions buyers ask before making contact. More articles won’t fix a service page that leaves people unsure about price, timescales, disruption, or suitability.

In each example, the first interpretation is plausible. Human context changes the conclusion. That’s why AI analysis should start a line of inquiry, not end it.

Dig deeper: How to build an AI governance framework for SEO

AI will nearly always find an improvement

The question shapes the answer. Ask a model to find problems, and it’ll find possible problems. Ask it how to improve a page, and it’ll normally suggest changes, even when leaving the page alone may be the safer choice.

Ask for improvements, and you may receive 27 of them, neatly organized under headings, before you’ve finished your coffee. The quantity creates an impression of certainty. It doesn’t tell you which change matters, whether the evidence supports it, or what the business risks by acting.

This doesn’t make the suggestions useless. It just makes them hypotheses. A marketer should check each one against search intent, commercial priorities, customer evidence, and the current performance of the page.

Sometimes the right decision is to make a change. Sometimes it’s to collect more information. And sometimes it’s to leave a successful page alone.

Remember, the model doesn’t carry the cost of a bad recommendation. The business does.

Fluent answers can still be inaccurate

The providers make this limitation clear. OpenAI’s Terms of Use say output may not always be accurate, should not be treated as the sole source of truth, and must be evaluated before it’s used or shared.

Anthropic’s Consumer Terms warn that outputs may contain material inaccuracies even when their detail makes them appear accurate, and advise users to confirm accuracy independently.

That wording matters because confidence is a feature of the writing style, not evidence that the conclusion is correct. A clear explanation can still be based on an incorrect calculation, a misunderstood column, or a source that doesn’t support the claim.

Treat plausible and verified as two different stages. Open any citation the model provides. Check that the source exists, is current, and supports the precise claim. Recalculate important figures.

If the recommendation affects budget, client communication, or a commercially valuable page, the checking should be proportionate to the risk.

Account type and privacy settings matter

Accuracy is only one concern. Marketers also need to understand what happens to the information they upload. The answer varies by provider, product, account, and settings, so a paid subscription alone shouldn’t be treated as proof that an account is suitable for client or proprietary information.

OpenAI says it doesn’t use business data from products such as ChatGPT Business, ChatGPT Enterprise, and its API to train models by default. On personal ChatGPT accounts, users can turn off “Improve the model for everyone.”

Additionally, OpenAI’s Data Controls FAQ says Temporary Chats aren’t used for training, are deleted after 30 days, and may be reviewed to monitor abuse.

Anthropic says inputs and outputs from commercial products such as Claude for Work and its API aren’t used for model training by default.

Anthropic’s consumer account guidance says chats may be used when the user permits model improvement, when a conversation is flagged for safety review, or through another explicit opt-in. Incognito chats aren’t used to improve Claude.

Keep in mind that these policies can change. Check the current terms for the exact product and account your team uses.

Before uploading client data, confirm that you have permission, remove information the analysis doesn’t need, and follow your organization’s policy. Personal or confidential information shouldn’t be included simply because it makes the prompt more convenient.

For UK organizations, the Information Commissioner’s Office provides guidance on applying data protection principles to AI and assessing risks to people’s rights and freedoms.

Remember, privacy isn’t a tick box hidden in the settings menu. It’s part of deciding whether the task should be given to an AI tool at all.

Dig deeper: The dehumanization of SEO: Displaced by code, ignored by communication

A practical review process for marketers

A sensible process keeps the efficiency while putting judgment back where it belongs:

• Decide whether the task is suitable for AI before uploading anything.
• Remove personal, confidential, and commercially sensitive information that the analysis doesn’t require.
• State the business question, time period, objective, and known changes around the data.
• Ask the model to separate observations, assumptions, and recommendations.
• Challenge the first answer by asking what information is missing and what alternative explanations fit the same pattern.
• Check calculations, sources, and factual claims independently.
• Speak to the people who understand sales, customers, stock, operations, and recent business changes.
• Make and record the decision yourself, including why you accepted or rejected the suggestions.

This process may feel slower than accepting the first polished answer. However, it’s still faster than acting on the wrong diagnosis, rewriting a valuable page unnecessarily, or explaining to a client why confidential information was uploaded without proper thought.

Keep the useful human

AI can reduce administration and help marketers examine more information. It can organize a messy set of notes, expose questions, compare possible explanations, and give a human reviewer a useful starting point.

But AI shouldn’t replace a straightforward conversation. Nor should it turn a partial export into an unchallenged business decision. The person using the tool remains responsible for what’s sent, what’s uploaded, and what the business does next.

A simple boundary works well: Use AI to organize information, test ideas, and identify questions. But keep humans responsible for relationships, interpretation, and decisions.

If two AI assistants are already emailing each other, the humans may as well put the kettle on. If, however, one of those assistants is also making the commercial decisions, it’s time to come back from the tea break.