Karolina Bendryk“We already use ChatGPT.”
This is one of the most common reactions I hear when discussing AI with AdTech companies, and in most cases it is a perfectly reasonable one. ChatGPT, Claude, and similar tools are already part of everyday work for AdOps, Product, Client Success, Sales, and analytics teams.
People use them to summarize reports, work with spreadsheets, prepare client responses, write SQL, structure documentation, or quickly explore possible causes behind a performance issue. For many tasks, there is no reason to make this more complicated.
The more interesting question starts somewhere else: how much work still happens before the team can ask AI a useful question?
We have already discussed the broader distinction between AI assistants and operational AI in one of our previous articles. Here, I want to look at a more practical issue we keep seeing in product conversations: what general-purpose AI solves well today, and where operational complexity still remains.
Look at what happens before the prompt
Imagine an AdOps manager notices that revenue for a group of publishers is down significantly compared with yesterday. Before they ask ChatGPT what happened, they first need to understand where the decline came from. They may start with a BI dashboard, break the change down by publisher, GEO, format, and demand source, compare bid rate and win rate, check QPS utilization, review recent configuration or manual changes made on the platform, and verify whether any newly enabled feature or optimization could have affected traffic or auction performance. They may also search Slack or Jira for a related issue. Only then do they have enough context to ask an AI model for help interpreting the situation.
The AI can be very useful at that stage. It can summarize the findings, challenge assumptions, suggest additional checks, or help explain the issue to another team. But it is worth noticing what happened before the prompt: the employee had already collected, filtered, and interpreted a large part of the operational context.
For an occasional investigation, that is completely fine. When similar investigations happen every day across dozens of accounts, the manual work before the prompt becomes more interesting than the quality of the final AI-generated answer.
Knowing AdTech is not the same as knowing the business
General-purpose models already have a strong understanding of programmatic advertising. They can explain why win rate may fall, what could create an SSP/DSP reporting discrepancy, or which variables usually affect publisher revenue. The limitation appears when the question depends not on general AdTech knowledge but on how a specific company operates.
Suppose a demand partner's bid rate falls by 12%.
For one business, that may be a meaningful deviation requiring investigation. For another, it may fall comfortably within the normal range. A third team may already know that this partner changes buying patterns during particular hours or for specific traffic segments.
The number is the same, but the operational meaning is not.
Useful interpretation may depend on historical behavior, account priorities, partner-specific rules, internal thresholds, configuration history, previous incidents, and commercial context. Much of that information exists somewhere inside the company, but rarely in one place.
This is why company context becomes so important once AI moves beyond isolated tasks.
“Analyze this file” is very different from “investigate this account”
Uploading a report to ChatGPT and asking it to analyze the data is already a strong use case. An operator can export a CSV and ask which publishers experienced the biggest CPM decline during the last seven days, where revenue shifted most sharply, or which traffic segments behave differently from the average.
The situation changes when the request becomes:
“Which active accounts are behaving abnormally today, and where should I investigate first?”
Now the system needs more than a file. It needs current data, an understanding of which accounts are active, a reasonable baseline for each one, knowledge of which metrics matter, and some definition of abnormal behavior.
If the next question is “Why did Publisher A decline?”, the investigation may require data on demand activity, win rate, bid rate, QPS, floors, traffic composition, connection status, and recent changes.
The LLM can help reason over these signals, but the larger challenge is making the right signals available and putting them into the correct context. That is a very different problem from spreadsheet analysis.
Reconciliation makes the limitation easier to see
Reconciliation is a particularly useful example because almost none of the individual steps are technically difficult. A specialist downloads an internal report, collects numbers from a partner, aligns the periods, compares totals, calculates discrepancies, identifies exceptions, and starts investigating them.
Any modern AI model can compare two prepared files. But the preparation still matters. Someone needs to know which sources to use, whether the periods and time zones match, what discrepancy threshold is acceptable, whether adjustments are expected, and which exceptions require action.
This is similar to a broader problem we described earlier. Automating one isolated action does not necessarily remove the manual workflow surrounding it. The opportunity is therefore not necessarily in making the comparison itself smarter. It is in reducing the number of mechanical steps required to get from raw operational data to the small number of cases that actually deserve human attention.
Prompt engineering should not become another operational process
Another pattern appears when teams become experienced AI users. They start creating their own prompt templates.
A senior AdOps specialist has one for revenue investigations. Client Success develops another for account reviews. Somebody creates a custom GPT with internal documentation. Another team maintains instructions explaining which KPIs to check and in which order.
All of these initiatives can improve results, but they also reveal an underlying dependency: the user still needs to reconstruct the company's operating logic inside the prompt.
For occasional work, that is manageable. For repeatable workflows, it becomes harder to scale.
A senior employee may know that a particular partner should always be compared against a seven-day median instead of yesterday. A new employee may not. One person may remember an account-specific exception while another misses it completely.
If these rules already exist inside the organization, it makes sense to think about how they can become part of the system rather than something each user has to remember individually.
At some point, “using ChatGPT” becomes an infrastructure project
There is also a predictable progression that we increasingly see. A company starts with individual ChatGPT or Claude accounts. Then it adds internal documents. Later, someone connects reporting data through an API. Users ask for alerts, permissions, history, and reusable workflows. Engineering adds monitoring, authentication, and access rules.
None of those steps is unusual on its own.
Together, however, they mean the company is no longer simply using a chatbot. It is gradually building an operational AI layer around an LLM.
This is also where the build-versus-buy discussion becomes relevant. The model is only one line in the total cost. Integrations, infrastructure, maintenance, internal engineering resources, evaluation, security, and continuous updates tend to become the larger long-term commitment.
That does not make internal development a bad decision. For some companies, owning this layer may be strategically important. The point is simply that the scope should be understood correctly. A ChatGPT subscription and an operational AI environment are not alternative prices for the same thing.
Access to data is not enough either
Even direct access to company data does not automatically make an AI system operationally reliable.
Imagine an AI assistant investigating a revenue decline while one reporting pipeline stopped updating six hours earlier. Technically, it has data. Practically, it is missing part of the picture. Operational systems therefore need to consider data freshness, source availability, permissions, validation, and uncertainty. Sometimes the right output is not another confident recommendation, but a clear indication that there is not enough reliable information to make one.
The same applies to actions. Recommending that an operator review a floor configuration and automatically changing that configuration are very different things. Different workflows require different levels of human oversight, and those boundaries need to be designed intentionally.
This is where the conversation becomes much broader than “which model gives the best answer?”
The interesting question is what remains manual
I do not think AdTech teams need to stop using ChatGPT or Claude. For research, writing, document analysis, ad hoc data work, technical questions, or brainstorming, general-purpose AI may already be exactly the right tool.
The more useful exercise is to look at what still happens around it.
How many systems does an employee open before they have enough context for the prompt?
How often do they export the same types of reports?
How much internal knowledge has to be explained manually?
How often does the same investigation follow approximately the same sequence of checks?
Those questions reveal where the real operational workload still sits.
In our own product work on TeqMate AI, this has been one of the most useful distinctions: the interesting part is often not whether AI can answer a question, but how much of the context required for that answer can be available without someone rebuilding it manually every time.
General-purpose AI has already removed a meaningful amount of repetitive work. The next stage is about deciding which parts of the operational process should remain manual, which context can be made persistent, and where AI can realistically become part of the workflow without taking control away from the people responsible for the outcome.
That is a much more useful question than whether a company “already uses ChatGPT.”
Want to see what this could look like in your own AdTech workflows?
Talk to the TeqBlaze team about TeqMate AI and explore how your existing platform data, company knowledge, and operational processes can be connected in one controlled AI environment.
Book a TeqMate AI demo.
