How We Introduced AI into Our Agency, and the Results We Saw

There’s been a shift in the way agencies operate, and it’s all down to technology and the way it’s being used and viewed by businesses. According to RBC Trends, businesses and companies are tending to view technology as part of their core operations – the heart of their business and its continuity. 

This explains why, over the past years, clients are requesting results that are measured more precisely, hypotheses tested faster, and why they are spending less on manual tasks. Agencies therefore are now expected to deliver faster. When it comes to the issue of hiring, simply increasing the number of hires to reduce and control operational processes is no longer sufficient or effective.

For us, AI was not a question of following a trend. It was about maintaining quality as the complexity of our work continued to increase. I am Eduard Lebedev, founder of WakeApp, and I approach this subject from inside the market. At WakeApp, we spent many years building mobile marketing expertise, working with international companies, and continuously restructuring our processes.

Where the Implementation Started

In academic terms, artificial intelligence systems are described as systems that perceive their environment, use available knowledge and select actions that help them achieve a goal. In business, this translates into several practical questions: What data does the system receive? What kind of draft should it prepare? Who checks the result? What happens if the result is wrong?

Here at WakeApp, we already had extensive experience under our belt working with major brands, international markets and constantly changing acquisition channels. All this allowed me to quickly identify the main risk: when a new technology is added on top of chaotic operations, it does not make the business faster – it simply multiplies the existing problems.

— Eduard Lebedev, Founder of WakeApp

Therefore, the first stage of our AI implementation process was similar to an agency-wide audit. We started by mapping the standard client journey:

  • brief
  • research
  • advertising variations
  • campaign launch
  • report
  • conclusions
  • next test

At every stage, we identified where specialists were genuinely making a decision, and where they were simply performing preparatory, manual work. We collected initial information, rewrote key points, organised information and identified weaknesses.

Having repeatedly seen the market change its rules almost overnight, such as channel changes, algorithms being updated, new, emerging advertising formats or audience behaviour shifts, in this sort of environment, agencies cannot wait six months for a perfect implementation plan. Using the above approach helped prevent ineffective launches and implementations. 

Where We Saw Results and Impact First

The most noticeable results appeared in processes that operate as a cycle:

brief → draft → review → launch → feedback → new hypothesis.

According to industry analysts, value does not come from isolated experiments. It comes from process redesign, role-specific training, a clear implementation roadmap, and KPI monitoring. Our own experience confirmed this almost exactly.

Working with Creative Materials

At WakeApp, we stopped treating generative models merely as sources of ideas. Instead, our agency uses generative models to prepare analytical drafts: breaking down an audience according to barriers, identifying risky claims, creating sets of possible messages, and immediately highlighting the restrictions that need to be considered.

McKinsey chart showing the types of content created using generative models

Alt: Eduard Lebedev, WakeApp, implementing AI in the agency business

My working tool stack includes ChatGPT Mobile, Midjourney, Pika Labs, and HeyGen Avatars, however, the list of services is less important than the workflow in which they are used. In fact, one of the lessons we learned from mobile campaigns is that user journeys are increasingly shifting into social networks, mini apps and messaging platforms. For example, in one of our fintech projects, promotion through Telegram bots integrated with a mini app demonstrated just how strongly campaign performance depends on the context of the channel. For advertising materials, this means that a single universal message is no longer sufficient. The motivation, format and entry point must be adapted to the specific environment.

We also trained the team to write prompts as compact, creative briefs. A good prompt should contain:

  • the objective
  • the target segment
  • the channel
  • the required format
  • the tone of voice
  • restrictions
  • the risk of misinterpretation
  • criteria for a strong response.

Analytics and Hypothesis Development

With regard to analytics, our objective was to shorten the path from a raw report to a useful business question. Previously, after a campaign ended, the team spent a great deal of time restructuring spreadsheets and restating facts that were already obvious.

Now, the preparatory stage is completed much faster: the system can group campaign combinations, identify anomalies, compare segments and propose questions for the next sprint.

In my experience, data alone rarely accelerates decision-making. When information is structured chaotically, managers end up debating the format of the report rather than discussing the next actions.

— Eduard Lebedev, WakeApp

This is why we analyse several factors in every analytical result:

  • what changed
  • where the interpretation could be wrong
  • what possible causes should be considered
  • what should be checked first
  • what result would confirm the hypothesis.

Training the Team: The Main Barrier

The most difficult part was not connecting new services but changing management habits.

A manager may say that they support AI implementation while continuing to assign tasks in exactly the same way as before. Observations from the developer community show that experts actively use new tools but are not prepared to trust their accuracy blindly. 

Stack Overflow data shows a low level of trust in the accuracy of AI-generated answers, meaning that manual verification remains essential.

Alt: Eduard Lebedev on automating business processes with artificial intelligence

We therefore established a clear rule from the beginning: the system would prepare the foundation while the specialists would remain responsible for the result, the associated risks, and communication with the business.

During the implementation process, our agency encountered three main barriers:

  • Resistance: “We were already managing without it, so why change the approach?”
  • Fear of replacement: “If the system can write and calculate, will my role disappear?”
  • Lack of skills: employees did not understand how to achieve stable and repeatable results.

To train employees, we used real tasks. We completed each task once through the standard workflow and once with an AI assistant, compared the quality and documented the errors. Over time, we developed an internal knowledge base, a review checklist and several operating rules:

  • do not upload sensitive information
  • do not accept an answer without validation
  • do not replace a real conclusion with an attractive formulation.

— Eduard Lebedev, Founder of WakeApp

We trained senior managers separately since they need to evaluate not only the final result but also the original inputs and the limits within which the answer can be applied.

What Changed in Terms of Efficiency

I do not think it is appropriate to promise severalfold growth without comparable conditions and a precise measurement methodology. As there are too many variables in agency work, like seasonality, acquisition channels, budgets, markets and quality of the brief, I considered it fairer to discuss the operational changes that became visible in practice.

Initial Versions Were Launched Faster

The most noticeable shift was the speed at which projects could begin.

Previously, the team spent a considerable amount of time preparing the initial structure: what we were testing, which options we would use, how we would explain the logic to the client. Now, a draft appears almost immediately, and the discussion begins with choices, obstacles, and risks.

We converted review stages into a structured list of questions:

  • What is supported by statistical evidence?
  • Which statements are assumptions?
  • What cannot be published without manual review?
  • What risks are created by the wording?

The Amount of Manual Work Decreased

We saw the largest reduction in manual work that goes into reports and meetings. This opened up more time for specialists to engage with the market, analyse competitors and think about further tests. We also reduced tasks that did not provide any additional value.

Decision Quality Did Not Improve Immediately

The first implementations did not produce a dramatic improvement in quality. In some cases, the opposite happened: we received more options and only some were useful.

The turning point came when we introduced formal evaluation criteria:

  • For advertising materials, the criterion was alignment with the target segment and the company’s restrictions.
  • For analytics, it was whether the hypothesis could be verified.
  • For reports, it was whether the next action was clear.

Once these criteria had been introduced, decisions became more structured.

Which Solutions Remained in Daily Use

Not every initiative survived practical testing. Ultimately, four areas became part of our day-to-day work.

1. A Library of Prompts and Templates

We created reusable prompts and templates for recurring tasks:

  • analysing a brief
  • preparing a summary
  • identifying weak points.

2. Preliminary Analytics

AI helps us identify anomalies, formulate questions and prepare the structure of a discussion before a specialist begins the deeper analysis.

3. An Internal Case Knowledge Base

We began documenting not only successful outcomes but also mistakes, including which prompts produced weak or misleading results.

4. Personalisation Scenarios for App Marketing

These include:

  • predictive analytics
  • dynamic advertising
  • behavioural segmentation
  • testing messages for a specific channel.

Our agency’s experience shows that value does not come from generation itself. It comes from connecting a hypothesis to an audience, a metric, and a clear next action.

What Did Not Work

A common mistake in AI implementation is the desire to automate everything at once. We went through this stage as well. When a tool can write, calculate, and summarise, it is tempting to place it into almost every part of the operation.

Several approaches did not work for us.

Tasks Without an Owner

When nobody is personally responsible for the result, the system does not improve the process. It simply makes accountability less clear.

Generation Without Sufficient Input

When the brief is weak, the answer is also weak. It simply sounds more confident.

Long Instructions That Nobody Uses

Extensive instructions may look impressive, but they have little value if employees cannot apply them at the pace of real agency work.

Another risk is believing in autonomy too early.

In the agency business, you cannot:

  • promise a client something that the product cannot deliver
  • use sensitive information without proper protection
  • transfer a conclusion from one market to another without validation.

Why AI Is a Tool, Not a Threat

I see one persistent mistake in discussions about artificial intelligence: the question is usually framed as whether AI will replace people.

In practice, what changes is not the existence of the specialist but their tasks and area of responsibility. Weak performers whose value was based solely on manual execution may indeed be partially replaced.

The impact of AI agents on task completion time, productivity, and quality

Alt: WakeApp automating marketing processes with artificial intelligence

Which skills will continue to be valuable?

  • defining the task
  • assessing risks
  • understanding the market
  • communicating with the client
  • interpreting statistics
  • using information ethically
  • explaining the reasoning behind a decision.

These competencies are not disappearing. On the contrary, they are becoming more visible and more valuable.

What Is Changing in the Agency Market

The first change is that speed is becoming more valuable. Businesses will expect teams to prepare reports, explain conclusions, and propose the next test more quickly.

Tasks in which developers most frequently use AI in their workflows

Alt: Eduard Lebedev, WakeApp, automating everyday work tasks

The second change is the format of the agency service.

Previously, agencies often sold a volume of work:

  • a certain number of creative materials
  • a certain number of reports
  • a certain number of working hours
  • a certain number of meetings.

Now, value is increasingly shifting towards the reliability of the overall system:

  • how hypotheses are developed
  • how conclusions are verified
  • how knowledge is accumulated
  • how the team reduces the probability of errors.

The third change is that training is becoming a permanent part of agency operations.

An agency cannot conduct a single workshop and consider the task complete. Tools change, use cases become more complex, and new restrictions related to security and data emerge. Rules must therefore be updated regularly, and real cases must be reviewed continuously.

The fourth change is the growing gap between agencies that simply add technology as another layer on top of their existing chaos, and agencies that genuinely redesign their processes.

Conclusions

Technology will not solve poor management. For a business to run efficiently and succeed, clear and complete information is needed, evaluation criteria and verification processes must be in place, and a process owner must be in charge. Simply adding a tool will only add to the chaos.

For me, implementing AI is a way to make agency work more manageable: we use AI to remove unnecessary manual work, to increase speed and efficiency, all the while improving the quality of discussions, and giving people more time for tasks that genuinely require their expertise.