AI brand strategy generator

Branding on Steroids: How to Create an AI Assistant for a Creative Agency

In this article, we will share the experience of creating an AI assistant for one of the leading creative agencies. The client approached us with the task of automating the process of generating reports to identify customer archetypes using artificial intelligence technologies.

How did the client used to work before?

The client creates a report for their customer based on the theory of archetypes. The final document is approximately 30 pages in A4 format. To gather all the necessary data and write such a report, the following process had to be followed:

At the beginning of the process, a brief is compiled based on an interview with the client.

Then the client sends out five questionnaires in Typeform, which are pre-created for them, considering their specifics. Each of these questionnaires is filled out by the company's management, employees, and their customers. The number of survey participants can exceed a hundred.

Based on the survey data, dominant archetypes are manually determined, which are then analyzed, and the report text is written.

Finally, all this is manually transferred to the report, which the client creates in Figma.

How does our development simplify and change the client's work?

The client logs into the Admin panel of our tool and creates a new client.

The system automatically generates links to the questionnaires, which can be copied and sent to the client.

All survey responses are stored in the database, moderated by AI. The admin panel indicates whether there are enough responses to start generating the report.

Next, it is necessary to harmonize and determine the dominant archetypes. This task is very large-scale, as there can be more than a hundred responses, but AI performs it with mathematical precision.

When there is enough data to generate texts and images for the report, the AI breaks the process down into subtasks and fills the database with the ready data.

Finally, the system creates a web page with the report and imports all the data into the CMS of the web page. Done!

If necessary, the system's work can be adjusted at any stage directly through the dialog box in the admin panel.

Getting down to business

The client insisted on using low-code solutions because their previous experience with "real programmers" did not meet their expectations. Of course, the issue was not the technologies but the executors, but we did not argue and got to work. Therefore, at this stage, we needed to choose suitable tools and determine how they would interact with each other. This requires a careful approach to system architecture.

Architecture

Choosing a data collection tool: We recommended using Formless.ai instead of Typeform. Formless.ai is a sub-product of Typeform, which conducts interactive surveys with the help of AI, structuring data in JSON format. Despite some limitations, the product is promising and looks impressive.

Report generation: We convinced the client that creating reports in the form of a web page with a custom design would simplify the process and make the results more accessible. The page design layout was developed in Figma, which allowed us to consider the necessary fields for the report, totaling 160.

Website development: For this task, we chose Webflow for its API management, design, and animation capabilities. Despite the limitation of 60 variables in the CMS, we configured the system to integrate all 160 variables.

Data coordination: To synchronize data between tools, we use Make.com, which coordinates data transfers.

Using artificial intelligence: For tasks requiring AI, we use the OpenAI Assistant API. It allows us to use the latest models, upload a database, conduct vector searches within this data, and supports referencing message threads via Thread ID.

Admin panel: We use GlideApps to create the admin panel. This tool allows for easy customization of any panel with a wide range of functions and attractive design.

Database: We use the built-in GlideApps database as the primary database as the data volume will be small. However, the tool provides the ability to integrate with various external data storages, from Google Sheets to PostgreSQL.

User Path

According to our concept, interaction with the system should look as follows:

1. Data collection: The agency's client fills out five questionnaires in Formless.

2. Data processing: Assistant API checks the responses for inconsistencies and structures the data.

3. Data storage: All collected data is then stored in the GlideApps database.

4. Review and correction: The user can manually review and harmonize archetypes using the button in the admin panel and the Assistant API functions.

5. Preparation for publication: Using Assistant API, the "raw" data is transformed (this is where the real magic happens) into structured content for Webflow, including formatting, HTML markup, and image generation using DALL-E.

6. Interactive management: The admin panel has a chat where you can ask Assistant API to change certain data, for example, rewriting an unsuitable paragraph.

7. Report publication: After final preparation and data synchronization with Webflow, the web page is ready for publication.

Stages of creating the system

1. Launching development in Webflow: This is the most labor-intensive stage. The layout, with several rounds of revisions, took about a month. When the CMS structure became clear, we exported it from Webflow in CSV format to accurately recreate it in the GlideApps database. Thanks to the convenience of importing CSV tables into GlideApps, the database preparation process went smoothly.

‍

‍

2. Developing questionnaires: Simultaneously with the design, we created five questionnaires on Formless.io, defined the prompts, and desired data structure output.

‍

‍

3. Configuring Assistant API: We write system prompts, upload necessary documents into the knowledge base, and test functionality in the sandbox. Assistant API can store the context of a specific client through Thread ID.

‍

4. Admin console development:

  - Creating the interface design in GlideApps.

  - The ability to add a client card and upload the brief as a file.

  - Generating links to Formless.io questionnaires with an integrated client ID to track responses.

  - Tracking the number of responses for each questionnaire.

  - Archetype harmonization and data generation buttons for the report.

  - All admin console fields exactly match the fields in Webflow, and each field has a menu to start a chat with Assistant API.

‍

5. Integrating all parts of the system through Make:

  - Receiving data from Formless in Zapier, then transferring it to Make. This sounds strange because Formless is still a raw product and does not have an open API, only integration with Zapier.

  - Archetype harmonization scripts, turning raw data into a report, and transferring data between Make, Assistant API, Glide, and Webflow.

  - Interaction through Glide Chat and Assistant API

‍

6. Generating images: We used Cloudinary for compiling images from archetypes, which we value for its ability to manipulate image parameters directly through the URL.

7. Fine-tuning prompts: Implementing prompt sessions using OpenAI's recommendations.

8. Project finish: Refactoring the Glide database, publishing on the client's domain, and creating documentation. The job is done!

Costs of services

Formless.ai: $59 per month, includes 250 surveys.

Make.com: $10 per month. The minimum plan with 10,000 operations per month is more than enough for us.

Webflow: approximately $80 per month.

GlideApps: $60 per month.

OpenAI: costs depend on usage; we usually did not exceed $20 per month.

Cloudinary: free

Results

Using the low-code stack of Formless.ai, Make.com, Webflow, GlideApps, OpenAI Assistant API, and Cloudinary allowed us to significantly speed up the development of the solution, simplifying the report generation process for the client, reducing the time to create a report to 15 minutes per report.

Moreover, using low-code tools increased the system's flexibility, making it easier to configure and modify later.

We are grateful for the opportunity to work on such an interesting project and look forward to future challenges!

Conclusion

Implementing an AI assistant in your business is not just a technological step forward but a strategic decision. It can drastically change the approach to handling routine tasks and increase the efficiency of the entire team.

We have detailed the process of creating such an AI assistant using the example of a specific task for a creative agency, but the key principles and approaches are universal and applicable in any industry.

‍