AI agents are increasingly taking on routine tasks: scraping competitor prices, managing social media accounts, testing advertising campaigns. However, any automation has a hidden cost — proxy traffic. We conducted three measurements on typical tasks to determine how many gigabytes an AI agent "consumes" in an hour and a month of operation, so you can accurately calculate your proxy budget and avoid overpaying.
Why Measuring AI Agent Traffic is Important
Most residential and mobile proxies are sold based on traffic volume rather than the number of IPs or time. This fundamentally distinguishes them from data center tariffs, where there is usually a fixed price per port. When it comes to manual work — one person scrolling through Instagram or checking prices on Ozon — traffic consumption is predictable and low. But when the task is taken over by an AI agent, the picture changes dramatically.
An AI agent works without pauses, often fully loading pages (with all images, scripts, trackers), can simultaneously manage dozens of sessions, and generate many more requests than a human in the same time frame. Without prior measurement, it's easy to end up in a situation where the purchased traffic package runs out in a few days instead of a month, causing automation to halt due to a lack of limits.
We decided to fill this gap and conduct measurements on three scenarios that are most commonly encountered by arbitrageurs, SMM specialists, and marketplace sellers using AI agents for automation through anti-detect browsers like Dolphin Anty, AdsPower, and Octo Browser.
Measurement Methodology: What and How We Measured
For the purity of the experiment, we used the same setup: the anti-detect browser AdsPower with connected residential proxies, a traffic counter at the proxy server level, and request logging through the built-in network monitor in the browser. The AI agent was implemented as a script based on a headless browser with an LLM module for decision-making (similar to a Playwright + GPT agent setup), run through the same proxy session throughout the test.
Each scenario was run 3 times at different times of the day to eliminate errors due to dynamic content loading and varying weights of advertising banners on websites. The final figures are the average values from the three runs. We recorded: the total volume of data transferred (incoming + outgoing traffic), the number of HTTP requests, the time taken to complete the task, and the traffic volume per "action" of the agent (one product view, one post, one ad launch).
An important detail: we did not block the loading of images and media because, in real tasks, the AI agent often needs to analyze visual content — screenshots of pages, previews of product cards, video thumbnails. This increases traffic consumption compared to text parsing via API, but more accurately reflects the real working conditions of most agents operating through a browser.
Scenario 1: Price Scraping on Wildberries and Ozon
The agent's task: to scrape 500 product cards in a specified category, record the price, availability, and rating, compare with competitor prices, and generate a report. This task is typical for sellers who monitor price dynamics on marketplaces in real-time.
Measurement result: processing 500 product cards took the agent 42 minutes and required 1.3 GB of traffic. In terms of one card — about 2.6 MB, which is quite a lot for purely textual data. The reason is that Wildberries and Ozon load the full set of product images, reviews with photos, and advertising blocks on each page, even if the agent only needs the price.
We tested optimization: we disabled the loading of images and videos at the browser level, leaving only HTML and JSON responses from the marketplace API. Consumption dropped to 340 KB per card — almost 8 times. However, Wildberries more frequently displayed captchas on "lightened" sessions without the standard resource set, forcing the agent to make repeated requests through a new IP, partially eating into the traffic savings.
When scaling to a full catalog of 10,000 products and daily data updates, traffic consumption amounts to about 26 GB per day without optimization and about 3.4 GB with media disabled. For scraping such volumes, data center proxies are better suited — they are cheaper per gigabyte and provide higher speeds, which is critical when handling a large number of parallel streams.
Scenario 2: Auto-Posting and Warming Up on Instagram/TikTok
The agent's task: to imitate the behavior of a live user — scroll through 15-20 posts in the feed, like 5-7 of them, leave 2 comments, view 3 Stories, and publish one post with a photo. This scenario is typical for SMM agencies that warm up new client accounts before launching ads or conduct organic promotion.
The measurement showed: one full cycle of warming up one account takes 8-11 minutes and consumes 180-240 MB of traffic. The main consumption comes from videos in Stories and Reels: even a short 15-second video averages 12-18 MB when auto-playing at default quality. Publishing one photo post with processing through a filter adds another 15-20 MB for upload and confirmation.
If the agency manages 30 accounts with daily warming, the total consumption will be around 6-7 GB per day, which amounts to about 180-210 GB per month for the entire pool of accounts. This is a significant figure that needs to be budgeted for proxies in advance — especially if a separate IP is used for each account to avoid chain bans in multi-accounting.
For this task, we recommend mobile proxies — Instagram and TikTok are noticeably less aggressive towards mobile IP addresses from operators, as the majority of real users access from them. This reduces the frequency of additional checks and captchas, which also increase traffic consumption due to repeated login attempts.
Scenario 3: Creative Testing in Facebook Ads
The agent's task: to log into 10 Facebook Ads Manager accounts, launch 3 ads with different creatives in each, monitor the moderation status every 20 minutes for 2 hours, and collect preliminary statistics on impressions and clicks. This is a typical scenario for arbitrageurs testing dozens of combinations simultaneously.
Ads Manager is one of the "heaviest" interfaces among all tested: just loading the dashboard of one account with statistics consumes 8-12 MB due to the large number of JS scripts, graphs, and Meta Pixel trackers. The full cycle — logging into 10 accounts, publishing 30 ads with images, plus 6 cycles of status checking — required 890 MB of traffic over 2 hours of continuous work.
In terms of a month, with daily work on such a volume of accounts, the consumption will be around 26-27 GB. If an agency or arbitrage team manages 50+ advertising accounts through an anti-detect browser like Dolphin Anty or Multilogin, the total traffic easily reaches 100-150 GB per month just for monitoring and launching campaigns.
Here, stability and "cleanliness" of IP are critical: Facebook aggressively bans accounts when there is a suspicious address change or when working with data center IPs, which are easily tracked as proxies. The optimal option is residential proxies with a stable IP assigned to each account: they appear as regular home connections and reduce the risk of blocking during AI agent activity.
Summary Table of Traffic Consumption
Below are the summary figures for all three scenarios for quick calculation of the required traffic volume when planning the work of AI agents.
| Scenario | Traffic per Action | Daily Traffic (Typical Volume) | Recommended Type of Proxy |
|---|---|---|---|
| Scraping Wildberries/Ozon (10,000 cards) | 2.6 MB / 340 KB without media | 26 GB / 3.4 GB without media | Data Center |
| Warming Up/Auto-Posting Instagram (30 accounts) | 180-240 MB per cycle | 6-7 GB | Mobile |
| Facebook Ads (10 accounts, 30 ads) | 890 MB for 2 hours | 10-13 GB (with 2-3 cycles) | Residential |
How to Reduce AI Agent Traffic Consumption
Measurements showed that a large portion of traffic is consumed by images, videos, and tracking scripts, rather than the useful data needed by the agent. Here are proven ways to reduce consumption without sacrificing automation quality:
- Disable media loading where possible. For pure price and product characteristic scraping, images are unnecessary — savings can reach 80-85%.
- Use APIs of marketplaces and social networks instead of full page rendering, if allowed by the anti-detect browser and platform rules — this reduces the request size significantly.
- Limit the frequency of status checks. In the Facebook Ads scenario, checking every 20 minutes instead of every 5 minutes reduces traffic consumption for monitoring by almost 4 times without losing data relevance.
- Cache static session resources — logos, interface icons, CSS files — so the agent does not download them again with each new launch.
- Set the video quality in Stories/Reels to the minimum for auto-viewing tasks if it does not affect the agent's main task performance.
A combination of these methods in our tests reduced overall traffic consumption by 40-60% depending on the scenario, while the speed of task execution by the agent even slightly increased due to the lower weight of pages.
What Type of Proxy to Choose for the Task
The type of proxy directly affects not only the cost of traffic but also the stability of the AI agent's operation. For mass scraping of marketplaces, where there are no strict checks on the "humanity" of IPs, data center proxies are optimal — they are fast, cheap per gigabyte, and well-suited for large volumes of similar requests.
For working with social networks and advertising accounts, where platforms actively combat automation and multi-accounting, priority is given to residential and mobile IPs. They appear as regular user connections, which reduces the frequency of captchas and blocks, thus indirectly saving traffic due to fewer repeated attempts and re-authorizations.
Practical advice: do not buy a single large traffic package "for everything" — separate tasks by proxy type according to their nature. This is both cheaper and more reliable for the stable operation of the AI agent in the long term.
Conclusion
Our measurements showed that the traffic consumption of an AI agent heavily depends on the type of task: scraping marketplaces without optimization can consume tens of gigabytes per day, auto-posting on social media can consume a few gigabytes for a pool of accounts, and working with Facebook Ads accounts can reach 10-13 GB with active monitoring. Understanding these figures helps to plan the proxy budget more accurately and avoid situations where traffic runs out prematurely.
If your AI agent is involved in scraping large volumes of data on Wildberries or Ozon, consider data center proxies — they offer the best speed-to-cost ratio for traffic. For warming up and automation on Instagram and TikTok, mobile proxies are more suitable, while for stable operation with advertising accounts and multi-accounting, residential proxies with a fixed IP per account are recommended.