AppsFlyer today put an Apple Ads agent into closed beta that changes keyword bids once a marketer approves them, the first agent in the company's line-up that acts rather than only advises. The mobile measurement company also opened a tool for building custom agents from a written description of a workflow, and added four capabilities to its MCP server that hand fraud figures, creative-level results, link creation and partner integration settings to outside AI tools. The release gives no pricing, no customer results and no explanation of how the Apple Ads agent arrives at the share of voice estimates that trigger its bid changes.

In Short

AppsFlyer, the company many app makers use to work out which ads led to an install, now lets marketers create AI agents that send scheduled reports and suggestions, and is testing one that changes Apple Ads bids after someone signs off. That matters to app marketers and agencies, because a tool that used to report on campaigns can now change what they bid on Apple's App Store ads. The safeguard is a person's approval before each change, but AppsFlyer has not said how the agent measures share of voice, which Apple interface it uses, or what any of it costs.

The agent that moves bids

The Apple Ads Share of Voice Agent is the item in today's release that changes what AppsFlyer does for its customers. According to AppsFlyer, the agent estimates share of voice for each keyword and geography, recommends a bid change when an advertiser is "off target," and then executes the change in Apple Ads directly from AppsFlyer. The company describes it as its first agent "that not only recommends - but takes actions." Changes are applied only once the marketer approves them, according to the release, and access runs through each customer's AppsFlyer customer success manager while the product is in closed beta.

The release does not define the measure. In advertising generally, share of voice is a brand's portion of all the advertising or visibility in its category over a set period, expressed as a percentage. On App Store search, that would point to how often an app's ads appear for a keyword compared with the opportunities available - but AppsFlyer does not say so, and it does not say where its estimate comes from.

That question is technical rather than rhetorical. AppsFlyer's core data describes what happens after an ad is seen or tapped: installs, in-app events and the media source credited with them. Apple runs the auction. Apple's newer Platform API includes insights on impression share and search-term popularity, as PPC Land reported when the Apple Ads Platform API debuted in September. Whether AppsFlyer's estimate is built on those Apple figures, on its own attribution data or on a model combining both, the release does not state. It also leaves open who sets the target, whether the target is a single figure or a range, and how large a bid change the agent may propose in one step.

The choice of interface has a deadline attached. Apple's documentation says the Platform API supersedes the Campaign Management API, which will be sunset on January 26, 2027 - 76 days after today's release. Any tool that writes bids into Apple Ads on an advertiser's behalf has to run on the new interface by then. AppsFlyer does not say which one its agent uses.

Keyword bids are not available everywhere in Apple Ads, either. Since February, Maximize Conversions has been open to all App Store advertisers running search results campaigns. In those campaigns an automated bidder sets bids for each relevant query against a target cost per acquisition, and Apple's documentation states that advertisers cannot add or manage keyword bids. On the face of that documentation, an agent that adjusts bids keyword by keyword would work on campaigns that still use manual bidding. The release does not mention Maximize Conversions or say how the agent treats campaigns that use it.

Bid levels carry real weight in the auction the agent would be adjusting. Apple began adding more ad positions to App Store search results on March 3, starting in the UK. An analysis of 132 UK keywords by ConsultMyApp, covered by PPC Land in March, found that the most relevant app took the top position on only 43.9% of keywords, and that Apple's auction order matched the relevance order exactly on just 7.6%. When the most relevant app loses the top slot more often than it wins it, an automated bid adjustment is not a minor setting.

Approval as the safeguard

AppsFlyer frames the release around trust. Agent Hub, the company says, is "establishing the architecture required for AI to run reliably," with transparent decision-making in every workflow so that teams can understand and evaluate the reasoning behind each recommendation. The release does not show what that transparency looks like - whether the agent exposes the queries, data and calculations behind a recommendation or only a written explanation.

Approval is the only control the release names for the bid agent, and it is described in a single sentence. Who in an account can approve? The release does not say. Nor does it say where the approval request appears - the AppsFlyer interface, Slack or email - whether several changes can be approved together, or whether a log records who approved what. Limits are not addressed either, such as a ceiling on the size of a bid change above which a person must sign off.

Other companies opening ad accounts to AI agents have drawn that line in different places. Google's open-source server for connecting AI assistants to Google Ads gained four write tools in version 0.0.6 on November 2, each gated by a two-step dry run and checksum; PPC Land's reading of the code found that the checksum proves the parameters did not change between the two calls, not that a person looked at them. X's server for AI agents, released on August 24, creates campaigns paused, with activation as a separate call. Supermetrics, which on November 3 let Claude build campaigns on six ad platforms, described campaigns paused by default, edits that could be reviewed before they applied and approval requirements set per account. Meta's Ads AI Connectors, in beta since April 29, let agents create and edit ads, ad sets and entire campaigns.

Each of those designs keeps a human in the loop on paper. The difference lies in whether the software itself refuses to act without a person, or whether the pause depends on an instruction the AI model is expected to follow. AppsFlyer's description does not say which applies to its agent.

Inna Weiner, VP Product at AppsFlyer, put the shift in broad terms. "Marketing teams are moving from manual campaign orchestration to agentic execution," she said. "With this release, AppsFlyer is giving marketers trusted, agentic tools that can analyze, recommend and act on their data to deliver clear value for growth."

Agents written in plain language

The second Agent Hub addition hands marketers the design of an agent. With Create an Agent, a marketer describes any marketing performance workflow in words, and AppsFlyer builds an agent from that description. According to the release, the agent pulls data down to the event level, analyses it and delivers the findings to Slack or email on a schedule the marketer sets. It can also run root cause analysis - tracing a change in a metric back to what caused it - and recommend a next action.

The verbs matter. The release describes custom agents as pulling, analysing, delivering, running analysis and recommending. Unlike the Apple Ads agent, they are not described as taking actions in ad platforms. AppsFlyer lists investigating fraud spikes, spotting creative fatigue and recommending Apple Ads bid changes as the kind of work agents can support.

This is a step away from the pre-built approach AppsFlyer took a year ago. On November 18, 2025, AppsFlyer released eight products, led by an Agentic AI Suite with pre-built agents that identified creative opportunities, surfaced daily insights, monitored configuration status and identified performance trends. Three weeks later, AppsFlyer data showed that 57% of agent deployments focused on technical automation, such as configuration checks and data-integrity validation, with 32% on business optimization. The same report analysed more than 10,000 queries to the company's AI Assistant and found that 60% were complex analytical questions.

The release does not say how many custom agents an account can run, whether they can draw on data from outside AppsFlyer, how often they can be scheduled, or whether they are available on every plan.

Four additions to the MCP server

MCP, the Model Context Protocol, is the open standard that lets an AI application discover and call tools in other software. AppsFlyer launched its MCP server on July 17, 2025, available at first through Claude, with an analytics chat for questions on return on ad spend, retention and lifetime value, visibility into audiences, and the ability to query the OneLink dashboard. By November 2025 the company said the server worked with Claude, ChatGPT, Cursor and VS Code. Today's release names no AI applications. It refers to marketers' own AI tools and to "any MCP client," and says all of the new capabilities are also available in AppsFlyer's AI assistant.

Fraud data from Protect360

The first addition makes Protect360 data available through the MCP server. AppsFlyer gives two example questions: a breakdown of fraud concentration by geography over the last 30 days, and which of an advertiser's custom validation rules drove the most blocks during a specific period.

Where mobile fraud sits has shifted. AppsFlyer's State of Fraud report in June, covering 106.4 billion installs across 246,000 apps, found that organic traffic accounted for 52% of fraudulent installs, with affiliates close to 40%. iOS fraud fell from 17.5% to 11.7% between the first quarters of 2025 and 2026, while Android stayed in a 14% to 15% range, and India accounted for nearly 28% of global Android fraud. A query by geography, of the kind AppsFlyer uses as its example, maps directly onto that last figure.

Creative results by asset

The second addition exposes creative-level spend, reach, conversion, retention and revenue cohorts to AI tools. AppsFlyer says the aim is to let teams identify winning and fatiguing creatives faster and build automation around them. The feed extends work the company started with the Creative Management Hub in November 2025, which added AI-based scoring and recommendations on top of its Creative Optimization suite. The release does not say which ad networks' creative data is included, or how a single creative is matched across networks that name and report it differently.

The third turns a plain-language campaign description into a fully configured OneLink, the AppsFlyer link that carries campaign parameters and sends users to an app or its store page. Nothing is created until the marketer approves. When the MCP server launched in July 2025, its link features were limited to queries within the OneLink dashboard, so creation is new. The release does not list which link settings the AI fills in - parameters, redirects, fallback pages - or whether a marketer can edit the proposed link before approving it.

Partner settings in conversation

The fourth surfaces a partner integration's configuration, attribution windows, postback rules and event mappings to any MCP client, so that marketers and support teams can diagnose setup problems in a conversation. Those settings decide which in-app events are passed back to an ad network and how long after an ad a conversion still earns it credit. They vary by partner and by channel: AppsFlyer applies a maximum 72-hour probabilistic window to app installs from connected TV ads, against 30 days in MNTN's own system.

The list of partners keeps growing. When ChatGPT Ads gained app attribution in seven markets in July, setting it up in AppsFlyer ran through Collaborate, Partner Integrations and Marketplace, across five tabs. The release does not say whether the MCP tool can change integration settings or only read them. Its own wording - surfaces, diagnose - points to reading.

Security not itemised

Every new data feed raises the question of how the connection is secured. PPC Land reported in July 2025 on tool poisoning risks in MCP for marketing platforms, in which instructions hidden in a tool's description can steer an AI model. AppsFlyer describes its infrastructure as "high quality, governed agentic AI infrastructure" without listing controls such as permission scopes, rate limits or logging of AI-initiated queries. Its November 2025 Enterprise-Grade Security Package added single sign-on, extended audit logs with API access, IP allow lists and role-based access control. The release does not say whether those controls extend to MCP and agent activity.

What the release leaves out

Beyond the technical gaps, several commercial details are missing. There is no price for any of the new tools and no indication of which AppsFlyer plans include them. The closed beta has no stated size, list of countries or date for general availability. No customer is named, and no result from the beta - a change in cost per install, share of voice or time saved - is offered. Create an Agent and the MCP additions are presented as available now, with no rollout schedule by market or plan.

The company's description of itself supplies the only figures: AppsFlyer says more than 15,000 businesses worldwide trust it, and calls itself the leader in mobile attribution for more than a decade.

A measurement company that now acts

AppsFlyer built its business as a mobile measurement partner: an independent company that attributes app installs and in-app events to ad campaigns and decides which ad gets the credit. That role depends on being seen as neutral between the networks it grades. In June, Google, Meta, Moloco and Unity agreed to minority investments in AppsFlyer at a $2.7 billion valuation, in stakes described as non-controlling and non-exclusive, a structure presented as keeping measurement neutral. Apple was not among the investors.

The company's first agent that acts sits on Apple Ads, a platform its own Performance Index placed among the iOS leaders on October 28, alongside AppLovin. Until today, AppsFlyer's agents reported, flagged and recommended. With the Share of Voice Agent, the same company that measures whether an Apple Ads campaign worked can also change what that campaign bids, once a marketer clicks approve. The release does not address how the two roles are kept apart, or whether agents for other networks will follow.

For app marketers, the practical changes are narrower than the talk of "agentic execution" suggests. Custom agents deliver analysis and recommendations on a schedule. The MCP additions let fraud, creative and integration data be queried from AI tools, and let a link be drafted from a sentence. Only one product - in closed beta, on one ad platform - changes anything in a live campaign, and it does so only after a person agrees. How much of that approval is enforced by the software, and how much rests on the process around it, is the question the release does not answer.

Timeline

Summary

Who: AppsFlyer, the mobile measurement company that says more than 15,000 businesses use it, with Inna Weiner, its VP Product, quoted. The changes affect app marketers, agencies and support teams that use AppsFlyer, and in the case of the bid agent, advertisers on Apple Ads.

What: AppsFlyer added Create an Agent, which builds a custom agent from a written workflow description and delivers scheduled analysis, root cause analysis and recommendations to Slack or email. It put into closed beta an Apple Ads Share of Voice Agent that estimates share of voice per keyword and geography, recommends bid changes when an advertiser is off target and executes them in Apple Ads after approval. Four MCP capabilities expose Protect360 fraud data, creative-level performance data, OneLink link creation with approval, and partner integration settings to outside AI tools and AppsFlyer's AI assistant.

When: Today, November 11, 2026. The Apple Ads agent is in closed beta with no general availability date; Apple's Campaign Management API is scheduled to be sunset on January 26, 2027.

Where: The release was issued from San Francisco and sets no geographic limit for the MCP additions or custom agents. The bid agent works on Apple Ads and is available through AppsFlyer customer success managers.

Why: AppsFlyer says marketing teams are moving from manual campaign orchestration to agentic execution, and positions Agent Hub as governed infrastructure for that shift. The release marks the first time one of its agents changes a live campaign, which raises questions it does not answer: how share of voice is estimated, which Apple interface carries the bids, how approval is enforced, and what the tools cost.