What is agentic media buying, and what can agents do today?

Section 01

A brand plans paid media per campaign and typically runs it weekly

Planning and setup happen once per campaign, while pacing, changes, and reporting repeat every week, and those three steps read account data or make routine edits.

Each of the five steps has a usual owner

• Step 1: Plan. The brand manager sets goals and a budget, and an agency strategist or in-house lead turns them into a media plan.

• Step 2: Set up and traffic. The buyer builds the campaigns, audiences, ads, and tracking on each platform. Trafficking means loading the finished ads and their tracking tags into each system.

• Step 3: Pace. The buyer or an analyst checks spend against the media plan, daily or several times a week. The State of PPC 2026 survey of pay-per-click (PPC) professionals, whose partners include PPC software sellers such as Optmyzr, reports that half of the 670 individual contributors and freelancers it asked track paid search budgets by hand¹.

• Step 4: Change. The buyer shifts budgets, adjusts target costs, adds negative keywords, and pauses ads that have stopped working, mostly in a weekly review.

• Step 5: Reporting. The analyst pulls each platform's results into one weekly summary for the brand.

Download: SVG, PNG. Wider than the screen: swipe sideways, or download the PNG.

For an example CPG brand, the plan starts at the retailer

Retail media is the ads a retailer such as Walmart or Amazon sells on its own sites and apps. Each retailer's ad business is a retail media network. For a consumer packaged goods (CPG) brand, such as a snack maker, the retailers set the promotion calendar. It shows when the brand's products go on sale at each retailer. The brand decides how much to spend on each retailer's ads around those dates and times its social and search pushes around the biggest promotions.

This example is an illustration, not a client or a real company. A snack brand is in the weeks before a spring promotion at Amazon, one of its retailers, with paid media on Meta, Google, and Amazon. An agency buyer runs the campaigns, and an in-house analyst builds the weekly summary. The buyer also plans a new Meta campaign for the promotion's launch. Each morning, someone checks whether the promotion's spend is on plan on all three platforms, which is Step 3. This Monday, Amazon is behind plan. The buyer weighs a budget shift between Amazon campaigns, from those not spending their share to those that can.

An AI agent calls tools until a task is done

An AI agent is a language model, the kind of AI model behind ChatGPT, Claude, and Gemini, given a goal and a set of tools. A tool is one named action the model can take in another system, such as pulling yesterday's spend or changing a campaign's budget. The agent picks a tool, calls it, reads the result, and decides the next call until the task is done.

Section 02

Automated bidding leaves budget allocation and goals to the buyer

Where a campaign uses the platform's automated bidding, the platform sets each bid, and an agent can check the settings the buyer still decides and draft edits to them inside one channel.

Buying moved from orders people negotiated, to auctions, to platform software that sets the bid. Direct sales run on insertion orders, signed orders for ad placements at a negotiated price. A 2017 academic survey of the field found that display ads were sold either in real-time auctions or in advance, and those sold in advance still mainly went through direct sales². Google's AdWords was selling search ads by auction by February 2002, and display ad exchanges, marketplaces that auction each ad view to bidding software, followed around 2005².

Google's and Meta's automation picks bids and placements

Google's Smart Bidding sets a bid in each auction from the chance that the ad view leads to a conversion³. In Performance Max, Google's campaign type that serves one campaign across all of its inventory, the advertiser defines the conversions it cares about and an optional return on ad spend (ROAS) or cost per acquisition (CPA) target, which Smart Bidding works toward³. Google says the system moves budget toward the inventory and placements it rates highest³.

On Meta, a campaign counts as Advantage+, Meta's automated campaign setup, only when Advantage+ audience, Advantage+ campaign budget, and Advantage+ placements are all on⁴. Such a campaign sets no placement targeting or exclusions, which leaves every available placement eligible⁴.

The media buyer still sets budgets, goals, conversions, and creative

• Budget allocation, meaning how the total splits by channel, campaign, and tactic
• Goals and target costs, such as a ROAS or CPA target
• What counts as a conversion
• Campaign structure, meaning how products and audiences split into campaigns
• Audiences and exclusions, where the platform still allows them
• Bids, only where the buyer still sets them, such as keyword bids in retail media or campaigns on manual bidding
• Creative, the images, video, and copy

Where automated bidding runs, an agent leaves bids alone, because the platform sees far more data about each auction than an outside agent can. Meta's developer documentation says significant edits send an ad set back into the learning phase, when Meta starts counting the ad set's conversions again from that edit⁵. Results swing until the ad set leaves the phase, so limit how often the agent edits each ad set.

Section 03

Agentic media buying is an agent calling tools under your login

An agent can do only what its tools exist for and what the login it runs under allows, whichever model sits behind it.

Three parts sit between a buyer's question and the ad account

• The API (application programming interface) is the platform's way in for software: most of the same account data and settings as the web screens, reached by code.

• The connector is a service that presents the API to a language model as a named list of tools, which the model can find and call on its own⁶. It follows the Model Context Protocol (MCP), a public specification for listing and calling tools⁶. Platform documents call a connector an MCP server, and an official connector is the platform's own.

• The AI assistant is the app you type into, such as Claude, ChatGPT, or Gemini. The language model runs inside it as the agent. The account data the tools return lands there, under the AI company's terms for storing and using your data.

Agentic media buying is an AI agent reading and changing your ad accounts through a connector, under a login your team sets up.

Example: One question to the agent takes four steps

On Monday, the snack brand's buyer asks the agent how the promotion is pacing on Google. Google's guide for its Google Ads connector describes what happens next⁷:

• The model picks a tool. It picks the connector's search tool, which queries account data.
• The connector runs the query. It sends the request to the Google Ads API.
• The results return. They come back into the model's context window, the working text the model reads before it writes anything.
• The model answers. It writes the result in plain language, such as which Google campaigns are behind plan.

A tool that changes the account takes inputs such as the campaign and its new budget. The platform applies the change when the call arrives, unless the connector or AI assistant first asks a person.

The tool list and the login set the limits

• With sign-in set up, every tool request to Google's connector runs under the signed-in person's identity⁷.
• The MCP specification lets a connector show only the tools that a login's permissions allow⁶.
• Google's API has no access settings of its own and follows each login's user role in Google Ads, so a login that should only read gets Google Ads' read-only role⁸.

Set every login's access tightly before an agent connects. An agent will try any tool its login reaches. The account owner, the one person at the brand who answers for that account's spend even when an agency runs it, gives every login the narrowest role that does its job. A system user, a business-owned login that is not a person, gets the same.

Use the platform's official connector where one exists. Use a vendor's connector only for work no official connector covers, such as several retail media networks at once. A vendor's connector can write through Google's API, which accepts changes⁸. Don't add one to make changes where the platform's own connector only reads, as Google's does. The buyer makes every change there anyway, so the vendor's write access saves only typing and adds a company that holds your login. Vet a vendor's login handling and data terms as you would an agency's.

Repeated checks run steadier as API scripts

An API script is a short program that calls the platform's API directly and runs the same steps every time, with no model involved. An agent's answers can vary. Anthropic, which makes Claude, says identical inputs may produce different outputs even with the model's randomness setting at zero⁹.

Move a daily check, such as pacing, from the agent to an API script once its answers match the platform's figures and its steps stop changing, where your team or agency has access to the platform's API. The script writes into the team's own reporting, so the team and the script share one set of numbers. Keep the agent for work that needs judgment, such as explaining why spend moved or trying workflows that combine several tools.

A lean team's best reason to keep a repeated check with the agent is upkeep: a script is code someone has to fix when campaigns are renamed or the API changes, and nobody on staff reads code. The agent can redraft the script when that happens, and the analyst checks the script's output against the platform's own figures rather than its code.

Section 04

Meta's and TikTok's agent connectors change campaigns, and Google's only reads

As of October 20, 2026, Meta's connector (in open beta) and TikTok's let an agent build and change campaigns, and no other platform in this table lets an agent change one yet.

Google and Microsoft let an agent read, and people make changes

Here the agent proposes new budgets for the snack brand's Google campaigns that are behind plan, and the buyer enters them in Google Ads. Google said in November 2025 that its own assistant inside Google Ads, Ads Advisor, can apply suggested changes once the advertiser reviews and approves them²⁰. An outside agent working through Google's connector can't.

Meta and TikTok let an agent build campaigns that spend money

Meta's connector creates every new campaign and ad set paused, and the AI assistant asks for confirmation before the tool ads_activate_entity switches one on and starts spending²¹. That question comes from the AI assistant, and a user can turn it off for a tool. Anthropic tells Claude users to click "Allow always" only for a connector and tool they trust to run unsupervised²².

A second tool, ads_update_entity, changes existing campaigns, ad sets, and ads²¹. Meta's help article says any action taken on your behalf requires "your authorization through the AI agent" without saying whether that means each edit or only the connection²³. The snack brand's agent builds the new promotion campaign through this connector. Switching it on spends money, so the buyer approves it first.

Meta lets a business set limits on what an agent creates and changes, in limited availability²⁴. Each one blocks one agent action on one ad account or product catalog, such as "any budget increase above 20%"²⁴. Meta's help article says Meta's connector enforces them, and a person with full control of the business portfolio, the Meta account that holds a company's ad accounts, sets them²³. A partner given access to an ad account, such as an agency, can set its own limits there, which cover only the partner's own people²³. Where an agency runs the agent, the account owner reviews the agency's limits too.

In Anthropic's published list of connectors for Claude, TikTok's connector lists tools such as smart_plus_campaign_create, which creates a campaign, and smart_plus_adgroup_budget_update, which changes an ad group's budget¹⁴. TikTok says agents can run whole workflows, from creating a campaign to adjusting bids, "without human intervention at every step"¹⁵. Its help page and launch announcement describe no approval step before a change goes live^{15, 16}.

For TikTok Shop ads, TikTok offers an automation that TikTok says "automates end-to-end campaign creation and optimization"²⁵. Its help page names no connector or API for it.

Expect a connector on every platform you buy on, perhaps read-only

Every connector in the table with a launch date came out in 2026, at least to a test group. Assume the rest will follow, to keep up with the platforms that have one and with the pressure on platform teams and agencies to adopt AI. Some may offer read access only and keep changes inside their own assistant, as Google does. Walmart Connect offers Marty, its own advertising assistant, in beta for Sponsored Search as of January 2026²⁶, and no connector.

Each new connector gets the same setup, with tight login roles and a read-only start. Until a beta ends, recheck the connector's tool list before giving the agent a new kind of task, because the platform can still change it. Meta's help article warns that "You may not have access to all of these tools and features yet"²³.

Section 05

Outside Amazon, retail media networks give agents read access or nothing

Of the retail media networks checked, only Amazon's official connector, an open beta for companies already on its ad API, lets an agent change campaigns, and Criteo's only reads^{27, 28}.

Amazon lets an agent create, change, and delete campaigns

Amazon's open beta is for Amazon Ads partners with active API credentials, meaning companies already set up on Amazon's ad API²⁷. Amazon's guide shows how to switch individual tools off in the AI assistant, such as the tools that delete a campaign or an ad³⁰.

The snack brand's shift between Amazon campaigns runs through this connector if the brand or its agency has access to Amazon's ad API. If not, the agent drafts it. The buyer then makes it in Amazon's console.

Skai's connector works across networks, and sales-data firms list no tools

A brand may also connect its ad software, or the sales-data firm that checks its ad spend against store sales:

• Skai, a software company for running ads across retail media networks and other channels, launched Agent Connect on September 22, 2026, with 34 tools³¹. Skai says agents can use them to change campaigns, bids, and budgets, with every proposed change subject to human review³¹. Its product page lists preview and approval controls as coming soon³². Ask how that review happens today.

• Circana (a sales-data firm) says its Liquid AI connects to AI assistants such as Claude and ChatGPT through MCP, and names no tool list, access terms, or status³³.

• SPINS (another sales-data firm) calls SPINS Foundry its data platform for agents and applications, but its September 2026 announcement names no connector or API to its sales data³⁴.

Section 06

Agent autonomy rises in five levels as human involvement falls

Any agent or pitch sits at one of five levels, from a person asking the agent questions to an agent acting alone while a person reads the record of its edits.

Researchers writing for the Knight First Amendment Institute at Columbia University name each level by the role the user takes, and treat it as a design choice separate from what the agent can do³⁵. An agent that can change campaigns can still run at Level 1, where it only answers. Here are their five levels:

• Level 1: You run the account, and the agent answers. You are the operator. You ask how the promotion is pacing on Amazon and act on the answer yourself.

• Level 2: You and the agent work through changes together. You are the collaborator. The agent suggests a budget shift, and the two of you revise it.

• Level 3: The agent does the work, and you give direction. You are the consultant and stay in the workflow. You set the goal, say which Amazon campaigns to favor, and agree to each step as the agent drafts and makes the shift.

• Level 4: The agent acts and asks you to sign off on consequential steps. You are the approver, out of the workflow until a consequential step. A consequential step here is one that spends money or is hard to undo, plus any kind of edit the account owner has not yet cleared for the agent to make alone. The researchers note that an agent may find it hard to tell reliably which steps those are³⁵. The shift between Amazon campaigns is one, and it waits for the buyer's sign-off.

• Level 5: The agent acts alone, and you watch what it did. You are the observer. You can read a record of what the agent did, but not steer it, and the only control is an emergency off switch³⁵.

Each level up needs more controls in place

Level 1 and Level 2 need read access and someone who checks the agent's numbers. Level 3 adds write access with a spend limit and written rules. Level 4 adds the approval step as a set gate, because you are no longer in the workflow to agree as the agent goes. Level 5 drops the approval step and puts a named person on the agent's record and the change history, the platform's record of each edit and who made it.

Download: SVG, PNG. Wider than the screen: swipe sideways, or download the PNG.

Section 07

Hand an agent pacing and QA checks first, then routine changes

Pacing and QA checks change nothing, so they show whether the agent reads your numbers correctly before it may propose a change.

Start with the daily pacing sheet and three QA checks

At the snack brand, the first job is the analyst's daily pacing sheet: spend against the media plan by platform, built by hand. The second is three QA checks:

• Setup: each new campaign's budget, dates, targeting, and tracking match the brief.
• Changes: the buyer's intended edits landed, per each platform's change history.
• Running campaigns: the agent tells the buyer about any campaign that no longer matches the media plan.

Give both to the agent under a read-only login, and have it explain why spend moved. Once its sheet matches the platform's figures and its steps stop changing, move the sheet to an API script.

The agent reads the same numbers your team uses

Meta says the API the agent reads has different defaults from Ads Manager, the buyer's screen, except that since June 2025 the API follows Ads Manager's attribution settings³⁶.

Give the agent one source of numbers: the team's own reporting, if it has one, where each metric is defined once and every number traces back to the platform's figure. Use the connector for what that source lacks, such as live settings, the change history, and edits. While the agent is read-only, the analyst checks each of its explanations against the change history and what the buyer knows happened. On a small team, the buyer holds that check. Once its explanations stop needing correction, the account owner lets the agent start drafting edits.

At Power Digital, an independent agency focused on data and measurement whose data team the author leads, agents today only read data and recommend, and a person makes every change in the ad platforms. The team is testing the action tools under tight controls.

Write the account's rules down before the first change

The account owner writes down:

• the spend limit for each campaign
• the campaigns the agent edits, and how often
• the campaigns it leaves alone, so no campaign has both a buyer and the agent editing it
• which changes need a buyer's sign-off
• the kinds of edit it can't make
• what it pauses, and what it never switches on
• who hears about each change

Those written rules reach the agent as standing instructions, the text the AI assistant gives it before every task. A language model doesn't always follow them. Set every limit the platform enforces too. On Meta, set its per-action limits where the account has them²⁴. Meta's account spending cap stops every campaign in the account once reached, so set it above planned spend³⁷. Where a platform enforces nothing, the login's role is the control.

Routine edits start with the buyer's sign-off on every one

Routine means an edit inside one channel that stays within the approved media plan, such as:

• adding negative keywords
• pausing ads that have stopped working
• shifting budget between ad sets that each have their own budget, never where the platform's automation sets that split

Each of these cuts delivery or moves money. At first the agent drafts it and the buyer signs off. The account owner clears a kind of edit for the agent to make alone only after the buyer's approvals show the agent drafts it as the buyer would and a test shows it works. Adding negative keywords and pausing ads can qualify, because each shows in the change history and is undone in one step.

A budget shift spends money, so it always needs sign-off. Each draft shows:

• the current and recommended settings
• the change in dollars and percent
• where the money comes from
• the review date, and how to undo it

Keep the sign-off even where a platform enforces limits, since a limit can't tell a good change from a bad one.

Every agent change is logged, and each kind of change is tested

Platforms record who changed what and when, not why. Meta's ads connector has a tool that reads the account's change history, which Meta calls the activity log³⁸. The change event report in Google's API reaches back only 30 days³⁹. The team keeps its own change log, which the agent writes as it acts: each action, its reason, and the settings before and after.

The analyst checks each routine edit against the metric it aimed at, such as spend on blocked search terms after a negative keyword. That shows nothing broke, not that the edit worked. When the check shows something broke, the buyer undoes the edit, and that counts against clearing that kind of edit, as a declined draft does.

Testing settles whether a kind of edit works. Google's experiment splits a campaign's budget or traffic between the original and a trial with the change⁴⁰. An experiment takes weeks, too long to run on every edit, so the team tests each kind of edit before the account owner clears it.

What the tests and the buyer's approvals show goes back into the written rules, as the marketing harness article describes. A kind of edit that failed its test goes on the list of edits the agent can't make.

Download: SVG, PNG. Wider than the screen: swipe sideways, or download the PNG.

IT, legal, and finance object first, and each gets a control

• IT and security ask where data goes and whether outside text could steer the agent. Anything it reads, such as a competitor's ad copy, enters its context window, and the model can follow instructions written there. Start read-only, keep agents that read outside text off write tools, and have IT check the AI company's data terms.

• Legal asks what claims an agent could publish. Keep copy and new ads behind legal's usual claims review.

• Finance asks who pays for media an agent wastes. The brand does, and Meta's terms from October 30, 2026 make the advertiser "solely responsible" for third-party software it uses or authorizes⁴¹. The controls are spend limits, one named account owner, and an agency contract naming what the agency covers.

Section 08

A person decides budget allocation across channels

Budget allocation across channels rests on margins and commitments no ad account holds, and its evidence arrives weeks later, so a person reviews and approves each shift.

A randomized lift test measures what ads caused

A platform counts each conversion it can tie to an ad view or click inside its attribution window. The window is the days after a view or click in which a conversion still counts⁵. That includes sales from people who would have bought anyway. A lift test, a randomized test that holds some people back from the ads, measures incrementality: the sales an ad caused that would not have happened without it.

Researchers compared 663 randomized experiments at Facebook with two methods that estimate effects without a test, using richer data than most advertisers can access⁴². For lower-funnel outcomes, those closest to a sale, the experiments' median lift was 5%, while the other two methods put it at 24% and 64%⁴². A platform's own count has no held-back group either, so it is no more reliable than those methods.

A move between channels compares counts each platform makes its own way, so it needs a lift test or a media mix model, a statistical model of how each channel's spend drives sales, calibrated with lift tests. Without either, the brand manager judges the shift on the brand's own sales.

Lift results adjust the target the agent checks against

If the brand has lift test results, a shift between tactics inside one channel is judged against a lift-adjusted target, the platform target corrected by those results. That matters most in tactics that reach people already likely to buy, such as retargeting and brand search, where the platform's own return on ad spend overstates what ads caused by the widest margin.

Incrementality factor can be defined as the incremental conversions a test found divided by the conversions the platform reported⁴³. With a factor below one, the CPA target on the platform's count sits below the real goal. A factor above one, as when a lift test reads store sales no platform sees, puts it above. In this example, a factor of 0.6 turns a $150 CPA goal into a $90 platform target⁴³. With a factor per tested channel and campaign type in its standing instructions, the agent checks similar campaigns against the adjusted target.

No connector sees the whole budget across channels

The best case for handing an agent the split across channels is that a media mix model already recommends one. The agent would only apply it faster. But the model and the lift tests behind it lag the spend. Its split predates the brand's current margins, stock, and commitments, such as a CPG brand's retailer agreements. None of those sit in an ad account, and the brand manager weighs them as they change.

When the agent drafts a shift from social to retail media for the snack brand, the brand manager decides. At some brands the buyer and the brand manager are one person. What matters is that a person reviews and approves every shift.

Download: SVG, PNG, MP4. Wider than the screen: swipe sideways, or download the PNG.PausePlay

Section 09

Make agencies and vendors prove agentic media buying on your account

Put six questions to any agency or vendor selling agentic media buying, then check the answers by running its agent read-only on your own account.

Ask each question, and look for this answer:

• Which connector. The platform's official one where it exists, with its tool list shown to you.

• Which level. One per task, such as Level 1 for reporting and Level 4 for budget changes, agreed between your team and theirs.

• Whose login, with which role. A dedicated login your brand owns, read-only to start, so the change history shows the agent, not a buyer. Meta says its connector accepts a system user, though not one with the admin role¹².

• Which limits stop a bad change. Limits the platform enforces, plus the written rules.

• Where every change is logged. The change history, the team's change log, and who reads them.

• Who covers a bad change. The contract names it.

Limits matter most for a scheduled task, which the AI assistant runs on its own at set times. Anthropic says a scheduled task in its coding tool, Claude Code, can use every tool of each connector the task includes, "including writes, without asking for permission during a run"⁴⁴. The same page says a scheduled task's connector actions appear as the signed-in user⁴⁴.

Run scheduled tasks under that dedicated login, with only the connectors each task needs. Keep that login read-only. Give it write access only when its role and the platform's own limits can block every kind of edit the account owner has not cleared.

Compare the agent's proposals with your buyer's own calls

No public case from a vendor, agency, or platform, as read in October 2026, shows an agent running retail or CPG campaigns with a result from anyone but the seller. Before you sign, have the person who checks the agent's explanations while it is read-only compare its proposals on your account with the platform's figures and with what the buyer would have done.

When one agency runs both the agent and the account, someone at the brand reads the comparison, because the seller should not grade its own work.

An agent costs a subscription, model usage, and people's time

• The subscription. The AI assistant's fee.
• Model usage. Every run reads account data into the model. Ask how the subscription bills or limits a scheduled run.
• People's time. Someone sets the agent up, reviews what it proposes, and reads the change history.

No source puts a monthly figure on an agent running an ad account. Ask your agency which tasks an agent now does, and what people's time its fee now pays for.

Marketing operations connects one read-only agent first

Marketing operations or an analyst sets it up with the buyer, and the account owner grants the read-only role, in this order:

• Pick the AI assistant and a business subscription whose data terms your team accepts. As of October 20, 2026, Anthropic says only the Owners and Primary Owners of a Claude Team workspace can add a connector by its web address²².

• Add the platform's connector. OpenAI's guide says that in ChatGPT this means entering the connector's address, setting up sign-in, and accepting a risk warning⁴⁵.

• Sign in read-only.

  • On Meta, grant ads_read rather than ads_management¹².
  • Where a platform offers no read-only permission, switch its write tools off in the AI assistant. That setting binds only that assistant, which makes it the weaker control.

• Give it the media plan, and ask for the daily pacing sheet.