The DSP Is Changing: What Happens When AI Agents Start Buying Ads?

For more than a decade, the basic logic of programmatic advertising has remained remarkably consistent.

An advertiser defines a campaign.

A DSP receives the campaign requirements.

The DSP evaluates impressions.

An algorithm decides whether to bid.

An exchange or SSP runs the auction.

The winning ad is served.

And then the cycle repeats — millions or billions of times a day.

But what happens when the software making those decisions is no longer just a bidding algorithm?

What happens when an AI agent can understand a campaign brief, research the market, discover inventory, evaluate opportunities, move budgets, negotiate deals, launch campaigns, monitor performance, and make decisions without waiting for a human operator?

That is the direction the advertising industry is beginning to explore.

And if it becomes mainstream, the DSP may stop looking like the DSP we know today.

The next evolution of programmatic advertising may not simply be AI-powered bidding.

It may be agent-powered media buying.

From automation to autonomy

It is important to understand the difference.

The advertising industry has used automation for years. A traditional DSP can already:

• optimize bids
• adjust budgets
• apply frequency caps
• select audiences
• predict conversion probability
• optimize toward CPA or ROAS
• pause underperforming campaigns
• shift spend between publishers
• use machine learning for bid prediction

This is automation. The system follows rules, models, objectives, and constraints defined by humans.

An AI agent introduces something different. Instead of simply asking:

“What bid should I submit for this impression?”

an agent could potentially ask:

“Given the advertiser's business objective, what should I buy, where should I buy it, how much should I pay, and what should I change next?”

That is a much bigger problem. The agent is no longer optimizing one decision. It is potentially orchestrating the entire buying process.

Imagine a different kind of campaign brief

Today, a media planner might create a campaign like this:

• Advertiser: Automotive brand
• Budget: $500,000
• Market: United States
• Objective: Qualified leads
• Target audience: 25–45
• Channels: Display + Video + CTV
• CPA target: $80
• Flight: 30 days

The DSP then translates these inputs into targeting, bidding, pacing, and optimization decisions.

Now imagine giving the same brief to an AI buying agent. The agent could potentially respond:

“Based on the objective, I recommend allocating 35% to CTV, 40% to premium video, and 25% to display. I found three supply packages with comparable audience quality. One has a higher CPM but stronger conversion performance. I recommend testing it with 8% of the budget.”

Then the agent could continue: inspect inventory, compare publishers, evaluate historical performance, check supply quality, evaluate deal terms, identify suitable PMPs, negotiate pricing where permitted, create media plans, allocate budget, activate campaigns, monitor performance, move budget, and explain why the changes were made.

The DSP becomes less of a dashboard. It starts becoming an execution engine for intelligent buyers.

The new programmatic stack

This could create a very different architecture.

Today

Agentic advertising

The difference is subtle but extremely important. That shift could fundamentally change where value sits in the ecosystem.

The DSP may become an infrastructure layer

This is perhaps the most important implication.

For years, DSP differentiation has often centered around bidder performance, audience data, optimization models, inventory access, reporting, UI, integrations, scale, and targeting capabilities.

But in an agentic ecosystem, some of those capabilities could become commodities. An advertiser might not care which DSP UI they use. They may simply tell their AI agent:

“Get me the best incremental conversions in the US while keeping CPA below $75.”

The agent could decide which buying infrastructure to use. That creates an interesting possibility:

The future DSP may be consumed by agents rather than humans.

Instead of a trader spending hours inside a DSP interface, an agent could interact with the DSP through APIs, protocols, tools, and structured capabilities. The UI becomes secondary. The APIs become critical.

APIs could become more important than dashboards

Traditional DSPs were designed around human workflows: log in, create campaigns, select targeting, inspect reports, adjust budgets, change bids.

An AI agent doesn't need a beautiful dashboard. It needs structured APIs, predictable responses, machine-readable capabilities, clear permissions, transparent pricing, inventory metadata, reliable execution, audit trails, explainable decisions, strong authentication, and policy enforcement.

In other words: the DSP of the agentic era may need to be designed API-first rather than UI-first. The interface isn't disappearing. But the primary user may no longer be a human.

What would an AI agent actually buy?

Today, the atomic transaction in programmatic advertising is often an impression opportunity. A bid request arrives. The DSP evaluates it. The DSP submits a bid. But an agent could operate at several levels.

Level 1: Impression

“Bid $1.72 on this impression.”

This is essentially AI-powered bidding. Not revolutionary.

Level 2: Audience

“Prioritize users with a high probability of purchasing an SUV within the next 30 days.”

Now the agent is managing audience strategy.

Level 3: Inventory

“Buy more inventory from publishers where incremental conversion probability is higher.”

Now the agent is optimizing supply.

Level 4: Deal

“This PMP provides better economics than open exchange inventory. Allocate $50,000.”

Now the agent is managing marketplace relationships.

Level 5: Media strategy

“Shift 15% of display budget into CTV because marginal CPA is lower.”

Now it is managing channels.

Level 6: Business outcome

“The campaign is unlikely to reach the advertiser's target profit at current economics. Reduce spend and recommend changing the offer.”

Now we're no longer talking about a bidding algorithm. We're talking about an autonomous marketing decision system.

The real change: buying intent instead of buying inventory

Traditional programmatic advertising is heavily inventory-centric. The ecosystem asks: What impression is available?

The future could increasingly ask: What opportunity can produce the desired business outcome?

Instead of Impression → Bid → Win, we move toward Business Objective → Opportunity → Decision → Transaction → Outcome. The impression remains important — but it becomes one component of a much larger decision.

What happens to the media trader?

This doesn't necessarily mean media traders disappear. Their role changes.

A trader may spend less time manually adjusting campaigns and more time defining business objectives, acceptable risk, budget boundaries, brand rules, supply preferences, audience strategy, performance thresholds, and experimentation rules.

Think of it as moving from Campaign Operator to Campaign Strategist + AI Supervisor. The human sets boundaries:

• “Do not buy below 70% viewability.”
• “Do not spend more than 20% with a single supply path.”
• “Keep incremental CPA below $90.”
• “Prioritize direct publisher relationships.”
• “Do not use sensitive audience attributes.”

The agent executes within those boundaries. This is a much more scalable model.

But there is a big problem: who controls the agent?

Autonomy introduces a completely different class of problems. Imagine an AI agent has access to a $10 million advertising budget. The agent makes a mistake. Who is responsible — the advertiser, the agency, the DSP, the model provider, the agent developer, the SSP, or the publisher?

This is not simply an optimization problem. It is a governance problem.

The industry will need mechanisms for authorization, spending limits, identity, agent verification, permissions, auditability, accountability, policy enforcement, human approval, transaction history, and dispute resolution.

An autonomous system cannot simply be given a credit card and told to “buy the best media.” The definition of “best” needs to be controlled.

The agent needs a budget, but also a constitution

A human campaign manager operates under company policies. An AI agent needs those policies encoded into its operating environment. For example:

Budget: $500,000
Maximum daily spend: $25,000
Maximum CPM: $18
Minimum viewability: 70%
Maximum frequency: 4 / 7 days
Allowed GEO: US
Preferred inventory: CTV, premium video, display
Excluded inventory: Unknown sellers
Performance objective: CPA < $80
Human approval required: Any single deal > $50,000

The agent can optimize aggressively. But it cannot violate the constitution. That layer could become one of the most important components of future AdTech infrastructure.

And then there is the seller side

The buyer side isn't the only part changing. If buyers have agents, sellers can have agents too.

Imagine a publisher-side agent that knows current floor prices, historical demand, audience composition, available formats, geographic demand, seasonal patterns, inventory forecasts, direct deals, and SSP performance.

A buyer agent asks:

“I need premium US automotive audiences at a target CPM below $12.”

The seller agent could respond:

“I have 18 million monthly impressions matching your requirements. Average viewability is 78%. CTV represents 35%. Available at $10.80 CPM through a guaranteed package.”

The two systems could potentially discover and transact with each other.

Buyer Agent ↔ Seller Agent

Instead of Human → DSP → SSP → Human, we could eventually see Buyer Agent → Marketplace → Seller Agent. That is a fundamentally different market structure.

This is why standards matter

AI agents cannot operate efficiently if every platform speaks a completely different language. A human can learn that one DSP says max_cpa, another says target_conversion_cost, and another requires optimization_goal = PERFORMANCE. An autonomous agent needs interoperability.

This is one reason industry work around agentic advertising standards is becoming important. IAB Tech Lab has been developing its Agentic Advertising Management Protocols (AAMP), with the goal of providing common structures for agent-based advertising workflows, including discovery, negotiation, transactions, and management.

The important point isn't the specific protocol. The bigger point is this: agentic advertising needs a common language. Without interoperability, we'll simply create hundreds of isolated AI agents trapped inside individual platforms.

OpenRTB won't simply disappear

Agentic advertising does not necessarily mean replacing OpenRTB. In fact, OpenRTB could remain extremely important. Think about the layers:

• Strategic layer — AI agent decides the business outcome it wants.
• Planning layer — agent discovers attractive supply sources and deals.
• Execution layer — DSP prices a specific impression.
• Auction layer — OpenRTB continues: bid request → bid response → auction → impression.

So the future may not be “AI replaces OpenRTB.” It may instead be: AI operates above and around existing programmatic infrastructure. The real-time auction still needs extremely fast infrastructure. An LLM does not belong in the critical path of every bid request.

The LLM should probably not be inside the 50ms bid path

If an auction requires a response within tens of milliseconds, sending every impression to a large language model would be impractical. The economics and latency simply don't make sense.

The future DSP may have two brains

Brain 1: Strategic AI

Responsible for campaign planning, budget allocation, channel strategy, supply discovery, deal selection, experimentation, analysis, recommendations, and explanations. Time scale: minutes → hours → days.

Brain 2: Real-time optimization

Responsible for impression scoring, bid pricing, conversion prediction, pacing, frequency, and auction decisions. Time scale: milliseconds.

The mistake would be trying to use the same AI system for both. The winning architecture may combine agentic AI with traditional high-performance ML.

What happens to DSP differentiation?

If every DSP exposes similar capabilities to agents, then some traditional differentiators may weaken. A buyer agent could compare economics, conversion rates, and viewability across DSP A, B, and C — without caring which dashboard looks better.

This could force DSPs to compete on economics, inventory quality, unique supply, data quality, execution reliability, transparency, latency, measurable outcomes, API quality, and agent compatibility. The brand of the DSP may matter less. The capability of the infrastructure may matter more.

The SSP also has to evolve

An SSP that simply returns bid requests may not be enough in an agentic ecosystem. Seller infrastructure could increasingly need to expose structured information: inventory characteristics, audience signals, pricing, deal availability, supply quality, publisher metadata, historical performance, format capabilities, geographic distribution, and commercial terms.

In other words: the SSP may need to become machine-discoverable. If an AI buyer cannot understand what you sell, it may simply move on to another source. See also what a good SSP needs to provide a DSP.

Supply quality becomes even more important

AI agents are extremely good at comparing large amounts of structured information. That could make opaque supply chains increasingly difficult to justify.

If two supply paths provide similar inventory but one has fewer hops, clearer ownership, stronger identity signals, better measurement, better economics, and stronger transparency, the agent can potentially identify the difference.

That means sellers.json, ads.txt, app-ads.txt, supply chain information, identity signals, measurement, and provenance could become even more important — not because an AI agent magically solves supply quality, but because machines can compare supply paths at a scale humans cannot. Related: why DSPs reject otherwise good traffic.

AI agents could change the meaning of “optimization”

Today, optimization often means: “Which bid should I increase or decrease?” Tomorrow, optimization could mean: “Should I buy this inventory at all?”

An agent could evaluate price vs. quality vs. performance vs. reach vs. incrementality vs. risk vs. supply path vs. business outcome. Optimization may move from bid optimization toward economic optimization — a much more powerful concept.

What happens to CPM?

CPM will not disappear. But its importance could decline as the primary optimization metric.

Imagine Publisher A at $4 CPM with a 0.08% conversion rate, and Publisher B at $8 CPM with a 0.25% conversion rate. A human trader might initially focus on the CPM difference. An agent can calculate the expected economic outcome. If Publisher B produces significantly more valuable conversions, the higher CPM may be irrelevant.

The agent doesn't necessarily want the cheapest impression. It wants the cheapest path to the desired outcome.

The new currency could be “outcome probability”

Expected Value =
  Probability of Desired Outcome
  × Value of Outcome
  − Media Cost
  − Risk

Now the question isn't “Is this CPM cheap?” It becomes: “Does this opportunity create enough expected value to justify the price?” That is closer to how financial systems evaluate opportunities than how traditional media buying often operates.

Creativity changes too

If AI agents are selecting media, they may also influence creative selection — matching creative to intent, format, and device. Eventually, an agent could coordinate Audience → Inventory → Bid → Creative → Landing Experience → Conversion instead of optimizing each independently.

What happens to agencies?

Agencies are unlikely to disappear overnight. But the value proposition could change. If AI can handle media planning, campaign setup, optimization, reporting, basic analysis, and budget allocation, agencies may need to focus more heavily on strategy, brand positioning, customer understanding, creative direction, business outcomes, experimentation, governance, and cross-channel orchestration.

The agency may become less of a media execution company and more of an intelligence and strategy layer.

The biggest risk: agent-to-agent advertising

Imagine millions of agents operating simultaneously. A buyer agent wants the cheapest qualified inventory. A seller agent wants the highest sustainable yield. A marketplace agent wants liquidity. An optimization agent wants performance. A brand agent wants customer growth.

All of them negotiate. All of them learn. All of them adapt. Advertising could start behaving less like a static marketplace and more like an autonomous economic system. That sounds exciting — and potentially dangerous — because when machines negotiate with machines, humans may not fully understand every decision.

New problems we haven't solved yet

1. Who is the agent? Can we verify that an agent is actually authorized to spend money?
2. Who controls the budget? Can the agent increase spend beyond its original limits?
3. Who is responsible for mistakes? If an agent spends $2 million incorrectly, who pays?
4. Can agents collude? Could multiple agents manipulate pricing or inventory?
5. Can agents be manipulated? What happens if malicious inventory metadata causes incorrect decisions?
6. Can agents explain decisions? If a campaign suddenly shifts $100,000 to a publisher, the advertiser needs to know why.
7. What happens when agents optimize against each other? Buyer and seller agents continuously pushing prices could produce unpredictable behavior.

Explainability becomes a commercial feature

Today, advertisers often accept “the algorithm optimized the campaign.” That answer may not be sufficient when an autonomous agent controls millions of dollars.

The future DSP may need a decision trace:

Budget moved: +$18,400
From: Exchange A
To: PMP B
Reason: PMP B delivered 23% lower CPA over the last 72 hours.
Confidence: 87%
Expected impact: +11% conversions
Constraint checked:
  Daily budget limit — passed
  Supply quality threshold — passed
  Frequency limit — passed

That kind of transparency could become a competitive advantage.

The winning DSP may be the most “agent-friendly” DSP

An agent-friendly DSP would offer:

• Machine-readable capabilities — the agent can understand what the DSP can do
• Structured inventory — inventory can be discovered programmatically
• Transparent economics — the agent can calculate expected value
• APIs everywhere — almost everything can be executed without a human UI
• Strong permissions — agents can only perform authorized actions
• Policy engines — business rules are enforced automatically
• Decision traces — every major action can be explained
• Real-time execution — once the agent decides, the platform executes immediately
• High-performance bidding — strategic intelligence does not compromise auction latency

What should AdTech companies build today?

1. Make your platform API-first — anything a human can do through the UI should eventually have a structured API where appropriate.
2. Make your data machine-readable — don't force an AI system to scrape dashboards.
3. Build a strong policy layer — budget limits, targeting rules, supply restrictions, and approval workflows as first-class features.
4. Separate strategic AI from real-time bidding — don't put an LLM into the critical auction path.
5. Invest in provenance — the better you describe where inventory comes from, the easier it becomes for machines to evaluate it.
6. Build decision transparency — if AI controls money, explainability becomes part of the product.
7. Think beyond the dashboard — ask: if an AI agent were our customer tomorrow, could it operate our platform without a human?

A possible future architecture

The longer-term stack could look something like this:

And underneath everything:

So, is the DSP dead?

No. The DSP is not going away. But the role of the DSP may change dramatically.

The DSP could move from being the place where humans operate campaigns to becoming the infrastructure that intelligent agents use to execute media decisions. That distinction matters.

The future might not be “AI replaces the DSP.” It may be: “AI changes what the DSP is.”

The DSP could become the high-performance transaction engine underneath an intelligent buying layer. And the platforms that adapt early may have a significant advantage.

The bigger shift

Programmatic advertising started by replacing manual media buying with automated auctions. Then machine learning replaced many manual optimization decisions. The next step may be replacing parts of the decision-making workflow itself.

The real question isn't “Will AI buy ads?”

AI will increasingly participate in buying decisions. The more interesting question is: How much authority will we give it?

Will AI only recommend? Will it optimize? Will it execute? Will it negotiate? Will it control budgets? Will it select supply? Will it choose creative? Will it decide where the next million dollars goes?

And eventually:

Will advertisers even think of themselves as “running campaigns,” or will they simply define business objectives and let intelligent systems figure out how to achieve them?

That is the real transformation. The next generation of AdTech may not be built around better dashboards. It may be built around better agents, better protocols, better APIs, better governance, and better economic decision-making.

The DSP isn't disappearing. It is becoming the infrastructure behind an increasingly autonomous advertising economy.

And for an industry built around milliseconds, the next competitive advantage may come from something that thinks in minutes, hours, and days.

That is the paradox of agentic advertising:

The smartest part of the system may not be inside the auction.
It may be deciding which auction is worth entering in the first place.

FAQ

Will AI agents replace the DSP?

No. The DSP is unlikely to disappear. Its role may shift from a human-operated campaign UI to high-performance buying infrastructure that agents call through APIs, while real-time bidding and OpenRTB continue to execute auctions.

What is agentic advertising?

Agentic advertising is when AI agents can understand objectives, discover inventory, allocate budget, negotiate deals, activate campaigns, and adjust strategy within policy limits — orchestrating media buying rather than only optimizing individual bids.

Should an LLM sit inside the OpenRTB bid path?

Generally no. Auctions need responses in tens of milliseconds. Strategic AI should operate on slower planning timescales, while traditional ML and bid optimizers handle real-time impression decisions over OpenRTB.

How do DSPs and SSPs prepare for agent buyers?

Build API-first platforms, expose machine-readable inventory and performance data, enforce policy and spend limits, invest in supply provenance, separate strategic AI from real-time bidding, and provide decision traces agents and humans can audit.

← All posts · What is a DSP? · What is an SSP? · What is OpenRTB? · Demand partners