Programmatic advertising AI tools: fit, limits and testing

What to take away

Named programmatic advertising AI applications

The market has moved on from vague categories. These six tools illustrate different approaches and buying constraints. Check current pricing with each vendor, because rates change.

The Trade Desk Kokai

Kokai is The Trade Desk's AI platform for bidding, forecasting and predictive clearance across CTV, display and audio. Public filings imply a take rate around 20% of client spend, so it suits agencies with enough media budget to justify the fee.

Minimum spend is not public; your agency or trading desk sets it in the contract, often as a monthly media threshold. Its main limit is that you buy into The Trade Desk's ecosystem.

Google Display & Video 360 automated bidding and Performance Max

Google's AI sets bids and allocates budget across YouTube, Display and Search. Performance Max carries no extra platform fee beyond media spend, but reporting granularity is limited. It suits advertisers with strong conversion tracking and Google Ads access.

Google does not publish a standard DV360 platform fee or minimum spend. Fees are negotiated, often as a percentage of media spend; minimum thresholds depend on your contract.

Amazon DSP

Amazon DSP uses Amazon's machine learning for bid optimisation, audience forecasting and creative selection. Pricing is negotiated or managed service, and access to Amazon shopper data is the draw. It suits retail and ecommerce advertisers, though the data advantage is weaker outside Amazon's ecosystem. There is no public minimum; the fee depends on your contract or managed service agreement.

Criteo AI Engine

Criteo's AI Engine powers predictive bidding and product recommendations across retail media and display. Pricing is often CPC, CPM or CPA, and it works best with a product feed and first-party commerce data. Criteo does not publish a standard minimum, so ask about thresholds. Managed service fees are negotiated and often tied to performance.

Scibids AI

Scibids, owned by DoubleVerify, is an AI decisioning layer that sits on top of DSPs such as DV360, The Trade Desk and Amazon DSP. It optimises bids and budgets against custom goals, and buyers often pay a percentage of media spend. It needs high spend and clean conversion data to train well.

No public minimum exists; the fee is usually a percentage of media spend and the vendor sets a spend threshold.

Moloco

Moloco provides machine learning for performance advertising and retail media. Pricing is negotiated and often tied to performance, and it suits mobile app marketers and commerce firms. The limit is that it focuses on measurable conversion events, not brand campaigns. No public fee or minimum is available; pricing depends on your performance targets and contract.

How to verify vendor claims

Creative claims remain hard to test. IAB UK published principles for creative best practice in programmatic video that set expectations for asset quality, but they describe standards rather than model performance digital video creative principles.

Forecasting claims are similarly slippery. A model trained on last year's auction data may underperform when inventory mix shifts. Ask for a backtest on a period the vendor did not use for training.

Reporting tools deserve the same scepticism. If a dashboard cannot show raw event counts behind its summary, you cannot audit it.

Agency planners weighing these tools against wider market shifts should read the programmatic advertising trends and outlook before committing budget, because platform incentives change faster than model quality does.

A decision table for buyers

What to write into the contract

Set a review point at 90 days. Compare the tool's output against a control group you kept running. If the uplift is inside normal variance, the case for renewal weakens.

Name the data inputs in writing. If the vendor needs first-party data, say who owns it and what happens to the model when you leave.

For a structured look at what goes wrong when these safeguards are missing, the programmatic advertising risk scenarios in England piece sets out the failure patterns agencies see most often.

How to size the spend

Take the fee and compare it with the margin it protects. A team paying £400 a month for a reporting tool that saves six analyst hours a month is buying time, not intelligence. That is still a valid reason, but write it down.

For bid tools, model the fee against expected media waste. If you spend £50,000 a month and the tool claims a 5% efficiency gain, that is £2,500 a month. Ask the vendor to show the calculation method, not just the headline percentage.

Illustrative example: a UK agency with a £50,000 monthly media budget paying a 20% platform fee pays £10,000 a month. A 5% efficiency gain is £2,500, so the fee exceeds the gain.

Common questions

Do I need AI to run programmatic campaigns?

No. Manual rules and placement lists still work at small volumes. AI applications pay off when you have enough events to train on and enough spend for small efficiency gains to matter.

How do I test an AI bidding tool before committing?

Run a limited test on one campaign or region. Compare cost per acquisition over at least four weeks before expanding.

What data does an AI bidding model need?

At minimum, impression-level auction data, your bid history and a conversion signal with a value attached. Without a value signal, the model optimises to the wrong target.

Can an AI tool replace a media planner?

No. It can shorten the loop between test and adjustment, but the objective, audience definition and budget still need a person to set them.