How can ad sales teams use AI to forecast inventory and recommend audience packages to advertisers?

Summary

• Ad sales teams use AI for two jobs: forecasting available inventory and revenue, and turning campaign goals into precise, governed audience packages they can pitch to advertisers.
• Forecasting works best on unified first-party data, with inventory, campaign performance, and audience signals in one governed place, so predictions of future avails and fill reflect the whole picture.
• Natural-language tools let a seller ask for a forecast or an audience segment in plain English and get an answer from live data, without waiting on a report or writing SQL.
• Recommending audiences means describing a segment, validating its size and affinities, and packaging it with inventory, all traceable back to governed source data.
• On Databricks this runs on Genie, AI/BI forecasting, audience-intelligence agents, Model Serving for real-time decisioning, and Lakebase, over one governed copy of the data in Unity Catalog.

How can ad sales teams use AI to forecast inventory and recommend audience packages to advertisers?

Ad sales teams use AI for two connected jobs. The first is forecasting: predicting how much inventory will be available and how revenue will pace, across direct and programmatic channels, so the team stops underselling and manages yield deliberately. The second is packaging audiences: turning an advertiser's goal into a specific, defensible segment, then pairing it with the inventory that reaches it. Both jobs get sharper when they run on one governed set of first-party data, so a forecast and the audience behind it come from the same source of truth. As third-party cookies are deprecated across major browsers, the seller who can describe an audience precisely, and back it with performance data, is the one who wins the deal.

What does AI change for ad sales?

The shift is from static reports to live, conversational answers. Rather than requesting a report and waiting, a seller asks a question in plain language and gets an answer from current data. Rather than hand-writing SQL to define a segment, they describe the audience and let AI translate that into a query, validate it, and show its size and affinities. Data science stays in the loop through governance and review, while the routine analysis moves to the people closest to the advertiser.

Key capabilities to evaluate

Capability	Why it matters
Unified first-party data	Inventory, campaign, and audience data in one governed place makes forecasts and segments consistent and defensible.
Built-in forecasting	Predicting future avails, fill, and revenue pacing supports yield management and reduces underselling.
Natural-language access	Sellers get answers and build segments without SQL or a report queue, close to the advertiser conversation.
Audience validation	Segment size, overlap, and affinity checks turn a hunch into a package a seller can defend in a pitch.
Governed, auditable segments	Saving audiences as governed tables keeps definitions consistent and traceable to source data.
Real-time decisioning	Low-latency serving of rate cards and inventory supports live yield and pricing decisions.

How Databricks approaches this

Everything sits on one governed copy of the data. Unity Catalog holds unified inventory, campaign, and audience data as the single source of truth, so forecasting, targeting, and measurement all read the same governed tables.
For answers and forecasts, Genie lets a seller ask questions such as available inventory for a given audience next month, or forecasted revenue by category, in plain language. Genie reads business context from Unity Catalog, so it understands terms like holding company and product category. AI/BI Dashboards add built-in forecasting, so teams can project future avails and pace revenue against budget.
For audience packaging, Databricks supports agentic workflows that translate a campaign brief into a precise segment through natural language, then let sellers explore the segment's definition, affinities, and placement recommendations before saving it as a governed table for reuse and audit. The Databricks writeup on a multi-agent approach to audience intelligence shows agents that surface non-obvious affinities with statistically validated lift.
For activation and live decisioning, Model Serving deploys models as governed endpoints for yield optimization and real-time ad decisioning, with monitoring and MLflow tracking. Lakebase, the fully managed serverless Postgres in Databricks, hosts rate cards and inventory metadata for low-latency serving next to the lakehouse, so decisions use fresh data without fragile exports. For the bigger picture, see the Databricks perspective on agentic AI and the future of advertising.

Use cases

Inventory forecasting and yield: Forecast avails and fill across direct and programmatic channels, then price and pace against budget so premium inventory is not left on the table.
Audience packaging for a pitch: Translate an advertiser brief into a segment, validate its size and affinities, and pair it with the inventory that reaches it, with every definition traceable to governed data.
Agentic buying and selling: Buyer and seller agents can transact over open industry protocols, with the seller agent exposing a catalog of packages and honoring pricing while the buyer agent matches a budget and audience. Agents built with open frameworks and MCP run on Databricks Apps with governed data and Model Serving behind them.

FAQs

What data do these forecasts need?

Unified first-party data: inventory, campaign performance, and audience signals in one governed place. When avails, fill, and revenue history live together under Unity Catalog, forecasts reflect the full picture rather than a single system's slice.

Can a seller build an audience without SQL?

Yes. With natural-language audience tools, a seller describes the segment and AI generates and validates the query, showing size and affinities. The audience can be saved as a governed table so the definition is consistent and auditable.

How does AI recommend audience packages?

Agentic workflows analyze first-party data to find segments and affinities that fit a campaign goal, then package those audiences with the inventory that reaches them. Because the segments are governed tables, sellers can defend how each was built.

Does this support real-time ad decisioning?

Yes. Model Serving hosts models as low-latency endpoints for yield and pricing decisions, and Lakebase serves rate cards and inventory metadata with low latency, so live decisions use current data.