Which songs should we invest in next?

An artist's catalog can keep earning after the artist has died. The team still has to choose where to spend time and money on the recordings that remain. Streaming statements are one source of evidence. I also want to investigate whether merchandise purchases associated with a body of work can help identify songs worth promoting.

In one earnings review, the supplied Spotify records covered different months in the two years. Restricting the comparison to the months present in both years changed the result.

I approach the work in three steps. First, I connect the available data through a unified API that checks the caller's access and preserves source identity. Second, I investigate the records and build domain tools that return a calculation with its coverage and original receipts. Third, I serve those tools through Google Cloud and MCP so the team can use the findings in its working environment and return campaign results for review.

For this example, the proposed action is a coordinated clipping campaign. The team would choose a group of recordings supported by the evidence and publish permitted short clips. We would follow the response through to proceeds the catalog owner can receive. The commercial test is whether the campaign produces additional cash flow after its costs.

An artist can begin with contacts and files in Google Workspace, then sign a record deal and acquire a new set of reporting systems. A merchandise store keeps purchase history. A promoter holds ticketing records. By the time the artist has a valuable catalog, years of business activity may be distributed across organizations with separate agreements and access rules.

Start with a royalty export you are permitted to use. Record which organization supplied it, which periods it covers, and what the amount fields mean. To investigate whether merchandise purchases are associated with the same body of work, find who controls those purchase records and establish the permitted use. The first dataset may already be available while the second requires a conversation with a partner. Keep both in the source register, with that difference visible.

Keep the original files in private storage and retain each source version used to prepare an answer. In the reference implementation, I used Google Cloud Storage and preserved references to original spreadsheet cells and statement rows. A corrected royalty export should create a new version. An analyst reviewing the earlier result should still be able to inspect the evidence used for that calculation. Preserve source dates separately from the date a file arrived.

At request time, resolve the caller's organization from its current authenticated membership. Check access to the selected evidence product and to any original receipts the caller requests. A catalog-earnings request can select a period and store, but those selections cannot confer permission to read another organization's records. OAuth provides the connection's authorization flow; the service still has to enforce membership and product scope.

The retained first-half groupings contained Spotify observations for all six months of 2024 and four months of 2025. April and June 2025 had no Spotify observations in those groupings. January, February, March, and May were present in both years.

Using every supplied month answers how the reported amounts in the available records differ between the two years. Restricting both years to the four shared months answers a different question about those selected periods. The decline was smaller in the second calculation. Keep both results and their month selections visible. Neither establishes that reporting was equally mature in both years, and an absent observation remains unknown rather than zero. The next task is to obtain evidence about reporting coverage before explaining the cause of the difference.

I used the bronze, silver, and gold stages of Databricks' medallion approach to make that investigation reproducible. In bronze, retain the original files. In silver, prepare the detailed records with explicit periods and source identifiers. In gold, publish the two calculations with their coverage and references to the contributing records. The implementation runs on Google Cloud and applies this pattern without a Databricks workspace or runtime.

Build the royalty review around the question an employee asks about a store or recording over a period. Return the calculated amount together with the selected months and source version. Let the employee inspect contributing records and request an original receipt.

Keep source discovery separate from the calculation. I used semantic search to locate relevant sources from their descriptions, then exact source references to select the records. Decimal arithmetic determined the amounts. The language model could explain the result while the calculation remained reproducible from the retained evidence.

Connecting music to merchandise requires another investigation. A product associated with an album could give you a reason to examine that body of work. Review the relationship between the product and the catalog before joining the records. A matching title can refer to different recordings, and a purchase does not establish which songs that customer listens to. Keep unresolved matches separate until you have evidence for combining them.

For the campaign example, inspect the available customer definitions and purchase history, then test whether they support a particular group of recordings and a relevant offer. Save the evidence for each proposed relationship so another person can review the selection. Use the missing evidence to decide what to request from the commerce or catalog team before selecting recordings.

An employee asks what changed in Spotify earnings over a selected period. The AI client requests a royalty review for that store and period through MCP, the Model Context Protocol. Before returning evidence, the service checks the caller's current organization membership and access to the selected product.

In the Google Cloud implementation, the request passes through an HTTPS gateway on Compute Engine to the service on Cloud Run. The runtime reads published evidence products from Cloud Storage and composes the review. Preparation of the source files happens before those products are published. An approved client receives the requested result and references to the evidence it may inspect.

The response should keep the calculated amount beside its period coverage and source revision. A follow-up can request an original receipt under the same access checks. Verify the amount against the published calculation and check the receipt against its original row or cell. Those checks establish the evidence behind the selected answer. Connecting the client and exercising its permissions is a separate part of acceptance.

In the worked implementation, selected private-data reviews were exercised in ChatGPT and the application. Hermes was exercised with an earlier synthetic version. For each client, test an ordinary question without supplying the tool name or expected figure, then follow the answer to its evidence. Include revocation in the access tests. Broader employee adoption still requires testing the work with the people who will use it.

The proposed next action is a coordinated clipping campaign. Thirty to fifty songs is an illustrative portfolio size. Select the actual recordings from the catalog and audience evidence, establish permission to publish the creative material, and assign someone to coordinate the campaign. The brief should explain why each recording was selected and where its clips will lead listeners.

Use a campaign ID to group the work. Give each creative version an identifier and each publication its own identifier. Retain the recording ID so the team can connect a clip to the intended recording. Record the publication date and destination alongside the spend.

Choose the baseline before launch. Where feasible, stagger publication or identify a suitable group of recordings that will not receive the campaign during the comparison period. Keep the same observation windows visible when comparing results. This is how the team would investigate whether a change followed the campaign rather than another release or event. A rise after publication alone does not establish an effect.

Collect results over months because listening activity and reported proceeds may arrive on different schedules. Preserve the reporting periods and source dates with each export. When a corrected statement arrives, retain the new version and reconcile the affected result. Matching calendar months still requires checking whether the reporting is sufficiently comparable for the question being asked.

Evaluate the proceeds the catalog owner can receive under the relevant rights share, after campaign costs. Include merchandise income only where that owner has an agreement to receive it. The commercial test is additional cash flow attributable to the campaign. Correct arithmetic and a working AI interface support the review, but they do not establish that outcome.

Use unresolved findings to request the next records from partners. A distributor may need to supply an updated statement; a commerce partner may hold the purchase history needed to assess a proposed audience. Agree on access and return those records to the same investigation process. The team can then reconsider the selection before committing to the next campaign.

I began by asking how a company could make its private data usable by employees and AI agents. Working through the catalog led me to a more specific commercial question. Streaming reports could show what the recordings earned, but I also wanted to understand whether purchases elsewhere in the artist's business could help us decide which songs to promote.

Merchandise associated with an album gives us something concrete to investigate. I would check which recordings the product refers to and examine the purchase history we are allowed to use. A verified link to that body of work gives the campaign team a reason to consider its recordings. We would still need to test whether those customers are an audience for the music. A purchase alone does not establish that.

The clipping campaign would put that selection to the test. I would track the selected recordings through publication and then through the reporting periods that follow. The commercial question is whether the catalog owner receives additional proceeds after campaign costs. Merchandise income belongs in that calculation only when the owner has a share of it. The review would need to distinguish earnings caused by the campaign from changes that would have happened anyway.

The next data request would come from that review. If we cannot follow a result because the records sit with a store or distributor, we know which partner to approach and what to ask for. We can bring those records back into the same system and reconsider the selection. I want to use the evidence around the songs to choose a campaign, then use its results to improve the next one.