A regional marketing lead asks a simple question: “Which product video improved consideration in Germany after we changed the opening shot?”
The agent can reach the ad account. It can retrieve performance in seconds. Yet it still may not know whether DE_final_03 is the 9:16 creator cut, the subtitled studio version, or a later export with a different offer. It may not know the attribution window, whether the creator licence is still active, or which hypothesis the edit was meant to test.
That is the new procurement problem. Access to platform data is becoming easier; decision-grade creative context is not.
The monthly deck still has a job. It aligns stakeholders, records a point-in-time conclusion and gives leadership a narrative. It should no longer be the only operational deliverable for an always-on, multi-market creative programme. Brands need a second layer: a permissioned, query-ready memory that connects platform metrics to creative identity, definitions, rights and approved decisions.
What changed in 2026—and what did not
On 3 August 2026, Snap launched its official Snap Ads MCP Server. It allows approved AI clients to query campaign performance, diagnostics and planning information in natural language. Snap's developer documentation is unusually clear about the control model: an organisation administrator approves the agent, individual members authorise access, permissions follow the user's account rights, and access can be revoked. At launch, the server is read-only; it cannot create, update or delete ads.
Pinterest announced Pinterest MCP on 17 June 2026, giving partners access to campaign, analytics, audience and keyword insights. Its developer materials describe OAuth and granular permissions rather than an unbounded connection.
On 22 September 2026, IAB Tech Lab introduced AAMP 3.0 with OpenProposal. The proposal is designed to make parts of the RFP-to-buy process agent-readable, connecting planning, proposals, transactions, execution and reporting. The public-comment period runs through 22 October 2026.
These developments show that agent-readable advertising workflows are moving from demos into real platform and standards work. They do not prove universal adoption. OpenProposal is still a proposal under review, and each platform exposes different capabilities and controls.
Queryable does not mean trustworthy
An agent can identify the highest-CTR video and still give the wrong business answer.
- Two exports may share a name but differ by market, subtitle, aspect ratio or offer.
- “Conversion” may use different windows or event definitions across platforms.
- A platform creative ID may not map back to the production master.
- The hook, voiceover, price and landing page may all have changed at once.
- A creator licence may have expired even though historical performance remains visible.
- Attribution can describe association without proving incrementality.
The valuable layer is therefore not a conversational dashboard by itself. It is a managed link between metrics, source records, creative lineage and human accountability.
For teams that have not yet stabilised creative identity, our earlier article on why every video needs a persistent creative ID is the right starting point. An ID is necessary, but it does not explain what the asset contains, what changed, who approved it or how it should be queried.
Deck, export or query layer?
| Deliverable | Best use | Main limitation |
|---|---|---|
| Monthly deck | Leadership alignment, narrative and signed-off conclusions | Every follow-up question restarts the analyst workflow |
| CSV or raw export | Audit, backup and ad hoc analysis | Definitions, lineage and permissions separate easily from the file |
| API/MCP query layer | Repeated questions, cross-account diagnosis and feeding learning into briefs | Higher governance and maintenance requirements |
This is not a replacement hierarchy. A mature handover may contain all three: a human-readable conclusion, a portable point-in-time snapshot and a controlled interface for recurring questions.
The ONCE Query-Ready Creative Memory
When buying an always-on creative and media service, require six connected deliverables.
1. Metric dictionary
For every decision metric, record the platform, definition, timezone, currency, attribution window, deduplication rule and refresh timestamp. A field called “revenue” is not a definition.
2. Creative identity and lineage
Connect the production master, edit, subtitle variant, market version, platform creative ID and ad-group ID. Every revision should say what changed—not simply append another final_v2 filename.
3. Approved question templates
Define the recurring questions the agent may answer, such as “compare first-three-second retention for the same claim in the UK and US” or “find versions that repeatedly reached the same audience in the last 14 days.” Each template should specify required fields, valid comparisons and forbidden conclusions.
4. Permission ladder
Separate four levels: read and query; recommend; prepare a draft change; write, publish or spend. Start at read-only. The right to retrieve a metric must not silently become the right to alter a campaign or move budget.
5. Refresh cadence and owner
Name the refresh interval, connection owner, approver for definition changes and person responsible for anomalous answers. An MCP endpoint without an operating owner is merely a faster way to become stale.
6. Exit snapshot and revocation
At the end of the relationship, the brand should receive the field dictionary, creative map, latest raw snapshot, query templates and permission register. Vendor and agent access should be revocable. Portability is a procurement condition, not an offboarding favour.
Begin with read-only, not “full automation”
A practical rollout is deliberately narrow:
- Connect one platform, one market and one well-named creative set.
- Test ten real operating questions and require every answer to expose its source fields and creative version.
- Have media, production and brand owners review the results together.
- Only then allow recommendation or draft generation.
- Authorise any action that changes delivery, publishes content or spends money separately, with human approval.
This is not anti-automation. It distinguishes reversible investigation from consequential action. Snap's read-only launch provides one clear implementation example. Other platforms may offer broader tools, but brands should still impose their own action-level controls.
Compare suppliers by cost per reusable decision
Interface fees alone do not describe the work. The economic unit should be:
Cost per reusable decision = (data preparation + field mapping + permission management + human review + ongoing maintenance) ÷ the number of approved decision questions answered without rebuilding the analysis
The denominator is not the number of agent responses. A response counts only when its definitions are explicit, its creative version is traceable, a responsible owner accepts it and it can change the next brief or media decision.
For a one-off campaign with few assets on one platform, a deck and export may be sufficient. For an always-on programme spanning markets, platforms and frequent creative iterations, a governed query layer has a better chance of reducing repeated analysis. If creative IDs and metric definitions are unstable, fix the data first. Connecting an agent will magnify ambiguity, not remove it.
A clause for the next procurement brief
Replace “provide an AI dashboard” with something testable:
The supplier will provide a revocable, read-only-by-default query interface. Every conclusion must be traceable to a metric definition, platform source, date range, attribution window and specific creative version. Write and spend actions require separate approval. On exit, the supplier will deliver a complete snapshot and access-revocation record.
Return to the original question. The goal is not to make an agent guess faster which video “won.” The goal is to let every answer travel back through the same chain: the original metric, the exact edit and the approval record that made it eligible for use.
The monthly deck explains the past. Query-ready creative memory prevents the next decision from starting with another spreadsheet reconstruction.
Sources and scope
The current platform facts in this article were checked against Snap for Business, “The Snap Ads MCP Server Is Now Live” (3 August 2026); Snap for Developers, “Ads MCP”; Pinterest Newsroom's Cannes 2026 announcement (17 June 2026) and Pinterest Developers, “About Pinterest MCP”; and IAB Tech Lab, “IAB Tech Lab Introduces AAMP 3.0 with OpenProposal” (22 September 2026). Platform sources establish only the scope of their own products or proposals. The six-part Query-Ready Creative Memory, permission sequence and cost model are ONCE analysis, not platform endorsements or forecasts of guaranteed product roadmaps.
