Making content is now cheap. Clearing it for use is not. The advantage goes to the brand that owns the record.
This is for the CIO, chief data officer and CMO, who now share responsibility for what AI makes.
AI has moved the cost of content from making it to answering for it. In a survey of 434 content professionals by Canto, a digital asset management (DAM) vendor, 75 percent said AI had increased their team’s content volume. One approved original now becomes dozens of versions across markets, channels and tools. In most estates nothing keeps one record of them all: what each version is, what it was made from, where it may run and how it did. So brands pay to make and review again what they already own, and struggle to find every live copy when a claim has to come down.
That makes the record the control point. Each vendor answers well for what happens inside its own tools. Agents, retailers and shopping assistants now act on content across all of them. The brand that owns the record can change any of those tools and still say what is live, where and why, across the tools and contracts it can verify.
I recommend that brands keep the tools that make content and own the record of what they made. This piece sets out what to own and what to rent, the architecture behind it (Figure 12), and a 90-day test that can prove me wrong. It is bounded to one product family, three markets, two tools and one agent, and judged on one number, the cost per cleared request. If your existing suite wins, stop there.
Carmakers learned to track every variant decades ago
A carmaker sells one platform as hundreds of versions. Since the 1981 model year every car sold in the US has carried a 17-character serial number, the VIN. The road safety regulator says it exists “to increase the accuracy and efficiency of vehicle recall campaigns”. Against that number sits a build record of what the car was built to. A recall finds the affected cars from the record, and a proven part can go to the next model, tested for what its new setting changes. Most content estates have neither: no ID every tool honours, and no record of what each variant was made from.
AI is moving the unit of content from the asset to the family
A brand used to manage a library: approved assets, each reviewed once and filed. Generative AI and agents change what is being managed. A master asset now spawns variants by market, then by channel, then by audience, and agents make more of them every quarter. Canto surveyed 434 content professionals in the US and UK in late 2025: AI had increased content volume for 75 percent of their teams.
In this piece, content means anything a brand publishes or approves to publish: images, video, copy and claims, banners, emails and newsletters, product and landing pages. A master is the approved original. A variant is any version made from it, by a person or an agent, for a market, channel or audience. A content family is a master and all its variants. A product family can contain several masters, and therefore several content families. A reference is an approved asset a person or agent may make new content from: a master, an approved product image, or a variant a person has approved for reuse.
Budgets have not moved with the volume. Gartner’s 2026 survey of 401 CMOs found marketing budgets flat at 7.8 percent of revenue, with 15.3 percent of that going to AI, and only 30 percent reporting mature AI readiness. Gartner’s warning is that CMOs are buying AI tools faster than they build the foundations; for content, the missing foundation is a memory of the family across its whole life.
Four everyday moments show where it breaks
The gap is easiest to see in four moments every brand meets. Each one asks a question the estate cannot answer from any single tool, and each costs money in a different place.
1. Launching in a new market. The approved German variants should give Spain a head start. Usually they do not, because nothing records which claims were cleared for which market. Regulated sectors show the cost most clearly. In medtech, Veeva, which sells review software to life sciences firms, found that 64 percent of more than 100 leaders name local re-review as their top barrier to market. RWS, a localisation vendor, surveyed 200 senior content leaders in April 2026: 86 percent said AI had sped up production, and 65 percent said it had slowed localisation. Without a record of what was already approved, AI makes more to check.
2. An agent building a page. An agent assembling a product page at run time needs the approved image, the claims allowed in that market and the rights still valid today, at the moment it asks. Real Story Group, an analyst firm that evaluates DAM and content platforms, puts answering requests from agents as well as people at the top of its 2026 DAM framework. In June 2026 research by WordPress VIP, a CMS vendor, 85 percent of enterprise leaders said publishing AI content without human review erodes brand trust. A record the agent reads before it acts narrows what people need to check. A brand can be held to what its agents say: in February 2024 a Canadian tribunal held Air Canada liable for bereavement-fare advice its chatbot gave against the airline’s own policy.
3. Learning what worked. Creative drives 49 percent of the sales lift from advertising, according to a study of nearly 450 consumer goods campaigns by NCSolutions, now part of the measurement firm Circana. The study dates from 2023; Circana still cites it, and I found no newer measure of the same kind. More variants do not mean better ones. In June 2026, the AI testing model of Kantar, a brand research firm, rated only 6 percent of more than 15,000 creator assets strong on both engagement and brand building. Yet results land against ad IDs, so a team with fifty variants cannot tell which master, recipe (the prompt, model and settings used) or claim produced the ones that sold. The link does not prove cause, but without it no test can be designed.
4. Pulling a claim. This is the stress test. A skincare brand withdraws a “clinically proven” claim on a Monday. Agents have made forty variants from the three masters that carry it, now spread across the DAM, web pages, emails in automated journeys, live ads, an agency’s tools and retailer portals. A recall is only as complete as the least-tracked tool the claim passed through. The pressure is real this year. From 27 September 2026, national rules implementing an EU directive ban generic environmental claims such as “eco-friendly” unless the brand can show recognised excellent environmental performance. Many member states had not passed theirs by then. Since 2 August 2026, the EU AI Act has required providers of covered generative AI systems to mark their output in machine-readable form, with some exceptions. Neither requires a record like this one, but both add evidence and disclosure duties to content that is already hard to trace. I have not found a public case where a recall failed because the trail broke between tools, so I treat this moment as insurance.
Across all four moments, more tools make it worse. In Canto’s survey, teams running two or more DAM systems reported more delayed launches and more wasted budget than teams with one.
Supporting exhibit: Figure 4, why AI slows the checking
Two more patterns run under all four moments: scale and reuse, and rights on AI variants. Real brands meet all six today, and some already solve part of one, but inside one system or for one kind of content. Set side by side, all six need the same thing: one record, enforced by the same rules and measured by the same numbers.
Supporting exhibit: Figure 5, the six patterns with a real case for each
| Problem pattern | A real case, and where it stops | What the record does | What enforces it | The number that shows it |
|---|---|---|---|---|
| Scale and reuse | A global consumer goods company makes every variant, label, pack and language of a product from one digital twin, but only for product imagery, and so far in one business group | Shows what is already approved before anyone makes more | Agents and people ask the record before they act | Cost per cleared request, explained by the reuse rate |
| Launching in a new market | Pharma companies keep approved claims in shared libraries, tagged by brand, market and language, but each library sits inside one platform | Holds each market’s clearance, so the next launch reuses it | Rules are held as data. Machines check first, and people approve by risk | First-time approval, and days from brief to live |
| An agent building a page | An airline was held liable for its chatbot’s advice, which contradicted its own policy (moment 2). The case shows who is accountable, but not that a record would have prevented the error | Tells the agent what it may use, at the moment it asks | Every agent goes through one gateway, and nothing is published without a current clearance receipt | Breaches caught at the gate |
| Rights on AI variants | A global fashion retailer plans digital twins of its models, each made with the model’s consent, so far in one programme | Passes limits to every variant and checks permission by how each was made | Every variant inherits its parent’s limits, and the gate checks how it was made | Breaches caught at the gate |
| Learning what worked | A global food company scores the creative features of each asset against sales. Its public account does not mention links to the master, recipe or claim | Joins results to the master, recipe and claim behind each variant | The record ID travels into each channel, and holdout tests show cause | Share of results joined to record IDs |
| Pulling a claim | The UK competition regulator required three fashion and grocery retailers to report regularly on how they keep their green-claim commitments. The public record does not say how they gather that evidence | Finds every live copy of a claim from one query | Stable IDs mapped to every tool, agencies reporting by contract, and scans of retailer pages | Complete history, with time to pull a claim as a side check |
The missing capability is a shared record of content across tools
Put the six patterns together and one capability is missing. For any piece of content and its variants, the brand needs to know what it is, what it came from, what it may do, where it went and how it did. It needs to know this across every tool, in a form people and agents can read before they act. The record has six parts: identity, context, rights, history, use and results. Identity is the hard one: stable IDs for each product, claim, rights contract and master, a unique ID per variant, the links between them, and a map to each tool’s IDs. Real Story Group’s content warehouse links variants to their source in the same way, with history and rights attached.
Data teams already work this way: lineage tools record which source and job produced each table, and catalogues hold one definition per term. The record applies the same discipline to content, with three differences. Content is copied into channels the brand does not run, its rights and claims expire, and its permissions depend on how each variant was made.
Commerce already does part of this for products: most branded retail products carry a global trade item number (GTIN). Many brands share product data with retailers through the Global Data Synchronisation Network (GDSN) of GS1, the body behind the barcode: certified data pools linked by the GS1 Global Registry. Content about a product should link to its GTINs, or to a product-group ID where one image serves several pack sizes; brand campaigns keep their own master ID. Some claims sit in the product data and others only on packs and in copy, so the record holds each claim once, with its approved wording in each market, and links it to both.
AI adds two questions to context: what a variant may be made from, and what it must match. The record holds the approved references for each use and audience, and the checks that define what good looks like. Video shows why: when each generation is made from the last one, small errors compound. The same holds for product content sent to retailers and ads assembled for each channel, though what counts as good differs: accuracy for a product page, and the standard plus results for an ad.
Storage, search inside one tool and most taxonomy sit mostly outside the record. They are about finding content, and today’s tools handle them, with two exceptions. The shared list of products, claims, markets and rights terms belongs in the record, because every tool must mean the same thing by “this claim in Germany”. Commerce makes this concrete. Marketplaces, social shops and search engines sort products into their own categories. Google assigns every product in Merchant Center a category from its own taxonomy, which it says keeps evolving. The brand does not fully control that category, so the record holds the map from the brand’s list to each channel’s. And search across the whole estate splits in two. “Has this been made before?” is a finding question, and a shared index across the tools can answer it. “Is it cleared for here?” is an answering question. Where content crosses tools, only a record that spans them can answer it, because it alone knows where each piece may run.
OpenArch, which the content platform vendor Hyland released as open source on 10 November 2026, shows the split. It leaves content in its source systems, builds one AI index over it and carries each repository’s permissions into AI search. So it can tell a marketer in Madrid that she may open the Spanish pack shot. It cannot tell her whether that shot may run in a paid ad after the model’s release expires in June 2027, or that legal withdrew a claim on it on Monday. That is the test I use: if a problem is about finding content, a tool solves it; if it is about answering for content, the record has to.
Supporting exhibit: Figure 7, finding content versus answering for it
The business case is the cost of clearing content for use
A business case built on avoided recalls is too thin and too rare to fund anything. Where content repeats across markets and channels, much of the money is in reuse. When AI makes a variant almost free to produce, the cost moves to everything after it: review, rights checks, translation and fixing errors in market. A request met from content already approved for that market skips most of that. One adapted for a new market still needs translation and a local check, which the record shortens by showing what was cleared before. As machine checks get cheaper this saving shrinks, so price reuse on your own review cost per variant. The reuse rate shows first whether the record works, then first-time approval and the days from brief to live (Figure 8). Each can be counted from systems most brands already run, once intake records the asset that closed each request and reviewers sample their time for two weeks. Above them sits one executive number, the cost per cleared request, defined in the 90-day test.
Supporting exhibit: Figure 8, the scorecard
| Stage in Figure 2 | The number | What good looks like | Owner | A bad reading means |
|---|---|---|---|---|
| Reuse, before create | Reuse rate: of asset requests logged in the intake tool this quarter, the share closed with an approved asset that existed before the request. Straight reuse and adaptation are counted separately | Reaches a target band set from the baseline, read with how reused assets perform against new ones | Content operations | Teams cannot find what exists, or do not trust that it is cleared. Too high a rate can mean stale creative |
| Create | Days from brief to live, per request | Falls as reuse and first-time approval rise | Marketing operations | Work is waiting on checks, not on making |
| Approve | First-time approval: the share of variants approved without a second round | Rises once rules are held as data and machines check first | Legal or regulatory | The rules are not in the record, or agents are not reading it |
| Adapt | Cost per live variant, read with total content cost: making, review, rights and translation, divided by variants that went live | Both fall, or total cost holds while live variants rise | Marketing operations, with finance | Agents are making cheap variants nobody uses |
| Distribute | Breaches caught at the gate: rights, market and claim limits stopped before publishing, against those found in a quarterly audit of a sample of live content | Nearly all caught at the gate | Brand and legal | Content is reaching channels around the checkpoints |
| Measure | Results joined: the share of live variants whose results come back to their record ID | Most live spend first, then all of it | Media and analytics | The record ID is not travelling into the channels |
| Throughout | Complete history: of variants found live in a scan of the channels, the share with master, maker, date and market recorded | Every new variant from day one; old ones as they are matched | Content operations | A tool or partner is making content outside the record |
Better history does not by itself improve marketing results; only an independent test can show that. So any gain from learning stays at zero in the case until a holdout test, which holds some variants back to compare, proves it. The oldest waste figure most people quote comes from CreativeX, a creative analytics firm: its finding that 52 percent of brands’ core assets were never activated was published in February 2024 from 2023 data. Unused content may also mean it did not fit local needs, which argues for better briefs as much as for reuse, so each brand should measure its own reuse rate.
Every vendor covers part of the lifecycle inside its own walls
If reuse is where the money is, the right test for vendors is whether the tools a brand already has can answer for content across all its tools, at AI volume.
Each camp moved in 2026. Aprimo, a Leader in the Forrester Wave for DAM in Q1 2026, launched agents that validate claims and enforce rules as content is used. Adobe now marks output from its supported generative tools with Content Credentials by default and keeps each source’s history when assets are combined. Databricks put images and video under its governance controls, and Snowflake gave business terms one approved definition and an owner. Veeva links pharma claims to materials, and FADEL, a rights management vendor, finds expired rights in live channels.
On my reading of what vendors publish, only one vendor is built for any of the four moments: Aprimo, for the agent building a page. Everything else is partial or left to another system, and on one identity across tools every camp is partial (Figure 9).
Supporting exhibit: Figure 9, the vendor grid
| Moment | DAM Aprimo | Suite Adobe | Structured content Sanity, Hygraph | Data platform Databricks, Snowflake | Record keepers Veeva, FADEL, product data |
|---|---|---|---|---|---|
| 1 Launching in a new market | Partial | Partial | Partial | Not its job | Partial |
| 2 An agent building a page | Core | Partial | Partial | Partial | Partial |
| 3 Learning what worked | Partial | Partial | Not its job | Partial | Not its job |
| 4 Pulling a claim | Partial | Partial | Partial | Partial | Partial |
| One identity across tools | Partial | Partial | Partial | Partial | Partial |
Corebuilt for it, in the vendor’s own material Partialshipped, but inside its own walls or for one job Not its jobleft to another system. Roadmap items are not counted.
Own, extend or buy: the line is where content crosses tools
Tools are getting better at working together, and Real Story Group’s advice is to start with the platform that already governs most of the content. I agree. Extending what you own is the option to beat.
| A. Extend the DAM or suite | B. Shared record across your tools | C. Shared record plus lakehouse | |
|---|---|---|---|
| What it is | Turn on the suite’s agents, history and Content Credentials, and connect the other tools to it | The record and history ledger, a gateway with rules service and internal content marketplace, and a catalogue, over the tools you have | Option B, plus open tables that join each variant’s history to its results |
| Across the four moments | Strong for content made inside the suite. Agencies, retailers and other tools still answer by email | Answers moments 1, 2 and 4 across tools, as completely as the tools report | Adds moment 3 at volume: which variants worked, by master, recipe and claim |
| Cost to build and run | Low to medium, for added suite licences and set-up | Medium to high. It grows with the tools to connect and the old assets to match, so price it per tool feed and per thousand assets matched | High, for data engineering, storage and analytics |
| Where it breaks | Content made outside the suite, and evidence held in one vendor’s format | Tools that report nothing, teams that define terms differently, and no agreed owner to run it | Proving value. A link shows what ran with what, but cause needs a test |
| Choose it when | One suite makes most content and holds the claims, and regulation is light | Several systems of record meet regulated claims, real people’s likeness or many markets | Variant volume is high and a holdout test shows learning changes what gets made |
Low Medium to high High relative cost to build and run.
In my view, option A done well goes further than it looks. If agents can only draw claims from the approved library inside the suite, every variant carries its claim from the start, and moments 1, 2 and 4 are largely covered. That holds until an agency’s tools, a retailer’s portal or a second DAM make or change content. History moves between rival vendors’ tools most reliably through open standards. So I would build the record on the W3C provenance standard and have it read Content Credentials, which use the Coalition for Content Provenance and Authenticity (C2PA) open standard for how a file was made. The line sits where content regularly crosses tools the suite does not run. Above that line, add a shared record; below it, buy the suite and stop. In pharma, where Veeva already links claims to materials, I would join the record to Veeva rather than rebuild what it links.
Supporting exhibit: Figure 11, when to buy the suite and stop
A lakehouse is one store where files and data tables sit together and can be queried as one, and it earns its place only in option C. Large DAMs already keep versions, IDs and similarity search for what they hold. The lake has to hold what the DAM cannot reach. That means metadata and fingerprints for variants made in every tool, a copy of the history ledger keyed to record IDs, and results by channel, in open tables any tool can query. It holds pointers, not second copies of files, so a withdrawn claim gains no new home. Estate-wide search belongs to the catalogue in option B; the lake adds analysis at volume. If a proposed content lake cannot say what it holds that the DAM does not, it is a rename. Moments 1, 2 and 4 need the record. Moment 3, joining the history of thousands of variants to their results a quarter at a time and testing what caused a lift, is analytics on open tables, which is what a lakehouse is for. I would add it only if a holdout test shows that learning which variants worked pays for itself.
Own the record agents read before they act
Options B and C share one design, shown in Figure 12. The enterprise owns the record and its history, access through the gateway and an internal content marketplace, the catalogue, governance and observability. Owning means the brand controls the definitions, decisions and evidence, and can move them; it does not mean building or hosting every part. The trade-off is a new run cost that grows with each tool feed and each thousand assets matched, so I would take it on only where the suite cannot answer. Vendors can supply the source tools, orchestration and processing; the people, agents and channels that use content are a mix of both. Source systems remain authoritative for the facts they own. The record carries shared IDs and policy-ready facts across tools, while the catalogue is the searchable index over them. It is the carmaker’s split: it owns the part numbers and build records, while suppliers make the parts. Workflow and agent orchestration tools move content from one tool to the next. They ask the gateway before any step that makes, approves or publishes content, and write back through it after.
The owned parts serve the six patterns. The record answers reuse and pulling a claim. The history ledger answers learning what worked and pulling a claim, and tells the rules service how each variant was made. The rules service answers the agent building a page and rights on AI variants, and the gateway enforces its answer. New markets need the record and the rules: clearance facts held in the record, checked by the rules service. A chief data officer will know the shape: content as found in each tool, resolved to one master ID, then made policy-ready, like the bronze, silver and gold layers of a data platform (Figure 12). The difference is that gold here is ready for a decision, not a decision. The rules service decides each time content is used: for this requester, use, market and date, and how the variant was made. The gateway enforces the decision, and the ledger records each governed change.
Approved content is offered back through an internal content marketplace in the access layer, and the gateway checks each use. Each set has an owner and a freshness date. Markets, agencies and agents request what they need, and use is counted to show demand and value.
I would run the record on five working rules.
Agents ask the record, through the gateway, before they act. An agent making a Spanish variant first asks which masters and references are approved for Spain, which claims are allowed there and which rights are still valid. It makes from those approved references, not from earlier variants, then writes back what it made and which references it used. That stops the record becoming an archive nobody reads.
Supporting exhibit: Figure 13, the record inside the work
When a partner does not report a change, the record shows that part as unverified. When an approval or a right expires, the rules service stops issuing receipts and the gateway refuses new uses. When a source system is down, the gateway treats the answer as unknown and stops the use. A channel that bypasses the gateway is outside preventive control and is found only by the live-channel scan.
Stable IDs, mapped to every tool’s own ID. Nobody rewrites the DAM. Old assets are matched on file fingerprints, image similarity and product codes, with a person checking doubtful matches. Matching faces to identify a model counts as processing biometric data, a special category under GDPR Article 9, so match on release IDs where you can.
Machines check first; people approve by risk. Once claims, rights and approved references are held as data, a machine checks every variant against them before a person sees it. What happens next depends on risk, and legal sets the tier in the record; the agent cannot change it. Routine content, such as a resized banner from an approved master, passes the automatic gate and is sampled. Medium-risk content goes to a person when the check flags it. Legally sensitive content, meaning regulated claims, consent-sensitive images and high-consequence statements, needs a person’s approval every time. A variant becomes a reference for later work only after a person approves it, so drift cannot compound. The WordPress VIP finding that 85 percent of leaders say unreviewed AI content erodes brand trust argues for keeping people in the loop; it does not show that reviewing only exceptions is enough. The gain is that reviewers spend their time where the risk is.
Limits pass down; permissions do not. A variant inherits every limit on its parent: expiry, market, claim, and whether the parent may be fed to a model. Real Story Group makes the same point: constraints must travel with the content. A variant made from several sources inherits the limits of all of them, and where they conflict, the strictest applies. Whether it inherits permission depends on how it was made. A crop sits inside the original release; a new AI image of the same model may not. That is the question in a model’s New York suit against the retailer Rainbow, which disputes the claims. The case is unresolved. People’s data, such as releases, sits in a separate store with retention limits, and the design needs a privacy review.
Supporting exhibit: Figure 14, how limits and permissions pass down
Results come back to the same IDs. Every published variant carries its record ID into the channel, so results return against the variant, its master, how it was made and its claim. Real Story Group’s DAM framework also feeds results back into the content system. Over time this shows which references produce variants that work, so what good looks like is measured, not only declared. Results choose which references to keep; they do not replace the check against them. Proving cause still takes holdout tests.
Rules on paper do nothing. Checkpoints on the paths the brand runs make them real. Every path the brand runs goes through the gateway, whether an agent, a person, a workflow or a batch job uses it. At making, approval and publishing, the rules service issues a clearance receipt for that one use: this content, this requester, this use, this market and date, and how it was made. The gateway checks it. Parent and record IDs show where content came from; they do not give permission. Contracts require agencies to report what they make and run. Retailers set their own terms, so their part is checked by scanning what they display. Where an agency will not report, its part is scanned and labelled too, and that gap limits option B as much as option A. These checkpoints are a proposed design, and the record’s guarantee ends where they end. When a partner stays silent, a right expires or a source system is down, the record flags the gap or the gateway stops the use (Figure 13).
Channels are now asked to answer for content too. Reuters reported in September 2026 that two US senators had asked the Federal Trade Commission to investigate whether Amazon’s and Walmart’s shopping assistants suppress “Made in USA” products and fail to flag falsely labelled goods. Amazon called the suggestion that it intentionally withholds country-of-origin information “fundamentally wrong”. If a retailer’s assistant can leave out or misstate what a product is, the brand’s scan has to read what each channel shows and says, as well as what the brand sent.
The record cannot promise completeness beyond the integrations and contracts it can verify. I would split ownership this way: the CMO sponsors the work, the CIO owns the gateway and integrations, legal owns the rules, and content operations runs the record with data engineers from the CDO’s team.
The record also has to be secured, because a store that every agent reads, and that lists every claim and right, is worth attacking. Access stays no wider than the source systems allow, each agent gets short-lived credentials for one task, and history is tamper-evident.
Supporting exhibit: the security model
Access is never wider than the source systems allow, and a variant made from several sources is shown only to people who may see all of them. Each agent gets short-lived credentials at the gateway, scoped to one task and market and traced to the person who authorised it. History is append-only and tamper-evident, for example through signed and chained entries or an existing audit service, with keys held outside the content team. Personal data sits in a separate store the history points to. The design must still allow lawful erasure without leaving identifying details in the history it keeps.
Decide in 90 days with the smallest test that can fail
This is an investment test, and it asks one question. Does owning a shared record make it meaningfully cheaper and faster to get content cleared for use than the best the existing suite can do, without losing effectiveness or control? Technical feasibility and operating results are the evidence. The savings land on different budgets and nobody owns the record today, so the case needs numbers a CFO can check. I would agree three things with finance before work starts: a budget cap, a time window and the margin that counts as a win.
I would take one product family in three markets, two tools that do not share IDs and one agent making variants. Run three scenarios, once with the best-configured suite and once with the record. A new market asks for content, and the team meets it from variants already cleared. The agent makes an email from allowed claims and approved references. Harmless test markers on internal pages are withdrawn, to time how fast and how completely every live copy is found.
The number is the cost per cleared request, with the completion rate reported beside it so it cannot improve by dropping hard work. Clearance has to stay complete, and reused content has to perform at least as well as new. If the record beats the suite by the agreed margin and the guardrails hold, I would extend it to the next product family. If it wins by less, use the suite. If neither repays its cost within the budget cap, stop. A successful pilot authorises only the next bounded slice, not the whole of Figure 12.
Supporting exhibit: the full test protocol
What to measure. The primary number is the cost per cleared request. Finance and content operations define the request unit before the pilot, including its scope, completion condition and treatment of multi-asset or multi-market work. Count the hours spent reviewing, checking rights, reconciling content and fixing errors in market, at a loaded rate, and add agency fees for that work. Licences, integration and upkeep are not in this number; they are counted once, in the decision rule. Divide by the requests fulfilled and cleared for their intended use. Scenarios 1 and 2 produce it; scenario 3 tests the completeness guardrail. It counts a request met by reuse as well as one met by a new variant, which a cost per variant would miss. Translation and production costs are tracked separately. Failed, unresolved and abandoned requests are recorded on their own, and the completion rate and open backlog are reported beside the number, by content type and risk, so it cannot improve by dropping hard work. Three numbers explain it: the reuse rate, first-time approval and the days from brief to live. Two guardrails sit over all of them. Clearance must stay complete: no in-scope live copy that an independent check can find is missed when a claim is pulled, and any channel the pilot cannot see is reported as a gap. Reused content must also perform at least as well as new. Ninety days gives early evidence on that, not proof, so the effectiveness check continues after the decision.
The smallest scenario that tests it. Take one product family, so a handful of masters and their variants, in three markets, so clearance and reuse matter. Use two tools that do not share IDs, such as the DAM and the email platform, because one tool alone proves nothing. Add one agent making variants, so the record is read before anything is made. The record itself stays thin. It holds master IDs mapped across the two tools, history written as each variant is made, market policy facts and approved references for the claims in scope, and one gateway check at use. Clearance is recomputed for that use. Legal signs off those claims before day one.
Three scenarios, compared fairly. In the first, the third market asks for content and the team tries to meet it from variants already cleared. In the second, the agent makes an email using only allowed claims and approved references, and reviewers count first-time approvals. In the third, several harmless test markers, published only to internal seed lists and test pages, are withdrawn one at a time. The team times how long it takes to find every live copy in both tools, and how complete the list is. Each scenario is repeated with enough matched requests to read the agreed margin. It is run with the existing suite configured as well as it can be (option A) and with the record (option B); last quarter’s logged requests give the baseline. Use requests of the same type and risk rather than the same requests twice, alternate which approach goes first, and log any earlier clearance either side reuses, so one cannot borrow the other’s work. The final quality check of the finished content is blind, so its reviewers do not know which approach produced each piece; approval work and costs stay visible and are measured. Finance sets the margin; an analyst checks that the matched requests are enough to read it.
The decision rule. Project the saving per cleared request to the brand’s real volume. Set it against each option’s added platform cost at the brand’s real number of tools. For the record, that is implementation, integration, licences, people and upkeep, priced per tool feed and per thousand assets matched. For the suite, it is any added licences and set-up. If the record beats the suite by the agreed margin and the guardrails show no early harm, extend it to the next product family. If the record beats the suite by less than the margin, use the suite. If neither beats last quarter’s baseline by enough to repay its cost within the budget cap, stop. Ninety days, including a few weeks of set-up, is enough for an initial investment decision on measured pilot evidence, provided source access and the in-scope claims are agreed before day one. One product family may give a directional result rather than proof, so the next family tests the savings. A successful pilot authorises only the next bounded slice; it does not validate Figure 12 as a whole. The test covers the brand’s own tools; retailers, marketplaces and shopping agents need a test of their own. Learning which variants worked, and the lakehouse behind it, stay out of the first 90 days, because they need a quarter of results and a holdout test.
Two builds fail this test however good the demo. A store of assets with search and a chat front end finds files but cannot say what each may do. A roadmap with its own approval screen builds one more tool, and the record is still missing. At the next renewal, I would put three questions to every DAM, suite and agency. Will you report every change, will you accept our master IDs, and will you carry our record ID into the channels so results come back.
What comes next: shopping agents and content made for one person
The 90-day test settles today’s case. Shopping agents will raise the stakes. Amazon says more than 400 million customers used its shopping assistant, Alexa for Shopping, in the year to October 2026. On 19 March 2026, Google added an optional catalogue capability to its Universal Commerce Protocol, an open standard for agent shopping. Where a retailer turns it on, agents can read its variants, stock and prices as they shop. When an agent reads a product page and buys on it, a wrong claim or spec in a product feed can be read as a promise made in the brand’s name. That is the risk the Air Canada case points to, extended to every listing.
Content made for one person will raise them again. When every email, page and ad is assembled for a single viewer, the family of variants has no practical limit, and no one can review each output. The limits still travel with the content. The record then has to hold the recipe as well as the output: the master, the prompt, the model, the rules and the claims that produced each piece. Machines check every output against the rules, and people check a sample. The record still logs each output, or its fingerprint, because the same recipe will not make the same output twice. Where rules require a person to approve each piece, as industry codes do for much pharma promotion, sampling is not enough. A DAM stores outputs. The record has to hold recipes too.
Each part of the enterprise needs the record for its own reason. Commerce needs product truth that shopping agents can read, media needs the record ID in the ad server, and regulated teams need history that stands as evidence (Figure 16).
Supporting exhibit: Figure 16, the record through each part of the enterprise
| Lens | The problem coming | What the record must add |
|---|---|---|
| Commerce | Shopping agents read and act on product content across the brand’s site, retailers and marketplaces | Approved product truth per market, keyed to the GTIN, mapped to each channel’s categories and tied to every listing, with approved product images and claims as the references for any AI variant |
| Media | Ads assembled at the moment they are served, with no person reviewing each one | Approved references and recipes, assembled ads checked by sample, and the record ID carried into the ad server so results return by variant |
| CRM | Emails and app messages made for one person | Recipes and rules, with outputs checked by sample |
| Service and AI search | Chatbots and AI search engines answering from the brand’s content | One approved source of product and policy truth they can read |
| Regulated sectors | AI-made materials under disclosure rules | History good enough to stand as evidence |
Carmakers did not handle variants by building one factory that holds every part. They gave every car a number and a build record, which let them recall a fault from the record and reuse a proven part on the next model. No law sets a cross-industry content identifier or lineage record comparable to the VIN, though some regulated sectors have their own record-keeping duties. Outside them, the brand’s contracts have to do it.
I would start with the 90-day test on one product family, and put the three renewal questions to every vendor now.
Keep the tools that make content. Own the record of what they made.
What this rests on
The evidence comes from vendor surveys by Canto, RWS, Veeva and WordPress VIP. Each shows correlation inside its own sample, and the Veeva figure is from a regulated sector. Gartner carries the budget point. NCSolutions’ 2023 campaign study, still cited by Circana, and the 2026 Kantar rating of creator content carry the learning point.
The rest is my reasoning, not a measurement: that the four moments share one missing capability, that the enterprise should own the record and keep the rest swappable, the scorecard and Figure 12’s architecture. I found no public measure of what a shared record returns across tools; the 90-day test is how a brand finds out.
The argument would be wrong if brands whose content crosses many tools moved reuse and review numbers by extending one suite alone. It would also be wrong if those numbers did not move once a shared record was in place.
Leave a Reply