Experts who describe, prove and explain what an organisation holds are too few for the work in front of them. Lantern shows that AI can take on the drafting and the research while a named person keeps the final say.
Author: Warren Kahn, Senior Solution Engineer, Microsoft
Version: 3.5.10, 12 November 2026
Written for: Architects and technical leaders at museums, libraries and archives, auction houses, banks, land registries and manufacturers with product passports, and the partners who build for them
Reading time: About 20 minutes
Notice. The museum in this white paper is the Musée National des Arts Réunis (MNAR), a fictional French national museum from a case-study scenario [1]. The demonstration collection is real: 501 records from the French national catalogue, Joconde, under the Licence Ouverte 2.0 [12]. I built and ran the system on Azure in September and October 2026. Every measured result in this white paper comes from that build, and so does every lesson in section 9. Measured results carry their date, and every other figure is labelled as an assumption or an estimate.
Executive summary
Many organisations rely on experts who must describe what they hold, prove where it came from and explain it to the people they serve. Curators do it for a collection, specialists for the lots in a sale, compliance officers for a client’s wealth, and registrars for a property’s title. The work is slow, and it rests on a few people. Generative AI can help with the drafting and the research. The risk is an invented fact published in the organisation’s name, or a decision that breaches someone’s rights and can’t be undone.
The museum I built Lantern for shows the scale: only 8% of its 600,000 objects are digitised with rich metadata, the cataloguing backlog is about a decade long, provenance research is largely manual, and visitors ask questions in a dozen languages [1].
This white paper sets out Lantern, a design on Azure that I built around that risk and measured in the museum. Any organisation whose experts do this work can reuse it. It has four parts, named by what they do:
- An assistant drafts the description, with a source for every fact. In the museum, the cataloguing assistant drafts the catalogue record.
- An assistant builds the item’s history from evidence, citing every link. In the museum, the provenance assistant builds the chain of owners.
- A named person signs before anything goes out. In the museum, the curator signs the record and the rights officer the rights decision.
- A guide explains in the reader’s language, by voice or text, opened from a code with no app. In the museum, this is the visitor guide.
I designed Lantern around three core principles: firstly, complete accuracy, with zero hallucinations; secondly, speed; and finally, ease of use, as the industries that would benefit most are often understaffed or have smaller technical teams. The assistants only make proposals. Four things carry over to any organisation: separate doors for staff and the public, one record of every decision that can’t be rewritten, templates that deploy in any Azure region, and the AI notice and machine-readable mark the EU AI Act asks for [4].
What I found while building Lantern for the museum:
- Cataloguing: an estimated 12.5 times less person-time per record (target: five times), from the measured review of the assistant’s drafts and stated times by hand. The pilot I propose measures it against the museum’s own figures.
- Provenance: an estimated 87% less research time (target: 60%). On five test objects with known histories, the assistant found 23 of the 25 known links in their chains of owners and invented none.
- Licensing: about €102,000 to €164,000 a year in gross revenue across the collection from the works that are licensable today (estimate). In the demonstration, a paid licence went from request to signed invoice.
- Running cost: €1,846 a month for the pilot, with the guide for the museum’s staff only, and about €3,100 a month for the production design (estimates at Azure’s list prices).
I recommend starting with cataloguing and rights together, in a pilot of 20,000 objects with the museum’s curators, and measuring every result against the museum’s figures before scaling up. For this type of solution, as with many others, a strong data set is critical for operability, and little else can be done before the objects are digitised. Once this is in place, more routes open and allow growth to occur naturally in the most pressing directions; for example, if the visitor experience is the most critical path, then that can be easily followed.
1. The problem, and why now
Museums, archives, libraries and heritage agencies share one problem: big holdings, small teams, and a legal duty to describe their collections and to protect them. Expectations are rising. In 2021 the Louvre put more than 482,000 entries online [11]. Since 2 August 2026 the AI Act has required that visitors are told they’re dealing with an AI, and that generated text is marked in a machine-readable way, with a transition to 2 December 2026 for systems already on the market [4]. French law now also asks museums to establish and make known the provenance of their works [3]. So I designed Lantern against invented facts and rights breaches first, for speed second, and for ease of use third. The case study set four outcomes for the museum: five times the cataloguing throughput, 60% less provenance research time, higher visitor satisfaction through guidance in many languages, and new licensing revenue [1]. I set every other target in this white paper myself.
2. Six design commitments, and the law behind them
Write these down before you pick any service, because each one rules out a design.
| # | Commitment | Law or policy (France) | Rules out |
|---|---|---|---|
| 1 | Curators keep authority over what the museum publishes | Law [2] | An assistant that writes to the catalogue or publishes |
| 2 | Rights are decided per asset, with text and image apart | Law [5, 6, 7] | One rights flag per object |
| 3 | Visitors aren’t identified, and are told they’re talking to an AI | Law [4]; policy under GDPR [8] | A conversation log tied to a person |
| 4 | Residency is stated exactly | Policy | Saying “everything stays in the country” when a global service is in the path |
| 5 | One auditable record of who proposed and who decided, when, and on what evidence | Policy, serving the provenance duty [3] | Records edited in place |
| 6 | The visitor channel is accessible | Law [9] | A guide that only works for some visitors |
In your jurisdiction, mark each one as law or policy. I’d enforce the legal ones through structure, such as permissions and separate services, and put policy into settings the institution can change.
3. How Lantern is built on Azure
Figure 1. Lantern’s production design for the museum, with its steps numbered.
No workflow engine or orchestrating agent runs the assistants. When a record is added or corrected, the assistant host, a single container app, picks it up within about ten seconds and runs the assistants one after another, each writing its own entry. The record itself coordinates the work.
Three of the assistants run as agents in Foundry Agent Service: the cataloguing, provenance and exhibition assistants. Each decides its own steps within its tools. The cataloguing assistant chooses which sources to look up and flags where they conflict, and the provenance assistant decides which leads in the archive text to follow. What they can write is fixed: proposals only.
Table 1. The main choices, and what I set aside.
| Component | Azure service | Why | Set aside |
|---|---|---|---|
| The record | Cosmos DB for NoSQL and its change feed [14] | Uneven payloads fit as they are; views rebuild without a broker | SQL with an outbox; Event Hubs plus a store [13] |
| Search | Two Azure AI Search services [15] | The line between public and research text is a legal one, so a permission enforces it | One index with a rights filter [25] |
| Assistants | One host on Container Apps; Foundry Agent Service for the cataloguing, provenance and exhibition assistants | The record coordinates the work; agent runs show in Foundry | One orchestrating agent |
| Model access | API Management [16]; Foundry’s AI Gateway in production [24] | One cap and one log for every caller | Each app calling models directly |
| Publication | Durable Functions [18] | A sequence that can undo itself | A transaction (none spans a database and an index) |
| Entrance | Front Door Premium for visitors; the museum’s network for staff | Managed attack rules, and staff traffic stays inside the EU Data Boundary [20] | Front Door for staff too |
The templates deploy to any Azure region (France Central by default), with a recovery copy in a second EU region. Models on Data Zone Standard deployments may process prompts anywhere in the EU, while data at rest stays in the chosen geography [21]. In my build, cataloguing and the guide ran on gpt-5.4. The production cost estimate puts the guide on the smaller gpt-5.4-mini.
What I measured it on. Everything I measured ran on a demonstration build of this design. The production templates add what a live service needs, such as staff sign-in with multi-factor authentication, private endpoints throughout and a write-once archive in a second region. They also put Foundry’s AI Gateway in front of the models, with one project for visitors and one for staff, each with its own token limit. The pilot I propose runs on these templates, with some availability extras scaled down.
4. Five design choices, and why I made them
4.1 Assistants propose, curators decide
If an assistant could write to the catalogue, every one of its errors would be published. In Lantern nothing reaches the public until a curator signs (Figure 2), and the system enforces that. The assistant host only lets an assistant write proposals, and an append-only role on the record blocks edits and deletions. Every proposed fact cites its source. Text, photographs and archive notes are screened for planted instructions when a record is saved, and in one live test (19 October 2026) a planted instruction was stopped. A checker reads every draft against its sources first, and each assistant works at a fixed level that its permissions enforce (Figure 3).
Figure 2. From proposal to publication. Every step is recorded with who took it and when.
Figure 3. Three fixed levels of independence.
I rejected the usual multi-agent design, a supervisor agent that calls the others and writes the result. It would need broad permissions, its steps would be hard to replay, and it would take decisions without a person. Letting the record coordinate keeps each assistant to one job and one permission.
The evidence. On a fixed test set of 20 records (21 October 2026), every proposed fact cited a source, all five known date errors were flagged, and 86.0% of compared fields matched the reference unchanged. A model judge reading 275 live fields (21 October) found 3 values that none of the sources hold, down from 10 before I tightened the cataloguing assistant. None could reach the public: the checker reads every draft against its sources, and nothing is published until a curator signs.
4.2 Keep research text away from the public guide
French law lets a museum keep in-copyright text for research only [5, 7], so it must never reach the index the public guide answers from. I use two separate search services. The guide’s identity has no role at all on the research one, and only the publishing sequence (section 4.5) can fill the public index.
I rejected the usual answer, a single index with a security filter. Filters are a documented way to control access [25], but one wrong filter would leak research text, and here the line is a legal one. Two services cost more than one, and I accept that for a line no query can cross.
The evidence. A boundary check on 13 and 14 October 2026 matched all 337 public documents to a publication on the record.
4.3 Use rules for legal decisions and models for language
A rights decision has to give the same answer tomorrow, and a rights officer has to be able to explain it to a rights holder in one sentence. That’s why the rights assistant is a decision table. It covers the term after the author’s death, anonymous works, works on deposit from other institutions, photographs of identifiable people [6] and out-of-commerce works [7], with separate licences for the text and the image (Appendix C shows five of the rules). The rights officer confirms or overrides each decision.
I use the same split in every assistant. A model reads and suggests, but rules make the decisions and a person signs them off. The checker’s model only advises. The exhibition assistant’s model groups works by meaning, but rules decide whether a work can travel. The provenance assistant’s model reads archive passages and proposes each link, and a curator decides.
I rejected the obvious approach, asking a model for the rights status with the record as context. The same record could get two different answers, and a rights holder is owed one answer with the rule behind it.
The evidence. In the live demonstration, the rules leave 66 of 501 works (13.17%) for a person’s judgement (22 October 2026), against a target of at most 15%. I first set it at 2%, from a run that a misreading had distorted (section 9). By 26 October 2026, three licence requests had gone through the demonstration end to end: the rules issued an open-access licence at once, a rights officer signed and invoiced a paid one (€490 before VAT), and one request was declined.
4.4 Check in one language, then translate
Each visitor question passes Prompt Shields first [19]. The guide then drafts its answer in English from the public index and checks that it’s grounded in the records it cites. Only then does it translate the answer, with the AI notice, a machine-readable AI mark and a source line, so one check scores all twelve languages the same way. If it can’t ground an answer, it says so and points the visitor to a person.
Visitors open the guide from one code at each gallery’s entrance, with no app and no new labels. They can type a question, or ask by voice in any of the twelve languages and hear the answer read aloud. Spoken words go through the same checks as typed ones. Lantern keeps no audio, and Microsoft doesn’t retain or store the audio its speech recognition processes [27]. In a check of all twelve languages, two rounds each (27 October 2026), 23 of 24 spoken questions were recognised, the language was detected 22 of 24 times, and the first sound of the answer came a median of 4.03 seconds after the question.
I rejected the obvious design, answering natively in each language, which gives twelve scores that can’t be compared.
The evidence. On 60 questions in twelve languages (26 October 2026), the guide answered 49, refused the 11 about things outside the collection, refused nothing it should have answered, and showed the AI notice on all 60. Foundry’s evaluators scored groundedness, how well an answer is backed by the records it drew on: all 48 answers they could check scored 4 or more out of 5 (mean 4.91), and the one left out has no record to check against. A rules-only citation check confirmed 43 of the 49, and the other 6 were correct when I read them.
4.5 Make every publication reversible
Only the publishing sequence can publish, and it needs a curator’s signature and a rights decision that clears the work. It follows the compensating-transaction pattern [17], so a failed step undoes the steps before it, and a rights withdrawal takes the work down again.
I rejected the usual content-management approach, a “published” flag on the record, because a flag can say published while the public copy says something else. The evidence: in one test (5 October 2026), a withdrawn work was gone from the public index within 60 seconds.
5. What I measured
I dated each target before its first run, and a miss stays on the record. The museum should supply the baseline for work by hand; my fictional museum has none, so I label my working assumptions. A p95 is the time within which 95% of answers arrive.
Table 2. What I measured (September to October 2026).
| Measure | Result | Basis |
|---|---|---|
| Cataloguing person-time (target: 5× less) | 12.5× less: a 0.4-minute check plus an assumed 2 minutes to complete, against an assumed 30 minutes by hand | Estimate: the review is measured; the 2 and 30 minutes are stated assumptions |
| Cataloguing accuracy | 86.0% of 143 compared fields unchanged; 5 of 5 date errors flagged (21 October) | Measured on a fixed test set of 20 records |
| Provenance research time (target: 60% less) | 87% less of a person’s time: up to 15 minutes to check the draft chain, against two hours by hand. Counting the assistant’s median of 67 seconds an object, 86.5% (22 October) | Estimate on five objects; the 15 minutes and the two hours are assumptions |
| Licensing | 59 of 501 licensable (20 October); about €102,000 to €164,000 a year across the collection | Counts measured; revenue estimated on published yields [10, 23] |
| Rights needing a person (target: at most 15%) | 66 of 501 (13.17%, 22 October) | Measured |
| The guide’s answers (60 questions, twelve languages) | 49 answered, 11 rightly refused, none wrongly refused; groundedness 4.91 out of 5 (26 October) | Measured |
| The guide’s speed | Half of answers within 2.5 s (26 October) | Measured |
| Visitors asking at once (target: p95 8 s at ten, no errors) | Twenty at once, with every model step on: p95 5.97 s, no errors (23 October) | Measured |
Two other results changed what I built. The provenance assistant’s model reading first got 22 of 26 links right (84.6%, 22 October) against my 90% target; after I fixed how it states each link’s role, a judge found all 31 of its links supported by their sources (23 October). And the smaller gpt-5.4-mini catalogued for US$0.0131 a record against US$0.0334 (21 October), but flagged only one of the five known date errors, so cataloguing stays on gpt-5.4.
Satisfaction is for the museum to measure once the guide opens to visitors, by comparing guide users with non-users.
6. Running it: ease of use, resilience and cost
Easy to run with a small team. The organisations that would gain most often have small technical teams, so I kept Lantern to managed Azure services, with no servers to patch. Its production templates take the Azure region as a setting and come with backup, monitoring and alerts built in. Staff work in one web app, the Workbench, reached only from inside their own network, and visitors need no app at all. Checking a draft took 0.4 minutes a record in my trial.
One part failing doesn’t take the rest down. A model outage loses nothing already done, the guide carries on without the research side, and staff keep working without the public index. The record, held in the event store, is the one component whose loss stops new decisions, so it gets the most protection: continuous backup, plus zone redundancy and a write-once archive in a second region in production. The restore procedure is documented, and the pilot I propose tests it against the museum’s own recovery targets.
The six-month pilot, with the guide for the museum’s staff only, costs €1,846 a month at Azure’s list prices (an estimate, within €1,673 to €2,168), about €11,000 in all (€11,076) [13, 22]. Lantern’s production design costs about €3,100 a month (an estimate of €3,129, within €2,873 to €3,565), for 20,000 records a year and 150,000 visitor questions a month [13, 22]. Search, the gateway, the guide’s model, Front Door Premium and Defender for Cloud make up 76.7% of it. A cheaper search tier, a smaller model for translation and more caching could save up to €759 a month without weakening security. With the guide on gpt-5.4, as I measured it, the month is about €3,950. On the live demonstration, a catalogue draft cost under three US cents in model calls and a visitor’s typed answer under one cent (US$0.0227 and US$0.0086, 20 to 22 October). The demonstration’s own bill for 1 to 20 October was US$312.26, within US$8.08 of my estimate for those days.
7. Beyond museums
I built and measured Lantern in a museum. The same four parts apply wherever experts must describe, prove and explain what an organisation holds. Table 3 shows what they become in five such places, and Appendix B lists more.
Table 3. Where the four parts apply.
| Where it applies | Drafts the description | Builds the history from evidence | Signs before it goes out | Explains to |
|---|---|---|---|---|
| Libraries, archives and broadcasters’ archives | The catalogue record for a book, a file or a recording | Where each item came from, and who holds its rights | The librarian or archivist | Researchers and the public |
| Auction houses and art dealers | The lot description | Past owners, sales and exhibitions | The specialist | Buyers and sellers |
| Banks | The client’s profile | The source of wealth, and who really owns what | The compliance officer | Clients, about what the bank needs and why |
| Land and property registries | The property record | The chain of title | The registrar | Owners, buyers and their advisers |
| Supply chains with product passports | The product passport | Where each material and part came from | The person responsible for the product | Buyers, repairers and recyclers, from the code on the product |
What changes from one organisation to the next is the data the assistants read, the rules the evidence must meet, and who signs. The platform stays the same: assistants that only propose, one record that keeps every decision, separate doors for staff and the public, and the same templates. Each organisation measures the gains against its own baseline, as the museum would in the pilot I propose.
8. From pilot to scale
- First month: agree the commitments, the event vocabulary (the kinds of entry the record can hold, such as a proposal or a signature) and the rights table, freeze the evaluation sets, and ask the museum for its baseline timings.
- Pilot: 20,000 objects, about 3% of the collection, with the museum’s curators. In the pilot the guide is for the museum’s staff, and it answers from the records the curators approve. Sample signed records each week, because a review that takes seconds may be a rubber stamp, and decide go or no-go on the museum’s figures at the end of June 2027.
- Scale: this is the step that clears the backlog, so price it separately. Add model quota, connect the collection management system, open the guide to visitors and measure their satisfaction, and publish through IIIF, the International Image Interoperability Framework.
What I would refuse to build: an assistant that publishes, a model that decides rights, a visitor profile, or one search index split by a filter.
9. What I learned building it
Read real records before you write a rule. A rule keyed on the word dépôt read a donor’s ban on lending, « interdiction de prêt et de dépôt », as a deposit, across hundreds of works. Sampling the records caught it.
Check what each link means. A judge that read what each provenance link claims found four with the wrong role, such as a place where a work was kept read as its owner. A check that only matched names passed all four. Once I fixed how the assistant states each role, the judge found all 31 links supported.
Ask for the customer’s baseline. I’m not a cataloguer, so timing myself would have proved nothing.
Fix the product when you miss a target. Cataloguing missed its speed target (p95 30.4 seconds against 20), so I gave each outside lookup its own time limit. The re-run took 18.4 seconds, with accuracy held.
Build the evaluation sets first. Freeze them before the first run, so every fix is proved on questions the assistants haven’t seen.
10. Limits, and what the pilot adds
- The pilot adds what a demonstration can’t: the museum’s staff, native speakers among them, reading the guide’s answers, a rights specialist sampling the rules’ decisions, a full accessibility test, and the museum’s data protection impact assessment, from a draft I wrote.
- The provenance assistant reads one outside database live, the national catalogue, and the museum adds more sources without code, on the Sources page of the Workbench, the staff’s web app. When a lookup runs out of time, the chain says that source was not consulted.
- Every assistant will sometimes be wrong. Lantern’s design assumes that. Each output carries its source, and the record shows which person made each call.
11. Conclusion
Experts who describe, prove and explain what an organisation holds are too few for the work in front of them. Lantern shows that AI can take on the drafting and the research while a named person keeps the final say. In my build, cataloguing took an estimated 12.5 times less person-time and provenance research an estimated 87% less, with a source for every proposed fact and nothing published without a signature. I would start with cataloguing and rights in a pilot, and measure every gain against the organisation’s own figures.
Appendix A. Where to start: a checklist
- Write the design commitments on one page, and mark each one as law or policy.
- Define the event vocabulary before any screen.
- Keep one append-only record, with assistants that can only propose.
- Use two search services and two identities for public and research text.
- Use rules for legal decisions and models for language.
- Put a checker between the assistant that proposes and the person who decides.
- Send every model call through a capped gateway, and write down the calls that bypass it.
- Freeze an evaluation set and date the targets before the first run.
- Ask for the institution’s own baseline.
- Screen content for planted instructions when it comes in.
- Turn on continuous backup and a write-once event archive in the first week, then test the restore in the pilot.
Appendix B. Where else the four parts apply
Section 7 shows five places where Lantern’s four parts apply. Table 4 lists more. In each, a guide would also explain the result to the people the organisation serves, and each would bring its own data, rules and signers.
Table 4. More places where the four parts apply.
| Where it applies | What the assistants draft | The history they build from evidence | Who signs |
|---|---|---|---|
| Artist foundations and catalogue raisonné committees | The entry for each work | Past owners, exhibitions and publications | The authentication committee |
| Natural history collections, herbaria and biobanks | The specimen or sample record | Where and when it was collected, and under what permit or consent | The curator or custodian |
| Archaeology services and heritage registers | The record of a site or find | Its excavation and custody history | The heritage officer |
| Photo agencies, image libraries and publishers | The caption and rights record for each image | Who made it, and who has held and licensed its rights | The rights manager |
| News organisations | The caption for a photo or video | Where it was captured and how it was edited | The picture editor |
| Universities | The record of a degree or other qualification | The study and assessment records behind it | The registrar |
| Insurers of art, jewellery and other valuables | The schedule of insured items, and each claim file | Proof of ownership, value and condition | The underwriter or loss adjuster |
| Trade finance | The review of each shipment’s documents | The chain of bills of lading, invoices and certificates | The document examiner |
| Lawyers in mergers, acquisitions and lending | The register of the assets and contracts in a deal | Who owns each asset, and how they came to own it | The responsible lawyer |
| Audit | The working papers | The evidence behind each figure | The engagement partner |
| Company and beneficial-ownership registers | The company’s record | Its owners and controllers over time | The registrar |
| Vehicle, vessel and aircraft registries | The registration record | Ownership and mortgage history | The registrar |
| Courts and police evidence stores | The exhibit record | Its chain of custody | The exhibits officer |
| Aircraft maintenance | The record for each part | Its history back to manufacture | The engineer who releases the aircraft to service |
| Medicines and medical devices | The batch or device record | Where every material and component came from | The person who certifies it for release |
| Food, farming and timber | The record for each lot | Its chain from farm or forest to the buyer | The quality manager or certifier |
| Mining, metals and gemstones | The record for each lot | Its chain from mine to market | The responsible-sourcing officer or assayer |
| Luxury goods, fine wine, classic cars and collectibles | The listing or certificate | Past owners, storage, servicing and restoration | The authenticator |
| Construction and building safety | The building’s safety record | Every design change, product and inspection | The person accountable for the building’s safety |
| Energy, water and transport infrastructure | The asset record | Its inspection and repair history | The responsible engineer |
| Software suppliers | The software bill of materials | Where each component came from and how it was built | The release or security lead |
| Teams building AI models | The record of each training dataset | Where each source came from, and under what licence | The data steward |
| Carbon and nature credits | The project record | Each credit’s path from issue to retirement | The accredited verifier |
Appendix C. Five rules from the rights table
These are five of the rights assistant’s rules for France, in the order they’re checked. The first rule that matches decides, and every decision names its rule. Your jurisdiction’s terms will differ.
| Rule | When the record shows | The decision |
|---|---|---|
| R1b | A photograph of an identifiable person | The record may publish; the image waits for that person’s consent [6] |
| R3 | An author who died 70 years ago or less, or is alive | In copyright: the record’s text may publish, and the image is blocked [26] |
| R4 | An author who died more than 70 years ago | Public domain: the work’s text and image are free, and the photograph’s licence is decided on its own [26] |
| R5 | An anonymous or uncertain work dated before 1850, or a named artist dead for more than 120 years | Public domain, whoever the maker was |
| R6 | Anything else | Undetermined: nothing publishes, and a person decides after a diligent search |
References
Web pages were read on 22 October 2026.
- Case-study scenario for the Musée National des Arts Réunis (fictional), 2026: the museum’s figures (600,000 objects; 8% digitised with rich metadata; 9 million visitors a year; a backlog of about a decade).
- Code du patrimoine, art. L. 441-2 (the missions of a musée de France); arrêté du 25 mai 2004 (inventory standards).
- Code du patrimoine, art. L. 441-2, as amended by Loi n° 2026-351 du 9 mai 2026; art. L. 451-2 (the ten-yearly récolement).
- Regulation (EU) 2024/1689 (the Artificial Intelligence Act), Article 50, as amended by Regulation (EU) 2026/1744; European Commission, Guidelines on the implementation of the transparency obligations for certain AI systems under Article 50 of Regulation (EU) 2024/1689, C(2026) 5054, 20 July 2026; European Commission, Quick Facts: Transparency rules for AI systems, updated 29 July 2026.
- Code de la propriété intellectuelle, art. L. 122-5-3 II (text-and-data mining for scientific research; Ordonnance n° 2021-1518).
- Code civil, art. 9 (private life and image).
- Directive (EU) 2019/790 on copyright in the Digital Single Market: Article 3 (text-and-data mining for scientific research) and the out-of-commerce works provisions; CPI art. L. 138-1 and following.
- Regulation (EU) 2016/679 (GDPR), Article 6(1)(e).
- Directive (EU) 2016/2102 on the accessibility of the websites and mobile applications of public sector bodies.
- Rmn-Grand Palais, Rapport d’activité 2022, p. 97: photo agency revenue of €3.5 million on about 1.5 million digitised images.
- Musée du Louvre, press release, 26 March 2021: The Musée du Louvre launches online collection database and new website (collections.louvre.fr, more than 482,000 entries).
- Joconde, the national catalogue of the musées de France, Ministère de la Culture, Licence Ouverte / Open Licence 2.0 (extract of 15 September 2026).
- The author’s project records, 2026 (not published): architecture decision records ADR-001 to ADR-014; the measurement plan; cost model v2 (19 October) and the pilot’s cost (23 October); revenue scenarios (19 October); the decisions log.
- Microsoft Learn, Work with the change feed (Azure Cosmos DB). https://learn.microsoft.com/en-us/azure/cosmos-db/change-feed
- Microsoft Learn, Connect using Azure roles (Azure AI Search). https://learn.microsoft.com/en-us/azure/search/search-security-rbac
- Microsoft Learn, AI gateway capabilities in Azure API Management. https://learn.microsoft.com/en-us/azure/api-management/genai-gateway-capabilities
- Azure Architecture Center, Compensating Transaction pattern. https://learn.microsoft.com/en-us/azure/architecture/patterns/compensating-transaction
- Microsoft Learn, Durable Functions overview. https://learn.microsoft.com/en-us/azure/durable-task/durable-functions/durable-functions-overview
- Microsoft Learn, Prompt Shields, Azure AI Content Safety. https://learn.microsoft.com/en-us/azure/ai-services/content-safety/concepts/jailbreak-detection
- Microsoft Learn, Services excluded from the EU Data Boundary (updated 13 April 2026). https://learn.microsoft.com/en-us/privacy/eudb/eu-data-boundary-excluded-services
- Microsoft Learn, Deployment types for Microsoft Foundry Models (Data Zone Standard). https://learn.microsoft.com/en-us/azure/foundry/foundry-models/concepts/deployment-types
- Azure Retail Prices API, prices read on 19 October 2026. https://learn.microsoft.com/en-us/rest/api/cost-management/retail-prices/azure-retail-prices
- J. Pekel, Democratising the Rijksmuseum, Europeana Foundation, 2014, pp. 5 and 12: €181,000 image revenue in 2012 on 125,000 works online.
- Microsoft Learn, Configure AI Gateway in your Foundry resources. https://learn.microsoft.com/en-us/azure/foundry/configuration/enable-ai-api-management-gateway-portal
- Microsoft Learn, Security filter pattern (Azure AI Search). https://learn.microsoft.com/en-us/azure/search/search-security-trimming-for-azure-search
- Code de la propriété intellectuelle, art. L. 123-1 (the author’s rights last for the author’s life and 70 years after the year of death).
- Microsoft Learn, Data, privacy, and security for Speech to text (read 28 October 2026). https://learn.microsoft.com/en-us/azure/foundry/responsible-ai/speech-service/speech-to-text/data-privacy-security
Legal notice
The views in this white paper are my own and do not necessarily represent those of Microsoft. It describes a design I built and measured. Lantern is not a Microsoft product, and nothing in this white paper is a commitment on the part of Microsoft. This white paper is for informational purposes only. MICROSOFT MAKES NO WARRANTIES, EXPRESS, IMPLIED, OR STATUTORY, AS TO THE INFORMATION IN THIS DOCUMENT.
Complying with all applicable copyright laws is the responsibility of the user. Without limiting the rights under copyright, no part of this document may be reproduced, stored in, or introduced into a retrieval system, or transmitted in any form or by any means (electronic, mechanical, photocopying, recording, or otherwise), or for any purpose, without the express written permission of Microsoft Corporation.
Microsoft may have patents, patent applications, trademarks, copyrights, or other intellectual property rights covering subject matter in this document. Except as expressly provided in any written license agreement from Microsoft, the furnishing of this document does not give you any license to these patents, trademarks, copyrights, or other intellectual property.
© 2026 Microsoft Corporation. All rights reserved. Microsoft and Azure are trademarks of the Microsoft group of companies. All other trademarks are property of their respective owners.