How to license operational data to AI labs, end to end

Begin with the work an AI lab could reconstruct. In Company A's case, a customer emails a service request, a team discusses how to handle it in Slack, and a Jira ticket records the action and confirmation. The buyer needs to understand the sequence and outcome. A ticket export that still points to inaccessible email and chat may contain plenty of text while omitting the answer to the buyer's question.

Rightmark's guide to company records for agent training explains this connection between a request, the steps taken, review and the result. Use it to describe a concrete task your archive can support. “Eight years of business data” tells a buyer much less than “service requests linked to the conversations and tickets that resolved them.”

Then define the product you are proposing. A fixed historical copy is different from a continuing feed, future refreshes or a runnable workplace environment. Company A is initially offering a fixed copy of eligible historical records for a defined AI use. It is not offering access to its live operations or promising to build a working simulation of its business. If a buyer wants either, the scope, work and price need another discussion.

Build one private inventory across the systems

Ask each system's administrator to fill out a row before producing a full export. The inventory should distinguish what the company remembers having from what it can actually retrieve.

Company A should not enter eight years for every tool merely because the business is eight years old. If Jira only retains the most recent three years, say so. If a migrated Slack workspace duplicates an older export, record the relationship. Combining two copies does not create twice as much independent history.

Here is a filled, invented inventory for Company A. Assume the administrators identify these retained periods and export copies. The counts below are teaching assumptions, not measurements from Rightmark or a customer's systems.

The Jira unit includes the issue record being counted, not a separate count for every comment. Attachments and comment history still need their own coverage checks. Uncompressed sizes have not been measured in this example; they remain unknown rather than being inferred from these counts.

Reconcile each tool separately, removing repeated copies first. In this invented quantity reconciliation, “excluded” and “unresolved” are non-overlapping categories within the remaining unique records. Unresolved items stay outside the proposed copy until someone resolves the relevant rights or scope question.

For Gmail, the calculation is 120,000 - 20,000 - 15,000 - 5,000 = 80,000. The same sequence applies to the other rows. “Candidate” means available for the proposed preparation and review in this scenario. It does not mean legally cleared, accepted by a buyer or already usable for training. Further exclusions or failed preparation can reduce the collection again.

Do not add messages and issues together and call the total training tasks. To count a connected workflow, Company A needs a separate rule: one distinct service request with the permitted initiating email, necessary chat decision and ticket outcome, with distinguishable actor roles. The three-system overlap here begins in October 2023, even though Gmail extends back eight years.

For a bounded inspection later in this guide, assume the owner authorizes sample S1 around 100 distinct candidate service-request IDs from that overlapping period, with their permitted records. In an invented test, 72 have all required links, roles and outcomes; 18 lack a required permitted link or attachment; 10 fail the actor-role check. Each request appears in only one result category, assigned at its first failure. That is 72 usable workflows for this particular inspection, not a yield estimate for all 80,000 emails or 12,000 issues. S1's selection and coverage still matter.

This inventory changes the scoping conversation: the company has eight years of some records, but not eight years of the complete three-tool workflow. Preserve that distinction when requesting a revised appraisal. Do not reduce the calculator's range by the duplicate share or multiply it by 72%. Neither is a price formula the model or a buyer has supplied.

Keep names and sensitive details out of the initial sales description. You can explain that an administrator has checked a retained period without emailing a list of employees or customer accounts. The detailed inventory stays controlled; the first appraisal can use a summary.

Check authority separately from technical access

An administrator may be able to export a system without having authority to license everything in it. Review the relevant company entity, signatory authority, earlier licenses, customer confidentiality obligations, employee notices and permissions, and third-party material. Decide which records are included, excluded or still under review. A platform subscription does not answer those questions for you.

For Company A, the proposed service workflow excludes unrelated HR and payroll records, privileged legal discussions, and health or children's information. A service ticket may still contain something that needs removal, even when the folder name looks harmless. Treat the scope as a record-level question where necessary, not just a list of approved tools.

Credentials, passwords, API keys and access tokens do not belong in a sample or delivered corpus. If the inventory exposes one, arrange revocation or rotation through the company's security process, as well as its removal from the proposed copy. Buyer access, if later needed, uses separately approved, scoped access. It does not require embedding a working credential in the data.

Record the basis for including each category and who resolved an uncertain item. Rightmark's provenance log guide is useful here: the record should connect the source, preparation and resulting copy. A well-kept log makes the evidence findable; it does not create rights that the company lacks.

The output of this stage is a defensible candidate scope. It is perfectly useful to discover that one tool or period must stay out. That finding lets you ask whether the remaining collection still supports a buyer's task before anyone pays to prepare it.

For company data monetization, the first number is a starting point for qualification. It is not a rate you can apply to every gigabyte. Rightmark's public operational estimator prices a company archive using factors including its retained years, team size, systems, language, scope and licensing route. Headcount and age are proxies for possible useful history; the export and review have to substantiate it.

The Q4 2026 Rightmark Price Index, published October 26, lists $185,000 to $1M+ per company for its operating-company segment with 30 or more staff and English records. It labels this Medium confidence and identifies the basis as Rightmark's summary of public company-data programs. This is a segment estimate, not a minimum offer for any company meeting those two attributes or a tariff shared by AI labs.

Use the index, pricing guide and estimator for different decisions

I use the index to establish the unit and the dated reference, then the data valuation guide to work out what makes one archive comparable with another. The guide, updated November 15, describes three levers: scarcity, structure and rights. For an operational archive, that means retained private history a buyer needs, relationships that keep the work intelligible, and authority to grant the proposed use. More files do not compensate for an unusable handoff or an incompatible earlier license.

The estimator then applies your declared inputs to a model. A buyer's quote comes after review of the description, eligible material, permitted sample and proposed rights. Compare the same copy, use, term and included costs at each stage. A listing, a model result and a signed price answer different questions; none becomes owner proceeds until the applicable fees, costs and payment conditions are understood.

Reproduce an estimate before comparing it with an offer

On November 15, 2026, the operational estimator returned $260,000 to $540,000 per company, Medium confidence, for the following Company A inputs. The selected route was an exclusive lump sum. This records the initial eight-year description, before the filled inventory above separates the shorter periods for Slack and Jira. It is not a recalculated offer for the surviving records.

Indicative estimate, not an offer. Final offers depend on the systems you connect, data size and quality, access terms and due diligence.

These inputs are a worked calculator example, not measured findings about a real archive. In particular, selecting “None” in a questionnaire does not establish that every message is free of sensitive information. Company A still needs the inspection and exclusions described above.

Save the inputs, date, unit and result together. If a later estimate changes, you can then distinguish a different model or route from a changed answer. A screenshot of a range without its assumptions is a weak basis for negotiating a final offer.

A smaller offer can be reasonable when the buyer is being offered a smaller or less useful asset than the initial description suggested. Conversely, an offer deserves scrutiny if the buyer asks for broader rights, additional datasets or expensive preparation while treating the original estimate as the full price for everything.

Compare asking prices for three different archives

These three profiles appear in Rightmark's public lot listings, checked November 15, 2026. They show how different the offered packages can be. The figures are indicative asking references, not buyer-agreed prices, completed transactions or owner proceeds. Counts and periods are reported listing information unless a measurement is specifically identified. They are not verified counts of the final deliverable.

The dollar signs above preserve the listings' notation; confirm settlement currency in the written quote. For all three lots, establish the licensed use and recipients, fee payer and fee base, preparation output and cost, acceptance conditions and payment trigger. The asking reference alone does not supply those terms or an actual buyer commitment.

The ten-person firm illustrates a smaller, non-exclusive package rather than the 30-plus-staff route discussed below. It does not demonstrate that a buyer has selected that firm. SE-A shows why an apparently larger exclusive price needs closer reading: a right with software-vendor exceptions is different from an unrestricted exclusive grant, and removing a tool can change what the buyer can reconstruct. Do not assume this lot qualifies for a whole-corpus exclusive request.

I would use these profiles to sharpen the questions about my own archive, not divide their prices by staff, years or gigabytes. Nor would I combine their lowest and highest asks into a market range. Each amount belongs to a particular collection and licensing scope.

See what a conditional buyer proposal would change

For a separate teaching example, use Archive B, an example; the company and the people are invented. It is not Company A or any lot above. Its permissions, offer and outcome are fictional. Assume that, after inspecting an identified permitted sample, an AI buyer proposes US$90,000 gross license consideration for a narrowed eligible snapshot. The assumed grant covers a defined AI-training use and recipient exclusively for 24 months, with retained-model and end-of-term treatment agreed separately. The offer depends on the promised delivery passing the agreed acceptance checks.

That amount is an assumption, not a Rightmark quote, typical price or discount calculated from the index. The decision is whether the selected records support the buyer's permitted task and whether the owner can grant that scope and deliver it for an acceptable return. A change in use, eligible material, exclusivity or preparation work calls for a revised written proposal.

The US$90,000 alone does not establish owner net or total buyer cost. Archive B's broker fee, payer and base, preparation bill, other owner costs and buyer implementation costs remain unknown in this illustration. Use the Company A comparison in step 3 to record them rather than entering zero. For the calendar in step 8, we will assume the parties settle their responsibilities before signing and the full US$90,000 license invoice is paid; that will show gross receipt, not profit.

Improve the asset a buyer can actually use

The first improvement is often evidence, not more records. Show the real coverage, distinguish original records from repeated exports, and demonstrate that the request, handoff and result remain connected. Resolve whether the company can grant the proposed use. Explain an exclusion before it becomes a surprise in the buyer's test.

In Company A's example, recovering a permitted missing ticket relationship could make a workflow usable. Adding another copy of the same emails cannot do that. Nor should the company create artificial historical records to make the archive look fuller. If a date or relationship is missing, identify it as missing unless a reliable source supports a correction.

Avoid commissioning a full cleanup simply to improve an indicative range. First agree which buyer task the work serves, what a small specimen needs to demonstrate and who will pay for the next stage. A collection that is immaculate but unsuitable for the buyer's use can still receive no offer.

Pro tip: Do not add the company-archive, chat, ticket and knowledge-base rows of a price index as if they were four independent licenses. If they describe overlapping material, ask which package and rights the quoted unit covers.

A good price is attached to a clear asset and a workable obligation. A larger headline can be worse if it buys a broader exclusive grant, leaves substantial costs with the owner or makes acceptance depend on an undefined test. Put those differences into the comparison before deciding that an offer is high or low.

Rightmark's operational page describes company-archive lump sums as usually exclusive for a term, often two years or more, with non-exclusive revenue share as an alternative. We also have a current buyer request requiring an exclusive license to the whole archive. That route is worth evaluating early: a buyer with that requirement may reject a proposal containing material carve-outs.

Check whether the archive fits the actual buyer request

As checked November 15, 2026, one current Rightmark company-archive request requires a whole-corpus exclusive license and at least 30 current or peak staff. Those are that request's filters, not a rule for every AI lab. The ten-person REC-CONSULT listing describes a different non-exclusive scope; its existence does not promise a buyer for smaller archives.

Exclusivity can give a buyer a reason to pursue a collection it could not obtain on the same terms elsewhere. For the owner, the corresponding cost is the licensing opportunities it gives up. I would negotiate the boundary before arguing about the premium. “Exclusive data” is too vague to value.

Specify the exact corpus and version, covered uses, term, permitted recipients, territory where relevant, and treatment of later updates. Discuss whether derived datasets and model uses continue after the exclusive period. Owning the original and continuing to use it in the business is different from retaining the right to license the same copy to another AI buyer. The end of an exclusivity period also does not, by itself, describe what must happen to an already trained model.

Company A might prefer a fixed historical cut, a defined use and a two-year restriction. A buyer seeking the entire archive may require a different boundary. Treat that disagreement as a commercial decision: accept a suitably priced scope, negotiate a permissible alternative, or choose another route. Do not promise a narrower package the buyer has already said it cannot use.

Compare owner proceeds and the buyer's full bill

Rightmark's published fee model says that on standard licenses we take nothing from the owner's payout and the buyer pays our fee on top. On premium and bespoke deals the owner keeps 75% to 85%, agreed in writing before any sale. These are our published routes, not industry-wide commission rates.

The operational page also says extraction, scrubbing, transfer and the fee are agreed in writing before signing. Ask what the proposed work includes, who performs it and who pays. A statement that the buyer covers preparation should not become an undefined promise of unlimited engineering.

Consider two invented proposals for Company A, in assumed US dollars. Both cover the same assumed data, rights and term. The amounts are conditional on the agreed acceptance and payment requirements being met; taxes, financing and currency effects are excluded. No other owner costs are assumed.

Proposal B has the bigger license headline but leaves the owner with $20,000 less and costs the buyer $20,000 more. The $70,000 commission is already inside its $350,000 consideration; adding it again would overstate the buyer's bill. The invented 10% in Proposal A is not a published Rightmark standard rate.

Ask what the broker is undertaking for that compensation. Finding a prospective buyer, qualifying the archive, coordinating permissions, arranging preparation and carrying delivery questions are distinct jobs. Assign the work you need. A capable owner and buyer may be better placed to transact directly when they can already perform it; an introduction alone should not be mistaken for a complete preparation service. Our broker-scope guide develops that decision.

Value the alternatives over the same period

Here is a second, separate invented model, again for Company A. Assume a two-year exclusive license would yield $300,000 net to the owner. Assume each completed, paid non-exclusive license on the same defined scope would yield $160,000 net during those two years. Assume no additional owner costs; exclude taxes and financing effects.

This arithmetic identifies what has to happen for the alternative to win. It does not predict that a second or third buyer will appear. In a real revenue-share offer, the price, commission base, allowed reuse and timing may also differ, so replace the assumptions with the actual proposal.

An exclusive offer can be attractive when the agreed net amount is worth the restriction and the owner values one transaction over uncertain additional sales. Non-exclusive licensing can be attractive when the owner has credible additional demand and wants to preserve that option. Neither choice removes counterparty risk: identify the payer, what delays or rejection do to the commitment, and what happens if the transaction does not complete. If you need a reservation to expire, negotiate its expiry rather than assuming one exists.

If you have the inventory but no comparable offer yet, use the operational archive estimator and request a written scope with the tools, teams, years and licensing route. That is enough to start the discussion without uploading the records.

A buyer needs evidence before committing to an archive. The owner needs to know what each access decision permits. You can meet both needs by describing the next action accurately, rather than sending a folder under a general promise of confidentiality.

For Company A, the first proposed sample permits ordinary inspection of its schema, completeness and relationships. It does not authorize training, fine-tuning, model benchmarking or creating embeddings. If the buyer later needs one of those tests, agree that use expressly before it runs.

An NDA may support ordinary inspection within its stated purpose. It is not useful to say that every inspection needs another contract regardless of the documents already in place. Equally, calling a file a sample does not grant all possible model uses. Read the applicable permission against the actual operation.

These functions can sit in one agreement or several. The objective is a coherent permission, not a target number of PDFs. Rightmark's agreement and permissions guide explains why confidentiality, appointing a broker and authorizing a buyer remain different decisions even when documents combine them.

Write a permission that someone can operate

For the invented Company A sample, a useful scope note would identify sample S1, the included service-workflow period and fields, the approved recipient, the ordinary inspection described above, and the allowed transfer and storage location. It would also state what reports may leave that location, when the copy and outputs must be retained or deleted, and who can approve another test. Use a named recipient or a confidential recipient designation the owner expressly approves, rather than treating an unidentified onward recipient as automatically covered.

The owner can approve an anonymous description for buyer matching without approving sample access. A sample approval does not authorize an ongoing connection to the live systems. Put each decision beside the action it permits so an administrator is not left interpreting a commercial email as an access request.

A broker's protection for an introduced transaction is another separate issue. It may affect the fee owed if that transaction proceeds. It does not by itself define which data is exclusive or what an AI lab can do with it. Check those obligations together before appointing multiple intermediaries or negotiating directly with an introduced buyer.

Bound the first preparation job

Before engineers begin, agree a specimen, a deliverable and a spending decision. For Company A, that could mean preparing the permitted email, chat and ticket records for one representative workflow and testing whether its relationships survive the exclusions. It is a feasibility exercise, not a promise to clean every record in eight years of history for the same price.

Name the person doing the work, the person verifying it, the payer and the person authorized to approve changed scope. Describe what the buyer will receive: a scrubbed specimen, a schema and a short report of the specified checks, for example. The cleaning-scope guide shows why supplying a tool and delivering a usable export are different assignments.

If a required relationship is broken, first ask what was agreed. Correcting a defect in the promised output differs from adding an excluded attachment category or enabling a new model test. The first may fall within correction terms; the others need a scope decision. Make that distinction before work continues, with the price and schedule consequences recorded. There is no useful universal promise of free corrections, replacements or refunds in place of the actual agreement.

I would keep the controlled source and the preparation copy distinct. Freeze or identify the source used for the job, within the company's actual retention obligations, then apply agreed exclusions and transformations to a working copy. This gives reviewers a way to establish what changed without turning the licensing exercise into a new policy of retaining everything forever.

Start with exclusions. Remove out-of-scope records and disallowed content before deciding how to transform the material that remains. Anonymization is not a reason to include an otherwise excluded legal or HR collection. Check attachments, quoted messages, filenames and diagnostic logs as well as the main text; the prepared package includes those surfaces if you send them.

Record the exclusion rules and resulting counts without copying the sensitive content into the exclusion log. Identify duplicate exports while preserving meaningful revisions where they are part of the agreed task. Normalize a timestamp or repair a relationship only when the evidence supports the change. A plausible invented completion date is still an invented fact in a historical dataset.

Preserve the relationships the task needs

Company A's buyer wants to distinguish the requester, the person who handles the work and the person who confirms completion. Replacing every name with the same generic label would make that test harder, even if it removed the names successfully.

The following is an invented miniature of Company A's prepared workflow, not real company data or a prescribed export schema:

The labels need to be consistent where the task requires continuity. They do not automatically need to connect the same person across every unrelated conversation in the company. Company A can test whether W1 is understandable without exposing a company-wide identity history. Any underlying identity map stays separately controlled on the owner's side and is not part of the delivery.

Replacing identifiers is not the same as establishing anonymity. The ICO's UK pseudonymisation guidance distinguishes pseudonymization from anonymization; keeping the identifying information separately does not automatically remove the data from data-protection law.

The ICO's effectiveness guidance also considers who could access the information and whether people can be singled out or linked across records. Both guidance pages are under review following UK legislative changes. Apply the relevant privacy assessment to the actual recipient and copy.

For Company A, the practical work has two tests. Can the permitted reviewer still reconstruct the service request? Does the proposed release expose material or connections that should not be there? Passing one does not settle the other. Our guide to identity links in company archives develops that tradeoff with a worked example.

Choose a preparation route someone can deliver now

An owner's own team can perform agreed preparation when it has the access, capacity and expertise. The buyer may perform specified work if its access, environment, outputs and responsibilities are expressly agreed. Rightmark also offers a specialist anonymization service inside the owner's systems, with before-and-after samples for owner review and buyer-funded scrubbing under the published offer. Put the particular deliverables and costs in writing whichever route you use.

The current Rightmark Clean Room has an important availability distinction. For operational archives it is early access: Archive Check is the first ready job, and later archive jobs are a pilot build. Archive Check counts locally. The broader sequence of sorting, cleaning, buyer scoring and signed receipts describes the pilot's design; it is not a promise that every archive can run that whole sequence as a finished self-service product today.

If you plan to use a pilot job, ask what is available for your systems, who operates it and what output they will demonstrate. If you need managed preparation, ask for the work order. Buying coordination or receiving a tool does not automatically assign an engineer to repair missing relationships in your export.

Separate a clean-room test from licensed delivery

A clean room controls an evaluation environment and its permitted outputs. Anonymization changes the proposed data. Delivery transfers the agreed licensed copy. One project can need all three, at different stages.

In Rightmark's described pilot fitting room, a buyer's scoring code runs with networking off and releases bounded, schema-checked numeric outputs that the owner sees before release. That design still needs a permitted test and an available implementation for the engagement. Company A's initial inspection-only permission would not silently expand to cover model scoring because a technical environment could run it.

Clean rooms also differ. AWS Clean Rooms documents a list analysis rule that returns row-level lists of overlap. Inspect the actual output permissions; the product category alone does not tell you that only aggregate results can leave.

Read-only access is another boundary worth checking: it may stop changes to a source while still allowing the reader to see or copy its contents. Ask what can enter, what operations can run and what can leave. Keep access permissions separate from the corpus; never supply working account credentials as sample data.

Once the final license is signed, a de-identified licensed copy may leave the owner's environment while the original archive and identity map remain owner-side. The pilot's promise about raw data staying put should not be read as a promise that no licensed dataset will ever be delivered. Keep those transfers separate in the plan: approved evaluation output first, agreed licensed copy later.

For Company A's invented owner-run path, the preparation environment is inside the owner's controlled systems. This custody table shows what moves; it is not a claim that the whole Clean Room pilot is available as a finished product.

Read the path in order: the source becomes a controlled preparation copy; an approved report or scrubbed S1 may cross to the buyer for its stated purpose; permitted findings return for a decision. Any model-scoring branch needs separate authorization before it runs. After the final license and release approval, D1 crosses to the buyer as the licensed package. The raw original and identity map remain owner-side throughout this illustrated path.

A managed or buyer-operated preparation arrangement can put different people at the controls. Agree its access and custody separately rather than assuming this owner-run diagram describes it. Likewise, S1 or a corrected S2 is still an evaluation version; neither automatically becomes D1.

Pro tip: Open the links inside the prepared export, using only the material the buyer is permitted to receive. A link working for an employee in the live workspace does not prove that the exported relationship works for the buyer.

A sample should answer a defined purchase question. For Company A, the initial question is whether its permitted service-workflow records retain the schema, completeness and relationships the buyer needs to inspect. A beautifully prepared conversation chosen by the owner cannot tell the buyer whether the relevant periods and teams contain similar usable records.

Define the population before selecting from it: tools, teams, dates, languages, record types and exclusions. Explain how the selection represents that population and where it does not. If missing periods or difficult cases are excluded from the sample, identify them. There is no universal sample percentage that makes those selection choices irrelevant.

Whole-archive exclusivity is a proposed commercial scope, not permission to transfer a full raw archive for evaluation. The buyer can first examine the permitted information and decide what further evidence it needs. The owner can approve another bounded test, negotiate the scope or decline.

A scan reporting no findings answers the scan's question. It does not, by itself, settle legal authority, prove anonymity or demonstrate that the workflow is useful. Equally, a buyer understanding the workflow does not establish that every field is permitted for delivery.

Make a failed sample produce a specific next decision

The following are illustrative failures within Company A's invented S1 inspection: 72 workflows passed and 28 failed the defined relationship, role or outcome check. These are examples of those failures, not additional cases to add to the count.

Suppose S1 contains a ticket whose linked thread was excluded. The buyer cannot reconstruct the reason for the action. The response is not to put the excluded raw thread back automatically. Review whether a permitted repair can support the task; otherwise narrow the promised use or accept that this part of the archive may not qualify.

In a second invented failure, the requestor and confirmer have both become the same generic person token. Correcting that transformation may restore the required distinction without revealing their identities. Check that correction against the applicable privacy scope too.

In a third, a ticket depends on an attachment that the owner cannot license. Ask whether the task remains usable without it. A note saying “attachment omitted” can document the limitation but cannot supply the missing evidence of the work.

For the invented correction round, suppose eight of the 18 missing-link cases can be repaired using permitted material, and six of the ten actor-role cases are corrected. S2 then passes the defined utility check for 72 + 8 + 6 = 86 of the same 100 requests. Ten missing-link cases and four actor-role cases remain unresolved. These are assumed results for this sample, with no claim that it represents the entire archive.

When the correction produces S2, retain the link between S1, the finding, the changed preparation rule and S2's result. Check whether the change was made only to that specimen or to the process that will generate the full delivery. As our sample evaluation guide explains, a repaired sample does not establish that the larger delivery has been repaired.

At the end of an evaluation round, record one decision: proceed to final terms, correct a defined defect, change the proposed scope, approve a different test, or decline. Assign an owner and any additional cost before another round. This makes the sample useful to both sides and prevents an undefined sequence of free experiments becoming the project.

The final agreement should reflect what the parties learned from the sample. Reconcile the data description, rights, exclusions, preparation, price and fees against the actual proposed delivery. If an excluded period reduced the useful archive, or the buyer now wants a model use that was absent from evaluation, resolve that difference before signing.

Identify the contracting counterparty, the permitted recipients and the party that owes payment. They may perform different roles. Specify any onward access rather than assuming the organization evaluating the sample is the only eventual recipient.

Ask counsel to review the proposed title and permission warranties, liability allocation, caps and exceptions, indemnities, governing law and dispute process against the actual asset and net proceeds. There is no single liability cap that fits every archive. A promise about material the owner has not reviewed is not made safer by a large headline price.

Put acceptance and payment into the same delivery specification

Write down what will be tested, who can accept it, what starts the review window and what happens if a criterion fails. Include correction and retest responsibilities, any additional charges, and how a changed request is approved. Then connect the decision to the invoice trigger and payment obligation.

For Company A, “buyer approves the data” is too vague. A useful specification identifies the permitted systems and periods, required relationships, exclusions, package format and checks that determine whether the delivered copy meets the agreed purpose. The parties still have to negotiate the actual thresholds, windows and remedies; none is supplied by a sample passing an earlier inspection.

Where an unresolved exclusion makes the promised task impossible, solve the scope or stop. Do not turn an unanswered requirement into a silent assumption that an engineer will fix after signature.

Deliver a manifest that identifies the copy

After signing, complete the agreed full preparation run and test the actual version being delivered. For Company A, call that delivery D1. Count it again after the full run: neither the 80,000 candidate emails in the inventory nor the 86 successful sample workflows is a final D1 delivery count. Use a package manifest that lets the recipient identify the files, understand their units and connect them to the agreed source and preparation policy.

Do not put a sample-looking hash in a manifest as a placeholder that could be mistaken for a real check. Generate the digest from the actual file. A matching digest establishes that two copies match; it does not prove ownership, privacy compliance, acceptance or payment. A signed manifest adds an attestation by its signer, whose scope still matters.

Review the delivery as a whole. Leave out original raw exports, identity maps, working credentials and out-of-scope records. Check the logs, path names and supporting reports too. A sanitized data file paired with a revealing diagnostic attachment is still a revealing delivery.

The export guides can help an administrator locate the relevant platform workflow. Test the actual method and permissions for your account; the delivery specification should describe the package the buyer will receive, not assume that a product's export button includes every needed relationship.

Record receipt without calling it acceptance

Use the transfer method the parties approved: a scoped connection the administrator can revoke or the agreed exports. Do not share a personal password to make the transfer convenient. Record which version arrived, when it arrived, its counts and integrity checks, and any discrepancies.

Resolve an incomplete upload as a delivery problem. Resolve a complete upload that fails an agreed utility requirement through the acceptance process. The distinction prevents a storage receipt from being treated as evidence that the buyer has approved the asset.

When access is no longer needed, remove it under the agreement and applicable retention requirements. Identify who retains the permitted licensed copy and which working copies, samples or temporary access sessions must end. Deleting an access token is not the same event as ending the license's permitted use.

Company A should maintain five separate statuses: selected for a transaction, delivered, accepted, invoiceable and paid. Each needs its own evidence. A buyer's encouraging message can move a conversation forward without creating an invoice right; an invoice can be valid without money having arrived.

A linked email, chat and ticket corpus may have one acceptance decision because the buyer needs the parts together. Three export folders do not automatically create three independently invoiceable batches. Partial acceptance can make sense when the contract defines separately useful collections, their prices and their individual triggers. The acceptance and payment guide explains that distinction.

If a finding blocks acceptance, identify its criterion, affected version and required response. Apply the agreed correction, retest, scope-change, rejection or dispute procedure. Log the revised decision when it occurs. Do not infer a refund, replacement, free preparation or immediate payout from the fact that one of those remedies would be convenient.

Use the right start and finish for each timing estimate

The operational page's public timing guidance, updated November 11 and checked November 15, 2026, separates the stages below. These are indicative clocks to confirm at scoping, not a promised completion date for Company A.

Do not add every row into a universal first-contact-to-cash forecast. The agreement clock already begins with the first email; transfer can occur within the period counted from signing to payment. Preparation and negotiation can overlap, and a revised scope can cause rework. Ask the parties to map the specific transaction instead.

Preparation, sample evaluation and acceptance need their own agreed clocks. I would add the following to that schedule rather than borrow a code-licensing turnaround or treat the transfer estimate as the entire preparation job.

A useful working schedule looks like this, with the dates and durations filled in by the people responsible:

Missing administrators, unclear rights, incomplete exports, broken relationships, unresolved privacy questions and requests for new model uses each need different remedies. Make the blocker specific. “Data cleaning is taking longer” does not tell the owner whether to find a file, approve a change or decline a use.

Put the stages on one calendar

Here is an invented planning example for Archive B. Its people, permissions, conditions, dates, durations, correction, offer and outcome are fictional. Day 0 is its first inquiry. All intervals below are assumed elapsed calendar days, not Rightmark service levels, observed averages or a forecast for the three public lots. Each exchange requires the relevant authorization; sending a raw archive does not start a useful clock.

This invented transaction takes 70 elapsed calendar days, or ten weeks, from inquiry to cash. Signing to cash takes 28 days, and transfer takes seven. Preparation and transfer are separate; the commercial track overlaps inventory and sampling. Adding the row lengths would count overlapping time twice. The total describes this example only and was not obtained by adding the public fastest estimates.

If S2 fails, rights remain unresolved, the buyer adds a model test or D1 fails acceptance, stop the affected stage and agree the next evidence, scope, work, payer and deadline. The day-70 cash date no longer follows automatically. A repair can be possible without having a predictable duration, and a new use may call for a new agreement rather than another round of the same test.

Finally, check later requests against the license. A refresh, a new recipient or a different use may need a new decision. The end of a temporary access session, the end of exclusivity and the end of permitted AI use can be different dates. Put those dates and obligations somewhere the owner can still find after the original deal team moves on.

• Worksheet instructions (Markdown)
• Archive b invented calendar (CSV)
• Archive b invented calendar (MD)
• Archive b invented offer (JSON)
• Archive inventory (CSV)
• Company a custody flow (MD)
• Company a custody flow (MMD)
• Company a invented inventory (CSV)
• Company a worked reconciliation (JSON)
• Delivery and payment clock (CSV)

• Delivery manifest template (JSON)
• Offer comparison (CSV)
• Sample decision log (CSV)
• Sample permission worksheet (MD)

The first useful step is small: identify the tools, retained years, people responsible for checking them and the material that must stay out. From there, build a price and process around the copy a buyer can actually use and the rights you are willing to grant. Request a written operational-data scope with those details and your preferred licensing route. You do not need to send the raw archive to start.