Skip to content

AI uses for records

Can it do X? Grading answers in a sales capability boundary dataset

By SourceX Editorial · Reviewed by Noah Loul ·

10 sections

Short answer

A sales capability boundary dataset records what happened when a buyer asked for something the product could not do: the capability on file that day, the seller's answer, who was consulted, what reached the contract and what later shipped. Grade the answer for accuracy and escalation, not persuasion. Base rates matter: of 1,056 GitLab feature requests labeled customer-raised in 2023, 18.3% were marked Complete within six months.

Grader rubric

Sales capability boundary

This page covers can it do X? Grading answers in a sales capability boundary dataset.Illustrative

Key takeaways

  • The unit is the capability request, not the call: one row per ask, with the product's capability on that date beside the seller's answer.
  • Label answers on a ladder from verified yes and documented workaround through logged request to unapproved commitment and overstatement.
  • A confident yes can create obligations: IFRS 15 counts promises implied by specific statements, and UCC section 2-313 treats affirmations of fact about goods as express warranties.
  • Of 1,056 GitLab feature requests labeled customer-raised in 2023, 54.7% were still open in October 2026 and 99 had slipped past at least one milestone.
  • Grade against the record of the day and the authority behind the answer, never the deal outcome.

What should a dataset record when a buyer asks for a capability the product lacks?

One row per request, holding the request itself, what the product could actually do on that date, the seller's answer, who was consulted before answering, what reached the signed documents and what happened afterward. The answer is the label. The aim is to teach and test an AI system to answer capability questions accurately and route them to the right people, not to sound convincing.

Call the moment a capability boundary: the point in a sale where a prospect asks for something outside the product as it exists, such as a security control, an integration, a report format, a deployment option or a volume limit. A revenue team's AI assistant meets it constantly while drafting replies, summarizing calls and suggesting next steps. A model trained on whatever closed will learn that confident yeses win.

Pages on promise tracking follow commitments after they are made. This one covers the step before: whether a commitment should have been made at all, and how to label the answer when it was.

Why is a confident yes the wrong thing to teach?

Because a seller's statement about a missing feature can become an obligation the company never approved, while the persuasion it bought shows up only as a closed deal. Three public references set the stakes, and none of them rewards fluency.

Revenue accounting comes first. IFRS 15 paragraph 24 says a contract's performance obligations may include promises implied by an entity's customary business practices, published policies or specific statements, if at contract inception they create a valid expectation that the entity will transfer a good or service. The IFRS Interpretations Committee's 2022 decision on a software reseller marks the edge of that test: pre-sales advice given before the contract began did not create an extra promise. A specific roadmap statement a customer relies on at signing is the kind of fact that test looks at, which is why finance teams care what sellers say.

Sale-of-goods law points the same way. Under section 2-313 of the Uniform Commercial Code, an affirmation of fact or promise by the seller that relates to the goods and becomes part of the basis of the bargain creates an express warranty, while a statement that is merely the seller's opinion or commendation does not. Whether UCC Article 2 reaches a software subscription depends on the contract and the court, but the label set should still separate statements of fact about capability from sales enthusiasm.

Large vendors already tell buyers not to purchase the roadmap. Oracle's standard safe harbor statement calls its product direction not a commitment to deliver any material, code or functionality, says it should not be relied upon in making purchasing decisions, and says it may not be incorporated into any contract. A 2017 Salesforce release preview deck told customers to make purchase decisions based on features currently available. A dataset that rewards answers those vendors' own disclaimers rule out is training the wrong habit.

How often does a requested feature arrive? 1,056 GitLab requests traced

Slowly and unevenly, and that base rate is what any roadmap answer should be graded against. GitLab runs its product backlog in public and asks sales and customer teams to record customer interest on the relevant issues, so its tracker shows what became of features customers asked for.

We pulled every public issue in the gitlab-org/gitlab project created in 2023 that carries both the customer label and the type::feature label: 1,056 requests. GitLab's handbook tells customer-facing staff to add a customer's interest to the issue, link the Salesforce account or opportunity and explain why the customer wants it. Outcomes were read on 2026-10-27, when every request had at least 1,030 days of history, from each issue's state and work-item status; GitLab's product development flow sets the Complete status after the feature is deployed to production and any needed verification is done.

Read the count with its limits. GitLab is one open-core company, so its delivery pace is not an industry norm; labels are as they stood at retrieval, so some issues may have gained the customer label after filing; and Complete is GitLab's own status, not an independent check that code shipped. The next evidence worth having is the same join inside a single company, with CRM requests linked to tracker outcomes, which few firms publish.

By then 424 requests (40.2%) were closed as Complete, 41 as Duplicate and 13 as Won't do, and 578 (54.7%) were still open, 439 of them in the New status. Only 135 (12.8%) reached Complete within 90 days of filing and 193 (18.3%) within 180; among those completed, the median wait was 220 days. Of the open requests, 421 sat in a milestone named Backlog and 42 in one named Awaiting further demand. A schedule was no promise either: 99 requests carry at least one missed:x.y label, which GitLab's automation adds when an issue's milestone expires before the work is done, and 40 slipped twice or more. The outcomes, with what each means for a seller's answer:

How often does a requested feature arrive? 1,056 GitLab requests traced
Outcome on 2026-10-27RequestsShare of 1,056What it means for a sales answer
Complete within 180 days of filing19318.3%A within-six-months answer would have held for fewer than one request in five
Complete after 180 days23121.9%Delivered, but later than many sales cycles
Closed as Duplicate or Won't do545.1%Folded into another request or declined
Open, status New43941.6%Never moved past intake in almost three years
Open, in triage or development13913.2%Still moving, not yet delivered
Slipped past at least one milestone (any outcome)999.4%Even scheduled work missed its first date
A rising step curve of the share of 1,056 customer-raised GitLab feature requests from 2023 marked Complete, by days since filing: 12.8% by 90 days, 18.3% by 180, 25.7% by 365, 33.6% by 730 and 38.2% by 1,030, beside a panel noting 54.7% still open.
Customer-labeled feature requests in gitlab-org/gitlab created in 2023, statuses read 2026-10-27. Source: GitLab public API; SourceX count.

Seven labels for the answer a seller gave

Label the answer against the capability record of its date, on a ladder that runs from verified yes to overstatement. The first four labels are accurate answers, the fifth is accurate only with approval, and the last two are errors whatever the deal outcome.

Two judgment calls recur. A workaround counts as accurate only if its limits were stated and someone who knows the product confirmed it before the answer. A logged request counts as escalation only if the link from the deal record to the product tracker exists, with a timestamp. Each label and the evidence it needs:

Seven labels for the answer a seller gave
LabelDefinitionEvidence requiredGraded as
Verified yesThe capability existed in the version the buyer would receiveRelease notes or documentation dated on or before the answerAccurate
Documented workaroundNo native feature, but a supported path met the need and was called a workaroundWorkaround documented or confirmed by a solutions engineer before the answerAccurate if its limits were stated
Clear noThe seller said the capability does not exist and named no dateAnswer text plus the capability record of that dateAccurate
Logged requestA clear no or workaround, with the request filed with productLink from the deal record to the product tracker, with a timestampAccurate, with escalation
Approved commitmentA delivery commitment authorized by product and written into the order or contractApprover role, approval date and the clause or order lineAccurate only if every element is present
Unapproved commitmentA date or promise given without product approval or left out of the documentsAnswer text with no matching approval or clauseError
OverstatementThe seller said the capability existed when it did notAnswer text contradicted by the capability record of that dateError

Who may turn a request into a commitment?

Only someone with authority over the roadmap, and the dataset must record who that was. GitLab's Customer Issues Prioritization Framework draws the line in public: customer-facing staff flag a request as Blocker when the product cannot be sold to the customer without it, Retention when an existing customer cannot stay without it, and Fast Follow only when GitLab team members have set delivery expectations with the customer. If a workaround lets the customer buy, the framework says, it is not a blocker, and general feature requests without commitments do not get the Fast Follow label.

The framework also builds in an escalation clock: a triage bot reminds team members when a flagged issue still has no milestone or planning label after five business days, while product teams keep the final say on priority. GitLab's product management guidance asks customer-facing staff to learn why the customer wants the feature, what problem it solves, which parts matter, whether workarounds were tried and how important it is before filing. Those five answers make good dataset fields. For declines, the feedback process for GitLab's agent platform asks product managers to choose one of 15 won't-do reasons, from out of scope and low priority to platform limitation and regulatory blocker, which is the kind of reason field that tells a model why a no was right.

Three fields carry authority in the dataset: the role that answered, the role consulted before answering (product manager, solutions engineer or none) and the role that approved any commitment, each with a timestamp. A commitment with no approver is an unapproved commitment, whoever made it.

The revenue decision evidence map for a capability request

Map each decision in the request's life to the system that holds its evidence, and decide before extraction which fields travel and which stay behind. The map keeps decision evidence and drops contact identities:

The revenue decision evidence map for a capability request
Decision pointSystem of recordFields to keepFields to remove or band
Request raisedCall notes, email thread or CRM activityRequest text, date, deal stage, product areaNames, email addresses and phone numbers of buyer contacts
Capability checkedRelease notes, documentation, product trackerVersion available, documentation link, date checkedInternal code names the label does not need
Answer givenEmail, call summary or proposalAnswer text, answer date, answering roleThe seller's name, replaced by a role code
EscalationProduct tracker, internal comment or chat channelLink to the request issue, consulted role, response timeCustomer names in internal comments
CommitmentOrder form, contract or statement of workClause or order line, approver role, approval datePrices, discounts and contract value, banded only if counsel agrees
OutcomeTracker status, release notes, CRM close fieldsDelivery date or decline reason, deal stage at closeAccount identifiers beyond a stable pseudonym

Could a won deal prove the answer was right?

No. A deal can close because of budget timing, a competitor's stumble or an internal champion while an answer inside it was an overstatement, and another can be lost after a perfectly accurate no. Grade the answer against the capability record of its date and the authority behind it, and keep the deal outcome in a separate field.

The GitLab base rates make the point concrete. Had a seller answered each 2023 request with a six-month delivery date, the tracker says the answer would have held for 193 of the 1,056. A model trained only on closed-won deals sees that the date helped close the sale; it never sees the 863 requests where the date would have been wrong.

Keep three outcome fields apart: whether the deal closed, whether and when the capability was delivered, and whether the customer later raised the commitment in a renewal or dispute. Each answers a different question, and none of them changes the label on the answer.

Illustrative: one request, three replies, three labels

The example is invented to show the labels and describes no real company. On day 12 of an evaluation, a prospect's security lead asks whether the product supports automatic user provisioning from the identity provider's groups. On that date the release notes show group sync is not available, and an open product issue tracks it with no milestone.

Reply A: Yes, we support that. Label: overstatement, because the capability record of day 12 contradicts it. Reply B: Not today; our solutions engineer can show you scheduled provisioning through the API as a workaround, and I have added your use case to the product request. Label: logged request with a documented workaround, provided the engineer confirmed the workaround and the link sits on the deal record. Reply C: It ships next quarter. Label: unapproved commitment, unless a product approver signed off and the order form carries the clause, in which case it becomes an approved commitment.

Suppose the deal closes after Reply C and the feature ships seven months later. The deal outcome is a win and the delivery is late. The answer remains an unapproved commitment, because nothing that happened afterward changes what the seller knew and was authorized to say on day 12.

Stripping names and prices before capability records travel

Sales records name people and expose commercial terms, and the labels above need neither. Replace buyer contacts and sellers with role codes, keep a stable pseudonym for each account so its requests can be grouped, and drop prices, discounts and contract values unless counsel approves coarse bands.

Operational teams already separate the two. GitLab's prioritization framework tells staff to record customer impact in internal notes, never to put customer names, revenue figures or contract values in issue comments, and to keep those details in the linked system such as Salesforce. A dataset extracted the same way inherits that separation.

Customer agreements may also limit reuse of call notes, proposals and order forms. Check the confidentiality and data use terms for each account before its records enter the set; removing names does not settle whether the records may be licensed.

Requesting capability-boundary records through SourceX

A team building or testing a sales assistant can describe this dataset as a requirement instead of hunting for it. SourceX does not keep a catalog of such records; it looks for them at U.S. companies once a buyer asks, working from the scope, volume, language, format, timing and licensing terms the buyer sets in the request form.

A useful request names the unit (one row per capability request), the label set above, the minimum share of rows with a capability record dated on or before the answer, the outcome fields required and an acceptance test, for example that two reviewers labeling a random sample agree on most rows and that every commitment label carries an approver field. Whether a supplier's records pass that test is decided on the actual data, not promised in advance.

Frequently asked questions

What is a capability boundary in a sales conversation?

It is the point where a prospect asks for something the product cannot do as it stands: a missing integration, control, report or limit. How the seller answers, and whom they consult first, decides whether the company takes on an obligation it did not plan.

Should an AI sales assistant ever mention the roadmap?

Only with information the company has approved for sharing, labeled as not a commitment, and without a date unless product has approved one that will be written into the order. Oracle's standard safe harbor statement, for example, says its product direction should not be relied upon in making purchasing decisions.

Can a sales rep's promise become a contract obligation?

It can matter even when it is not in the signed text. Under IFRS 15, specific statements can imply promises if they create a valid customer expectation at contract inception, and under UCC section 2-313 an affirmation of fact about goods can create an express warranty. Counsel and accountants decide how a particular statement is treated.

How long do customer feature requests take to ship?

It varies by company, and few publish the data. In GitLab's public tracker, 18.3% of 1,056 feature requests labeled as customer requests in 2023 were marked Complete within 180 days of filing, and 54.7% were still open in October 2026.

What fields does a sales capability boundary dataset need?

One row per request with the request text and date, the capability record of that date, the answer and answering role, the consulted and approving roles with timestamps, any contract clause, and later delivery and deal outcomes. Contact identities are replaced by role codes and accounts by stable pseudonyms.

Sources

  • GitLab's Customer Issues Prioritization Framework defines Blocker, Retention and Fast Follow labels, says a workaround that allows purchase is not a blocker, reserves Fast Follow for set delivery expectations, adds a five-business-day triage reminder and bars customer names, revenue and contract values from issue comments. Source
  • GitLab's product management guidance asks customer-facing staff to link the Salesforce account or opportunity when sharing customer feedback and to ask why the customer wants a feature, what problem it solves, which parts matter, whether workarounds were tried and how important it is. Source
  • GitLab's engineering workflow says scheduled automation moves open issues past an expired milestone to the next one and labels them missed:x.y, adding missed-deliverable when the Deliverable label is present. Source
  • GitLab's product development flow maps former workflow labels to statuses including Complete, Duplicate and Won't do, and sets Complete after the feature is deployed to production and any needed verification is complete. Source
  • GitLab's feedback framework for its agent platform lets sales team members submit feature requirements from sales conversations and has product managers decline requests with one of 15 won't-do reason labels. Source
  • The public GitLab REST API returned 1,056 gitlab-org/gitlab issues created in 2023 carrying both the customer and type::feature labels, with state, dates, labels and milestones. Source
  • The public GitLab GraphQL API returned each of those issues' work-item status, such as Complete, Duplicate, Won't do or New. Source
  • IFRS 15 paragraph 24: performance obligations may include promises implied by customary business practices, published policies or specific statements that create a valid customer expectation at contract inception. Source
  • The IFRS Interpretations Committee's 2022 agenda decision found pre-sales advice already given before contract inception created no valid expectation of a further good or service from a software reseller. Source
  • UCC section 2-313: an affirmation of fact or promise by the seller relating to the goods that becomes part of the basis of the bargain creates an express warranty; mere opinion or commendation does not. Source
  • Oracle's safe harbor statement says its product direction is not a commitment to deliver any material, code or functionality, should not be relied upon in making purchasing decisions and may not be incorporated into any contract. Source
  • A 2017 Salesforce release preview deck's safe harbor told customers to make purchase decisions based on features that are currently available. Source

Tell us what data you need

Share scope, volume, format, timing and licensing needs. Datasets are sourced on request.

See if your company qualifies