Systems and records
Grounded resolution when a support screenshot contradicts the customer
By SourceX Editorial · Updated
10 sections
Short answer
Turn each disagreement into a conflict task. Quote the customer's claim, describe what the attachment actually shows, classify the conflict, name the system record that decides it, and fix the expected agent action, which is often to clarify rather than to pick a side. Then label the outcome only after a maturity window that catches reopenings and follow-up tickets, so the task rewards grounded resolution rather than a fast close.
Record timeline
Grounded resolution
Key takeaways
- A screenshot that contradicts the customer's words is a distinct task type, and most ticket datasets erase it by keeping only the final reply and status.
- Conflicts fall into a few types, such as version mismatch, paraphrased error, stale capture and wrong account, each with a typical controlling record.
- The image is evidence of one screen at one moment; capture time, account and environment must be confirmed from system records.
- Rebuild what the agent saw from comment-level attachments and the audit trail, including inline images that default exports can omit.
- Outcome labels need a maturity window longer than the solved-to-closed delay, joined to follow-up tickets and repeat contacts.
- Screenshots need their own privacy review, because text redaction never touches what is inside an image.
How do you build support tasks where the screenshot contradicts the customer's description?
Build each one as a conflict task, not as an ordinary ticket. A conflict task pairs the customer's words with what the attachment actually shows, names the kind of disagreement, records which evidence controls and why, and fixes the expected agent action and the outcome that was later confirmed. The point of support attachment grounded resolution is that the right answer depends on reading both the text and the image, and on knowing when neither is enough.
Most ticket datasets lose this. The resolution code says solved, the agent's reply quietly corrects the customer, and nothing records that the screenshot and the description disagreed. A model trained on that history learns to paraphrase the customer or to restate the agent's final reply, not to notice the conflict. Research benchmarks show why this matters: the SWE-bench Multimodal paper built software tasks from issues that contain images such as screenshots and found that systems that did well on text-only issues struggled with visual problem solving.
Three rules keep the tasks honest. The expected answer must be grounded in evidence the agent could see at the time. The customer is never labeled as wrong, only the claim as unsupported or contradicted. And the outcome comes from what happened after closure, not from the status that ended the conversation.
Which contradictions between words and images show up in real tickets?
Contradictions fall into a small number of types, and each type has a typical controlling source and a typical correct move. Classifying them first keeps the task set balanced instead of filled with the easiest case.
| Conflict type | What it looks like | What usually controls | Expected agent move |
|---|---|---|---|
| Version or environment mismatch | Customer says latest version; screenshot shows an older build or a test environment | Version or environment string in the image, checked against account or log data | Confirm the environment, then answer for the version actually in use |
| Paraphrased error | Customer describes a payment failure; screenshot shows a permissions error code | Exact error text in the image | Work from the code shown and say why it differs from the description |
| Different screen or step | Customer describes step 3; screenshot shows a settings page from step 1 | Customer's stated goal, with the image as evidence of where they are | Locate the user in the flow before giving steps |
| Stale capture | Screenshot predates a fix or a configuration change | System records and change history for the relevant time | Ask for a current capture or check logs before acting |
| Wrong account or user | Screenshot shows another workspace, tenant or user name | Account records tied to the requester | Clarify which account is affected; do not act on the other one |
| Partial or cropped image | Error banner cut off, or key field hidden | Neither, until the missing part is seen | Request the specific missing element, not a generic new screenshot |
| Attachment from another case | Image belongs to a merged or forwarded thread | The originating ticket and its timeline | Separate the cases and answer the requester's own issue |
| Description right, image misleading | Intermittent fault; screenshot taken when it worked | Customer report plus logs over the reported window | Investigate the window the customer describes |
Why the attachment does not automatically win
The image is strong evidence of what a screen showed at one moment, not of what is true now or for this account. Treating every screenshot as ground truth creates a different bias, where a model learns to dismiss customers whenever a picture is attached.
Time is the usual weak point. Screenshots are often taken before the ticket, sometimes days before, and many carry no reliable capture time. When an attachment is a photo from a camera or phone, its Exif metadata may help, but Exif's DateTimeOriginal tag records local time without a zone; the standard added a separate OffsetTimeOriginal tag for the offset from UTC in version 2.31, so older files may lack it. Record the capture time and its source in the task, and mark it unknown rather than inferring it from the upload time.
Scope is the other weak point. A screenshot can show the right product but the wrong tenant, a staging site or an administrator view that the customer does not normally see. The task should name the system record that confirms which account and environment the image shows, and the expected answer should depend on that record when the two conflict.
The multimodal evidence conflict task record
This is the asset: one record per conflict, with every field filled from the ticket history or marked unknown. It is built to be used as is for an evaluation set or a supervised task, and the source column assumes a Zendesk-style helpdesk; map it to the equivalent objects in another system.
At a glance
The multimodal evidence conflict task record
| Field | Content | Where it comes from | Rule |
|---|---|---|---|
| customer_claim | Verbatim span of the customer's description | First public comment from the requester | Quote; never summarize the claim |
| attachment_ref | Comment id, file name, content type and inline flag | Attachment objects on the comment | Include inline images explicitly; they can be missing from default lists |
| attachment_observation | What the image shows, written by a reviewer | Human review, with OCR text kept separately | Describe, do not interpret; note crops and blur |
| conflict_type | One value from the conflict taxonomy | Reviewer classification | Second reviewer on a sample; disagreements logged |
| controlling_evidence | The record that settles the conflict, or none | Account data, logs, change history or the image | State why it controls; none is a valid value |
| policy_version | Support policy or macro version in force | Help center or macro history at the ticket date | Expected answer follows the version in force then |
| expected_action | Clarify, act on image, act on description, escalate or refuse | Agent reply and internal notes, reviewed | Must be justifiable from evidence available at the time |
| resolution_label | Confirmed fixed, workaround, not reproduced, reopened or unknown | Status history, follow-ups and later contacts | Set only after the maturity window |
| privacy_status | Cleared, redacted or excluded, with reviewer | Privacy review log | Excluded tasks keep their id and reason |
Tracing attachments, corrections and handoffs in the helpdesk history
A conflict task is only as good as its reconstruction of what the agent saw, in what order. In Zendesk, attachments hang off individual comments: each attachment object carries a content type, a file name and an inline flag, and inline images are excluded from a comment's attachment list unless the request asks for them with include_inline_images. Each comment is also public or an internal note, which matters because an internal note may record the conflict even when the public reply does not.
Corrections and handoffs live in the ticket's audits. Zendesk describes audits as a read-only history of every update to a ticket, where each audit holds one or more events such as field changes, comments and notifications. For extracting change history at scale, the same documentation points to the incremental ticket event export rather than repeatedly paging the list-all audits endpoint, which can skip records. Use the audit trail to place each attachment and each status change on one timeline, then cut the task's inputs at the moment the agent made the decision being labeled.
Other helpdesks have their own equivalents: case feeds and email messages in a CRM service module, issue history and attachments in a service desk. The principle is the same: rebuild the evidence timeline from history records, not from the ticket's final state.
What proves the customer's problem stayed resolved after the ticket closed?
Proof comes from what happened after closure: no reopen, no follow-up ticket and no repeat contact about the same problem within a defined window. Closing status alone proves only that the conversation ended.
Zendesk's defaults show why the window matters. A standard automation closes a solved ticket four days after it was set to solved, and administrators can change that delay to between one hour and 28 days. A solved ticket reopens when the customer replies, but a closed ticket cannot be reopened; a later reply creates a new follow-up ticket instead. Through the API, a follow-up created from a closed ticket carries via_followup_source_id pointing to the original, which lets you join the two.
So set a maturity window longer than the solved-to-closed delay, collect follow-ups by their source link and repeat contacts by requester and product area, and only then assign resolution_label. A conflict that the agent resolved by trusting the screenshot, followed a week later by a follow-up ticket describing the same fault, is evidence that the description was right.
Worked example (illustrative): the export button that was there all along
This case is illustrative; the product, ticket and timings are invented to show how the record is filled.
A customer writes that the export button disappeared after the latest update and attaches a screenshot. The image shows a reports page with no export button, a version string two releases old in the footer and a workspace name that differs from the requester's main account. An internal note from the first agent says the customer has two workspaces. The second agent replies that exports in that workspace are disabled by an administrator setting and links the help article for that setting.
The record is filled as follows. conflict_type is wrong account or user, combined with version mismatch. controlling_evidence is the account record showing the setting disabled in the second workspace. expected_action is clarify, then act on the image, because the image identified the real workspace. The ticket is solved, then closed four days later, and no follow-up arrives within the 30-day maturity window chosen for this illustration, so resolution_label is confirmed fixed.
The exception teaches as much. In a second illustrative ticket the screenshot shows the export button present, and the agent closes the case as not reproduced. Ten days later a follow-up ticket from the same requester reports the same fault with a timestamped error. The description was right and the image was a capture taken when the fault was absent, so conflict_type becomes description right, image misleading, and the expected action in the task is investigate the reported window, not close.
Screenshots carry other people's data: the privacy pass before tasking
Screenshots routinely contain people who are not the requester: other users' names, email addresses, customer records in a list view or a colleague's chat. Text redaction does not touch any of that. Before a ticket becomes a task, the attachment itself needs review.
Helpdesk tools help but have limits. Zendesk's native redaction in the agent workspace can remove text, attachments and images from ticket comments, and the help center notes that redaction cannot be reversed, that merge notes cannot be redacted and that side conversations remain unredacted on the external channels where they were delivered. Record privacy_status per task, exclude any task whose controlling evidence cannot be shown without third-party data, and keep the exclusion reason.
Rights are a separate check from privacy. Being able to open a ticket in the helpdesk is access, not a right to license its contents; customer-written text and customer-supplied files may also be covered by the company's terms of service or customer contracts. De-identifying an image does not settle who may use it, so counsel reviews the applicable terms before a dataset leaves the company.
When to drop a conflict task, narrow it or ask for review
Not every conflict makes a fair task. These conditions call for dropping the ticket, narrowing what the task asks, or sending it to a customer operations lead or privacy reviewer.
- The controlling evidence is gone: logs past retention, an account since deleted or an attachment missing from the export. Keep the task only if the expected action is to ask for evidence.
- The image cannot be cleared for privacy without removing the very detail that creates the conflict.
- The conflict turns on a policy that changed during the ticket and the version in force cannot be established.
- The resolution has not matured: the ticket is solved but not closed, or the maturity window has not passed.
- Reviewers disagree on the conflict type after discussion. Record the disagreement and leave the task out of scored sets.
- The ticket involves account security, payment details or legal threats. Route to the owner of that process rather than tasking it.
Attachment-grounded support histories in a licensing review
Support histories where attachments, internal notes and follow-ups are intact are the raw material for this kind of task, and they are often the records companies are asked about when licensing support data. Support teams that license attachment-rich ticket histories through SourceX have the rights review, contracting, data preparation, delivery and payment handled for the records they select. A review of screenshot-heavy tickets looks at exactly the questions above: whether attachments survived export, whether outcomes can be confirmed after closure and what the privacy pass removes.
A company can check whether it qualifies through a short assessment with no data upload. Any records considered later go through rights, privacy, quality and buyer-fit review before an agreement is signed.
Frequently asked questions
Should a support model trust the screenshot or the customer's description?
Neither by default. The screenshot shows one screen at one moment, and the description shows what the customer experienced. The right answer depends on which system record confirms the account, version and time, and sometimes the right action is to ask for the missing piece.
How do you tell whether an image was needed to resolve a ticket?
Ask whether a trained agent reading only the text could have reached the confirmed resolution. If an error code, version or account name appears only in the image, the image was needed. Conflict tasks are a subset where the image and text point in different directions.
Why do inline images go missing from support ticket exports?
Some helpdesk APIs list inline images separately from attached files. In Zendesk, inline images are excluded from a comment's attachment list unless the request includes them explicitly. Check that both kinds are present before building tasks.
How long should you wait before labeling a support ticket resolved?
Longer than the helpdesk's solved-to-closed delay, and long enough to catch follow-up tickets and repeat contacts. In Zendesk the default delay is four days and can be set up to 28 days. Pick a window that fits your product's usage cycle and apply it to every task.
Can screenshots with customer data be used in AI training tasks?
Only after a privacy review of the image itself and a rights check of the applicable customer terms. Text redaction misses names and records inside images. Exclude tasks where the conflict cannot be shown without third-party personal data.
Sources
- The SWE-bench Multimodal paper evaluates AI systems on software issues that include images such as screenshots and reports that top systems struggle with visual problem solving. Source
- The Exif standard defines DateTimeOriginal and a separate OffsetTimeOriginal tag holding the offset from UTC of DateTimeOriginal. Source
- Zendesk ticket comments carry attachments with content type, file name and an inline flag, inline images are listed only with include_inline_images, and each comment is public or an internal note. Source
- Zendesk ticket audits are a read-only history of all updates to a ticket, and Zendesk recommends the incremental ticket event export for capturing change data. Source
- Zendesk closes solved tickets after four days by default, adjustable from one hour to 28 days, and a reply to a closed ticket creates a follow-up ticket. Source
- The Zendesk Tickets API creates a follow-up for a closed ticket through the via_followup_source_id parameter. Source
- Zendesk agent workspace redaction can remove text, attachments and images from tickets, cannot be reversed and has documented gaps for merge notes and side conversations. Source
See if your company qualifies
A short company assessment. No data uploads are needed.