The brief arrives looking finished. It has a tidy summary, five recommendations, a confident timeline, and headings that make it easy to skim. Ten minutes into review, somebody asks where the customer evidence came from. Nobody knows. Another person notices that the recommendation quietly assumes a budget the team does not have.
The document took 12 minutes to generate and an hour to untangle.
This is one of the less obvious ways AI creates work. Its output can look complete before it is useful. The polish invites a light review, while the missing context, judgment, and evidence travel downstream to the next person.
Teams often respond by tightening prompts or adding another approval step. Both can help. Neither answers the more basic question: have we agreed what good work looks like here?
A short quality-calibration meeting can answer it. This is not a recurring review of everything AI produces. It is a focused session held once for a repeated kind of artifact, then revisited when the work changes. The aim is to create a shared bar that people and AI tools can actually use.
A small study makes the case for saying the standard aloud
On October 27, 2026, Atlassian's Teamwork Lab published results from a pilot involving 31 employees across five workplace teams. Participants spent 60 minutes identifying what weak AI-generated pages looked like, what made a page good, which principles they shared, and how they would maintain quality.
Immediately after the workshop, all participants agreed that they understood what their team considered a high-quality page, up from 62 percent beforehand. Two weeks later, 67 percent agreed that their team's AI-created pages met a high standard, compared with a 40 percent baseline. The researchers explicitly note that the results were not statistically significant because the pilot was small. That caveat matters. This is a promising practice to test, not a universal law. Atlassian Teamwork Lab
The most useful idea is also the simplest: quality becomes easier to produce and discuss when a team makes it observable.
"Make it better" is not a standard. "Put the recommendation first" is clearer, but still incomplete. Better standards explain what the artifact must enable, what evidence it needs, where judgment belongs, and what the recipient should be able to do next.
Calibrate one artifact, not all knowledge work
Do not invite the whole company to define good work in the abstract. Choose one artifact that is produced often enough to matter and reviewed inconsistently enough to cause friction.
It might be a customer-call summary, project brief, incident update, interview scorecard, launch plan, or weekly executive note. Bring three examples: one that helped the work move, one that looked polished but created follow-up effort, and one that sits somewhere between them.
Remove confidential details, but keep the examples real. A fictional perfect document encourages vague preferences. A real artifact reveals practical disagreements: whether executives need source links, whether an action item requires a due date, or whether a customer quote can support a product decision without broader evidence.
The meeting needs the people who create the artifact and the people who must act on it. If only authors attend, the standard may optimize for ease of production. If only recipients attend, it may demand a level of detail that makes the artifact too expensive to create.
Ask four questions that expose the hidden work
Give the group 10 quiet minutes to inspect the examples before discussing them. Independent reading reduces the chance that the first senior voice becomes the team's standard by default.
Then work through four questions.
1. What decision or action should this artifact make easier?
A useful document has a job. A customer-call summary might help an account team prepare the next conversation. A project brief might let a sponsor approve scope. An interview scorecard might support a hiring decision without forcing the panel to reconstruct the conversation.
Name that job in one sentence. If the group names three unrelated jobs, it probably needs three views or a more explicit hierarchy. Trying to serve every reader equally is how short artifacts become long and still leave the important person searching.
2. What must be traceable?
List the claims that require a route back to evidence. Numbers need sources. Quotations need attribution. Decisions need the meeting or owner that authorized them. A recommendation should distinguish observed facts from interpretation.
Traceability does not mean attaching a transcript to every sentence. It means a skeptical reader can check consequential claims without starting an investigation.
3. Where must uncertainty remain visible?
AI often smooths the rough edges that tell a careful reader where judgment is still needed. Decide how the artifact should mark assumptions, conflicting evidence, missing input, and low-confidence attribution.
For example, a sales summary might separate "customer stated" from "account team inferred." A project update might label a date as committed, estimated, or blocked. These distinctions are small, but they prevent a plausible sentence from hardening into false certainty.
4. What proves the handoff is complete?
Define the minimum state in which the recipient can act. An action item may need an owner, an accepted due date, and the reason it matters. A decision may need the chosen option, the alternatives rejected, and the condition that would reopen it.
This is where many polished drafts fail. They describe what happened but leave the next person to convert description into work.
Turn preferences into a compact quality contract
End the meeting with a one-page quality contract. Keep it short enough to sit beside the prompt, template, or automation that creates the artifact.
For a customer-call summary, it might say:
- Purpose: prepare the account team for the next customer commitment.
- Required evidence: direct customer statements are attributed; product and pricing claims link to an approved source.
- Judgment boundary: separate what the customer said, what the team inferred, and what still needs confirmation.
- Complete handoff: every next step has an owner who accepted it, a date or trigger, and enough context to act.
- Do not include: generic conversation summaries that do not affect the account.
Notice what is missing: instructions to sound intelligent, comprehensive, or professional. Those words invite cosmetic compliance. The contract describes usefulness.
Assign one person to maintain it, and set a review trigger rather than a weekly calendar event. Revisit the standard when recipients repeatedly ask the same follow-up question, the artifact changes audience, a new data source becomes available, or the workflow gains a new AI tool.
Review the system, not the person
When weak output appears, compare it with the contract before blaming the author or the model. Was the standard available? Did the source material contain what the standard required? Did the workflow leave enough time for human judgment? Did an automation remove an uncertainty label while reformatting the document?
This keeps feedback specific. "The recommendation has no evidence link" is actionable. "This feels like AI slop" is mostly a verdict.
It also protects against a subtler failure: making the reviewer responsible for reconstructing the author's thinking. The creator of an AI-assisted artifact still owns the result. AI may draft, organize, or surface omissions, but the handoff is not complete until someone has checked it against the shared bar.
The best quality-calibration meeting eventually makes itself unnecessary. People internalize the distinctions, prompts improve, templates carry the right fields, and reviewers stop debating the same preferences. One hour has created a small piece of operating infrastructure.
Caspi can support that work where the standard meets the live conversation. Its real-time recap, suggested questions, contextual chat, proactive flags from connected tools, post-call action items, and persistent meeting memory can help a team notice missing evidence, clarify uncertainty, and preserve accepted next steps while the relevant people are still present. The copilot can help carry the contract through the meeting. The team still decides what good means.