.webp)
Atlassian Insights: Your Metrics Need a Dictionary
Archived
Imagine asking three teams how many customers are active and receiving three different answers. Each answer is correct. Each team means something different by "active." Add AI, and those answers arrive faster. Progress, of a sort.
That is why the most interesting part of Atlassian Insights is not the conversational interface. It is the opportunity to agree on what the conversation is about.
Atlassian announced Insights on October 7, 2026 as an evolution of Atlassian Analytics, combining AI-assisted analysis with a shared data foundation. The useful question for platform owners is whether it can make answers easier to defend, not merely easier to produce.
What Atlassian announced, and what to verify first
Insights brings together a data catalog, business glossary, dashboards, and Rovo-powered analysis. Atlassian describes a connection between organizational work context and structured data so people can ask analytical questions in natural language.
The announcement describes a phased rollout for paid Atlassian Cloud customers, beginning after Team '26 Europe. That is not a statement that every customer already has access. Check your organization's rollout status and administrator settings before planning around the new experience.
There is a second distinction: availability of the app is not availability of every capability. Atlassian's Insights feature matrix lists the glossary, catalog, dashboards, and Rovo-powered analytics across Standard, Premium, and Enterprise, while data lineage and Atlassian Data Lake access are Enterprise-only.
Data lineage describes where data came from and how it is connected or transformed. Do not build an evaluation around that feature without checking the required plan. Likewise, confirm the connector and query capabilities needed for your actual sources rather than assuming all editions are interchangeable.
Atlassian's Insights product page also says Rovo Insights uses Rovo credits. Include expected analysis activity in the pilot's cost review. Our guide to Atlassian usage-based pricing explains the broader budgeting questions. Included access and unlimited consumption are different promises.
A glossary should settle arguments, not collect words
Atlassian's glossary documentation describes terms with definitions, owners, synonyms, related entities, and supporting documentation. Those are useful building blocks. They still need decisions behind them.
Consider an illustrative service-management question: "Did we resolve requests faster last month?" Before evaluating any AI-generated answer, agree on the reporting definition.
Does "resolved" mean the first resolution or the final resolution after reopening? Are you measuring calendar time or working time? Does the population include every request, only incidents, or only requests completed during the month? Which time zone defines the month?
These are choices for your organization, not universal Jira Service Management rules. A mathematically correct answer to the wrong definition is still the wrong management answer.
For a pilot, write a compact definition record containing the business question, calculation, included population, exclusions, reporting period, source, and accountable owner. Add one ordinary example and one edge case. Store details in the glossary and linked documentation where appropriate; this is a recommended governance record, not a claim that every item is a dedicated Insights field.
Then ask the owner to approve it. The difference between a glossary and an operating agreement is someone being responsible when the definition changes.
This is the practical next step in closing the gap between AI and business context. More context should mean more usable meaning, not simply more documents.
"Metadata only" is not the whole data-handling story
Catalog discovery and query execution are different operations. Atlassian's data-management documentation says catalog reads do not copy table contents. Executed analyses can nevertheless leave retained SQL, limited result data, conversation previews, and CSV result files.
The same documentation distinguishes source connections using a person's credentials from connections using administrator-configured integration credentials. Avoid reducing that model to "every query always runs as the individual user."
For your review, identify which credential each connection uses, what that identity may read, who can access resulting analyses, and what happens to retained results. Include exports and conversation content in the assessment. A restricted source does not remove the need to examine the life of its output.
Our Atlassian permissions audit discussion provides useful background. For Insights specifically, validate the documented connection model and test the actual setup. A broad permission slogan is not an access test.
Pilot one question you can already answer
Start with a decision that matters and a small, approved dataset. Choose a question with a manually verified answer. The point is to establish a baseline, not to give the new tool an easy demonstration.
Our earlier guide to Jira dashboards executives can actually use starts from the decision a dashboard should support. Apply that same discipline to conversational analytics.
Use three acceptance tests:
- A known answer. Ask a precisely defined question. Compare the result with the approved baseline. Inspect the population, date boundaries, exclusions, and calculation, not just the final number.
- An ambiguous question. Ask about an undefined measure such as "fast resolution." Your acceptance criterion should require clarification or an explicit, reviewable assumption. Do not count a polished guess as a pass.
- A restricted-data request. Using an authorized test account, ask for data that account should not receive. Confirm the expected denial or approved limitation. Treat unexpected disclosure as a reason to stop the pilot.
These are proposed acceptance criteria, not claims that Insights will pass each test automatically.
Next, change one condition at a time: the reporting month, a reopened request, a missing value, or the user's approved access. Check the source freshness as well. Save the question, expected result, observed result, and review notes so another person can reproduce the evaluation.
Track the effort required to verify an answer alongside the effort required to obtain it. Saving five minutes of report building is not a win if it creates twenty minutes of reconciliation. Those numbers are an illustration, not a benchmark.
Existing Analytics users should compare the experiences
Atlassian's Analytics-to-Insights transition guidance says existing dashboards, charts, and embedded Smart Links remain available. It also documents limitations in the Rovo-powered experience around creating legacy Analytics dashboards, modifying custom tables, and creating or changing scheduled exports. Existing scheduled exports continue to run.
That distinction deserves attention before anyone describes the change as a frictionless replacement. Inventory the reporting features your teams actually use. Verify the experience that supports them. Retaining existing work is not the same as preserving every editing workflow in the new interface.
I would prioritize a pilot where useful questions are delayed by repeated interpretation work and where someone can own the definitions. I would wait where existing reporting already supports decisions adequately, or where no one is prepared to resolve disagreements about the measures.
The Avaratak Take
Atlassian Insights is worth evaluating as an opportunity to make business definitions executable and reviewable. The mistake would be treating the glossary as cleanup to perform after the AI rollout.
Our recommendation is to fund the definition work as part of the pilot. Give a business owner responsibility for meaning and a technical owner responsibility for implementation. Require a clear path to challenge an answer, correct a definition, and retest the affected questions.
Judge success by whether people reach a defensible decision with less verification effort. Count unresolved definition disputes and repeated corrections alongside adoption. More generated answers can be a sign of curiosity; they are not, by themselves, evidence of better decisions.
A dictionary will not run your business. But agreeing on the words is a sensible place to start before asking AI to explain the numbers.
Product details checked October 16, 2026. Hero photo: Felicia Montenegro on Unsplash, used under the Unsplash License.