The Backlog Problem Is Not New, But The Solution Is
Every engineering team has a backlog they cannot work down. The half-life of a Jira ticket past 90 days is essentially infinite, it sits there, stale, indefinitely. Backlogs grow because the rate of intake exceeds the rate of close, and humans cannot read 2,000 tickets to figure out which 100 are worth doing this quarter.
AI-assisted backlog grooming is a workflow that uses a coding agent (plus a classifier) to do three things humans cannot: read every ticket, propose a disposition, and execute on the ones that need execution. Done well, it converts a backlog from a source of guilt into a source of throughput.
This is the playbook.
The Four Dispositions
Every ticket sorts into one of four buckets:
Do. The ticket is actionable, valuable, and in scope. The AI either takes it or queues it for human pickup based on complexity.
Defer. The ticket is actionable but not now. Pushed to a future quarter with a deferral reason. Reviewed when that quarter arrives.
Convert. The ticket is too big or too vague to be a ticket. It needs to be split, scoped, or turned into a doc. Routed to a tech lead with the AI's proposed split.
Close. The ticket is stale, duplicate, no longer relevant, or describes behavior that has since been fixed by something else. Closed with a reason.
In a typical 2,000-ticket backlog, the disposition splits look like 12% Do, 28% Defer, 18% Convert, 42% Close. The 42% Close is the most surprising number. Most backlogs are 40-50% dead weight.
The Classifier Stage
A cheap LLM (we use Haiku) reads each ticket and proposes a disposition with reasoning. The classifier has access to:
- The ticket text.
- The ticket's age and last-update date.
- The labels and components.
- Whether the originating reporter is still on the team.
- For "Close" candidates, whether any merged PR mentions the ticket.
The classifier outputs a disposition, a confidence, and a one-line reason. At the team's confidence threshold (we recommend 0.85), high-confidence Close and Defer dispositions can be auto-applied. Below threshold, a human reviews.
Time to classify 2,000 tickets: under 30 minutes. Cost: under $20. The human review of the under-threshold subset takes 4-6 hours. Total: a day of effort to triage a backlog that had been ignored for years.
What The AI Should Not Decide Alone
Three categories that always go to humans:
Anything labeled "customer." A customer-reported issue gets human disposition regardless of age or vagueness. The reputational cost of auto-closing a customer ticket is too high.
Anything in a security label. A stale security ticket might be stale because it was fixed, or stale because it was forgotten. The AI should not guess. Security team reviews.
Anything from a current team member. If the reporter is still on the team and the ticket has not been updated, the AI proposing "Close" is rude. Better to surface the staleness to the reporter and let them decide.
The Convert Workflow
When the AI proposes Convert, it also proposes the split. Example:
Original: "Improve the API."
AI's split:
- "Add pagination to the /users endpoint." (Do, AI-pickup candidate.)
- "Add bulk-update support to /orders." (Do, requires product input.)
- "Document the rate limiting rules." (Do, AI-pickup candidate for docs.)
- "Decide whether to deprecate v1 endpoints." (Convert further: needs an RFC.)
The original ticket gets closed with a link to the four replacements. Each replacement is independently triagable. This is the highest-leverage AI move on a vague backlog, it does the disambiguation work that humans avoid because it is unrewarding.
The Do-AI-Pickup Subcategory
Of the Do tickets, the AI proposes which ones it can pick up itself. Criteria:
- Single-service scope.
- Clear acceptance criteria (or the AI can derive them).
- Confidence above the AI's pickup threshold.
- Not in a category the AI has historically struggled with.
In our data, about 30-40% of Do tickets meet the criteria. The rest go to human assignees, sometimes with an AI-generated implementation plan attached for context.
Running The Grooming On A Schedule
Three cadences:
Daily intake triage. New tickets get a disposition within an hour of creation. Closes and defers proceed immediately at high confidence; everything else waits for the human triage queue.
Weekly backlog grooming. Stale Do tickets get re-evaluated. Tickets that have been Do for >30 days without pickup get a "still relevant?" comment. Defer tickets coming up on their deferral date get surfaced.
Quarterly bulk grooming. A full backlog pass with the classifier on everything. Catches the dead-weight tickets that accumulated and re-evaluates Defer tickets whose deferral date has passed.
What Surprised Us
Two patterns we did not expect:
The classifier is better at "close" than humans. Humans get sentimental about old tickets they created. The classifier does not. Auto-close acceptance rates land around 90%, humans agree with the classifier on what to close.
Convert is the highest-value disposition. "Improve the API" sat in the backlog for two years because nobody wanted to do the splitting work. The AI splits it in 30 seconds, and suddenly there are four actionable tickets, three of which are picked up within a week.
Risks
Three things that can go wrong:
Auto-closing real work. Mitigation: confidence threshold, explicit categories that bypass auto-close, and a 14-day "undo" window where any closed ticket can be reopened by the original reporter with one click.
Backlog churn. If every ticket gets bulk-edited weekly, watchers get fatigued. Mitigation: batch comments, suppress notifications for AI-driven changes by default, and give users a way to subscribe to AI commentary explicitly.
Loss of institutional memory. Some "stale" tickets are actually important context for future work. Mitigation: closed tickets are never deleted, and the AI's disposition reasoning is preserved in the ticket history.
Where To Start
Pick the worst-tended project in your tracker. Run the classifier on it. Read the first 50 dispositions and tune the prompt. Approve the high-confidence auto-closes. Read the convert proposals. Watch your backlog shrink by 40% in a week.
The hardest part is not the technology. It is letting go of the idea that the backlog itself has value just by existing.
For the agent's pickup integration, see Jira, GitHub, and Azure DevOps integration. For the engineering manager's perspective on rollout, see engineering manager playbook.
Frequently asked questions
How can AI help clear a large engineering backlog?
AI does three things humans cannot: it reads every ticket, proposes a disposition, and executes the ones worth doing. A cheap classifier reads each ticket's text, age, labels, and reporter status, then outputs one of four dispositions (Do, Defer, Convert, or Close), with a confidence score and a one-line reason. High-confidence closes and defers can auto-apply while everything below threshold goes to a human queue, converting a backlog from a source of guilt into a source of throughput.
What percentage of a backlog should actually be closed?
In a typical 2,000-ticket backlog the disposition split lands around 12% Do, 28% Defer, 18% Convert, and 42% Close. The 42% close rate is the most surprising number, but most backlogs carry 40-50% dead weight, stale, duplicate, no-longer-relevant tickets, or ones describing behavior already fixed by something else. Auto-close acceptance rates around 90% show humans largely agree with the classifier on what to close.
Can AI automatically triage Jira tickets safely?
Yes, with guardrails. Set a confidence threshold (around 0.85) so only high-confidence closes and defers auto-apply, keep closed tickets undeletable with the disposition reasoning preserved in history, and offer a 14-day undo window where the original reporter can reopen with one click. To suppress notification fatigue, batch comments and mute AI-driven changes by default. These mitigations protect against auto-closing real work and against backlog churn.
Which backlog tickets should AI never decide on its own?
Three categories always route to humans: anything labeled customer, because the reputational cost of auto-closing a customer issue is too high; anything in a security label, since a stale security ticket may be fixed or merely forgotten and the AI should not guess; and anything reported by a current team member, where the right move is to surface the staleness to the reporter rather than propose Close. Everything else can flow through the confidence-gated queue.
How does AI split a vague ticket into actionable work?
When the classifier proposes Convert, it also proposes the split. A ticket like 'Improve the API' becomes several scoped tickets (add pagination to /users, add bulk-update to /orders, document the rate-limiting rules, decide whether to deprecate v1), each independently triagable, some flagged as AI-pickup candidates. The original is closed with links to its replacements. This disambiguation is the highest-leverage AI move because it does the unrewarding work humans habitually avoid.
How often should you run AI backlog grooming?
Run three cadences: daily intake triage so new tickets get a disposition within an hour, weekly grooming that re-evaluates stale Do tickets and surfaces upcoming defer dates, and a quarterly full pass with the classifier on everything to catch accumulated dead weight. Of the Do tickets, about 30-40% meet the criteria for the agent to pick up itself. For how the agent connects to your tracker, see connecting Jira, GitHub, and Azure DevOps.
EnsureFix Customer Success
The EnsureFix customer success team works with engineering teams post-deployment, capturing the playbooks and metrics that separate successful rollouts from stalled pilots.