September 29, 2026
by
Jason Dugdale
A Support Ops Guide to Inventing Work
Explore with AI
This post was inspired by Sujith Jay Nair's article A Staff Engineer's Guide to Inventing Work, and adapts the idea of finding and justifying a team's next piece of work to Support Operations teams.
When an AI agent handles a support conversation, the customer still experiences every wrong answer, repeated step, and missed escalation. They may also report a bug or ask for a feature that would make their work easier, without anyone on your team seeing the request.
Support Ops remains responsible for finding opportunities to improve the experience and deciding which ones deserve attention. The work might be a knowledge correction, a change to the agent's instructions, or a product fix that removes the need to contact support in the first place. Before any of those tasks can enter the backlog, someone has to recognise the problem and understand why it matters.
"Inventing work" means finding an opportunity to make a positive change for your support team or customers, then building the case for acting on it. A difficult conversation, or example of a missed escalation pasted into Slack makes a problem visible, but recurring issues in conversations handled entirely by AI can go unnoticed. Finding the right work increasingly depends on recognising patterns your team would otherwise miss.
Read signals from across support
Operational metrics give you a place to start. AI resolution, escalations, reopens, backlog, and cost per conversation can show where performance has changed, with a rise in escalations prompting a closer look at the requests your agent struggles to handle.
Those numbers help you choose where to look, but understanding the cause takes more context. More escalations could mean the agent is struggling, the requests have become more complex, or a revised rule is correctly sending more cases to a person. Likewise, a higher AI resolution rate can coexist with a growing product problem if the agent answers each question without anyone noticing how often customers ask it.
Your support team adds detail that a dashboard can miss. We've historically relied on teammates reporting common requests and friction in internal processes to guide improvements to tooling and the customer experience. As AI support agents take on more of the volume, those informal reports cover a smaller share of what customers encounter, so Support Ops needs another way to keep track of emerging problems.
Find the patterns your team might miss
Customers describe what needs attention every day, often without submitting a formal bug report or feature request. A question about exporting data from several workspaces at once might reveal a missing capability, while repeated complaints about a payment step or error message could signal a product failure that needs investigation.
Each request can receive a useful answer while the underlying need remains. If customers repeatedly ask for the same workaround, the agent may "resolve" every conversation and still leave the team unaware of an opportunity to improve the product. Recognising that pattern gives Support Ops something to take to Product and Engineering, alongside the operational fixes it already owns.
Manually reading conversations can uncover these patterns, but it also creates another recurring task for Support Ops. Automatic recognition helps keep customer needs visible as volume grows, including the problems that never reach a human agent or a feedback board.
Build the case for one change
Finding a pattern gives you a candidate for work; deciding whether to act still requires judgment. Before turning it into a task, consider its scale, impact, and fit with the team's current priorities:
| Question | What to consider |
|---|---|
| How often does it happen? | Affected conversations and distinct customers in a defined period, with the total volume for context. |
| What is the impact? | Failed tasks, repeated effort, incorrect guidance, or a missing capability that affects how customers use the product. |
| Is it changing? | Whether the problem is new, growing, or recurring, accounting for changes in traffic and the types of requests. |
| How confident are we? | Whether the reports describe the same problem and whether it affects several customers or workflows. |
| What would a fix take? | A proposed change, an owner, the effort required, and dependencies. |
| Why does it matter now? | Customer impact, risk, and the company priorities the change would support. |
A serious account-access failure may need immediate action even if it has happened only once, while a frequent but minor inconvenience may wait behind a more consequential fix. The questions help make that trade-off clear without turning prioritisation into a formula.
Turn a pattern into work
Consider a hypothetical pattern in cancellation requests. The AI agent repeatedly gives pause instructions to customers who ask to cancel a subscription, leaving some to repeat their request and others to leave without a clear next step. Once the team recognises the recurring misunderstanding, it can decide where a correction belongs.
The help content might leave the distinction unclear, or the agent's instructions might direct it to offer a pause even when the customer asks to cancel. If the answer is correct but customers cannot complete the steps, the problem may belong with Product instead. Identifying the cause gives the task an owner and keeps the team from correcting the wrong part of the process.
CraftCX helps teams discover this kind of work by automatically identifying signals from customers, classifying them, and monitoring for patterns. Support Intelligence recognises bug reports, feature requests, documentation gaps, and other recurring customer needs, and flags possible incidents. Teams can see which problems are new, growing, or recurring, including requests they might otherwise miss as AI handles more support.
Support Quality adds monitoring of the agent's performance; helping teams find recurring failures in resolution, customer effort, and handoffs. Together, these capabilities give Support Ops a broader view of what needs improvement without relying on someone to notice every issue manually. The team still decides how to respond, taking account of the cause, the cost of a fix, and company priorities.
Check whether the change helped
Suppose the team from our example finds that the help content leaves the distinction between pausing and cancellation unclear. Before correcting it, record how often the agent gives pause instructions for cancellation requests over a defined period, so there is a baseline for checking whether the change worked.
After the change, test the agent against the requests that exposed the problem and monitor whether the same misunderstanding continues. Record your test case as an eval you can run in the future to help detect and prevent regressions. The useful result is that customers receive the right instructions and no longer have to repeat their request, which a rise in the overall AI resolution rate cannot establish on its own.
Compare similar periods, accounting for changes in traffic and the types of requests. A fall in affected cases means more when the issue also becomes less common among cancellation requests. If it persists, check whether the agent is using the revised guidance or whether another part of the workflow still needs attention.
The same loop applies to a product bug or missing feature. Recognise the recurring need, decide what deserves work, and watch whether customers continue to report the problem after the change. Support Ops gains a steady source of opportunities to improve the service, even when the conversations that reveal them never reach a teammate's inbox.