Five users at work spent this week struggling to hear the people on the other end of their calls. All five use a softphone — in our case Mitel’s MiCollab, a VoIP client that turns a PC and a USB headset into a desk phone. Everyone else with the same softphone was fine.
If you’ve run an IT operation, you know that shape of problem. Five users is too many for a bad headset and too few for a broken phone system. The cause could plausibly live in the network, the phone server, the client software, the audio drivers, the headsets, or something each of those five people did to their own configuration. Nothing is obviously on fire, and everything is a suspect.
Most of my readers in IT already know where this is going, because they know what tools like Claude Code can do. This story is for everyone else — the people whose mental model of AI is still a chat window that answers questions and drafts e-mail.
The swap test
We worked the problem conventionally for a day and a half, alongside our telecom VAR. This morning the VAR suggested the classic move: take one of the affected users and sign them in on a different PC. The problem didn’t follow them. Same account, same extension, same servers — clean audio on the other machine. One test cleared the user’s configuration and the entire back end, and parked the problem squarely on the PCs.
Real progress. Also where the hard part starts. A modern Windows workstation is thousands of moving parts — OS builds, drivers, update histories, security agents, vendor clients — and “something on these five PCs” is a big haystack.
Fifteen minutes
So before anyone started the ritual of reimaging machines, I opened Claude Code — Anthropic’s AI tool that runs in a terminal and executes real commands rather than just chatting — and gave it a plain-English assignment: this PC has the audio problem; this nearly identical PC doesn’t; compare them and tell me why.
It ran from my desk, under my account, with my permissions, reaching both machines remotely the same way a human admin would. Then it ground away for about fifteen minutes: OS builds, full update histories, driver and software inventories, system and application event logs, the softphone’s own debug logs on both machines, and a day-by-day timeline of what changed on the broken PC in the week before the complaints started. It saved the raw evidence to a folder as it went, so none of its conclusions had to be taken on faith.
Then it wrote me a report. Here is the heart of it, edited only to remove machine names and a few internal details:
The two machines are NOT running the same softphone. [The broken PC] runs MiCollab 9.7.117.1 (an Aug 2023 build, installed Aug 2024) while the working machine runs MiCollab 10.1.8.1 (Jun 2025 build). The old 9.7 client on the HP began hanging outright this week (Application Hang events on Jun 11 at 11:24 AM and Jun 12 at 10:11 AM), immediately following the June 2026 Windows servicing wave (cumulative KB5094126 + .NET 8.0.28 + WebView2/Edge 149 + a new Microsoft Voice Clarity audio-effects driver […]). The working Lenovo received the same updates […] and is unaffected with the newer client.
Recommended fix: upgrade MiCollab [on the broken PC] to 10.1.8 (parity with the working machine) and update Jabra Direct (6.15 → 6.26).
What it found
In plain English: the broken PC was running a softphone client almost two years older than the one on the working PC — version drift nobody had noticed, because the old client worked fine right up until it didn’t. June’s Windows updates, which included a brand-new Microsoft audio-processing component, landed on both machines. The current client took the changes in stride; the 2023-era client started hanging, and the hangs were sitting in the event log with timestamps matching the week’s complaints.
Just as useful was what the report ruled out, with evidence attached: the network, the firewall, USB headset dropouts, a wrong default audio device, Windows patch differences — the two machines were on identical builds. Ruling things out is half the value of troubleshooting, and it’s the half that gets skipped when five users are waiting and everyone is in a hurry.
It also flagged two problems nobody asked about: a wireless card on the broken PC that has been logging a hundred-plus corrected hardware errors a day since early May (on a desktop that’s wired anyway), and a bad printer driver crashing a spooler component every weekday morning. Neither caused the audio issue. Both are real. Both came free.
A member of my team upgraded the softphone on the affected machines to the current client, updated the stale headset software the report flagged, and the five users were back in business.
More than a chatbot
Two things deserve credit here, and only one of them is software. The swap test that isolated the problem to the PCs is pure 1990s methodology. The AI didn’t replace that discipline — it picked up the baton from it. But once the question became “what, on this machine, out of everything installed on it,” fifteen minutes of tireless evidence-gathering beat what was otherwise heading into day three.
The configuration matters as much as the model. Claude Code worked under my credentials, touched exactly the two machines I pointed it at, and left an evidence folder any admin can audit. An AI agent with hands needs what a new hire with admin rights needs: scoped access, supervision, and a paper trail.
We keep having to remind ourselves of this at work, so it’s worth saying plainly: AI is useful for far more than answering questions and drafting e-mail. Given real access and real constraints, it solves real problems. This week it was five phones.