Stop Telling People to Communicate Better. Instrument It.

Most companies can tell when internal communication has gone sideways. They can feel it before anybody can prove it.

Too many channels. Too many people copied. Too many messages that sound like conclusions but are really one person's bad morning. Criticism with no analysis under it. Threads that should have turned into a ticket, a document, a decision, or a root cause analysis, and instead turned into a mood.

Then leadership stands up and says the thing that never works.

“We need to communicate better.”

That sentence has never fixed a single channel. If communication matters to how the work gets done, instrument it.

The Tool Is Not the Villain

Slack is the easy example, so people blame Slack. The same mess shows up in Teams, in email, in the project tracker, in the support queue, in comments on tickets. Anywhere people talk about operational work, the same rot sets in.

The reason is simpler than the tooling debate. Companies treat communication as conversation that evaporates, when it is actually a running log of how the organization operates. AI is what makes that log readable.

Every Message Carries More Than Words

Read a week of any busy channel and you'll find evidence of:

• real problem solving
• complaints with nothing under them
• the same issue escalated for the fourth time
• criticism aimed at a person instead of a problem
• disagreement that's making the work better
• negativity that's making it worse
• engineering concerns nobody closed
• decisions that got made and never written down
• good work that deserved a thank-you and didn't get one
• frustration on its way to becoming a management problem
• the same confusion, in three departments, for six months

Management usually learns about all of this through anecdote. Somebody walks into a meeting and says engineering keeps dropping the ball. Somebody else says that team is doing great. Now the executive is trying to decide which story sounds more credible, which is a coin flip wearing a suit.

Running a company on competing anecdotes is a terrible operating system.

Ask Better Questions Than “Is This Person Negative?”

Generic sentiment scoring won't get you there. You do not want software deciding who's a positive person and who's a negative one. That's simplistic on a good day and dangerous on a bad one.

Score observable behavior instead. For any message that claims something about the work:

• Was the statement factual?
• Was any evidence attached?
• Was the criticism aimed at a problem or at a person?
• Did it name expected behavior versus what actually happened?
• Did it suggest a next step?
• Did the issue land anywhere durable?
• Was the language professional?
• Did the message move the work forward?

Someone can be relentlessly critical and communicate beautifully. Some of the most valuable people I've worked with were the ones willing to say out loud that something was broken. I want more of them, not fewer.

Criticism is fine. Criticism with no evidence, no ownership, and no next step is the expensive kind.

Put a Weekly Scoreboard on It

We already measure revenue, uptime, tickets, defects, and customer satisfaction down to the decimal. Communication quality gets a pep talk.

A weekly report, generated by AI and read by humans, can track things like:

• how many channels or groups are actually active
• channels doing the same job as another channel
• channels that have grown strangely large
• conversations happening in the wrong place
• threads that should have become tickets or documents
• complaints repeating week over week
• concerns raised and never closed
• recognition and de-escalation worth noticing
• decisions made in a thread and never documented
• claims made without evidence, and claims made with strong evidence and a real recommendation

None of that is surveillance for sport. The point is to see the shape of the organization before the shape becomes a personnel meeting.

What usually turns up is not a villain. It's plumbing. One department has twice as many channels as anybody else. One channel has quietly become the place people go to complain. Five people keep raising the same concern because nobody ever converted it into a decision. One manager generates half the company's escalation just by leaning on the CC line.

You were looking for a difficult employee and you found a broken operating system.

Turn Anecdotes Into Documents

This is the part I'd build first.

A complaint typed into a channel should not become organizational truth just because it was typed loudly. Somebody writes “engineering keeps messing this up,” and instead of letting it sit there and season everyone's opinion, AI turns it into a structured concern on the spot.

I sometimes describe the goal as making stupidity expensive. Not disagreement, and not complaints. Tossing around loosely formed accusations is what's cheap, and cheap things get overproduced. Ask for evidence and analysis the moment somebody makes a serious claim, and behavior changes within a couple of weeks. People still raise hard issues. They just bring receipts.

Watching One Person, Without Playing Psychologist

Sometimes management already suspects that one person's communication is doing damage. AI can help there, carefully.

Do not build a secret “is Jane negative?” score. Build a monitor that applies the same professional standard you'd apply to anybody, including yourself, and watch for patterns:

• repeated personal criticism
• accusations with nothing behind them
• hostile or dismissive language
• escalation without any attempt at resolution
• the same complaint, never converted into analysis
• an unusually high share of the company's conflict

Then show management the evidence, not the number. A useful flag reads: this message was flagged because it contained a personal accusation, no supporting evidence, and no proposed next step. A useless one reads: Jane has a negativity score of 73. One of those you can act on and defend. The other one is amateur psychology with a dashboard.

The System Can Also Speak Up in the Moment

Some organizations won't want to wait for a weekly report. When a message shows the smell, the assistant can answer right there in the thread.

For an engineering concern with no analysis under it: “This looks like an engineering concern. I've drafted a concern document from your message. Add the expected outcome, actual outcome, evidence, impact, and requested remediation before this gets treated as settled.”

For a message that got personal: “There's a legitimate operational issue in here, but the wording is about a person rather than the behavior. Restate it as observable behavior, evidence, impact, and requested action.”

Compare that to a manager quietly going desk to desk asking everybody to be nicer. The bot teaches the standard in public, consistently, without anybody losing face.

Catch People Doing It Right

If the monitor only ever finds problems, you've built a police force and called it instrumentation. It should be just as good at spotting people who:

• talk a hot thread back down
• bring evidence without being asked
• convert a complaint into analysis
• help another team out of a hole
• catch an issue early
• write the decision down
• credit a colleague
• ask the clarifying question everybody else was avoiding
• turn vague worry into a next step

Those behaviors belong in the same weekly report as the flags, on the same page, with names attached.

The Build Is Not Exotic

Most platforms already expose enough to do this. The architecture is about as plain as it gets: communication platform, event stream, AI evaluation, rules and scoring, intervention or document creation, weekly report for management.

For Slack, that's the events API and a small application. For Teams, Microsoft Graph inside the tenant you already run. For email, shared operational mailboxes or the specific threads where company work actually happens. The model can be OpenAI, Claude, whatever your security review already approved, or two of them side by side while you're testing.

The model matters much less than the rubric. A vague prompt produces a vague system that flags everything and convinces nobody. Write down the scoring criteria, worked examples of good and bad, thresholds, escalation rules, and what counts as evidence. Then run it over months of historical threads and read the results yourself before you let it say a word out loud.

Three Modes, In That Order

1. Observe. The system scores and stays quiet. This is where you find out whether your rubric is any good, and it usually isn't on the first pass.
2. Coach. It starts replying when quality drops below the agreed standard, asking for evidence, documentation, or a more professional restatement.
3. Escalate. Only once it has earned trust does it notify managers, open formal concern documents, or trigger workflow.

Skipping to the third mode is how these projects die. A handful of false alarms and the whole thing becomes the joke in the channel it was supposed to fix.

Mechanics for the Machine, Judgment for the Humans

The AI flags, summarizes, scores, converts anecdotes into structured concerns, spots patterns, and produces evidence. Humans handle discipline, employment, and every other consequential call about a person's career.

The AI is not the manager. It is the instrumentation.

Blur that line and you'll deserve everything that follows.

The Deeper Point

Most companies don't have a communication problem. They have an uninstrumented communication system. They're asking people to behave better without giving them immediate feedback, agreed rules, or any way to tell whether things are improving.

We stopped making that mistake in engineering years ago. Nobody tells a team to please make the system more reliable. We measure uptime, error rates, latency, failed deployments, and defects, and then we argue about the numbers instead of about each other.

Internal communication deserves the same courtesy. Define what good looks like. Measure it. Turn anecdotes into structured work. Reward the people willing to say something is wrong. Make unsupported accusations cost something. Let the machine handle the mechanics, and let humans manage the humans.