The AI Quality Gap in Internal Comms

The problem with AI in internal comms is not that it writes badly. It writes with fluency, and fluency is the trap. A draft that reads finished invites you to skip the read that matters.

The AI quality gap is the distance between a message that reads right and a message that is right. Closing it is a review discipline now, not a writing one. The bottleneck moved. Drafting an all-staff update used to take an afternoon, so the afternoon was where communicators spent their judgment. A tool returns that same update in twenty seconds, and the judgment has nowhere to go unless you put it back on purpose. Most IC teams have not built that habit yet.

What actually breaks when nobody checks the AI draft

Four failures show up again and again once AI drafts start landing in the queue, and none of them look like bad writing on the surface.

The first is fabricated specifics. A model asked to write a benefits reminder will supply a percentage, an effective date, a plan name, or a quoted line from your CHRO, and it will do all of that in the same confident register whether the detail is real or invented. The sentence is clean. The number is wrong. Employees act on the number.

The second is voice flattening. Feed a tool your rough notes and it returns competent corporate prose, sanded down to the register every company sounds like. Nobody flags it in approvals because nothing is incorrect. But the message stops sounding like your organization, and people register that shift before they can name it. Trust erodes a quarter-percent at a time.

The third is lost nuance. The reason behind a policy change, the exception for one region, the caveat that made a hard message fair, these are the parts a model trims when it optimizes for a clean read. The draft gets tighter and less true at the same time.

The fourth is confident tone on shaky ground. Ask AI to announce a restructure or a benefits cut and it will often deliver the news in an even, upbeat cadence that misreads the room. The words are polite. The effect is tone-deaf, because the machine has no idea what this message costs the person reading it.

The review pass most IC teams skip

Reading an AI draft for typos is not reviewing it. The spelling will be perfect. Reviewing means verifying every claim the draft makes and restoring the judgment the draft flattened out. That is a different job, and it takes a different kind of attention than proofing your own writing.

The rule that holds the pass together is this. Read the draft against its source, not against your ear. When you edit your own work, your ear is a decent guide because you know what you meant. When you edit a model's work, your ear is the enemy. The draft was built to sound right. If you review by feel, you will approve fluent inventions all day.

Read for facts first, prose second

Before you touch a comma, run a verification pass. Every number, date, name, dollar figure, and policy detail gets checked against the source document it should have come from. Highlight each specific claim, then confirm it against the brief, the HR doc, the approved leadership quote, the real deadline.

If a claim has no source, it does not ship. Treat any unsourced specific as fabrication until someone proves otherwise, because that is the safer default and it costs you nothing but a Slack message to the person who owns the fact. The alternative is publishing a plausible number to five thousand people and finding out later the model made it up.

Read for what is missing

Fabrication is the failure you can see. Omission is the one you cannot, and it is worse. A model drops content without a trace, so nothing on the page tells you a caveat is gone. The only way to catch it is to compare the draft back to the brief and the source and ask what got left out. Which exception disappeared? Which reason for the change? Which group this does not apply to? The gap you have to hunt for is the dangerous one.

Read the draft aloud for voice

Read it out loud. Fluent-but-generic is the AI tell, and your ear catches it faster than your eye. Does this sound like your organization, or like every organization? Where the draft went smooth and anonymous, put back the specific verb, the plainer word, the one line only your team would write. You are not polishing here. You are re-signing the message in your own hand.

The governance habits an AI-drafting team needs

Individual review discipline holds up only if the team builds a few habits around it. Four are worth setting now, before AI-assisted volume becomes normal and the shortcuts calcify.

Label provenance. Mark which drafts in the queue were AI-assisted, so a reviewer knows to run the full verification pass rather than a quick proof. An unmarked AI draft is the one that slips through as if a person had already vetted every claim.

Name the reviewer. Every AI-assisted draft gets a human owner of record before it publishes, the same standard you would hold for any message going to the whole company. No draft reaches the all-staff list with no name attached to whether the facts are true.

Keep the source trail. The draft links to the document, data, or approval it came from. Verification is only fast if the source is one click away, and a habit that is slow is a habit that gets skipped under deadline.

Set a stop rule. Some categories never publish on an AI draft without a named subject-matter check: compensation, benefits, safety, legal, and any quote attributed to a leader. These are the messages where a fabricated detail does real damage, so they get the strictest gate regardless of how clean the draft looks.

Closing the AI quality gap is a review skill

Drafting tools keep getting better at fluency. The newest ones, including tools like Cerkl Broadcast, write cleaner first drafts every release. None of them get better at knowing what your CFO approved, which caveat your legal team needs, or how the news will land on the warehouse floor. That gap does not close on its own, because it is not a writing problem the model can solve. It is a judgment problem that belongs to you.

So the teams that win with AI will not be the ones drafting fastest. They will be the ones who turned review into a discipline, so the speed never costs them trust. Build the pass now, while the volume is still manageable and the stakes of getting it wrong are still small. The draft that reads right will always be easy to produce. The message that is right is still your job.