The Cost of a Miss: Comms With AI in the Govern Phase

There's no better way to kick off a deep dive than with an uncomfortable arithmetic problem. Here goes...

If a Create-focussed agent helps your team produce three times the content it used to, someone is now reviewing three times the content. The drafting step got an agent. The reviewing step, in most teams, did not. So the volume problem the last article ('The Volume Problem: Comms With AI in the Create Phase') described does not disappear when a draft is finished. It moves one phase downstream and lands on whoever signs the work off.

That is the Govern phase, and it is the part of the Comms With AI Operating System teams are least keen to talk about. Strategise is interesting. Create is visible. Govern is the bit that feels like brakes. It is approvals, compliance, fact-checking, tone audits, accessibility, the careful and extremely unglamorous work of making sure nothing goes out that should not have. Few build an AI strategy because they are excited about approval workflows.

However, Govern is where AI-assisted communications either hold up or quietly fall over. It is the phase where AI introduces new ways to fail before it offers any way to cope with the extra load. And it is the phase where the argument for using agents is less obvious – and therefore more interesting – than it first looks.

[If you are joining the series here, the overview introduces the Operating System and its five phases, Part 2 goes deep on Strategise, and Part 3 covers Create. You can find the full series here. This fourth article is about the control layer that protects everything the first three phases produced.]

The Governance Gap

Communications has always had a governance bottleneck. Approval was regularly slow and painful long before anyone had heard of an AI agent. Legal needed to see the claims. Brand needed to see the tone. A senior leader needed to see anything that touched reputation. The work queued, and the queue was the price of not making expensive mistakes in public.

AI did two things to that queue at once, and only one of them was helpful.

Unhelpfully, it multiplied the volume of work arriving at the gate. More drafts, more formats, more variants, all produced faster than the review layer was built to absorb. A governance process calibrated for one press release a week strains badly when it meets a dozen.

Additionally, it has changed the kind of error that can reach the gate. A human writer under deadline pressure makes predictable mistakes. They get a date wrong, they over-claim slightly, they reach for a tired phrase. A communications professional reviewing that work knows what to look for. An AI agent fails differently. It will state a fabricated statistic with total confidence. It will attribute a quote to a real person who never said it. It will cite a source that does not exist, formatted immaculately. It will drift, paragraph by paragraph, toward the bland statistical centre of its training data, so that the tone is not wrong exactly, just no longer anyone's.

These are not failures of effort. They are failures of a different shape, and therefore, a review process designed to catch human error is not set up to reliably catch them. That is the governance gap: production capability scaled, the volume and the failure modes both changed, and the checking layer stayed roughly where it was.

The hard truth of the AI-in-comms story is that most teams have upgraded Create and left Govern alone. That is the most dangerous possible configuration. It is faster production feeding an unchanged, now overloaded, gate.

What the Govern Phase Covers

Govern is the control layer of the Operating System. It holds everything that stands between a finished draft and a published asset the organisation is willing to defend:

• Approval workflows: routing, sequencing, sign-off, status tracking, the coordination of who sees what and in what order
• Tone and brand-voice verification: checking that a piece sounds like the organisation, consistently, across volume
• Claims substantiation: fact-checking, source verification, making sure every assertion can be stood behind
• Accessibility and inclusive language: reading level, alt text, plain language, contrast, jargon
• Compliance and regulatory review: sector rules, legal exposure, disclosure requirements, including disclosure of AI use itself
• Crisis preparation: pre-approved response frameworks, holding statements, and the governance done in advance so it does not have to be done under fire

The thread running through all of it is that Govern is where reputational risk is priced. Every other phase produces work. Govern decides whether the organisation can live with it.

The 25/75 Rule for Govern

Across the CWAI Operating System, the working rule is roughly 30% AI, 70% human. Govern sits below that average, at 25% AI, 75% human.

This ratio shifts across the five phases:

• Strategise: 20/80 – research and synthesis AI, judgement human
• Create: 35/65 – drafting AI, voice and nuance human
• Govern: 25/75 – systematic checking AI, reputational judgement human
• Monitor: 40/60 – pattern detection AI, interpretation human
• Transform: 20/80 – readiness assessment AI, change leadership human

The reason Govern keeps the human share high is not sentiment. It is asymmetry. In Create, a weak agent draft costs you an hour of editing. In Govern, a missed claim costs you a correction, a news cycle, and a quantity of trust that takes far longer to rebuild than any agent ever saved. The maths of governance is lopsided. The downside of a miss dwarfs the upside of a fast pass, so the phase has to be run conservatively by design.

The instinct is that high stakes mean keeping AI out. Govern is reputational; therefore, it's the last place you would let an agent near. This instinct isn't entirely mistaken: it is right that humans must own the reputational calls. However, it is backwards about where governance actually fails.

Governance failures, in practice, are rarely failures of judgement. They are failures of consistency. The claim nobody thought to check. The alt text nobody added. The terminology that drifted across a campaign it was drafted by three people. The disclosure line that was in the template and got cut. These are not hard calls anyone got wrong. They are routine checks a tired human skipped at five o'clock on a Friday, or due to the umpteen other reasons that they were pressed for time.

A notable upside for AI in 2026, is that consistency is the one thing an agent is reliably good at. (Something you would not have said even a year ago.) It does not get tired. It does not skip the boring check on the forty-first asset. It applies the same standard to the all-staff email as to the flagship report. So the 25% an agent does in Govern is not the judgement. It is the floor. It is making sure the routine, checkable, consistency-dependent work actually gets done every single time, which frees the human 75% to spend its attention on the calls that genuinely need a person.

The risk is not putting agents into Govern. The risk is upgrading Create and leaving Govern as it was.

Where Agents Change Governance Work

Four applications are worth examining in detail, because they are where the change in the Govern phase is most tangible.

1. Claims substantiation and fact-checking

This is the application that matters most, because it addresses AI's most dangerous failure mode directly.

A claims-substantiation agent does something a human reviewer rarely has time to do properly. It reads a draft and extracts every factual assertion in it, one by one: every number, every named source, every attributed quote, every "studies show", every comparative claim. For each one, it asks a blunt question. Where did this come from, and would a sceptic accept the answer? Claims with a solid source are passed. Claims with a weak source are flagged. Claims with no source, or a source the agent cannot verify, are escalated for a human to resolve before the piece moves.

The principle underneath this is worth stating plainly, because it is the single most important governance design decision in AI-assisted comms. The agent that drafts a piece must not be the only agent that checks it. If the same system, with the same prompt and the same blind spots, both writes and verifies, its errors are correlated. It will confidently wave through its own fabrications. Effective claims governance uses a separate checking layer, ideally a different model with a different instruction set, whose entire job is to be the sceptic. Separation of duties is an old governance idea. It applies cleanly here.

This is also where the interview with NOAN's Neal Mann on this site is worth revisiting. His argument, that most AI-powered communications is built on sand because it has no verified fact layer underneath it, is precisely the problem a substantiation agent exists to address. The agent does not create the fact layer. It is the discipline that refuses to let a piece proceed without one.

The Claims Substantiation Checklist on CommsWith.AI is built to structure this work, so the agent's output is a usable list of resolved and unresolved claims rather than a vague reassurance.

2. Tone and brand-voice auditing

Part 3 described how a voice-trained agent can help produce content in a specific spokesperson's or organisation's voice. The same voice profile has a second job, and it belongs in Govern.

A brand-voice audit agent takes a finished draft and checks it against a defined voice profile: the organisation's characteristic register, the words it uses and the words it avoids, the level of formality, the rhythm. It flags drift. It catches the paragraph that has slid into corporate neutral, the sentence that is technically on-brand but tonally off for this particular moment, the piece that reads as though it could belong to any organisation in the sector.

This matters more as volume rises. When a team published one considered piece a week, voice consistency was held by the simple fact that one or two people touched everything. When a team publishes at agent-assisted volume, no single person reads all of it, and voice consistency stops being automatic. The audit agent is how you hold a standard you can no longer hold by hand.

It is worth saying that this is not a new idea on Applied Comms AI. The "Will My Boss Hate This?" stakeholder review tool built earlier on this site was, in effect, an early governance agent: a structured check that ran a draft against the predictable reactions of the people who would have to approve it. That is Govern work. The Tone and Style Checker and Brand Voice Audit Checklist templates carry the same logic into a repeatable form.

3. Accessibility and inclusive-language scanning

This is the lowest-risk, highest-consistency win in the entire phase, and it is routinely skipped.

Accessibility checking is exactly the kind of work humans do badly and agents do well. It is systematic. It is rule-based. It is tedious. It is the same set of checks on every asset: is the reading level appropriate for the audience, is there meaningful alt text on every image, is the language plain where plain language is needed, is jargon defined, does the structure work for a screen reader, is the inclusive-language standard met. A human reviewer under deadline pressure will do this properly on the important pieces and quietly let it slide on the rest. An agent does it on all of them, closer to identically, in seconds.

4. Approval workflow orchestration

The fourth application is the least discussed and, for many teams, the one that removes the most friction.

Most of the pain in approvals is not the reviewing. It is the coordination around the reviewing. Who needs to see this. In what order. What does each reviewer actually need: legal wants the claims and the contracts, brand wants the tone and the visual treatment, the executive sponsor wants the strategic risk and does not want a copy-edit. Chasing. Reminding. Version control. Knowing what is stuck and why.

An orchestration agent handles that coordination layer. Given a piece and a workflow definition, it assembles the right review pack for each reviewer, so legal is not wading through brand notes, routes the piece in the correct sequence, tracks status, and flags what is overdue. It does not approve anything. It removes the administrative drag that makes approval feel slow even when the actual reviewing is fast.

The Approval Workflow Mapper and Content Approval Tracker templates are designed to define the workflow first, in plain terms, so that what the agent orchestrates is your actual governance process and not an invented one. Map before you automate. An automated bad workflow is just a bad workflow that now runs faster.

A Test Your Governance Should Pass

Before a piece is published, run it against the following five questions. These are the questions that decide whether a governance process is real or decorative.

1. For every type of risk in this piece, can you name who is accountable?

Factual risk, legal risk, brand risk, reputational risk. If the answer is "the team reviewed it", no one is accountable. Governance without named ownership is a group of people each assuming someone else caught it.

2. Is every factual claim traceable to a source a sceptic would accept?

Not a source you trust. A source someone who wants you to be wrong would still have to concede. If a claim cannot meet that bar, it is an opinion, and it should be written as one or removed.

3. Would this piece survive being read aloud by your most hostile stakeholder?

Pick the specific person: the journalist with a grudge, the regulator, the board sceptic. Read it as they would. If a sentence hands them a weapon, you want to know now, not after publication.

4. Is your use of AI disclosed where disclosure matters?

This series discloses its own AI use deliberately. Your organisation needs a clear position on when AI assistance is disclosed, to whom, and in what form. "We never decided" is not a position. It is an exposure.

5. If this went wrong in public, is your response already drafted?

For routine content the honest answer is that it will not go wrong, and that is fine. For anything sensitive, anything making a strong claim, anything on a contested topic, the response should exist before the piece does. If it does not, the piece is not finished.

If the work passes all five, it is ready to publish. If it fails any of them, the fix belongs in Govern, before the piece goes out, where the cost of closing the gap is measured in minutes rather than news cycles.

Where To Start

If Govern is the phase your team has under-invested in, which it very likely is, because production is where the visible excitement sits, three starting points will pay back quickly.

Start with accessibility and claims checking. These are the lowest-risk, highest-consistency agent applications in the whole Operating System. The work is rule-based, the agent does it identically every time, and the failure you are preventing, an unsupported claim or an inaccessible asset, is both common and avoidable. If you do one thing in Govern, make it this.

Build one brand-voice profile and run it as a standing check. Not a document describing your tone in adjectives. A working profile, built from real pieces that sound right, used by an audit agent on outgoing content. It is the only practical way to hold voice once you are publishing at agent-assisted volume.

Map your approval workflow before you automate any of it. Write down who reviews what, in what order, and why, in plain language. Most teams discover their real workflow is not the one they think they have. Fix the workflow first. Automate it second. An orchestration agent should accelerate a good process, never entrench a bad one.

The templates in the Govern phase on CommsWith.AI are built around this logic. Each sits at the point where systematic agent checking meets human reputational judgement, with the split held where it belongs.

Browse the full Govern phase on CommsWith.AI.

About

Applied Comms AI is the practical guide for communications leaders navigating AI, grounded in hands-on experimentation, workflow transformation and real-world implementation. Read the full AI Agent series here. CommsWith.AI is the companion template and resource library for communications professionals using AI, and the Govern phase library covers approval workflows, tone and brand-voice verification, claims substantiation, accessibility and crisis preparation. Both sit alongside Faur, a communications consultancy pioneering practical AI expertise for organisations ready to implement at scale. If your team is working on the Govern phase and needs bespoke support, whether governance design, claims and compliance workflows or crisis-readiness work, get in touch at michael@faur.site or connect with me on LinkedIn.

This article is part of the AI Agent series published on Applied Comms AI. The series maps to the Comms With AI Operating System: Strategise, Create, Govern, Monitor, Transform.