The 20 roadblocks between comms and useful AI – and what to do about them

Intro

Ask a room of comms pros what stops them getting value from AI and you tend to hear the same one or two answers: it makes things up, or the lawyers will not allow it.

Both are real. Both are very, very far from the whole picture.

It's critical to recognise the issues inherent in current AI use for communications, not least since adoption has effectively plateaued. Muck Rack's State of AI in PR 2026 puts generative AI use at 76%, up just a single point on the year before, and Cision's Inside PR 2026 has 91% using it in their workflow (the two count slightly different things).

In 2026, almost everyone is making some use of the current raft of AI tools. This means the interesting question is no longer whether comms teams will adopt AI: it's why so much individual use turns into so little organisational value.

One clue comes from Ethan Mollick. Writing about a study of people using agentic tools, he noted something that should give every comms leader pause:

"What actually mattered was not the profession of the user, but their expertise. The more domain experience someone had, the more successful they were in using Claude Code in that domain. And, even more interestingly, the more useful output they got from Claude from each prompt."

His context was software, not press releases. But the direction he draws is general: we are moving from a world where non-experts use chatbots to fill gaps, to one where experts use AI to get real work done. The differentiator is not the tool. It is the expertise, judgement and organisation around it.

That reframes the roadblocks. Very few are really about the models. They are about people, knowledge, governance and structure. What feels like one or two AI problems quickly becomes 20, as you'll see below. (I could easily have added a half dozen more, but there is only so much time in the day, both for you and I...)

The reassuring part – and the reason this reads as a map rather than a warning – is that for almost every one the way through is already emerging.

So here they are, in four handy groups! Most are things I have run into myself, so alongside each I have added what I actually do about it. Client examples are anonymised throughout.

Group one: skills, craft and workflow

1. The scarce skill is not prompting

Everyone assumes the missing capability is prompt-writing. That may have been true a year or two back, but now it's judgement: knowing when to trust an output and when to distrust it. AI is the single biggest skills shortage the profession reports (CIPR, 36%), but the real gap sits deeper than the interface.

• What I do: Pair people on live work, have each critically analyse the other's AI output, to discover the misses and where it bluffs. You cannot lecture judgement into someone; they build it by being wrong a few times with a safety net.
• The way through: Treat this as guided practice, not a one-hour course. Design deliberate reps where getting it wrong is cheap, and review them together so the judgement compounds across the team rather than living in one person's head.

2. The training vacuum

Most people are using tools nobody trained them on, so they self-teach on live client work. Only 43% of organisations offered AI training in 2026, up from 35% in 2025 and 21% in 2024 (Muck Rack). Not-for-profit teams lag furthest.

• What I do: I self-taught the same way, badly at first, so now I keep a shared prompt library that means nobody on a project starts from a blank page – hello Comms With AI! It is the single cheapest thing that has improved output quality. If you have an enterprise tool, ensure there are shared project folders that new users can study and rely upon, with updatable foundational documents and instructions at their root.
• The way through: Budget protected time for hands-on training, and capture what works in a shared library so one person's learning becomes everyone's baseline rather than being rediscovered five times. You can turn the AI on the shared library itself, through canny use of document management, so this becomes part of its own knowledge base – now you're really cooking 🍳

3. Tool sprawl

The market is flooded. A 2023 CIPR report drew on a dataset of more than 10,000 marketing and PR tools, and the number has only grown. Evaluating and bedding in tools is a job in itself, so teams either freeze or scatter across shiny things that never stick.

• What I do: When I built a small internal review tool, the very first thing the AI handed me was an error message. Even with AI's help, the boring parts still ate the afternoon, which cured me of chasing every new launch. I keep my own stack deliberately small, and for every tool which requires a subscription, I look to ensure at least 3 months of continuous use (ideally 6 or 12 for those I use most often) before reviewing. This is still a far quicker turnover than would be traditional for reviewing your software suite, but it provides a manageable buffer to prevent having your head turned and testing too many near-identical tools on a daily basis and coming out none the wiser.
• The way through: standardise on a few vetted tools, ideally embedded in software you already run, and set a high bar before anything new earns a place. Boring and mastered beats novel and half-used.

4. AI can make more work, not less

The promised time saving evaporates when generic first drafts need heavy rewriting, or when volume balloons downstream. Individual gains are real (USC's 2025 Relevance Report has 73% saying they work faster) but they rarely net out at team level.

• What I do: My own test of a deep-research tool lost most of a day to the AI trying to reach a restricted page, then returned work with minor factual errors and missing context. I now budget for the checking, not just the drafting.
• The way through: Measure time including review and rework, not just the shiny first draft, and build a verification step in by default so the saving is real rather than borrowed from your future self. Compare and contrast on your first go, and ensure to review your process on a semi-regular basis.

5. Brand voice drift

Generic AI prose flattens distinctive voice, so heavy use risks making every organisation sound the same. That is the precise opposite of what comms exists to do – you can feel this directly whenever your eyes glaze over reading generic pseudo-thought leadership LinkedIn slop – and the rewriting it demands eats the time saved.

• What I do: I keep a short, brutal voice brief and a set of "never write like this" examples that I feed the model up front, then also build this into the checklist for after. Then I allot myself a decent amount of time for my own review and amendments before it goes anywhere near a client.
• The way through: Train a model on house style, keep a strong human editing layer, and ensure you sample outputs regularly for drift, because the flattening is gradual and easy to stop noticing.

This group comes back to one habit: pair each prompt with a worked example and a review step, kept where the whole team can reach it. That is the thinking behind the Create library.

Group two: governance, legal and ethics

6. The policy vacuum

A large share of comms professionals still work with no house rules on what is allowed, what to disclose, or what data is off-limits. 28% of teams had no AI policy at all in 2026 (Muck Rack), which pushes every risk decision onto individuals. As Ben Verinder told us, "there hasn't been a session I've done in the last three years where the head of comms hasn't gone, 'I did not know you were doing that.'"

• What I do: I start clients with a concise, accessibly written policy – not a manual – because a page people actually read beats a framework that quickly withers away into irrelevancy in a shared drive.
• The way through: Write a short use-case policy covering disclosure, data limits and a human sign-off point. Offer an amnesty for any shadow use already happening (and it will be happening), so you are governing reality rather than a fiction.

7. Data, IP and confidentiality

Comms handles embargoed announcements, client-confidential material and personal data. Feeding that into third-party tools without governance is real exposure, and the ownership picture is murkier than most assume.

• What I do: I look to keep anything confidential out of consumer tools entirely, and am up front with clients about which parts of a deliverable were AI-assisted, partly to protect exactly this.
• The way through: Set clear data rules, client-consent protocols and vendor vetting, and decide ownership and disclosure at the briefing stage rather than discovering the problem at handover.

8. The honesty gap

Too few are actually talking about their AI use. Around half of in-house teams never ask their agencies about it, and around half of agencies say no client has ever asked, a figure unchanged in two years. Ben Verinder's phrase for it was exact: "there's a big gap in honesty and transparency around AI use, both from in-house teams and consultancies."

• What I do: I raise AI use before the client does (admittedly easier given my public work), and I use it to mark out where we are not using it, which tends to build more trust than staying quiet would.
• The way through: Put AI on the agenda in the contracting conversation and agree a disclosure standard both sides can live with. Silence is not neutral; it is where the risk accrues.

9. Regulatory overhang

A moving patchwork makes leaders in regulated fields cautious about anything published. The EU AI Act's transparency obligations apply from 2 August 2026, and an agency can never move faster than its most cautious client's legal team.

• What I do: Map which rules actually touch a given piece of work, rather than letting a general nervousness become a blanket "not yet" that ends up costing more than the risk it avoids.
• The way through: translate the real obligations into your specific comms use-cases, and build compliance into the workflow rather than bolting it on afterwards, so caution is targeted rather than total.

10. Ethical objections you cannot override

A meaningful minority avoid AI on principle: training data, plagiarism, energy and water use. Among those who steer clear, 56% call it overhyped and 41% call it risky (Muck Rack). Top-down mandates create resentment, not buy-in.

• What I do: Take the objection seriously rather than trying to win the argument. Even if I don't fully agree with every concern, the bulk of mainstream arguments require serious consideration, and agreed means of collectively moving forward and addressing them. (Not that this will happen and/or be easy, but it's a conversation that we all ought to be engaged in and a part of.)
• The way through: Name the ethics openly in your policy, give people a real say, and win consent through transparent framing and hands-on development rather than compulsion. A quiet objector becomes a slow adopter, not a convert, if you steamroll them. They may even become your biggest roadblock, should you try to manage their concerns away without due recognition and consideration.

The through-line here is governance you will actually use: a short policy, clear data rules and a named sign-off point beat any framework nobody reads. The Govern library is a place to start if you would rather not draft from scratch.

Group three: trust, accuracy and risk

11. Hallucination in published comms

Comms is on the record – what's on the internet stays on the internet (for the most part...). A fabricated statistic or invented quote does not stay private; it becomes a public, attributable error under your name. Even strong models still slip: yes, they're currently low single figures on easy grounded tasks, but 6% to 33% in specialist domains like law and medicine.

• What I do: I look to treat every AI-supplied fact as unverified until I have seen the source myself, and I get the tools to provide the relevant direct links to check (tools like Perplexity already known for a decent level of attribution).
• The way through: Make human verification a non-negotiable step, not a nice-to-have. The rule is simple: never publish an AI claim you have not personally checked, however confidently it is phrased.

12. The mess is often yours, not the model's

This reframes the last one. Much of what looks like hallucination is the system reasoning over contradictory inputs: 25 versions of the positioning, pricing that disagrees with itself, strategy buried in a deck nobody can find. Neal Mann put it plainly to us: "you cannot put AI on top of a mess of business knowledge and expect accurate results. The knowledge has to be right."

• What I do: before blaming a tool, I look at what I fed it, and more often than not the contradiction was already in the client's own materials. I've had many, many thoughts over the years – and a wealth of work which isn't clearly named and signposted – and as it turns out many of these things do not align...
• The way through: Fix the knowledge layer first. A clean, single source of truth does more for accuracy than any prompt trick, and it is the one investment that pays back across every tool you will ever use.

13. High-stakes moments where AI can harm

In a crisis, a hallucinated link or a missed one-off fact does not merely waste time; it points the whole response in the wrong direction. Crisis expert Amanda Coleman warned us that AI monitoring "can be subject to hallucinations which can derail the approach by putting the focus in the wrong place."

• What I do: Keep humans authoritative on situation assessment in a live crisis, and, in Amanda's words, I "build the prompts library when you're calm" rather than experimenting under fire.
• The way through: Define explicitly where AI is and is not allowed near live crisis work, and test your tools on low-stakes issues first so that when it matters you are deploying proven workflows, not improvising.

14. The trust penalty, even when you disclose

Telling people you use AI does not protect you if they object to how you use it. Undisclosed or unendorsed use carries a measurable trust penalty with stakeholders. Disclosure alone is not a shield. As Ben Verinder put it, "it's not sufficient just to say, 'hey, we're doing this, guys.'"

• What I do: Treat winning acceptance for how we use AI as its own small comms job, with its own audience and message, rather than a line in the small print. Be open and honest, which has informed how Applied / Comms With AI has operated since day one.
• The way through: Socialise your AI use with the stakeholders it affects, explaining the how and the why, and give them somewhere to push back. Consent you have earned survives scrutiny; consent you assumed does not.

15. Agents are becoming your spokespeople

A customer-facing agent speaks for the organisation all day, every day. Ask a well-built one about a live controversy and a blank "I can't answer that" is its own reputational event. Speaking at a recent Comms With AI webinar, Elif Güvençer captured it well: reputation is now built one token at a time, and these agents "need media training, the difficult questions asked from a reputational lens before they ever go live." (Our full write-up follows later this month, you can read her Two Clocks framework here.)

• What I do: Testing. Testing. TESTING. When considering an external-facing agent, I aim to run it through the awkward questions a journalist or most ardent antagonist would ask, before it is anywhere near the public.
• The way through: Bring comms into AI governance, and treat any external-facing agent like a spokesperson, with lines to take, escalation routes and a reputational stress-test before launch.

Across this group the move is the same: put human verification where it counts, and catch drift and false signals before they reach the public. That is what the Monitor library is for.

Group four: organisation, culture and value

16. Pilots that never scale (hi agents)

Almost every team has a pilot that worked: an agent that drafts the newsletter, a workflow that turns a report into a week of social in an afternoon. Six months later it has stopped, because the one person who built it got busy or moved on. Only 12% of AI users in PR have adopted agents at all (Muck Rack).

• What I do: Ask this question: if your most AI-fluent person left tomorrow, would the capability survive? If the honest answer is no, you have a person, not a capability. This is the test at the heart of our recent Pilot Trap piece.
• The way through: Build the operating model, not just the pilot. Write the workflow down, share it, put it in more than one pair of hands, and make it something the team owns rather than a clever person's side project.

17. Nobody can prove the value

When benefits stay at the individual level and go unmeasured, leaders cannot see a business case, so mandate and budget stall. Comms has always struggled here, and Elif Güvençer's diagnosis, from the same webinar mentioned above, is sharp: the barrier is not bandwidth or capability, it is vulnerability, the discomfort of dismantling your old value proposition "visibly under scrutiny" and building a new one.

• What I do: On a recent client plan, much of the work existed to head off problems before they surfaced, which runs straight into the old difficulty of proving the worth of a crisis that never happened. I now name that dynamic openly with clients rather than hoping the value is obvious.
• The way through: Measure team-level outcomes, not personal time saved, and treat AI as organisational change to be evidenced. Reopening the value question is uncomfortable, but the discomfort is the work.

18. The leadership perception gap

Leaders think the organisation is far more AI-ready than the people doing the work do. In Cision's Inside PR 2026, one in three executives called their organisation "extremely agile"; only 14% of employees agreed. Mandates then land on teams with neither the workflows nor the capacity to deliver them.

• What I do: Try to get leadership and the frontline reporting the same reality – 'singing from the same hymn sheet' is the phrase which often comes to mind – because most of the friction I see comes from a gap between the two, not from the tools.
• The way through: As Ben Verinder puts it, treat this as "a change programme, not an IT programme." Ground the strategy in what the team can actually do today, and close the perception gap before you set the ambition.

19. Sign-off becomes the bottleneck

The human review chain turns into the constraint, so the speed AI promises never actually arrives. In the same Cision study, 63% named team size and organisational design, and 53% named slow approvals, as the biggest brakes on agility. Small teams simply lack the capacity to redesign around AI.

• What I do: Look at where work waits, not just where it is made, because the queue in front of legal or the CEO is usually the real ceiling on speed.
• The way through: Redesign sign-off for AI-assisted volume, with lighter-touch review for lower-risk outputs and human attention reserved for what is the genuinely sensitive. Create sped up; Govern has to speed up with it, or the value leaks out in the gap.

20. Deskilling the next generation

Practitioners themselves fear that heavy AI use will stop juniors learning the craft the profession runs on. More than three in four PR professionals worry about exactly this (Muck Rack). If AI does all the early reps, the judgement that separates good comms from average never gets built.

• What I do: Have juniors do the thinking first and use AI to pressure-test it, not the other way round, so the tool sharpens their judgement instead of replacing it.
• The way through: Protect the training pathway deliberately. Use AI to augment junior work rather than to skip the practice that turns juniors into the experts Mollick is describing. The reps are the point.

This last group is the least technical and the hardest: organisational change, measured at team level and owned by more than one person. The Transform library is where that operating-model work lives.

And once you've navigated all that...

Twenty roadblocks is a lot to put in front of a reader, I know. Initially I had intended five or 10, but all those above felt too important to reduce down.

Also, playing them out together makes the point clearer than any single one could: the barriers to useful AI in communications are overwhelmingly human and organisational, not technical. The models are improving faster than we can track. The bottleneck has moved to us, to our expertise, our knowledge, our governance and our willingness to change how we work.

Which is why Mollick's line at the beginning matters so much. The value does not go to whoever has the newest tool. It goes to whoever brings the most expertise to it. That is brilliant news for communications, a profession built on judgement, on knowing an audience, on the difference between accurate and true. The roadblocks are real. So are the ways through. The teams that pull ahead will be the ones that treat this list not as reasons to wait, but as a to-do list.

What have we missed? If you have hit a roadblock that is not here, or found a way through one that is, reply and tell us. We are documenting this in the open.

If you want to take any of this further, the method behind this piece is set out at Comms With AI: a free template library organised by phase, a free AI-readiness diagnostic and a free planning tool, Plan, that turns a specific need into a brief and a prompt, and a paid Consult if you want a second pair of eyes on a governance or deployment decision. All of it is built to work inside the tools your team already uses.