Nick QuirkMost agencies got the AI adoption sequence backwards. They started with client-facing communication — the emails, the reports, the check-ins — because that's where the volume is. Then they wondered why response rates dropped and renewal conversations got harder.
The internal work is where AI earns its keep: pulling keyword clusters, drafting scopes, summarizing 40 pages of crawl data. Client communication is different. It's the one surface where the client is actively evaluating whether you're paying attention. Getting AI in client communications wrong doesn't produce bad writing — it produces writing that's fine, generic, and quietly erodes the thing you're actually selling.
Here's where to draw the line, based on what actually goes wrong.
The short answer: nine situations where AI should stay closed
Do not use AI to write:
- Any message delivering bad news — missed deadline, blown budget, flat results
- Responses to a complaint or an unhappy client email
- Anything involving money: scope changes, rate increases, invoicing disputes
- Strategic recommendations where you're asking the client to make a decision
- Apologies
- Renewal and upsell conversations
- Anything containing numbers you haven't personally verified
- Communication with a client who is already skeptical of your value
- The first three months of any new relationship
The pattern underneath all nine: these are moments where the client is buying judgment, not information. AI is excellent at information transfer and structurally incapable of accountability.
Bad news is a trust transaction, not an information transfer
An SEO agency I know missed a content deadline on a 14-piece cluster for a SaaS client. The account manager, under pressure, asked ChatGPT to draft the explanation email. It came back polished: acknowledged the delay, referenced "unforeseen resourcing constraints," proposed a revised timeline, closed with appreciation for their patience.
The client replied with one line: "Who wrote this?"
They'd read the same email structure a hundred times. What they wanted was the specific reason — that a senior writer had left mid-project and the agency chose not to hand the cluster to a junior rather than ship weak work. That's a defensible answer. The AI version buried it under corporate padding because AI defaults to hedging when the input is vague.
Bad news emails work when they contain three things AI reliably strips out: a specific cause, an admission of what you'd do differently, and a concrete date. Write those three sentences yourself. It takes four minutes.
The escalation rule
Set a hard trigger in your team's process: the moment a message contains an apology, a revised date, or a number lower than what we forecast, no AI touches it. Make it a rule rather than a judgment call, because the moments when people reach for AI shortcuts are exactly the moments they're stressed and least equipped to judge.
Numbers you haven't verified
This is the failure mode that costs actual money. AI-generated client reports hallucinate figures — not wildly, which would be easy to catch, but plausibly. Organic sessions up 23% when the real number is 12%. A ranking improvement attributed to a technical fix that shipped after the ranking moved.
A client catching one wrong number in a report doesn't think "AI error." They think "these people don't check their work," and they start auditing everything you've ever sent. That's a six-month trust rebuild triggered by one hallucinated percentage.
The safe version isn't "never use AI for reporting." It's AI never generates numbers — it only formats numbers you supply. Pull the data from GA4, Search Console, or your rank tracker, paste the verified figures into the prompt, and let AI handle structure and narrative flow. We wrote a full breakdown of this workflow in our guide to AI client summaries from project data — the core principle is that the AI reads from your project record, not from its own memory.
The causation trap
Even with correct numbers, AI writes causal claims it can't support. "The 18% increase in organic traffic was driven by our internal linking improvements" — maybe. Or maybe a competitor got hit by a core update, or seasonality, or a PR mention. AI will assert the flattering explanation because that's the pattern in every marketing report it trained on.
Overclaiming feels good for one month and terrible for six, because the client now expects that lever to work again. When it doesn't, you've got nothing to say.
Strategic recommendations
If you're asking a client to spend $40,000 on a site migration, or to kill a content program that isn't working, or to change their primary conversion path — write that yourself.
Recommendations carry implied accountability. When you say "we should deprioritize the blog and move budget to programmatic landing pages," you're staking your reputation on it. AI can't stake anything. And it shows: AI-drafted recommendations hedge constantly, present four options with balanced pros and cons, and end without a clear call. Clients read that as "they don't actually know."
The useful split — which we cover in more depth in automated insights vs. human strategy — is that AI can surface the pattern and a human decides what to do about it. AI finds that 60% of your ranking pages have thin internal link equity. A strategist decides whether that's this quarter's priority or a distraction from a fundamentally weak commercial page set.
The first 90 days of a relationship
Early on, the client is building a mental model of who you are. Every email is evidence. AI-flattened prose in month one teaches them that your agency sounds like every other agency — which makes it much harder later to charge a premium for being different.
Write your own onboarding messages, your first audit summary, your first strategy call recap. Once the client knows your voice, they'll read subsequent messages through that lens and small efficiencies won't register. Before that, everything registers.
Clients who are already skeptical
If someone has questioned an invoice, asked "what did we actually get for this month," or gone quiet after a flat results report, they're in evaluation mode. Every communication is being scored.
Generic language in that context reads as evidence for the prosecution. The recovery play is aggressively specific: name the pages, the dates, the exact changes, what you tried that didn't work. AI writes around specifics unless you feed it every one — at which point you've done the work anyway.
Where AI genuinely belongs in client communication
To be clear, the answer isn't abstinence. AI is legitimately good at:
- Meeting recaps from your own notes. You supply the messy bullets, AI structures them. Verify the action items.
- Routine logistics. Scheduling, asset chasers, access requests, reminders about overdue client-side approvals.
- First-pass report narrative around verified data — with a human editing pass before it ships.
- Tightening something you wrote. Draft it yourself, then ask AI to cut 30% without changing meaning. This preserves your judgment and your voice.
- Translating technical findings into plain language. Explaining hreflang conflicts to a marketing director is a genuine AI strength.
- Prep, not delivery. Anticipating objections before a QBR, drafting talking points you'll then rewrite.
Roughly 70% of your client message volume is Tier 1 or Tier 2 work where AI saves real hours. The 30% that isn't happens to be the 30% that determines whether the retainer renews.
Build the policy into your workflow, not your team's memory
Rules that live in someone's head get abandoned under deadline pressure. The agencies that get this right encode it: report templates that require a "data verified by" field, a client-facing message tier tagged on the task itself, an approval step before anything in Tier 3 leaves the building.
This is a tooling question as much as a policy question. If your project system holds the actual delivery record — time logged, deliverables shipped, audit findings, phase status — then AI-assisted summaries pull from verified facts rather than inventing them, and your team spends its writing energy on the messages that matter. It's one of the reasons we built PeakKR's AI features around project data rather than open-ended generation, and it's worth checking how your current stack handles this if you're comparing agency PM tools.
The one-question test
Before you open a chat window for a client message, ask: if the client knew AI wrote this, would they feel differently about it?
Nobody cares that AI formatted a meeting recap. Everybody cares that AI wrote the apology, the recommendation, or the explanation of why results are down. That reaction is the line.
Quick checklist before you send
- Does this message contain bad news, an apology, or a revised deadline? → Write it yourself
