How to actually use Claude (or Gemini) for marketing operations
The artificial intelligence (AI) conversation in marketing has been dominated by "how to write blog content with AI" and "how to generate ad copy with prompts". That's the obvious stuff.
The real value of AI in marketing operations is less obvious. It's less about replacing creative work and more about speeding up the routine middle of the job: audits, summaries, briefs, standard operating procedures (SOPs), analysis, pre-meeting prep.
Here's how we use Claude (and occasionally Gemini, plus the odd other platform) every week at Sector 106. What it actually speeds up, where it gets in the way, and the specific workflows that are worth the effort.
What AI is genuinely good at (in marketing ops)
Six things, in rough order of how much time they save us.
1. Summarising long, messy documents. Audit reports, sales call transcripts, competitor websites, policy docs, regulatory updates. Take 30 pages, produce a one-page summary with the key actions pulled out. If we feed Claude a 14,000-word client transcript and ask "what did the client commit to and what did we commit to", we get a clean list in 20 seconds.
2. Analysing structured data and finding patterns. Drop in a comma-separated values (CSV) file of order data, ask "what's the relationship between average order value (AOV) and discount code usage". You get a real answer with caveats and follow-up questions. We did this with the Balmonds 12-month order export (22,832 rows). We found the subscription pattern (subscribers spend 3.9x more than one-time buyers) faster than we would have from spreadsheet pivots.
3. Drafting structured documents. SOPs, briefs, post-meeting notes, project plans. Anything with a known structure. We give Claude the messy notes from a meeting plus a template, and get a tidy first draft to edit. Three hours of work done in 30 minutes.
4. Stress-testing arguments. Before a board meeting or a client presentation, we paste in our argument and ask "what's the strongest counter-argument?" Then we either fix the argument or we're ready when the counter-argument lands in the room.
5. Quick creative starting points. Most of the output is unusable. The 5–10% that works is faster than waiting for it to come from one head. Good for ad copy variants, email subject lines, campaign names, and product naming. It's also useful for practical image jobs, like extending a photo's background to fit a different format, and for rough creative ideas a designer or copywriter can then take and make their own. Treat it as a starting point for a professional, not a finished piece.
6. Translating between technical and non-technical language. Explaining attribution windows to a founder. Writing up a Shopify migration plan for the marketing team. Turning a paid media report into board-ready commentary. AI is fast at this. Faster than us, sometimes.
What it's bad at (and don't trust)
The flip side. This matters just as much, because the failure modes are where AI users get caught out.
1. Maths. Genuinely bad. Claude and Gemini will confidently invent numbers when calculating percentages, ratios, or projections. Always verify in a spreadsheet or with a calculator. We've had AI report a "12.5% increase" that was actually a 4.2% increase. The plausible-but-wrong number is the dangerous one.
2. Real-time data. Trained on a snapshot, no access to the live web (without specific tools). It doesn't know what your latest Klaviyo flow performance looks like. It doesn't know what's currently on a competitor's site. Don't ask it to "check" anything.
3. Brand voice. Will default to AI voice without strong constraints. Em dashes everywhere. Triadic stacking. "Delve", "navigate", "tapestry", "leverage". Researchers have tracked how words like these spiked once these tools became widespread, and readers now spot them. You have to give it explicit anti-AI instructions or it sounds like every other AI-written piece on the internet.
4. Strategic decisions. Has no actual judgement. Will produce confident-sounding strategic advice that's based on average-of-internet thinking. Useful as a sparring partner. Useless as a decider.
5. Anything requiring real-world taste. Visual decisions. Subtle brand voice calls. Gut feels about clients or campaigns. AI can generate options. It can't tell you which one is the right one for your business.
6. Naming things you'll publicly attach yourself to. Names that come out of AI sound like AI. Use it for working titles, but get a human to settle the final name.
Six specific workflows that work
Concrete is better than abstract. These are the workflows we actually run weekly.
1. Audit summarisation
We feed Claude an audit document (Klaviyo, Google Ads, Meta, Shopify, full search engine optimisation [SEO]) along with a template asking for: top three findings, prioritised recommendations with effort and impact estimates, expected revenue range. Claude produces a tight one-pager from a 30-page audit in two minutes.
We then edit. Claude adds adjectives. We cut them. Claude smooths over uncertainty. We reinstate it. The final document is a hybrid: Claude's structure, our judgement.
2. Competitor analysis
Paste in a competitor's homepage, About page, and pricing page. Ask: "What is this brand's positioning, who's the target customer, what are the proof points, and what's their pricing strategy?" Then: "What gaps exist in their messaging that a competitor could exploit?"
The output is rarely brilliant on its own. It's usually a useful starting point that picks out the obvious things and lets us focus on the rest.
3. Monthly performance report writing
We give Claude the raw numbers (in a structured table), the bullet-point findings, and the client's brand voice guidelines. Claude drafts the commentary. We rewrite anything that drifts into agency-speak.
For Nailberry's monthly month-to-date (MTD) report, the AI-drafted commentary section is now where 60% of our writing time used to go. We use the saved time on the actual analysis instead.
4. SOP drafting from messy notes
After we figure out a new process (campaign launch, weekly key performance indicator [KPI] report, escalation flow), we paste the rough notes into Claude and ask for an SOP in standard format: trigger, owner, steps, escalation, output. Claude formats. We review.
The discipline of forcing notes into SOP shape often reveals gaps in the process. AI is what makes us actually write the SOP down rather than carrying it in our heads.
5. Brief expansion
Client sends a two-line brief: "We want to launch X product for Mother's Day." We expand it into a campaign plan (target audiences, channels, creative angles, key messages, timeline, budget split, KPIs) and use that as the discussion document for the kick-off call.
The expanded brief is wrong about half the time. That's fine. The discussion is faster because we're correcting a draft rather than building from scratch.
6. Pre-meeting prep
Before a client call, we paste in: previous meeting notes, the most recent KPI report, any unread email threads, the agenda. Ask Claude to produce a "pre-meeting brief" with: open issues, key numbers, items the client is likely to raise, items we want to surface.
A 15-minute prep that used to take 45.
Going further: scheduling, artifacts, code checks and memory
Once the basics are working, a few of Claude's features take this further. These are the ones we use most.
Scheduled tasks. Claude can run a job on a schedule, so work that used to depend on someone remembering now happens on its own. We use it for recurring jobs: a Monday summary of last week's numbers, a daily check on a metric that flags us if it moves too far, a weekly pull of new search terms to review. You set it up once and it runs. A person still checks the output before anything reaches a client, but the gathering and the first draft happen without us.
Artifacts. An artifact is a small, self-contained page Claude builds that you can save and open again later, and it pulls fresh data each time. Instead of a static table that's out of date by lunchtime, you get a page you keep coming back to: a campaign tracker, a weekly numbers view, a simple dashboard. Good for the regular "where are we right now" question, without rebuilding the report by hand every time.
Code checks. A lot of marketing work has code in it now. Google Apps Script for automations, Looker Studio formulas, tracking tags, Liquid in Shopify themes, the rules behind your filters. Claude is good at reading that code, finding the bug, and explaining in plain terms what it does. We use it to check a tracking setup before it goes live, or to work out why a Looker Studio field is returning the wrong number. Test it properly before you trust it, but it's faster than staring at the code on your own.
Memory files and skills. This is the part that gets better for each client over time. Claude can keep memory files: notes about a client that carry over between sessions, so you don't re-explain the brand voice, the targets, the tech stack, and the dos and don'ts every time. You can also build skills: saved instructions for a job you run often, so the way you want a monthly report written or an audit run is captured once and reused. The more you give it (the brand guidelines, past reports, naming conventions), the closer the first draft gets to something you'd actually send.
One honest point: you are not retraining the model itself. You are giving it a growing, client-specific memory and a set of house rules to work from. That is what makes it sharper for each client over time, not some change to the underlying AI.
Rules for using AI well
Six rules we apply across every workflow.
1. Always provide context. Who's the audience, what's the goal, what's the brand voice, what should be excluded. The first time you use AI without a system prompt or context, you'll get generic. Set context, get specific.
2. Always edit the output. AI gets close. It's rarely right first time. Editing is where the value is added, and where the brand voice is reinstated. Anything we ship has been rewritten at least once.
3. Never trust the maths. Verify every percentage, ratio, and projection. Always. Without exception.
4. Build a reusable prompt library. The same five or six prompts run weekly. Save them. Improve them. Don't retype context every time.
5. Don't outsource judgement, only execution. AI doesn't decide. AI accelerates the execution of decisions you've made.
6. Watch the brand voice drift. AI will erode brand voice over time if you keep accepting AI-default language. Cross-check against your style guide. Cut "delve", "navigate", "leverage", "tapestry", "robust", "in today's fast-paced world" on sight.
Common mistakes
Things we see other operators do that don't work.
Treating it like Google. AI doesn't know facts. It produces plausible language. If you ask "what's the average return on ad spend (ROAS) for skincare brands in the UK", you'll get a confident-sounding number with no source. Don't quote AI numbers without verification.
Trusting the first output. First drafts are first drafts. Iterate two or three times before settling.
Letting it write in its own voice. The output is recognisable as AI-written if you don't edit it. Readers can tell. Clients can tell.
Using it for strategic decisions. AI can be a sparring partner. It cannot replace a senior person who's run the channel for 20 years.
Not citing internal sources. When AI produces a recommendation, ask "what is this based on?" If the answer is "general best practice", treat it as a starting point, not a conclusion.
Storing client data in tools without a data processing agreement (DPA). Check the tool's data processing terms before you paste in customer data. Claude Team and Enterprise come with a DPA and don't train on your data; the Free and Pro tiers don't. Other providers' paid business plans are usually similar, but check each one. Don't paste personal data into a free consumer tier.
How Sector 106 uses it
Concrete examples from real client work in the last quarter.
Nailberry trend reports. Two market deep-dives produced for Nailberry: UK Nail Polish Ecommerce Trends 2026 (16 pages) and UK Nail Treatments Ecommerce Trends 2026 (18 pages). Claude assisted with synthesising market data, drafting consumer behaviour sections, and producing initial structures. We did the strategic recommendations and the brand voice editing. Three days of work compressed to one and a half.
Balmonds Klaviyo audit. The 13-flow audit for a different client (Rainbird, but the same pattern applies for Balmonds) used AI to summarise the analytics data per flow into a comparative table. Five hours of manual work compressed to one. The seven recommendations themselves came from human judgement.
Weekly KPI reports. Automated screenshot generation + AI-drafted commentary section + human edit + branded portable document format (PDF) wrapper. End-to-end automation for two clients now (Nailberry and one other). Saves us four hours per client per month.
Daily client briefings. Pre-meeting prep on every recurring call. AI produces the structured prep doc, we add the strategic commentary.
Per-client setup. For each client we keep a memory file (brand voice, targets, tech stack, dos and don'ts) and a few saved skills (how we run their audit, how their monthly report is written). The first draft gets closer to right every month, because it's working from their real context and our house rules, not generic defaults.
Tools we actually use (and why)
Claude (Anthropic). Primary tool. Best at long-document analysis, brand voice work when properly constrained, and code-adjacent tasks (Apps Script, Looker Studio formulas). Claude Pro is $20/month; Anthropic bills in US dollars, so it lands around £20 in the UK. Team and Enterprise cost more per seat and add a DPA.
Gemini (Google). Our main second option, useful where Claude is harder to access or where Google's own tools help (anything close to Google Workspace, for example). Google AI Pro is $19.99/month. We test other platforms from time to time, but Claude and Gemini do most of the work.
Midjourney. For image generation when a client brand needs custom imagery. We use it for hero images, ad creative, section backgrounds, and extending image backgrounds to fit new formats. Plans run from $10/month for Basic to $120/month for Mega.
Tools we tested and don't use: Jasper (over-engineered for what it does, brand voice features less reliable than Claude with good prompts), Copy.ai (output quality lower), Notion AI (fine for note summarisation, weaker for analysis).
The honest test
Three questions:
Do you have three to five reusable prompts saved somewhere your team can find them?
Do you have a written rule for what AI is and isn't allowed to do in your business?
Did the last AI-drafted document you shipped get edited before it went out?
If any answer is no, your AI workflow is going to get sloppy fast. Build the discipline before the volume. AI speeds up good processes and bad ones alike.
Want help building AI into your marketing operations?
Our Strategy & Operations service includes setting up AI workflows, prompt libraries, scheduled tasks, and a per-client memory and skill set so the tools get sharper for your brand over time. Plus the rules to keep them honest. Senior operator-led. 28 years of experience deciding what should be automated and what shouldn't.
Or read more: Strategy & Operations →
