The AI Stack I Actually Use for Client Marketing

Who this is for: marketers and small-firm owners who want a working AI setup without buying every new tool. I'll explain the jargon as I go.

Quick glossary (terms used below):

Building an effective AI stack for marketing isn't about owning the most tools — roughly half of all software licenses go unused, and AI is making the pile bigger. The fix is a smaller stack you actually master. I run two layers: Perplexity Max and Comet for cited research and browser tasks, Claude Pro for long-form writing, and a local Qwen model for confidential work, all wrapped in a routing policy and a verify-everything rule. Here's how it runs in real client work, and why owning more tools usually makes people worse at the work, not better.

Roughly half of all software licenses sit unused — one analysis puts it at 50%, another at 51% of enterprise SaaS. AI made the habit worse because a subscription is cheap to start and easy to abandon. Surveys put unsanctioned "shadow AI" use somewhere between 50% and 78% of workers, much of it without oversight or training.

That's the problem. Buying a tool feels like progress. Getting good at one is slower and more boring, but it's the only part that pays. The enterprise dollar figures don't apply cleanly to a small shop. The personal version does: a drawer full of logins you half-learned and a monthly bill you no longer question.

The tools in my stack fail on me every single day. That's not a confession that my setup is bad. It's what the research says to expect.

AI agents — the browser automation I lean on — fail roughly 70% of the time on multi-step office tasks run autonomously, not simple Q&A, in a Carnegie Mellon study. Stanford's 2026 AI Index still has agents missing about one in three attempts on structured benchmarks. And on legal questions, the Stanford RegLab study Large Legal Fictions found general-purpose models hallucinate between 58% and 88% of the time when asked specific, verifiable questions about federal court cases.In plain English: when AI is asked to chain together several office tasks on its own — open a site, find a file, copy data, paste it into a report — it gets the whole sequence right less than a third of the time.

The rate changes with the task. The lesson does not. Confident, wrong output is normal.

The math explains why. A step that's 95% reliable, chained twenty times, drops to a 36% end-to-end success rate. Multi-step automation doesn't stay reliable because each individual step looks fine. You need someone in the loop who knows what “right” looks like.

For me, the failures show up most often in product information and specifications. An AI tool will grab the wrong model year, mix trims, repeat an outdated spec, or turn a dealer-page detail into a broader claim it cannot support. That is exactly the kind of mistake that looks harmless in a draft and expensive once it is live.

The fix is now a rule: any product spec, finance number, towing capacity, warranty detail, feature claim, or model-year statement gets checked against the manufacturer's page, the dealer's current listing, or the document the client supplied. If I can't verify it there, it doesn't go in the piece. This is why I don't treat citations as decoration. For dealer work especially, the citation is how you keep a confident sentence from becoming a wrong one.

Perplexity Max is my research desk because it cites sources inline, so checking a claim takes a click instead of a second round of digging. With the failure rates above, the citation trail is the product.

Max also gives me a rotating bench of model families in one workflow — Perplexity's Sonar, OpenAI's GPT, Google's Gemini, Anthropic's Claude, Moonshot's Kimi, and NVIDIA's Nemotron. Model Council lets Max users run one question through three frontier models at once and compare where they agree, where they split, and what one caught that the others missed. I don't want six subscriptions. I want a small number of tools I know how to use.

Comet is an agentic browser. It acts across pages instead of just showing them. One daily example: an active client Facebook page generates more comments than anyone wants to monitor by hand, so Comet runs a pass that flags the negative ones, rates severity, extracts the terms, and adds them to Facebook's negative-keyword filter. I used to do versions of this manually.

Most of what Comet handles is not impressive demo material. It writes CSS fixes for client sites, finishes whatever technical task I have open in a tab, and clears one-off jobs that used to break my day. Max lets me pick the model driving the agent when a job needs more reasoning than speed. My heavier recurring use is client SEO: auditing pages and formatting content for traditional search, Google's SGE, and answer engines.

The real jobs, by tool: Perplexity Max pulls the questions buyers actually ask before a purchase — for a powersports or marine dealer client, that might mean financing, towing capacity, winterizing, or warranty — and verifies a spec or finance figure with a citation before we put a number on screen. Comet runs a site audit and exports every page missing a meta description or H1, then reformats a blog draft for answer engines. Claude Pro drafts a five-email sequence from a pillar transcript in the client's voice and turns an approved outline into a full script with B-roll notes. Local Qwen summarizes internal numbers or vendor pricing and cross-checks facts in a near-final post, both off the cloud.

Perplexity meters Max in credits, not a raw token count. The current plan page lists 10,000 monthly credits plus 35,000 bonus credits for Perplexity Computer, along with access to advanced reasoning models. That means I can run the browser agent, use heavier models, and do deeper research without rationing every task. Tokens still exist underneath, but for this workflow credits are the number that matters. Longer jobs and heavier models spend faster.

Say we're producing a pillar video for a marine dealer client — a buyer's guide to choosing a first pontoon — and cutting it into a month of content. Perplexity Max handles the cited research on what buyers ask and what competitors are publishing. I keep it in a dedicated Perplexity Space for that client, where I can upload their own sources, set instructions, and keep threads organized.

Comet runs the answer-engine formatting and the page audit. Claude Pro drafts the long pieces — the script and email series — inside a Project that holds the client's voice across the batch. When a job is a defined, multi-step build rather than a chat, I hand it to Claude Cowork, which works across files and apps on the desktop to return a finished deliverable while I move on.

Claude inside Perplexity is Claude in a cited search workflow. Claude Pro is the workspace I use when I'm building something longer. Projects, Artifacts, and Cowork matter when I'm working on a script, email sequence, or rewrite that needs context to stay intact across the whole job.

That is not paying twice for the same thing. It's paying for different work surfaces. I use Claude-in-Max when I want a quick cited answer. I use Claude directly when I'm living inside the draft.

I run a free AI model on my own laptop (Qwen3.5-35B-A3B) because some client material should not go to the cloud: sales figures, vendor pricing, internal notes. It stays on my machine.

It runs fast on a normal laptop because of a design called Mixture-of-Experts — in plain terms, only a small slice of the model wakes up for each word, instead of the whole thing. That means I get serious AI horsepower without needing a data-center GPU. It is my private workhorse and a second check before client work goes out.

A data-routing policy decides what goes where. Public or shareable work can go to Perplexity and Claude. Client work gets identifiers removed before it touches the cloud. Sensitive numbers stay on local Qwen or don't go near AI at all.

A knowledge base keeps tasks from starting cold. Research, templates, and client voice guides live in Perplexity Spaces and Claude Projects. Every output that reaches a client is checked by a person first. That last rule matters more than the model choice.

For me, yes. I estimate the stack saves me 8 to 10 hours a week. I pay for the higher Max tier instead of the cheaper one so I don't run out of usage in the middle of a project. (Cheaper AI plans cap how much you can do per hour or per day; once you hit the cap, you wait or get cut off.) That kind of interruption breaks the steady focus that creates the savings in the first place. The higher tier gives me more usage and access to the best models, so I can run the browser agent, push long research, and keep going without counting every query.

But I save those hours because I assume the tools are wrong until I check them. Someone who buys the same subscriptions and trusts the output will not get my result. They'll get a faster way to ship a wrong price, a bad spec, or a confident sentence that should have been cut.

This is overkill for light use. If you're a solo operator doing occasional AI work, a $20 plan and discipline beat Max plus a local model you won't maintain. The stack earns its cost at real volume, when interruptions cost more than the subscription, and only if verification is part of the work.

This works because the stack is small. A few tools, used every day, split between cloud and local, with rules around them. The advantage is not the subscription list. It's the reps, the skepticism, and the habit of checking before anything leaves the building.

The tools are easy to copy. The way you use them isn't.

What AI tools do you actually need for marketing work?
Fewer than most people buy. A small, mastered stack — one research tool, one agentic browser, one writing tool, and a local model for confidential work — beats a pile of half-learned subscriptions. Roughly half of software licenses go unused.

Is Perplexity Max worth $200 a month?
It is worth it at real volume, when you need higher usage limits, multiple frontier models, and fewer interruptions during long work sessions. For light use, Perplexity Pro at $20 plus discipline is the better value.

Why pay for Claude Pro if Claude is already in Perplexity?
Claude in Perplexity is a cited search workflow. Claude Pro is a drafting workspace with Projects, Artifacts, and Cowork for longer, multi-step work.

How often do AI tools make mistakes?
Often enough that verification has to be part of the workflow. A Carnegie Mellon study found autonomous agents fail roughly 70% of the time on multi-step office tasks, and Stanford found general-purpose models hallucinated 58–88% of the time on specific legal queries.

Why run an AI model locally instead of in the cloud?
For confidentiality. A local model like Qwen3.5-35B-A3B keeps sensitive client data off third-party servers, and its Mixture-of-Experts design helps it run on local hardware.

About the author

Bob Gonsalves writes The Small Firm AI Playbook for professional small firms — law, accounting, consulting, agencies, and other knowledge-work practices — on how to actually use AI for client work. The Playbook covers workflows, tools, verification habits, and the failures worth talking about. Bob is relentlessly AI-curious and uses AI tools daily to ship better work in less time — from cited research and SEO audits to long-form drafting and confidential analysis on a local model. The focus is practical: where AI saves real hours, where it costs more than it gives back, and how a small firm can use it without burning budget or trust. Connect on LinkedIn or reply to this newsletter to compare stacks.