How to choose an AI marketing agent: 8 criteria
AI marketing agents promise to do the work of a marketer. Demos all look good, so judge them on what happens with your real data, your channels and your money. Here are eight criteria, the exact question to ask a vendor for each, and how Toby answers them, limits included.
The 8 criteria
For each one: why it matters, the question to ask the vendor, and Toby's answer.
- 1
What data it can actually read
An agent that only reads your homepage can only give generic advice. The useful ones read your analytics, search data, revenue and the tools you already work in, and let you add the ones that are not on their list.
Ask the vendor
Which of my tools can you read today, without a waiting list or a platform review? What happens with a tool that is not on your list?
How Toby answers
Ready-made connectors for Google Analytics 4, Search Console, Stripe, Shopify, WordPress, Webflow, Notion, Gmail, X, Rewardful, HeyGen, Slack, Discord and meeting notetakers (Fireflies, tl;dv, Fathom). Any other tool through a remote MCP server, its API (Toby reads the docs and proposes actions you tick) or a Zapier, Make or n8n webhook. Meta Ads and Google Ads account connection waits for those platforms' reviews. With nothing connected, Toby shows no numbers rather than invented ones.
- 2
Evidence behind every recommendation
“Post more on LinkedIn” is not a recommendation. A good one says what it is based on, so you can check it in two minutes: a figure from your data, a sentence from your site, a gap found in the audit.
Ask the vendor
For this recommendation, show me exactly where each claim comes from. What stops the AI from inventing a number?
How Toby answers
Each card lists its reasons and the evidence used (your data, your site, the audit), an impact given as an order of magnitude, the effort and the agent in charge. A second, critical model reads every recommendation before you see it and keeps it, rewrites it or drops it; an unsupported number or feature is a reason to rewrite or drop.
- 3
Deliverables you can use as they are
The time saved is in the work, not the advice. Check whether the first step is a finished piece you could paste today, written in your voice and with your product names.
Ask the vendor
When you recommend something, what do I receive: a to-do item, or the finished first deliverable?
How Toby answers
Every recommendation comes with its first deliverable already prepared: the email sequence, the new title tags, the ad copy, the reply draft. It draws on your Brand Brain (six editable documents written at sign-up), so the voice and positioning stay consistent.
- 4
Approval and autonomy you control
An agent that can post, email or spend on its own is useful only if you decide where. Look for per-agent settings, clear limits on what “autopilot” covers, a way to stop everything, and a log of what was done.
Ask the vendor
What can your agents do without asking me? Can I set that per agent? Is there one button to stop everything?
How Toby answers
Three levels per agent: Suggest only, Ask before publishing (the default) and Autopilot. Autopilot only applies when both the agent and the company are on Autopilot, is limited to safe actions on your own channels (blog posts, Webflow and Notion drafts, X posts, Gmail drafts) and never covers ads or sending emails. A global Pause button stops the team, and every action is logged.
- 5
Channels covered, and what really gets published
A long list of channel logos often hides the fact that most of them produce drafts you post yourself. That can be fine, as long as you know it before you buy.
Ask the vendor
For each channel, do you publish directly today, or prepare drafts that I post myself?
How Toby answers
Twelve agents cover SEO and GEO, Meta Ads, Google Ads, influence, affiliates, Reddit, X, LinkedIn, social media, UGC video and data. Direct publishing today: WordPress articles, Webflow drafts, Notion pages and X posts with your own X API keys (paid per use by X). LinkedIn, Instagram, Facebook, TikTok, YouTube and Reddit: ready-to-post drafts until each platform approves the app; you post them and click “Mark as published”. Ads: plans and drafts that you launch.
- 6
Results measured after the fact
Most tools stop at the suggestion. Ask whether the agent comes back to say if it worked: that is what makes the next recommendation better than the last one.
Ask the vendor
After I apply a recommendation, how and when do you tell me whether it worked?
How Toby answers
Seven days after you mark a recommendation as done, the agent compares the related metric over the last 7 days with the 7 days before and posts the result in #team; it is kept in Toby's memory. Without analytics connected, it says the result is not measurable yet instead of guessing. When you dismiss a recommendation, the reason you give is remembered too.
- 7
Costs you can predict
Look at three costs: the subscription, the AI usage (some tools bill credits or tokens), and third-party fees such as API access or video credits. A runaway agent should not be able to run up a bill.
Ask the vendor
What will I pay per month, all in? Is AI usage capped? Which third-party costs are on me?
How Toby answers
One plan at $99 a month per company, with all 12 AI agents and every integration included; 7-day free trial, cancel anytime. A daily AI cost cap per company is checked before each AI call. Third-party costs stay yours: the X API is paid per use, and HeyGen renders use your HeyGen credits (each render needs your approval).
- 8
Security and data isolation
You are handing over keys to your site, analytics and sometimes your mailbox. Ask how those keys are stored, what the agent may call, and how content it reads is handled.
Ask the vendor
How are my tokens stored? Can your agents call any address on the internet? Is my data separated from other customers'?
How Toby answers
Tokens are encrypted with AES-256-GCM and never sent back to the browser. Data is isolated per company. Outgoing calls only go to public addresses on each connected tool's own domain. Emails and meeting transcripts are read as data, never as instructions. You can disconnect any tool at any time.
Red flags in a demo or a trial
Precise numbers for a business it has never connected to. If the tool quotes your traffic or your conversion rate before you connect anything, those numbers are made up.
Publishing switched on by default. Posting or emailing in your name should be something you turn on, agent by agent, not something you discover.
Recommendations without sources. If you cannot see where a claim comes from, you cannot tell good advice from confident guessing.
Autopilot on ad spend without caps. Spending money is the one action that should always ask first.
Pricing you cannot work out. Credits, seats and add-ons that make the real monthly cost unclear before you start.
A simple 7-day trial plan
Day 1. Paste your URL and read the audit and the Brand Brain. Correct anything wrong: this is the base for everything else.
Day 2. Connect your analytics and one revenue source. Check that the numbers match what you see in those tools.
Days 3 to 5. Apply one or two recommendations whose first deliverable you can use as is. Dismiss the others with a reason, and see whether the next ones get better.
Days 6 and 7. Ask the CMO agent to prioritize next month, and decide whether that list is one you would have written yourself, or better.
Where to go from here
If you are still deciding which kind of help you need (an agent platform, a general assistant, an agency or a person), our overview of AI CMO tools compares the six categories. If you already have a name in mind, the side-by-side pages below compare Toby with Okara, ChatGPT, an agency and hiring a CMO.
Frequently asked questions
What is an AI marketing agent?
Software that uses AI models to carry out marketing work with some autonomy: analyzing data, deciding what to do, preparing content or campaigns, and sometimes publishing. The best ones work from your real data and ask for approval before anything goes out.
Should an AI marketing agent publish on its own?
Only where you decide, and only for low-risk actions. In Toby, Autopilot is opt-in per agent, limited to drafts and posts on your own channels, and never covers ads or sending emails. Everything else waits for your approval.
One agent or a team of agents?
A single assistant can do many things if you steer it. A team splits the work by channel and adds a lead that ranks priorities across them. Toby uses a team of 12: its AI CMO delegates to 11 specialists who answer in the same thread.
How long should I test an AI marketing agent before deciding?
A week is enough to judge the quality of the audit, of the first recommendations and of the deliverables. Measured results take longer: Toby reports on each applied recommendation 7 days after you mark it done.
What happens to my data if I stop?
In Toby, you can disconnect each tool from Integrations at any time, which removes the stored token. Your company data stays isolated from other customers. See the Security page for details on hosting and storage.
Other comparisons and guides
Give your business a marketing team that knows what to do next
Create your account, paste your URL and get your growth audit, your Brand Brain and your first three recommendations, already prepared.
7-day free trial, then $99/month per company. Cancel anytime. Already have an account? Sign in