Every marketing platform now has AI in the product name, the pricing page and the demo. Some of it is a language model bolted onto a form builder. Some of it changes how the work gets done. Telling the two apart before you sign a contract is the whole job of this guide.
The buyer's mistake is to evaluate AI features in isolation: how good is the copy, how clever is the chatbot. The useful question is what data the AI can see, what it is allowed to do, and whether its output lands where your team already works. Those three questions decide whether an AI marketing tool earns its licence or becomes another tab.
Start With the Job, Not the Category
AI marketing software splits into a few jobs, and most vendors do one of them well. Be clear which job you are hiring for before you compare products.
- Content production: drafting, editing and adapting copy, images and video at volume, governed by a brand voice.
- Data and targeting: enriching records, scoring leads, identifying accounts in market, building segments.
- Personalisation: adapting web pages, emails and journeys to the visitor or account.
- Conversation: agents that answer visitors and customers, qualify them and hand off.
- Orchestration and insight: agents that plan and run campaigns, and reporting that explains what moved pipeline.
A CRM-native platform such as HubSpot covers several of these with Breeze; standalone tools go deeper on one. Your shortlist should name the job first and the vendors second.
The Seven Things to Check
1. What data can it actually read?
An AI feature is only as good as the context it is given. Ask exactly which systems the tool reads: the CRM, the website, product usage, the knowledge base, the warehouse. Then ask whether it reads them live or from a copy that was synced last night. A personalisation engine that cannot see lifecycle stage will personalise on guesswork.
2. Where does the output go?
If the tool's work stays inside the tool, someone has to copy it out. Insist on write-back: content into your CMS, scores and enrichments onto the contact record, agent conversations logged in the CRM. In a HubSpot shop, that means a native integration or one that writes to the contact, company and deal objects, not a CSV export.
3. What is it allowed to do without a human?
Agentic features act. Check the permission model: can you define what the agent may send, change or spend on its own, and what needs approval? Can you see a log of what it did and why? A vendor that cannot show you an audit trail is asking you to trust a black box with your brand.
4. How does it hold your brand voice?
For content tools, ask how brand rules are encoded: a style guide the model follows, approved terminology, banned phrases, tone examples. Then test it with a brief your team would actually write and read the output critically. Fluent is not the same as on-brand.
5. How is it priced, and what happens at scale?
AI pricing is moving from seats to usage: credits, resolutions, conversations, tokens. Model your expected volume and ask for the price at three times that volume. A tool that is cheap in the pilot and expensive in production is a common surprise.
6. What does the security review look like?
Where is your data processed, is it used to train the vendor's models, what certifications hold, and can you control retention? Your security team will ask; get the answers before the pilot, not after.
7. How will you know it worked?
Agree the metric before the pilot: pipeline influenced, resolution rate, time saved per campaign, conversion lift against a holdout. If the vendor cannot help you measure it, that tells you something.
A Scorecard You Can Reuse
| Criterion | What good looks like | Red flag |
| Data access | Reads CRM, site and product data live | Works from uploads or nightly syncs |
| Write-back | Updates CRM records and CMS natively | Export to CSV |
| Permissions | Configurable actions, approvals, full audit log | "The AI decides" |
| Brand control | Style guide, terminology, examples enforced | A tone dropdown |
| Pricing | Clear usage model, predictable at 3x volume | Credits you cannot forecast |
| Security | SOC 2, data residency, no training on your data | Vague answers |
| Measurement | Holdout or attribution in the CRM | Vanity metrics in the vendor dashboard |
What the Checks Look Like by Category
The seven checks apply everywhere, but the questions that expose a weak product differ by job. These are the ones that separate the shortlist in each category.
Content tools
Ask to see the brand-voice setup, not the output. How are terminology, banned phrases and tone examples stored, and does every user's generation run through them? Ask how the tool handles facts: does it cite, does it flag numbers, does it invent? Then ask how content gets to your CMS and whether the tool keeps a record of what was AI-drafted, which you will need for review and, increasingly, for disclosure.
Data and targeting tools
Ask where the data comes from, how often it is refreshed and what the match rate is on your own records. Run a sample of a hundred contacts through it and count how many come back correct. Ask what a scoring model was trained on and whether you can see the features it uses; a score you cannot explain to sales is a score sales will ignore.
Personalisation tools
Ask how a visitor is identified and which data the personalisation reads: CRM stage, intent, firmographics, past behaviour. Ask how experiments are run and whether a holdout is built in. Ask what happens to a page when the tool is slow or down; personalisation that blocks rendering costs you more than it earns.
Conversational agents
Ask what the agent is grounded in and what it does when the answer is not there. Ask to configure a refusal and a handoff yourself during the demo. Ask for the resolution rate on a comparable customer, and how it is calculated, because vendors define resolution differently. Then ask to read a hundred real transcripts, not the curated ones.
Orchestration and reporting
Ask which systems the agent can act in and which it can only read. Ask how attribution is calculated and whether it matches the model in your CRM; two attribution models in one company is a standing argument. Ask what the agent does when a campaign underperforms, and whether you can see the decision it made.
The HubSpot Angle
If HubSpot is your CRM, the evaluation has a shortcut and a trap. The shortcut is Breeze: the Customer, Prospecting and Data agents, Copilot and the AI features in Marketing Hub already run on your records with no integration, and for many jobs they are good enough to make a standalone tool unnecessary. Check them first. The trap is the marketplace listing: an app that says it integrates with HubSpot may only pull contacts out, never write back. Ask specifically which objects and properties the integration writes, and whether activity is logged on the timeline. Our guide to Breeze agents on HubSpot covers what the native agents do and cost.
A Worked Example
A B2B software company with a six-person marketing team wanted faster content and better lead follow-up. Two vendors pitched an all-in-one AI marketing suite. Against the scorecard, the suite scored well on content and poorly on data access: it read HubSpot nightly and wrote nothing back. The team instead activated Breeze for prospecting and data enrichment, added a content tool with a proper style guide and CMS publishing, and set a single pilot metric for each: reply rate on AI-assisted sequences against a holdout, and hours per article. Twelve weeks later both numbers were defensible, the finance team could see them in HubSpot, and the suite was never bought. The point is not the tools chosen; it is that the decision came from the checks rather than the demo.
Run the Pilot Like a Test, Not a Trial
A free trial shows you the interface. A pilot shows you the outcome. Pick one workflow with a measurable result, run it for four to six weeks with a holdout, and involve the people who will use the tool every day. Compare against your baseline, not against the vendor's case study.
Write the pilot down before it starts: the workflow, the metric, the baseline, the holdout, the reviewer, the end date and the decision rule. A pilot without a decision rule becomes a subscription.
Ask yourself before signing:
- Which job is this tool hired for, and does it do that job better than what we have?
- Does its output land on the CRM record without anyone copying it?
- Do we control what it can do on its own?
- Will the price still make sense at production volume?
INSIDEA
Ready to put AI to work in your business?
Practical automation and AI workflows built on top of your stack, not bolted on.
Buy Around the CRM
Most of the value in AI marketing software comes from context, and the CRM is where the context lives. INSIDEA, an Elite HubSpot Partner and the AI-first growth operating system for modern businesses, helps teams choose AI tools that fit the HubSpot record, activate Breeze where it does the job, and connect standalone tools so their work shows up in pipeline reporting. If you are shortlisting now, we can run the evaluation with you against your data rather than the demo. For a view of specific tools by job, see our guide to AI tools for B2B marketing.
How INSIDEA works
