When to trust AI in your marketing ops, and when not to
The most useful question I have been asked this year was from a CMO who had inherited a Salesforce instance and a HubSpot instance that did not talk to each other cleanly. She asked me, plainly, "Where am I supposed to be using AI in this stack, and where am I supposed to be staying away from it?"
It is a better question than most of what gets asked, because it presumes the answer is not uniform. It is not. AI belongs in some places in a marketing operation and absolutely does not belong in others. The decision is not about how powerful the model is. It is about three properties of the work itself.
Reversibility
The first question to ask of any task before letting AI run it: if this is wrong, how hard is it to undo?
Drafting a blog post is highly reversible. The post does not exist until somebody clicks publish, and a human is going to be in the loop on that click. The cost of a bad draft is the time it took to read it.
Sending an email to a hundred thousand contacts is not reversible. Once the email is in the inbox, it is in the inbox. If the segmentation was wrong, the wrong contacts now have an impression of your brand that you cannot retract. If the link was broken, the bounce rate is now permanent in the historical data. Recovery is partial at best.
Updating a CRM field on twenty thousand records is somewhere in between. It is technically reversible if you have a backup, but the operational disruption while you reverse it is real. Sales notice. Reporting goes sideways. Trust drops.
The rule we use: AI gets to operate freely on highly reversible tasks. AI gets to propose on partially reversible tasks. AI gets nowhere near irreversible tasks without a human approval step that is honestly engaged, not rubber‑stamped.
Observability
The second question: when the AI is wrong, will you notice?
A wrong blog post is observable. A human will read it before publication. The error surfaces immediately.
A wrong segmentation rule is sometimes observable. If the segment is named and reviewed in the next campaign, the error surfaces fast. If the segment is buried inside a multi‑step automation, the error may run for months before anybody notices a population delta.
A wrong attribution model is barely observable. The dashboard still renders. The numbers still look reasonable. Nobody knows the model is misallocating credit until somebody compares the dashboard to the CRM directly, which most teams do once a quarter at best.
The rule: AI gets to operate on observable tasks. AI does not get to operate on tasks where its errors will hide for a long time. If you cannot define how you would notice a failure within a reasonable window, you do not yet have the instrumentation to deploy AI in that part of the stack.
Boundedness
The third question: is the scope of what the AI can affect well‑defined, or could a single bad output cascade through the system?
A drafting task is bounded. The output affects one document.
A scoring rule update is bounded if the score affects only one routing decision; unbounded if the score feeds three other systems that compute their own decisions on top of it.
A workflow that writes to a custom CRM field that is consumed by sales reporting, BI dashboards, and a sync to a separate system is highly unbounded. A bad write cascades. A bad write that runs on a schedule cascades repeatedly.
The rule: AI gets to operate on well‑bounded tasks. Unbounded tasks need a human in the loop or a hard scope limit on what the AI can write to.
Putting the three together
In our practice, we end up with roughly this division of labor on most engagements.
AI draws drafts, summarizes sales calls, generates first‑pass segments for a human to review, prototypes dashboards, scores intent on incoming behaviors, suggests next‑best‑content, and rewrites copy against a voice guide. All of these are reversible, observable, and bounded.
Humans (or rules engines that humans wrote and audit) handle the final segmentation decisions before a send, the actual send, the lead routing logic, the scoring threshold that determines hand‑off to sales, the workflow design, and any write to a field that more than one downstream system consumes.
This division is not based on what AI is capable of doing. It is based on the operational cost of being wrong. Capability is increasing every quarter; the operational cost of an irreversible, unobservable, unbounded mistake is the same as it was in 2018 - high enough that you do not want to find out the hard way.
Two specific places to be careful
Two specific patterns are getting teams in trouble right now and are worth naming.
Letting AI run reactivation campaigns against dormant lists. This is unobservable (you will see deliverability damage two weeks later, not at the time of the send) and partially irreversible (a sender reputation hit takes months to recover). I have written about the downstream consequences for deliverability and the topic is also covered in our DMARC piece. Either is worth reading before you let an LLM rebuild your reactivation strategy.
Letting AI define lead scoring rules without a human review of every change. This is unbounded (the score feeds routing, dashboards, and forecasting) and partially observable (you may not see the error until a sales hand‑off goes sideways). AI can suggest scoring changes. A human owner has to approve them and document the rationale.
The honest takeaway
There is no general answer to "should I use AI here." There are three specific questions: how reversible is this, how observable is this, how bounded is this. If the answers are good, use the model. If any answer is bad, do not.
Most of the AI‑in‑operations work we get hired to clean up failed at least one of those three questions. None of the cleanup is mysterious; it is just unglamorous.
If you want a read on which parts of your stack are safe to AI‑enable and which are not, that is what an operational audit gets you. The starting point is here.
Philip Easley-Bosley is the founder of Tactical Marketing and a thirty-year expert marketing consultant. His path to founding the firm ran through sales and marketing leadership, years inside Act-On Software consulting with thousands of clients as Lead Marketing Automation Strategist, and a consistent priority on training and team building that a linear career could not have produced. He sets strategy, owns the architectural calls on every engagement, and writes about marketing operations, automation, and the discipline of building systems that hold up on Monday morning.