The biggest AI challenge in marketing is not access to a model or automation tool. It is controlling the path from data input to AI output to business action. Before a system can change a CRM record, publish copy, or influence outreach, a team should be able to explain what triggered the process, which data it used, what result is acceptable, and who handles an exception.
Start with one bounded, reversible task. Define the permitted inputs, expected output, validation rules, destination, failure behavior, and named owner before enabling automation. Technical risks concern data and reliability; ethical risks concern fairness and accountability; organizational risks concern skills, ownership, and adoption.
Let AI propose a bounded result; let a defined validation gate and accountable owner control the business action.
What is the biggest AI challenge in marketing?
The central challenge is operational control. A marketing team may ask AI to classify an inquiry, summarize a brief, or extract a value from a CRM record, but the model should not be treated as the final business decision-maker. The workflow must limit what it can see, restrict what it can return, test the result, and decide what happens when the result is missing or unexpected.
A good first use case has four characteristics: the input can be limited, the output can be expressed as a small contract, a wrong result is reversible, and a person owns the exception path. For example, classifying a form message into product_question, pricing_request, support_issue, or other is usually easier to control than allowing AI to decide eligibility, consent, billing status, or an irreversible customer action.
The operating principle is simple: do not give AI a consequential action until the team can identify the trigger, permitted data, acceptable output, validation gate, destination, and exception owner.
Why AI projects stall: the risks behind the numbers
Survey results can help a team decide what to investigate first, but they do not explain why a particular company’s project stalled. HubSpot’s accessible 2025 AI survey summary reports that 42% of surveyed marketers said data privacy concerns prevented their team from adopting new AI tools, 39% cited training and time investment, and 35% cited too many similar tools that do not connect. These figures should be attributed to HubSpot’s summary and not extended with additional percentages that have not been verified against the underlying report. Read HubSpot’s survey summary.
Use those barriers as discovery prompts. Identify the data class involved, the tool and workflow owner, the path between systems, the time required to train reviewers, and the outcome the pilot should improve. A concern about disconnected tools, for example, calls for a map of the current data path before another application is added.
Choose a task AI should handle, and one it should not
Use explicit rules when a decision depends on structured data, a fixed threshold, consent, suppression, billing, legal status, or an irreversible action. If a short auditable condition fully expresses the decision, AI adds uncertainty without adding useful interpretation. A suppression rule should block outreach; an AI classifier should never override it.
AI is more useful for bounded interpretation of unstructured material, such as classifying an inquiry, extracting themes, or drafting a summary. The result should remain advisory until the workflow checks it. Send other, an empty response, a null value, or an invalid category to a person or exception queue rather than guessing.
Prefer a staging field or review queue over direct publication or a consequential record change. Consider both the cost of a wrong result and how easily it can be corrected when selecting the task.
| Decision condition | Better control | Example | Fallback |
|---|---|---|---|
| Structured rule or threshold | Deterministic condition | Block outreach to a suppressed contact | Log the rule result and route policy conflicts to the owner |
| Unstructured text needs interpretation | Bounded AI classification or extraction | Label an inbound inquiry | Send null, invalid, or ambiguous results to review |
| Content needs a first draft | AI-generated draft in staging | Prepare an email from an approved brief | Save unpublished copy for reviewer approval |
Design the workflow before enabling automation
A usable workflow makes each handoff visible: trigger or source, permitted input, bounded AI task, structured result, validation and decision gate, destination action, and named exception owner.
HubSpot documents AI workflow actions that can analyze, summarize, categorize, or populate properties for enrolled CRM records, with results available to later workflow steps. Availability depends on settings, permissions, credits, and subscription. HubSpot’s Data Agent custom prompt action has a 3,000-character limit, uses a model that is not connected to the internet, and can fail with null values when credits are insufficient. A null response should be treated as a failure and should not overwrite a trusted value. See HubSpot’s AI workflow documentation.
HubSpot documents filter-based, event-based, scheduled, webhook-based, and manual enrollment triggers. Records enroll only the first time they meet enrollment criteria by default, and re-enrollment must be enabled. Enrollment settings govern workflow behavior, not duplicate safety across systems. The product documentation establishes capabilities and constraints, not factual accuracy, legal compliance, or production reliability.
For inquiry classification, a practical sequence is:
- A form submission creates or identifies a CRM record and supplies a stable submission ID.
- The workflow passes only permitted fields, such as message text and country, to the AI action.
- The action proposes one category from the approved enumeration.
- Deterministic conditions reject null, malformed, overlong, or out-of-list responses.
- A valid proposal goes to a staging property or triage queue. The CRM owner handles conflicts, credits failures, and unexpected output.
Test representative records before activation, including empty messages, long messages, ambiguous requests, prohibited data, and duplicate submissions.
For a content draft, an automation can pass an approved brief to an AI action, return defined fields, and send the draft to a reviewer. Zapier documents structured output options in AI by Zapier and a Human in the Loop step that can pause a Zap for a reviewer to approve, decline, or edit. These are documented capabilities, not an official end-to-end publishing template. Configure decline and timeout behavior deliberately: for consequential publishing, stop the run or create an exception task rather than silently continuing. See AI by Zapier’s documented capabilities and Human in the Loop approval options.
A proposed inquiry result could look like this:
{
"source_event_id": "form_submission_1042",
"record_id": "contact_381",
"category": "pricing_request",
"ai_status": "proposed",
"prompt_version": "inquiry-classifier-v1"
}This is an illustrative contract, not a vendor schema. The source event ID identifies the incoming submission, while the CRM record ID identifies the contact. A validation step should reject a missing ID, missing category, null response, or category outside the approved list before a production-field update.
Teams aligning CRM properties, workflow ownership, and system-of-record decisions can explore HubSpot systems consulting.
Protect data and make the review gate real
Set approved-tool and data rules before sending customer or internal information to an AI service. Select only the properties needed for the task, and involve privacy, security, and compliance owners in defining prohibited inputs, retention expectations, access, and approved destinations.
IBM’s 2025 research found that 13% of surveyed organizations reported breaches involving AI models or applications. Among those organizations, 97% said they lacked proper AI access controls. The 97% figure applies to organizations reporting AI-related breaches, not to all organizations. Read IBM’s research announcement.
Zapier AI Guardrails can check for PII and other risks, but Zapier warns that false positives can occur and that Guardrails are not a standalone defense. Its documentation describes temporary retention in logs and Zap runs. A detector therefore does not mean information was never transmitted or retained; check the platform’s handling terms and your approved-data rules. Review Zapier AI Guardrails’ limits and data handling.
For a sensitive CRM change, configure a reviewer to approve, edit, or decline, and set a timeout path. Decline or timeout should stop the write or create an exception task, not count as approval. Preserve the original AI output separately from any reviewer-edited result so the decision can be audited.
Validate output, preserve provenance, and prevent duplicates
Structured output is not validated output. Even when an AI action is configured to return specific fields, deterministic checks should confirm that required values exist, types are correct, categories are allowed, and text fits the destination. Keep the previous trusted CRM value if the action returns null, fails, or supplies an invalid category. Record the failure and route it to the workflow owner.
A proposed contract might include ai_status with values such as pending, approved, rejected, or failed; an enumerated ai_label; a stable source_event_id; a prompt_version; and a run identifier. Keep proposed output in staging. Promote it to the business field only after the decision gate passes.
A record enrolling once in a workflow can still be duplicated by retries, concurrent writers, or another workflow. Derive a stable event key, enforce destination-level uniqueness, or use a transactional upsert where supported. A lookup followed by a create is not a race-safe control.
Keep records at the correct grain:
- Source-event record: one row per incoming form submission, webhook delivery, or other source event, keyed by source system and stable event ID. Multiple deliveries remain distinguishable when the source supplies distinct IDs.
- AI-run observation: one row per execution or attempt, with a unique run ID, source event ID, prompt version, status, validation outcome, and error where relevant. Retries create separate run observations.
- Business-state record: the current approved CRM value, not a dump of every model attempt.
For concurrent processing, use a stable idempotency key such as marketing_form:submission_1042:classification-v1. Enforce uniqueness on that key at the destination or use an atomic upsert. If the destination cannot provide either control, use a durable idempotency table or serialized queue. Do not use a timestamp, contact name, or URL alone as the key.
If measuring citations or AI visibility, keep those grains separate too. A citation record represents one citation attached to one AI run, while a daily aggregate represents a defined reporting period, query set, brand, model variant, and aggregation version. Any mention rate or share metric must state its denominator. Do not treat a per-run observation as a team-wide rate.
Address skills, trust, and accountability
Training is rollout capacity, not a task to postpone until after procurement. Teach reviewers the permitted-data rules, tool limits, validation steps, exception paths, and how to record corrections. Assign a workflow owner for prompts and failures, a business owner for the destination field or content, and a reviewer authorized to approve consequential changes.
Pew found that 52% of surveyed U.S. workers were worried about the future impact of AI in the workplace. That is not a finding that 52% expected to lose their own jobs. The World Economic Forum projects, across multiple global trends rather than AI alone, 170 million new roles and 92 million displaced roles by 2030, for a net increase of 78 million. It also says nearly 40% of skills required on the job are expected to change. Read Pew’s findings and the WEF projections.
Explain which tasks are changing, what decisions remain with people, and where employees can get support. Before launch, name who can approve a CRM write or publication and who resolves a timeout, policy block, invalid output, or disputed value.
Pilot, measure, and expand only when the controls work
Choose one low-risk, reversible task and record a baseline. Compare time saved or throughput with error rate, review rate, exception volume, and downstream correction work. Log one observation per AI execution, including run ID, prompt version, status, and validation result. Aggregate by a defined task and period rather than presenting a single run as a team-wide outcome.
Review examples of rejected or edited output to improve the prompt, rules, or training. An approval rate alone does not prove accuracy or business value. Expand only when the owner can explain the measured benefit, permissions, exception route, duplicate handling, and rollback plan.
- Confirm permitted inputs and test records.
- Test null, malformed, overlong, and out-of-list outputs.
- Name the reviewer, workflow owner, and exception owner.
- Define a stable event key and destination uniqueness or upsert control.
- Record a baseline and track correction work by task and period.
- Document how to stop, recover, and roll back the workflow.
Teams evaluating bounded AI roles alongside automation controls can explore AI agent services as an optional next step.
Frequently asked questions about AI challenges
Which marketing tasks should remain deterministic?
Use explicit rules for consent, suppression, fixed thresholds, legal status, billing, and irreversible actions. Use AI for bounded interpretation when unstructured input needs to be understood and the result can be checked.
Does PII detection mean data is not retained?
No. Detection is an additional control. Check whether information is transmitted and how the platform processes and retains it.
Does HubSpot re-enrollment prevent duplicates?
No. Re-enrollment controls documented workflow enrollment behavior; it is not a universal exactly-once guarantee or a race-safe write mechanism.
Can AI output be published or written directly to a CRM?
Only when the task’s risk, validation, permissions, and review design justify it. Use a human gate for sensitive, consequential, or reputationally important changes.