Where to use AI in marketing, and where to leave it out

What makes an AI marketing use case useful?

A useful AI marketing use case improves a defined job after the cost of checking its output is included. It should have an owner, an input, an expected result and a way to recognise failure.

“Use AI for content” is too broad. “Group these approved customer questions into themes, retaining the original wording and source” is a task. The second instruction makes it possible to check whether the grouping helped, whether questions disappeared and whether the system invented a theme.

Different AI tasks need different systems. Predicting demand from transaction records is a statistical modelling problem. Producing an outline belongs in a generative content workflow. Sending a campaign involves permissions, customer preferences and operational controls. Putting all these under one adoption target hides the decisions that matter.

For a small business, a useful starting boundary is internal work that is reversible. A poor outline can be discarded. A wrong public offer or an unsolicited message can travel beyond the team before anyone corrects it. Reversibility does not eliminate risk, but it changes the cost of learning.

Why should teams avoid a blanket AI rollout?

Performance depends on the task and the people doing it, so an improvement in one workflow does not establish an improvement everywhere.

Dell’Acqua and colleagues’ 2023 field experiment describes a jagged technological frontier: apparently similar professional tasks can sit on different sides of a model’s capabilities. The practical implication is to evaluate actual work rather than infer ability from a polished demonstration.

Brynjolfsson, Li and Raymond’s 2023 customer-support research also reports differing productivity effects across workers. It is evidence about a particular deployment, not a forecast of the savings any marketing department will achieve.

These studies support experimentation with clear limits. They do not support a universal replacement ratio or a promise that any tool subscription pays for itself. A local-language campaign, a pricing claim and a desk-research summary have different failure modes.

The business question is therefore simple: does this use of AI improve this decision or deliverable under our actual conditions? That question is more useful than asking whether the team is using enough AI.

How do you choose and test a use case?

Choose a task by comparing value, verifiability, consequence and review effort, then run it against the existing process.

1. Describe the job before selecting the tool

Write the input and desired output in ordinary language. Identify what the team currently finds slow. If the bottleneck is missing product information, a writing tool may only make unsupported copy appear sooner.

Use a representative brief, including the awkward exceptions. A campaign with perfect source material is a useful demonstration, but an incomplete brief tests whether the workflow recognises uncertainty and asks for missing evidence.

2. Classify the consequence of an error

Separate suggestions from actions. An internal list of headline options needs editorial judgement. A system that changes a live price or sends a customer message needs additional controls. Ask what could happen before someone notices, who would be affected and whether the action can be reversed.

Do not use the classification as an excuse to stop all experimentation. Use it to place review where it matters. A team may allow automatic formatting while requiring explicit approval for any new product claim.

3. Build a small evaluation set

Collect typical inputs, difficult inputs and examples that should be refused or escalated. Decide in advance what counts as correct. For research organisation, check coverage and source preservation. For copy variations, check factual accuracy, distinctness, tone and whether the offer remains unchanged.

Evaluate the final output after editing, but also record the editing needed. Otherwise a skilled reviewer can conceal a weak system by repeatedly repairing it, while the business mistakes the result for automation.

4. Compare the whole workflow

Include preparation, tool use, checking, correction, coordination and handover. Record errors that reach the next stage. A faster first draft may still be useful, but only if it leaves enough time for the harder editorial work.

Avoid measuring adoption through the number of prompts or generated assets. Those are activities. A better measure is accepted work completed at the required quality, with a clear record of the resources used.

5. Set a stopping rule

Decide when the pilot should pause. Repeated unsupported claims, unreliable handling of private data or unmanageable review queues are reasons to redesign it. Stopping a poor use case protects the good ones from losing credibility.

Where does AI fit in an Indian marketing team?

AI fits where the team can check the language, facts and context of the output. It fits less well where those checks are absent.

For an Indian retailer, potential tasks include organising product information, drafting alternative descriptions from approved specifications and classifying support themes. These are proposed uses, not claims of proven revenue improvement. A staff member must still verify sizes, prices, availability and regional terms.

Translation deserves particular care. A fluent sentence can contain the wrong relationship, register or implied promise. Ask a competent language reviewer to assess the actual customer meaning. Back-translation is a useful diagnostic, but agreement between two machine outputs is not proof that local readers will understand the message correctly.

Healthcare, lending and other consequential categories need subject specialists. Marketing staff should not use a model’s confidence as evidence for a treatment benefit, eligibility rule or legal position. Where qualified review is unavailable, reduce the scope to administration and approved information.

When should a pilot become a regular workflow?

A pilot should become routine only after the team can reproduce acceptable results and handle exceptions without improvisation.

In week one, record the existing process and build the evaluation examples. In week two, run AI-assisted work alongside that process. Review both ordinary outputs and failures. This is a suggested sequence, not a universal timetable: a seasonal business may need a longer observation period.

Before expanding access, document which tools are approved, what data may enter them and who can authorise consequential actions. NIST’s 2023 AI Risk Management Framework organises risk work around governing, mapping, measuring and managing. A small team can apply that logic without turning it into a large bureaucracy.

Recheck the workflow after changes to models, source material, staff or customer segments. A successful pilot is evidence for a defined process under defined conditions; it is not a permanent exemption from quality control.

Who should own the decision?

The marketing owner should choose the business use case, with editorial, data and technical owners responsible for the controls that affect their work.

One named person should be able to explain why the system is in use and stop it when necessary. Vendors can provide documentation and support, but accountability for the brand’s message remains with the organisation publishing it.

Reviewers also need capacity. Assigning approval to someone who cannot inspect the evidence produces a ceremonial sign-off. Reduce the volume or simplify the task until the review is credible.

My view

I would start with tasks that help people think more clearly: organising evidence, checking a brief for missing information and exploring alternative explanations. These uses can improve the work even when nothing is published automatically.

The trade-off is that bounded adoption looks less dramatic than a fully automated campaign. It may also reveal that some tasks are better handled by templates, a spreadsheet or a clearer brief. That is a useful result. The objective is better marketing, not a larger collection of AI workflows.

An AI use case earns its place when the checked result is better than the process it replaces.

Start with one work sample

Choose a recurring task, preserve the current output and write down what an acceptable result means. Run the proposed workflow on that same input. Compare the finished work, the checking effort and the remaining uncertainty before expanding its role.