The real question is not whether an agent can move quickly. It is whether it can act safely, explain what it did, and stop when the stakes rise. AI agents are moving from suggesting campaign copy to taking actions: building audience segments, preparing creative variants, scheduling posts and shifting media budgets. That promise is compelling for marketers managing many products, regions and languages. But access to a workflow is also access to its mistakes. A wrong audience rule can exclude customers; a confident but unsupported claim can become a public ad; a small budget adjustment can compound across channels.
Indian campaigns show both the opportunity and the care required. Cadbury Celebrations’ Shah Rukh Khan-My-Ad initiative used machine learning and local business data to create thousands of personalised ads, while Instamart’s “Ajay Infinite” film used AI as a creative device. Those are examples of designed, bounded use—not proof that an autonomous agent should be handed every campaign control. Before granting access, marketers need to check the permissions, inputs, limits and human checkpoints around the work.
Five Checks Before You Connect AI:-
1. Define the job—and give only the access it needs: Write down the exact task: draft variants, flag underperforming ads, or prepare a schedule. Then map every connected tool and permission. An agent that drafts copy may need a brief and a read-only view of prior ads; it does not automatically need permission to publish, change billing, export a customer list or edit account settings. Separate “view,” “prepare” and “execute” permissions, use a dedicated account where possible, and start with a sandbox or low-risk campaign. Cadbury’s localised ads are a useful reminder that the meaningful unit of work may be a particular store and geography—not unrestricted access to the whole brand account.
2. Know what data enters the agent, and where it goes: List the data sources the agent can read: customer attributes, conversion events, CRM notes, creative files, or campaign reports. Ask the vendor what is retained, whether inputs may be used to improve a model, who can access logs, and how deletion works. Remove fields that are not needed; use aggregated or pseudonymised data for analysis where practical. Check consent, purpose and retention with your privacy and legal teams before connecting personal data. India’s Digital Personal Data Protection Act frames processing around lawful purposes and individual protection; consult the Act and current rules for your specific use, rather than assuming an AI tool changes existing obligations.
3. Set hard limits on money, audience and claims: Define what the agent may change and set ceilings: daily spend, bid movement, campaign count, geography, frequency and audience eligibility. Put prohibited claims and required disclosures in the brief, then test edge cases—sale dates, pricing, product availability, regional language and sensitive categories. Instamart’s “Ajay Infinite” film used AI to multiply a recognisable performance for a specific sale message; it is a reminder that the tool can amplify an idea while the brand still owns the offer’s accuracy and the portrayal’s context.
4. Keep a person in the approval path for consequential actions: Decide which steps can run automatically and which require a named reviewer. Publishing new creative, changing a material claim, targeting a sensitive group, or moving significant budget should trigger human approval. Show the reviewer the source brief, proposed change, audience, spend impact and reason—not just an “approve” button. Reviewers should be able to edit or reject, and the agent should not treat silence as consent. A good checkpoint is specific: approve this asset for these markets and dates, not “let AI handle the campaign.”
5. Make actions traceable, reversible and easy to stop: Before launch, check that every action records who or what initiated it, which inputs and instructions were used, what changed, and who approved it. Agree on alert thresholds, a kill switch, rollback steps and a human owner for incidents. Run a limited pilot, compare the agent’s output with the existing process, and inspect failures as well as average performance. Expand only when the team can explain results and recover quickly. Scale should follow evidence—not the number of tasks an agent can technically complete.
Key Takeaways:
1. Grant agents the narrowest access needed for one defined task.
2. Keep personal data, budgets and public claims inside clear limits.
3. Require human approval, complete logs and a tested stop button.
Giving an AI agent access is not a single yes-or-no decision. It is a set of choices about what the system can see, what it may change, and when a person must step in. Indian brands have demonstrated how technology can help campaigns become more personalised or imaginative; the valuable lesson is not “automate everything,” but design the work so that creativity and accountability grow together.
Start with one contained workflow, a limited permission set and a clear success measure. Keep approvals for decisions that affect customers, spend or brand trust. Record actions, rehearse a rollback and review the pilot with the people who run the campaign. If the agent can explain its changes and the team can safely reverse them, widen access in measured steps. If it cannot, the next improvement is not more autonomy—it is a better-defined workflow.




