Most Indian marketing teams asking about agents do not have an engineer to spare. That is workable. The constraint is not building the thing, it is choosing a task narrow enough that no engineering is required.
Pick the task by failure cost, not by ambition
The instinct is to automate the biggest time sink, usually content production or campaign reporting. Both are bad first projects. Content has a public failure mode. Reporting touches numbers people make decisions on.
Better first tasks share three traits: output is internal, correctness is checkable in seconds, and being wrong costs a few minutes. Categorising inbound leads by industry. Turning sales call notes into structured CRM fields. Drafting internal briefs from a campaign spec.
The four-week shape
Week one, collect fifty real examples of the task done correctly by a human. This is the only genuinely tedious part and skipping it is why most pilots fail. These become your test set.
Week two, write the instruction and run it against the fifty. Score how many it got right. Anything below eighty percent means the task is underspecified rather than the model being weak. Tighten the instruction and rerun.
Week three, connect it to one real data source. Read-only. No writing back yet. Watch it work on live data with a human checking every output.
Week four, allow it to write, keeping a human approval step. Only remove that step once you have a month of clean output.
The three decisions that matter
Where does the output go? An agent writing directly to production systems on day one is how you get an incident. Route to a review queue first.
What is the budget ceiling? Set a hard monthly cap. Agents that loop are the most common source of surprise bills.
Who owns it when it breaks? Name a person. Agents that nobody owns get quietly abandoned after the first bad week, and the team concludes agents do not work when what happened is nobody was watching.
What this costs
For a narrow internal task on a mid-tier model, running costs are usually small enough to disappear into an existing software budget. The real cost is the fifty examples and the person watching output for a month. Budget attention, not licence fees.