India’s AI Spending Soars. Governance Misses the Meeting

01The money arrives first

India has bought AI a very expensive seat at the table. The person responsible for its conduct appears to be running late. ServiceNow’s September 8, 2026 research reports annual enterprise AI investment growth of 119% in India. Only 22% of enterprises, meanwhile, report having testing, auditing and risk-assessment processes in place. The cheque clears faster than the rules are written.

Different measures, one spending decision. Source: ServiceNow · September 2026.

That is a striking mismatch, though the smaller number needs care. It measures the reported presence of particular processes. It does not certify that 22% have mastered governance, or establish that everyone else has no controls whatsoever. A committee can exist without exercising useful judgment; a capable team can exercise judgment without satisfying every survey category.

The research was conducted with ThoughtLab and surveyed 4,500 senior leaders globally, including 350 in India. Its maturity rankings use a proprietary index. These are vendor-sponsored survey findings, rather than an independently audited census of Indian business. They offer a view of reported practice. They cannot tell a buyer which product will repair it.

Read the research announcement and methodology.

02An agent is hired. Authority is withheld.

The next pair of figures gives the story its tension: 54% of Indian organizations report deploying AI agents; only 11% report autonomous workflows. Installing an agent and giving it permission to finish a consequential task are different decisions.

Consider an illustrative purchasing request. An agent can read a message, locate a supplier and draft an order. Completing the purchase requires more: the current contract, the right budget, an approval threshold and permission to write into the purchasing system. If any of those is missing, the employee gets a polished suggestion and another job to do.

The two percentages are separate survey measures, not a conversion funnel tracking the same agents through successive stages. Their difference cannot reveal how many deployments failed. It does show why counting agents is an inadequate description of how an enterprise works.

Keeping a person in the loop may also be the correct design. A routine request can travel quickly; an exception deserves a pause. Autonomy becomes useful when authority is specific enough to delegate. An agent that knows when to stop can be more valuable than one that is always eager to proceed.

03The missing context has a price

Only 18% of Indian organizations report replacing fragmented legacy systems with an integrated platform. And 74% flag data accuracy, access and management as needing improvement. Those findings point beneath the model, toward the records and connections on which a task depends.

In our purchasing example, the supplier could have one name in procurement and another in finance. The contract might be current in one repository and obsolete in another. An agent cannot settle those disagreements merely by composing a more elegant sentence. Someone has to establish which record governs the decision.

That is a plausible explanation for why a larger AI budget may produce more activity without more completed work. The survey does not prove that fragmentation caused a particular failed deployment. But the operating problem is concrete: a workflow crossing departments needs dependable context at every handoff.

Integration is also broader than buying a single platform. Enterprises must decide who owns each record, how changes travel and what happens when systems disagree. A shared interface can make a mess easier to see. It takes accountable owners to put the mess in order.

04Five controls, five different jobs

ServiceNow has a commercial answer to this problem. Its AI Control Tower product page describes discovery, security, governance, observation and measurement across enterprise AI. These functions address different needs; the survey supplies no comparative proof that ServiceNow performs them better than alternatives.

Discovery establishes an inventory of agents and models. Governance connects lifecycle and risk controls to their use. Observation follows performance through metrics and log traces. Security covers identities, access and intervention, including stopping an agent. Measurement tracks adoption, realized value and return on investment. These are ServiceNow’s stated capabilities.

The buyer’s questions follow from the gaps. Can discovery identify who owns an agent that touches a disputed record? Can governance preserve evidence of tests and approvals? Can observation reveal where a workflow stalls? Can security limit the action before it becomes costly? Can measurement distinguish a completed order from a generated draft?

Each question needs its own demonstration. A neat inventory does not establish sound testing. A log does not establish appropriate access. A fast response does not establish a useful result. Treating all five as one reassuring word—governance—allows the weakest control to hide behind the strongest.

05Ask it to finish an untidy request

The supplied platform link now presents ServiceNow Otto, its AI assistant for work. ServiceNow says Otto coordinates requests across systems and departments, works with approval chains, logs actions and involves people when judgment is required. Its description connects governed execution with the promise of completed work.

For a buyer, that promise becomes testable in a deliberately awkward demonstration. Use a request with a conflicting supplier record, an unavailable approver and a spending limit. Watch where the system retrieves context, asks for a decision, refuses an action or hands the case to a person. Then inspect what evidence survives.

This is an evaluation approach, not a result reported by the survey. It puts the advertised capabilities in contact with the buyer’s actual operating conditions. A demonstration staged on immaculate data answers a much easier question.

Bring the people who own the process into that room. Procurement can recognize an invalid order; finance can identify a budget problem; security can judge excessive access. The person presenting the assistant should not be the only person deciding whether it succeeded.

06The next cheque needs a receipt

A useful next step is to choose one bounded workflow and define what completion means before spending more. For the purchasing request, that could be a valid order entered in the correct system, with the necessary approval and an accessible record of the decision.

Track how often that happens, how often people must repair the result and what it costs to operate. Review exceptions alongside successes. An enterprise that measures only requests answered can celebrate while its employees quietly finish the work elsewhere.

India’s investment jump demonstrates appetite. The governance findings raise a harder question about the work that appetite can support. The answer will emerge in ordinary transactions: the request that reaches the right owner, the permission that expires, the exception that stops before it causes damage.

The next budget meeting should begin with one of those transactions on the screen. Follow it from request to receipt. When the agent finishes, ask the process owner to show that the job is done.