Budget season forces a useful question: which parts of the AI conversation are going to change how the organisation works next year, and which are noise? From where we sit — building agents for our own operations and deploying them for universities, governments and enterprises — five shifts look durable enough to plan around.
1. From copilots to agents that own a job
The first wave of enterprise AI put an assistant beside the employee: draft this email, summarise this meeting. Useful, but hard to measure, because the person still owns the work and the time saved leaks away.
The second wave gives an agent a defined job — first-line support, invoice follow-up, weekly MIS reporting, lead qualification — with a stated metric and a stated point where it hands work back to a person. The planning implication is organisational rather than technical: someone has to own each agent the way a manager owns a role, and headcount plans will start to include roles that are partly or wholly held by software.
2. Open standards for connecting agents to systems
Until recently, every connection between a model and a business system was custom integration work. Open protocols for exposing tools and data to agents — the Model Context Protocol is the most widely adopted — are turning those connections into reusable components.
For a CIO this changes the build-versus-buy question. The valuable investment is a well-governed layer that exposes your core systems — ERP, CRM, HRMS, document stores — to agents with the right permissions and full logging. Build it once and every future agent, from any vendor, can use it safely.
3. Governance moves from policy document to infrastructure
Regulation is catching up with automated decision-making. In India, the Digital Personal Data Protection Act makes retention, consent and purpose limitation enforceable obligations, and similar rules apply in most markets our customers operate in.
The organisations that handle this well will stop treating AI governance as a policy PDF and start treating it as infrastructure: every agent action logged with who approved it, retention enforced automatically rather than by reminder, and access scoped per agent the way it is scoped per employee. Expect audit teams to ask for agent activity logs in 2027 the way they ask for access logs today.
In practice this looks less like new technology than like discipline applied to AI: in-country hosting with in-country disaster recovery, round-the-clock security monitoring, and periodic penetration testing by a CERT-In empanelled auditor. Agents that touch that data should inherit exactly the same regime.
4. Smaller, private models for the everyday work
The most capable frontier models will keep improving, and they will keep being the right choice for complex reasoning. But a large share of enterprise AI work — classifying, extracting, routing, summarising — runs well on much smaller models, including open-weight models an organisation can host itself.
That matters for two reasons. Cost: running the routine steps on small models makes high-volume use affordable. And control: for data that cannot leave your infrastructure, a capable private model removes the hardest blocker to adoption. Plan for a portfolio of models, not a single provider contract.
We already see this in assessment: AI marking of handwritten answers runs on a domain-tuned model rather than a general one, and institutions that need it can have the question-generation platform deployed privately with an isolated schema per institute.
5. Paying for outcomes, not effort
When an agent does the work, pricing by the hour or by the seat stops making sense. We expect more AI engagements — including ours — to be structured around the metric the agent moves: tickets resolved, documents processed, collections recovered, hours of reporting removed.
Assessment already works this way: onscreen marking is priced per page evaluated, and generative AI features are metered in credits consumed rather than seats. The logic will spread to agents that do general business work.
For buyers this is good news with one condition. You need a baseline. The organisations that measure the process before the agent arrives will be able to negotiate on outcomes; those that do not will be buying on faith.
What to do this quarter
None of these shifts requires a large programme to prepare for. Three steps put an organisation in a strong position for next year:
- Pick three processes and baseline them — volume, time, cost and error rate, measured now
- Map which core systems an agent would need to reach, and who approves that access
- Put one agent into production against one metric, with draft-and-approve guardrails, and learn what operating it actually involves
The advantage in 2027 will not go to the organisations with the most AI pilots. It will go to those that learned, early, how to run agents as part of the business.