AI GCC India: What the 2026 Mandate Changes

What does an AI mandate mean for an AI GCC in India?

It means the centre is funded to build and operate AI systems, not to staff existing processes. Hiring, data access, evaluation and seniority all follow from that brief, and each differs from a cost-led captive.

According to the NASSCOM-Zinnov 2026 report, roughly 80% of new GCCs launched in 2026 name AI and machine learning as their core mandate. The same report counts 2,117 GCCs in India generating USD 98.4B and employing about 2.4 million professionals in FY2026.

Reuters has described the same turn, with the GCC model shifting from cost to capability as AI and talent strains bite. Cost arbitrage gets an AI GCC in India approved. A model that behaves in production is what keeps it funded.

Why does an AI mandate change who you hire first?

Because AI work is built by small, senior-heavy teams rather than large pools of juniors. The first ten hires decide whether the centre ships products or runs a succession of pilots that never reach users.

A cost-led captive scales by adding people to a defined process. An AI capability center scales by adding judgement: someone has to decide which model, which retrieval design and which failure rate is acceptable. That judgement is scarce, and the supply of GCC AI talent in India is contested by every other centre with the same mandate.

The practical consequence is sequencing. Hire for architecture and evaluation before volume, and accept that the first quarter will look thin on headcount.

Which roles come first in an AI capability center?

Start with an engineering lead who has run AI systems in production, not only built demos. Add two or three ML or LLM engineers, one data engineer, and a platform engineer who owns deployment and monitoring.

Then name an evaluation owner. This is the role most centres forget, and it is the one that stops the team from arguing about quality by anecdote.

Product ownership usually stays with the parent, close to the business problem. Splitting product from engineering across time zones works only if the product owner is available for a daily overlap.

What should the seniority mix look like?

Invert the usual pyramid at the start. A first cohort that is mostly senior engineers sets patterns, review standards and the evaluation suite that later juniors inherit.

The trade-off is cost per head and key-person risk. If two people hold all the architecture knowledge, one resignation stalls the centre.

Some parents cover the gap with fractional senior leadership while permanent hires land. Viithiisys offers CTO-as-a-Service on a fractional basis for exactly this stage, scoped after a discovery call.

How does data governance change for an AI capability center?

Data access becomes the critical path. An AI team that cannot reach governed, representative data in week one will spend its first quarter on synthetic examples that do not predict production behaviour.

Settle four questions before the first engineer starts:

• Which datasets may cross the border, and which must stay in the parent's region?
• How is personal data masked or removed before it reaches a model or a vector store?
• Are prompts, responses and retrieved documents logged, and who can read those logs?
• What do the model vendors' terms say about retention and training on your inputs?

A common failure is a fully staffed centre waiting months for a data-sharing agreement. The fix is dull: a named data owner in the parent, a classification scheme, and a data engineering pipeline that delivers governed extracts to the offshore side, not ad hoc file transfers.

What evaluation infrastructure does an AI engineering offshore team need?

An evaluation suite that exists before the first release and is owned by the parent. Without it, quality is judged by whoever built the system, and regressions arrive silently with each prompt or model change.

Evaluation built late tends to become a rationalisation of what the system already does. Evaluation built first becomes the specification.

Treat it as infrastructure, with versioning, an owner and a budget. The tooling side overlaps with MLOps: pipelines, monitoring, model registries and rollback.

What belongs in the first evaluation suite?

Begin with a golden dataset of 100 to 300 real cases, labelled by someone who understands the business outcome. Add regression checks that run automatically on every prompt, model or retrieval change.

Layer a human review sample on top, because automated judges drift and need calibration against people. Track cost per request and latency alongside accuracy, since a correct answer that costs too much or arrives too slowly still fails.

Be honest about limits: LLM-as-judge scoring is useful but biased, and no suite covers every edge case.

Where does evaluation live across two time zones?

Put the suite in a repository both sides can run, with results visible to the parent's product owner without asking. The offshore team writes the system; the parent holds the pass criteria.

That separation matters most with an AI engineering offshore team, because the people building the system are the ones least able to doubt it. A weekly review of failed cases, with both sides present, keeps the overlap hours useful.

How do the main build models compare?

There are four common shapes: an offshore development centre, build-operate-transfer, a captive GCC, and a dual-shore extended team. They differ mainly in who owns the team and how soon the parent can start.

An ODC stays vendor-owned and vendor-managed, while a captive is client-owned; BOT is the bridge between them, as Morgan Lewis describes it. Entity setup, compliance and leadership hiring make a captive a multi-quarter commitment, so many parents test the AI work with an extended team first.

Why is Mohali on the shortlist for an AI center of excellence in India?

Because tier-2 cities now carry a real share of new GCC activity, and Mohali has a graduate pipeline, technology parks and a state government actively courting centres. Bengaluru is no longer the only serious answer.

According to NASSCOM-Zinnov 2026, nearly a quarter of India's new GCC units in the past year went to emerging cities beyond the metros, Mohali among them. The Government of Punjab partnered with Zinnov as State Partner at Zinnov Confluence 2026, with a session titled "The Location Playbook: Mohali as North India's Emerging GCC Hub", reported by Business Standard and ANI.

For an AI center of excellence in India, the question is whether a smaller city can staff senior roles. The honest answer is that it can staff a focused team, but a 200-person AI division will still draw on several cities.

What do the tier-2 numbers actually say?

EY India's analysis puts cost of living in tier-2 cities 10-35% lower than the nearest tier-1 location. That is a cost-of-living comparison, not a salary figure.

EY also observes attrition up to 10 percentage points lower in tier-2 cities. For an AI team, that matters more than rent: retention protects the accumulated context of evaluation sets, data quirks and architecture decisions.

EY's list of tier-2 cities with GCC operations names Chandigarh among them. Business Today reports the Chandigarh tricity produces over 40,000 fresh graduates a year, with technology parks including Rajiv Gandhi IT Park, Quark City and Panchkula IT Park.

What does Punjab's policy offer GCCs?

Invest Punjab describes an Industrial and Business Development Policy 2026 covering IT, ITeS, data centres and GCCs, targeting roughly Rs 75,000 crore of investment.

For GCCs, the policy offers an employment subsidy of Rs 7,500 per employee per month, plus rental and capital subsidy. Check eligibility conditions and duration with the state directly, because incentives carry conditions that headlines omit.

What does Viithiisys' AI delivery work show?

It shows AI features built into products people already use, by an engineering team working from Mohali with a Canadian front office. Viithiisys has run that dual-shore model since 2007, across 500+ projects and 6 countries.

Four builds are relevant to AI work:

• Conscious Chemist (consumer and beauty): a skin-analysis tool and an AI assistant wired into the website and CRM. The client confirmed a +15% result.
• Fitelo (health and wellness): a voice-first coaching layer inside the existing app and coach workflow.
• Milo (mobility and fleet): driver communication and daily fleet operations on one platform.
• Vizitor (workplace SaaS): a seed-stage idea built into a security-first workplace platform.

Only the Conscious Chemist figure is published; the other three have no published metric. The pattern across all four is AI wired into existing workflows and data, which is where AI development succeeds or stalls.

How does a dual-shore team work in practice?

Delivery sits in Mohali, in the Chandigarh tricity, with a client-facing presence in Markham, Ontario. A team of 20+ engineers, designers and strategists works against overlap hours with US, UK and Canadian clients.

The same shape suits a small AI capability center: a compact senior team close to the work, with a front office in the parent's time zone. Earlier clients have included Paytm, Snapdeal, IKEA, Nestle, Shiprocket and Vikram Solar.

For a fixed-scope first step, Moonship ships a working MVP in 30 days against a fixed scope, which gives a parent real evidence before committing to a larger structure.

How should you start an AI mandate in India?

Start with one product-shaped use case, one evaluation suite and one data owner, then decide on the ownership model with evidence. Structure follows proof, not the reverse.

A workable order is: pick the use case with the clearest business owner, write the evaluation set, secure governed data access, and only then staff the team. Delay entity and captive decisions until the first release has run against real users.

If you are unsure where AI fits in your current operations, a broken workflow assessment identifies the processes where the work stalls and whether AI is the right fix. For a specific build or team question, contact Viithiisys and the scope is set after a discovery call.