Forward-deployed AI consulting for global capability centres
Global capability centres build technology for a parent organisation that owns the business outcome. A centre often has strong engineers, good tooling and a mandate to deliver AI. What it often lacks is a clear answer to a harder question: who owns an agent once it is built, and who answers for its decisions when the parent asks.
I am Mayur Patil, based in Nashik, India. I work with capability centres on forward-deployed AI consulting, and I am the consultant who sets up the operating model and the first working system, with the centre's engineers alongside.
The ownership gap
A pilot in a capability centre can look successful. It produces a demonstration, the parent is pleased, and the work moves to a second use case. The gap appears later, when the first system has to run every day, change when the source systems change, and answer for a decision it made. If the centre cannot name the owner, the system drifts or is switched off.
Forward-deployed delivery treats ownership as a deliverable. Each system ends with a named owner inside the centre, a runbook that covers the known failures, an evaluation set the owner can re-run, and recorded decisions the owner can explain. The handover pack term describes what the owner receives.
Serving business units in the US and Europe
Many centres serve business units whose working day does not overlap with theirs. The practical answer is a deliberate overlap for decisions, a written handover to the parent's owners, and a runbook that the parent's team can use without a call to the centre. For a system that answers customers, the approval steps and the stop control must work outside the centre's working hours, and the operating model should say who holds that responsibility.
The rules that apply to a system depend on where its data and its users are. Each country page summarises those rules and links to the sources: United States and Europe cover the two markets that most often sit behind a centre's business units, and India covers the data protection rules that apply to the centre's own processing.
Relevant work
The organisations I have worked with and can name are listed on the about page. Among them are Tredence, Persistent and Walmart, and I have worked with Wipro and Infosys on capability programmes. I do not describe the detail of that work beyond what the clients have agreed to share.
What an engagement looks like
The usual first step is a readiness review on one workflow, which is two weeks and ends in a written recommendation. If the centre decides to proceed, the next step is either an architecture sprint, which decides the operating model for the first system, or a forward-deployed delivery engagement that builds it with the centre's engineers.
The engagements page describes each shape. I do not publish prices, and every engagement is quoted on its scope.
Training the centre's own team
Some centres want their own engineers to build the next systems. For them, the FDE training page covers team programmes built on the centre's own tools and use cases.
What the 2026 numbers say
The Nasscom-Zinnov India GCC Landscape report released in July 2026 counts 2,117 GCCs in India, reported as employing about 2.36 million people, with revenue just under $100 billion in FY26. Coverage of the same report says that nearly three quarters of India's GCCs have the potential to evolve into portfolio or AI-led hubs over the next five years. Those numbers describe the choice in front of a centre: to remain a delivery function, or to become the place where the parent's AI decisions are made and owned.
The ownership question is therefore the first one a centre should answer. A centre that builds an agent for a parent's business unit has to say who runs it, who answers for its decisions, and what happens when the source systems change.
The three shifts a centre makes
From pilots to owned workflows. Many centres have pilots that work in a demonstration. The shift is to choose one workflow, agree its owner, and ship it with a runbook. A single owned workflow teaches more than ten demonstrations.
From central delivery to business ownership. A centre serves several business units. The shift is to place the system in the unit that uses it, with the centre keeping the standards, the templates and the evaluation approach.
From manual checks to measured systems. An agent changes, and so do the systems it reads. The shift is to keep an evaluation set, re-run it after every change, and report the results to the owner. The regression eval term describes that step.
What a centre needs before its first agent
Five things should exist before the first agent runs. A named business owner. A data map that says which personal data the agent reads, and why. A record of the controls: which actions need approval, and who gives it. An evaluation set agreed with the owner. And a handover plan that names the person who will run the system after the consultant has left.
Three rules shape the data map. The Digital Personal Data Protection Rules were notified on 13 November 2025, and most of their obligations apply from 13 May 2027. The India AI Governance Guidelines, released on 5 November 2025, are non-legislative, and they set the principles a parent's risk team will ask about. For a centre that serves a regulated parent, the parent's own regulator sets the stricter standard, and the centre designs to that.
Bengaluru and Hyderabad
Bengaluru and Hyderabad are the two cities where most of these decisions are made in India. Bengaluru is home to more than 880 GCCs, and according to Zinnov it is home to one in three of the new GCCs added in FY24. The Bengaluru consulting page covers the work there.
Hyderabad hosts more than 355 GCCs, according to the same Zinnov analysis. The Telangana government's AI mission, T-AIM, is an AI-specific organisation set up with NASSCOM to drive the state's AI framework, which has six pillars including governance, ethics and privacy. The Hyderabad consulting page covers the work there.
How I measure a centre's readiness
I measure a centre on four things. The share of its AI work that has a named owner. The share of its agents with an evaluation set. The share of its changes that re-run the evaluation before release. And the cost per completed task of its agents, measured without human rework. The cost per completed task term explains the last measure.
Questions a centre should ask any partner
A centre should ask a partner four questions before it signs anything. Who will run the system after the partner leaves, and how will that person be trained? What evidence will the partner leave: an evaluation set, a runbook, recorded decisions? How will the partner work with the centre's own engineers, so that they learn rather than wait? And how will the work be reviewed by the parent's risk team, in the parent's time zone?
Working with a parent in another time zone
A centre that serves a US or European business unit has to handle an overlap of working hours. The practical arrangement is a short overlap for decisions, a written handover to the parent's owner, and a runbook that the owner can use without a call. For a system that answers customers, the approval steps and the stop control must work outside the centre's working hours, and the operating model says who holds that responsibility.
How to start
Send an enquiry with the centre's size, the business units it serves, the workflow that matters first and who owns it today. I reply with a suggested starting point.
Common questions
What does a global capability centre need from forward-deployed engineering?
Most capability centres build software for a parent organisation that owns the business outcome. The gap is ownership: who runs an agent after it is built, and who answers for it when the parent asks. Forward-deployed delivery closes that gap with a named owner and a runbook.
Which operating model suits a capability centre?
There are three common patterns: a central pool assigned to projects, embedded teams inside one business unit, and a hub and spoke design where a small central group sets standards and places teams in business units. The right choice depends on how many business units the centre serves and how much standardisation the parent requires.
How does the work fit with business units in the US and Europe?
Through overlap hours and a clear handover to the parent's owners. The engagement sets the working hours for each side, and the decisions and the runbook are written so that the parent's team can run the system across time zones.
Do you work with a centre's existing engineers?
Yes. Delivery is built alongside the centre's engineers, who take the system over at handover. The aim is that the centre can run what it builds without me.
What is the first step for a capability centre?
A readiness review on one workflow, which ends in a written recommendation to proceed, reframe or pause. Starting with one workflow keeps the decision small enough to make, and it shows whether the operating model needs to change before scaling.