Tech Mahindra Puts SpaceX’s Grok Build Into Enterprise Software Engineering
India’s fifth-largest IT services firm will use the coding agent on legacy modernisation. The deal is notable; the evidence for the productivity it implies is smaller than the marketing.
Tech Mahindra, India’s fifth-largest IT services firm, will put SpaceX’s coding agent into its engineering and application modernisation work. The agent, Grok Build, comes from SpaceXSI, the AI division of SpaceX, and the collaboration covers legacy modernisation, software delivery and developer productivity across technology, telecom, banking, manufacturing, healthcare and retail. The deal makes Tech Mahindra the first Indian IT major to work with SpaceXSI, following its existing pacts with Google’s Gemini and Microsoft, and it landed on the same day SpaceXSI opened its partner network to more than 130 companies.
The announcement is easy to read as another AI partnership, and it is one. It is also a useful test of a question the industry keeps answering with marketing: does AI coding actually speed up enterprise software work? The evidence, which now runs to randomised trials, telemetry studies and hundred-thousand-developer analyses, says the gains are real but far smaller than the vendors claim, and that the places they appear most clearly are narrow.
What the two companies announced
The collaboration combines Tech Mahindra’s engineering, enterprise transformation and AI consulting with Grok Build, and it targets application modernisation, code quality and AI-assisted engineering across the software lifecycle. Tech Mahindra will build enablement and certification programmes, joint customer engagements and industry-specific solutions, and it frames the offering around governance, security and compliance for mission-critical systems. The release names no customer, no pilot and no commercial terms, and it quantifies nothing. Sumit Kumar Popli, president of Tech Mahindra’s technology, media and entertainment business, describes enterprises moving beyond isolated AI experiments, while SpaceXSI’s head of global partnerships, Toni Adams, argues that the missing ingredient is expertise in rolling AI out across large engineering teams without breaking what works.
What Grok Build is
Grok Build is a terminal-based coding agent, launched in early beta in May and open-sourced in July. It runs on SpaceXSI’s Grok 4.7 model and works like the other agentic coding tools: it plans before it edits, delegates work to parallel subagents, reads and searches large codebases, runs commands and tests, and integrates with repositories through git. For enterprises, the important part is the governance surface. Grok Build reads policy from a managed configuration file that an administrator can pin so users cannot override it, supports single sign-on through corporate identity providers, runs tools in a sandbox, and offers zero data retention at the team level. It also reads a subset of Claude Code’s managed settings for organisations running both. That list matters, because the objections enterprises raise to coding agents are rarely about code generation and usually about permissions, auditability and where the code goes.
The productivity evidence is more modest than the pitch
The claim that AI coding transforms delivery has outrun the measurement, and the measurement is now substantial. DX, which studied pull-request throughput across more than 400 companies from November 2024 to February 2026, found that AI tool usage rose 65 per cent while median throughput rose 7.76 per cent, with an average of 13.1 per cent. Its conclusion is blunt: a five to fifteen per cent gain is what the current generation of tools delivers, against vendor marketing that promises three or ten times. MIT Sloan researchers, working with more than 100,000 GitHub developers, found a larger effect earlier in the pipeline, with coding activity rising up to 180 per cent as teams adopted autocomplete, then synchronous agents, then asynchronous ones. But the gain attenuated as work moved downstream: only 50 per cent more projects and 30 per cent more releases, and no corresponding rise in app downloads or reviews.
The bottleneck moved rather than disappeared. METR’s randomised trial is the most uncomfortable data point. In early 2025 it found that experienced open-source developers took 19 per cent longer with AI tools, and that they believed the tools had made them 20 per cent faster. Its 2026 follow-up could not produce a reliable estimate, because too many developers refused to work without AI, but its raw figures point to a speedup of around 18 per cent for returning participants and 4 per cent for new ones, with confidence intervals that straddle zero.
Where AI coding does work
The evidence points to a specific pattern: the gains concentrate where the work is well defined and voluminous, and where a human bottleneck does not sit immediately downstream. Legacy modernisation fits that description, which is precisely what Tech Mahindra is selling. Morgan Stanley’s internal tooling illustrates the scale: agents read legacy code, gather context and generate the specifications developers previously spent months reverse-engineering, which the bank says saves more than 280,000 developer hours a year. That is the kind of task where a coding agent earns its place, because the alternative is expensive, repetitive and low-judgment. The harder question is what happens after the specification lands, when someone has to review the generated code, test it against decades of undocumented behaviour and release it into a regulated system.
The partner-network context
Tech Mahindra’s announcement arrived the same day SpaceXSI launched its partner network, which now includes Snowflake, EPAM, Sonar, Endor Labs, Atlan, AtScale and others, with more than 130 companies in an early-access phase. The structure of that network tells you where the money is expected to come from: transformation partners who deploy the platform, and marketplace partners who plug it into what enterprises already run. For the large services firms, the calculus is defensive as much as offensive. If coding agents shrink the volume of billable development work, the firms that resell and integrate them keep a seat at the table; the ones that do not may find their traditional pricing under pressure. That is why every major services provider now has a relationship with a model vendor, and why the model vendors compete to sign them.
The shift matters most for the services firms themselves. For two decades, the large Indian providers built revenue on headcount: more engineers, more billed hours. If a coding agent does a meaningful share of the work, that model faces a squeeze, and the industry’s answer so far is to move toward outcome-based pricing and platform-led delivery. Tech Mahindra’s framing, which pairs a tool with enablement and certification programmes and joint customer success, reads as an attempt to sell transformation rather than hours. The risk is that the same tool lets a competitor bid the work lower, and that the value migrates to whoever owns the model and the data, which is not the services firm.
The governance problem
None of this removes the risk that matters most to a bank or a hospital. Research on AI coding assistants consistently finds that generated code carries vulnerabilities at higher rates than developers expect, particularly in security-critical domains, and that developers over-trust the output. Grok Build’s enterprise controls address the plumbing, such as sandboxing, policy pinning and retention, but they do not answer the question a risk committee will ask: who reviews the code before it ships, and how do we know the agent did not quietly change something it was not asked to touch? The answer has to sit in the delivery process rather than in the tool, which is exactly the expertise Tech Mahindra claims to bring.
What buyers should ask
For a chief information officer weighing a similar arrangement, the questions are practical. Which workloads will the agent touch first, and are they the ones where the evidence supports a gain? What is our baseline today, in pull-request throughput and cycle time, so that any improvement becomes measurable rather than asserted? Who reviews AI-generated code, and does the review process scale when the volume rises? How do we pin the agent’s policy, log its actions and prove to an auditor what it changed? What happens to our data, and can we verify zero retention? And if the productivity gain lands at five to fifteen per cent rather than three times, does the business case still hold?
What it signals
This is what the AI transition looks like inside IT services: a major provider pairing its delivery capability with a frontier model vendor, packaging governance for regulated buyers, and promising measurable outcomes it has not yet measured. The honest reading is that the partnership is sensible and the marketing is ahead of the evidence. Coding agents do help, especially on legacy estates where the alternative is manual reverse engineering, and the enterprise controls Grok Build now exposes are the ones buyers were asking for. But the industry’s own data says the gain is a modest percentage rather than a multiple, that it moves bottlenecks rather than removing them, and that someone still has to own the code that ships. Tech Mahindra has bought itself a credible tool and a place in the new distribution chain.
Whether its clients get the productivity it implies will depend on what it does with the process around the tool. One thing is already clear from the timing. The alliance was not a standalone announcement but a node in a distribution network that SpaceXSI switched on the same day, and Tech Mahindra joined it alongside EPAM, Snowflake, Sonar and others. In this phase of enterprise AI, being early to a partner programme matters more than the underlying model, because buyers are not shopping for models; they are shopping for someone to make the models safe and useful inside their estate. The services firms compete for that role, and the model vendors compete to have them.