•THE MISSION Frontier models and autonomous agent frameworks have moved out of the lab and into the P&L. The binding constraint is no longer model capability — it is deployment. Most enterprise AI programmes stall in the gap between a pilot that impresses a steering committee and a system the business can actually run. The Frontier Deployment Consultant exists to close that gap. You will embed inside client environments — their tenancy, their data, their security review, their release process — and take a workflow from hypothesis to a production agentic system that keeps running after you leave the room. This is a builder-advisor role in the literal sense. You will hold the boardroom with structured business thinking and write the production code that makes the system real. Candidates who have only ever done one of the two will not be a fit.
•HOW THE ROLE OPERATES • Embedded, not adjacent. You work inside the client's environment on their infrastructure and release cadence. Expect 50–70% onsite presence, including travel across India, GCC hub cities and client HQ geographies where the engagement requires it. • Pod delivery. FDCs deploy in small pods — typically one Senior FDC plus two to three FDCs and engineers — with access to the practice asset library, PwC's Responsible AI framework and the GCC operating-model IP base. • Outcome-owned. Engagements are measured on time-to-value, reliability, adoption and realised ROI. Delivery milestones are an input, not the score. • Real enterprise conditions. You will work under client security clearance, on client-issued devices where mandated, and inside regulated, data-resident and occasionally air-gapped environments.
•WHAT YOU'LL DO 1. Diagnose and shape value Partner with CEOs, COOs, CDOs, CTOs and business unit heads to find the workflows where agentic execution changes unit economics rather than shaving minutes. Size the prize in terms a CFO will act on: cost-to-serve, cycle time, throughput per FTE, revenue leakage, total cost of ownership. 2. Build 0→1 at startup velocity Form the hypothesis, diagnose the workflow with structured consulting toolkits, and put a working, interactive prototype in front of the client in days. The proof is running software, not slideware. 3. Architect production agentic systems Design and implement multi-agent architectures across brownfield estates: planner-executor patterns, agent handoffs, durable state and checkpointing, retries and cancellation, human-in-the-loop escalation. Model Context Protocol (MCP) and agent-to-agent (A2A) integration. Hybrid and graph-based retrieval over knowledge-graph and semantic-layer foundations. 4. Own the evaluation and run layer Build the evaluation harness before the system ships — golden datasets, regression suites, task-completion and toolcall accuracy, groundedness and safety scoring, LLM-as-judge where it is defensible. Instrument tracing end to end. Define and hold service-level objectives for accuracy, latency, cost per task and containment rate. 5. Engineer for the enterprise, not the demo CI/CD for non-deterministic systems, infrastructure as code, containerised deployment into client infrastructure, canary and rollback strategy, runbooks, incident response, and a documented handover that leaves the client's own team selfsufficient. 6. Make it safe and defensible Agent identity and credentialing, least-privilege tool scoping, secrets management, prompt-injection and OWASP LLM Top 10 defences, PII and DLP controls, data residency, full audit trail. Apply PwC's Responsible AI framework and map obligations to the DPDP Act 2023, sectoral regulation (RBI, IRDAI, HIPAA) and EU AI Act risk tiers where in scope. 7. Manage the room Translate non-deterministic behaviour, evaluation results, latency and cost trade-offs and governance constraints into commercially sound executive narratives. Own the working relationship with client CTO, CISO and CFO stakeholders. 8. Operate AI-native Set the benchmark for AI-accelerated delivery on the account — Cursor, Claude Code, GitHub Copilot, agentic test generation, evaluations running in CI — and pull client engineering teams up to it. 9. Codify the IP Abstract deployment patterns into reusable assets: reference architectures, deployment blueprints, evaluation harness templates, governance artefacts, runbooks and role maps that travel across accounts and industries. WHAT YOU WILL BE HELD TO Accountability How it is measured Time-to-value First production workflow live within the mobilisation window agreed at engagement start Reliability Agreed SLOs on accuracy, groundedness, latency and uptime held after go-live, not just at demo Adoption Active usage against the target user base; workflow containment and fall-backto-human rate Unit economics Cost per task inside the agreed envelope; demonstrable TCO improvement against the baseline Self-sufficiency Client team operating and releasing the system without PwC in every change Asset contribution At least one reusable, documented asset contributed to the practice per engagement
•WHAT YOU'LL BRING — TECHNICAL Depth is assessed on demonstrated production work. Named tools are illustrative; equivalent experience counts. Core engineering — required • Python to production standard: typing, testing, packaging, profiling. Plus one of TypeScript/Node, Go or Java. • API and service design, microservices, async and event-driven patterns, Docker and Kubernetes. • Working depth in at least one hyperscaler — Azure, AWS or GCP — including networking, identity and cost controls. • Ability to architect hybrid pipelines that route across LLMs, SLMs and traditional ML components based on task fit, cost and reliability requirements Agentic build layer — required • Demonstrated depth in at least two of: LangGraph, Claude Agent SDK, OpenAI Agents SDK, AutoGen, CrewAI, Semantic Kernel, Pydantic AI. • Model Context Protocol (MCP) server and client implementation; agent-to-agent (A2A) interoperability patterns. • Multi-provider abstraction, model routing and fallback (LiteLLM or equivalent). • Long-running and durable execution (Temporal or equivalent), agent memory and state (Mem0, Zep or custom), checkpointing and replay. Retrieval and knowledge — required • Ingestion and chunking pipeline design, embedding lifecycle and re-indexing strategy, reranking, hybrid and graph retrieval, RAG evaluation. • Vector stores: Qdrant, Pinecone, Milvus, pgvector or Azure AI Search. • Knowledge graph and ontology work — Neo4j, GraphRAG, semantic layer design — is a strong differentiator. • Document AI and structured information extraction - extracting structured data from unstructured and semistructured sources (contracts, invoices, forms, scanned records, PDFs) using OCR, layout-aware models and document AI platforms (e.g., Azure Document Intelligence, AWS Textract, Google Document AI) Evaluation, observability and LLMOps — required, and assessed first • Evaluation harness design: golden datasets, regression suites, LLM-as-judge, adversarial and edge-case sets, statistical treatment of non-determinism. • Tracing, prompt and version management, drift and hallucination detection — LangSmith, Langfuse, Arize Phoenix, W&B Weave or equivalent. • Guardrail frameworks, red-teaming, and continuous evaluation running inside CI. • Traditional NLP, SLM fine-tuning, quantization and distillation to verify, ground or cross-validate LLM outputs Enterprise AI platforms — required • Production experience on at least one of: Azure AI Foundry, AWS Bedrock, Google Vertex AI, Databricks, Snowflake Cortex. • Working familiarity with agent capabilities embedded in enterprise suites — Salesforce Agentforce, ServiceNow AI Agents, Microsoft Copilot Studio, SAP Joule. Integration and data — required • Systems of record: SAP, Oracle, Salesforce, ServiceNow, Workday. Integration middleware: MuleSoft, Boomi or equivalent iPaaS. • Event streaming (Kafka), CDC and ELT patterns, data quality, lineage and contract testing. • Comfort with undocumented legacy APIs, mainframe-adjacent estates, and incumbent RPA estates (UiPath, Automation Anywhere, Blue Prism) that agents must displace or coexist with. Production operations — required • CI/CD, infrastructure as code (Terraform, Helm), monitoring and alerting, structured logging, incident management and on-call discipline. • Release strategy for non-deterministic systems: canary, shadow mode, staged autonomy, rollback. Security and Responsible AI — required • Identity and access for non-human actors: OAuth2/OIDC, service principals, workload identity, short-lived credentials, secrets management. • Least-privilege tool scoping, prompt-injection and data-exfiltration defence, OWASP Top 10 for LLM Applications, PII and DLP handling, immutable audit trails. • Working knowledge of NIST AI RMF, ISO/IEC 42001, EU AI Act risk tiering and India's DPDP Act 2023. Inference economics — required • Token cost modelling, prompt and context caching, model cascading and routing, batch versus real-time tradeoffs, small-model substitution and distillation. WHAT YOU'LL BRING — CONSULTING AND COMMERCIAL • Structured problem solving. Issue trees, hypothesis-driven diagnosis, MECE decomposition — used to narrow scope fast, not to decorate a deck. • Commercial instinct. You design to ROI, TCO, adoption and change effort rather than to novelty, and can construct and defend a business case. • Executive presence. You can debate integration detail with client engineers in the morning and lead a board-level transformation session in the afternoon. • Change and adoption design. Role redesign, human-in-the-loop operating models, controls and assurance — the work that determines whether the system is actually used. • High agency. Ambiguous problem statements, undocumented systems and unhelpful vendors are the normal operating condition, not an escalation. THE EVIDENCE BAR We assess demonstrated deployment, not CV pattern-matching. Shortlisted candidates should be able to evidence all of the following: • At least two LLM or agentic systems you personally built that reached production and were used by people outside your own team. Internal demos, vendor lab builds, hackathon projects and proofs of concept that never entered a production environment do not qualify. • Code you have written in the last twelve months. You will be asked to walk through it line by line. • An evaluation harness you designed — and a specific failure it caught before it reached users. • A production incident on an AI system you owned: what broke, what you changed, what it cost. • A business outcome you can quantify and defend under CFO-grade questioning. • Willingness to remain hands-on. FDCs spend at least half their time building; Senior FDCs at least 30%. TWO LEVELS, ONE ROLE Level Experience Time building Owns Frontier Deployment Consultant (Manager) 5–7 years 50–70% Workstream outcomes, system build, evaluation harness, client engineering relationship Senior Frontier Deployment Consultant (Associate Director) 7–10 years 30–40% Account-level outcomes, pod leadership, architecture and governance standards, CXO relationships, practice assets BACKGROUND WE LOOK FOR A degree in computer science, engineering, mathematics or another quantitative discipline, or equivalent capability demonstrated through work. We deliberately do not filter on institution. Non-linear profiles are actively welcomed — technical founders, staff engineers who moved into client-facing work, consultants who learned to ship. An MBA is neither required nor an advantage on its own. Preferred, not required • Delivery experience inside a GCC or global business services environment. • Depth in one industry — BFSI, life sciences, energy, telecom or industrial products — and process fluency in S2P, O2C, R2R, claims, KYC/AML or clinical operations. • Deployment inside a regulated or data-resident environment. • Process mining or workflow discovery (Celonis or equivalent). • Open-source contribution, published technical writing or conference work in the agentic space. HOW WE SELECT • 1. Portfolio walkthrough (45 min). One production system you built, end to end, including what you got wrong. • 2. Live build (2–3 hours). A timeboxed agentic build against a messy, partially documented data source. AI coding assistants permitted and expected. • 3. Evaluation and failure design (60 min). You are handed a working agent, asked to design its evaluation harness, then asked to break it. • 4. Brownfield integration case (60 min). Architecture and trade-off reasoning against a legacy estate with real constraints. • 5. Executive simulation (45 min). Present outcome, risk and cost narrative to a mock CXO panel. • 6. Practice fit. Conversation with practice leadership on how you work and what you want to build. WHY JOIN PwC • Front-row impact. Shape the core AI strategies and operating fabric of leading global enterprises and Global Capability Centers, at the point where decisions actually get made. • Both halves of the job. The strategic reach, industry breadth and executive access of top-tier consulting, with the speed, autonomy and modern stack of a frontier AI team. • Leverage. What you build once becomes practice IP deployed across accounts and markets — your work compounds instead of ending at go-live. • Compounding learning. Senior practice partners, exceptional peers, and a direct hand in PwC's point of view on autonomous enterprise systems. PwC India · Agentic AI Practice · Frontier Deployment Consultant
Mandatory Skill sets:
GenAI and Agentic AI Engineer
