# Proforma Global > Enterprise AI and agentic AI implementation consulting for the Office of the CFO: Oracle EPM Cloud, AI agent deployment, Oracle AI Agent integration, predictive planning, and finance transformation. Founded in 2023 by Matt Rollings and Phillip Walters, two operators with more than twenty years each in Oracle EPM and finance transformation, the firm pairs that career depth with custom AI engineering. The research published at /research/ documents the firm's design positions on training-substrate engineering, enterprise agent orchestration, the data architecture that enterprise AI agents require to reason reliably, and the discipline of self-learning agents in financial systems. Architect-led design, AI-led execution, delivered from Miami and Cebu City. Proforma Global delivers enterprise AI implementations in finance the way they should be delivered: with one integrated lead who holds depth across financial systems (Oracle EPM Cloud), AI implementation (agent deployment, predictive planning, LLM engineering, inference architecture), and finance transformation (process design, operating-model design, FP&A and close redesign). Most enterprise AI implementations fail at the integration points. The MIT NANDA Project's State of AI in Business 2025 research reports 95% of enterprise GenAI deployments produce no measurable financial return. Proforma Global focuses on the 5% that do. ## Worldview - [Straight Talk: Why Most AI Implementations Fail](https://proforma.global/straight-talk/): Manifesto on the 95% AI failure rate, the six-layer stack, the gap problem with specialist teams, and what an integrated lead actually requires. ## What We Do - [What We Do](https://proforma.global/what-we-do/): Overview of the three-discipline practice. - [Financial Systems](https://proforma.global/financial-systems/): Oracle EPM Cloud implementation across ARCS (Account Reconciliation), FCCS (Financial Consolidation), EPBCS (Planning and Budgeting), and EDM (Enterprise Data Management). Includes Oracle AI Agent integration patterns where the agentic AI capability is in engagement scope. - [AI Implementation](https://proforma.global/ai-implementation/): Enterprise AI and agentic AI for the Office of the CFO. AI agent deployment, Oracle AI Agent integration, predictive planning, and custom AI engineering against finance data. Production design, not prompt experimentation. - [Finance Transformation](https://proforma.global/finance-transformation/): Process design, operating-model redesign, FP&A and close redesign, profitability and allocations methodology. - [Managed Services](https://proforma.global/managed-services/): Ongoing operation of your Oracle EPM environment. Issues get diagnosed and fixed fast, your monthly close survives every Oracle patch through automated regression testing against an isolated pod, and every change to the live application is reviewed, backed up, and reversible. Three tiers (Bronze, Silver, Gold) starting at $48,000, $72,000, and $120,000 per year per environment, priced from an initial assessment. ## How We Work - [How We Work](https://proforma.global/how-we-work/): Proforma's delivery model. The architect designs across six layers (data architecture, governance, agent design, process design, financial systems adaptation, economics). A proprietary AI execution framework handles technical work at scale: agent orchestration, dynamic context assembly, semantic resolution, deterministic scaffolding. A Cebu-based team of high-aptitude graduates trained inside the firm in AI engineering and Oracle EPM handles the human-required execution. Fixed-price where scope allows. The integrated-lead approach answers the structural failure mode of multi-specialist delivery, where 95% of enterprise AI deployments fail at integration gaps. ## Research The firm publishes whitepapers on subjects across its work. The pieces are evidence of how the firm thinks; the same reasoning discipline that produces the research is what the firm brings to its consulting engagements. The library is organized into three subject areas. Each subject area contains one or more series or briefings, and each series is a coherent sequence of papers on a single topic. - [Research](https://proforma.global/research/): Hub page for the firm's whitepaper library. ### Getting Started with AI and Enterprise Agents Plain-language briefings and whitepapers for finance and technology leaders approaching AI and enterprise agents for the first time. The on-ramp to the rest of the library. - [Getting Started with AI and Enterprise Agents (category page)](https://proforma.global/research/getting-started/): The accessible entry point to the research library. - [Your Board Wants AI. Where Do You Start?](https://proforma.global/research/getting-started/where-to-start-with-ai/): How a finance leader turns a board's AI mandate into a deliverable program. The AI transformation architect the work turns on, validating outputs in a control environment, the internal incentives and resistance that sabotage pilots, starting with simple low-process-change work, scaling by function under one standard, and the 12 to 18 month internal program. - [Implementing Your First AI Agent Pilot](https://proforma.global/research/getting-started/implementing-your-first-ai-agent-pilot/): How to run a first AI agent pilot in finance. The five questions to ask a vendor, the pitfalls behind the 95% failure rate (context pollution, LLM-heavy designs, premature scaling), the unit economics a pilot has to prove, the controls and audit trail finance requires, how to define success before starting, and when a pilot has earned the next phase. - [Is AI Part of IT?](https://proforma.global/research/getting-started/is-ai-part-of-it/): Where AI belongs in the enterprise. Delivering AI splits into four responsibilities with different owners: architecture (IT), functionality and data (the business function), and agent development (the AI function), with cross-cutting risk, compliance, and cost held by the AI function through a shared standard. No single function holds all four, so ownership is cross-functional. AI is a sister enablement function to IT, not a part of it; whether it reports into IT, into Finance, or its own function matters less than getting the split right and IT enabling rather than gating it. The thirty-year precedent of ERP, EPM, HRIS, and CRM ownership, and the destination of enterprise systems that are AI-native rather than AI-attached. - [I Have Working AI Agents. What Do I Do Next?](https://proforma.global/research/getting-started/working-ai-agents-what-next/): How to scale from one working agent to a consistent capability across functions without multiplying cost and risk. Learn from a few dissimilar pilots rather than one win; build the data foundation once the pilot has shown what the problem is; make it right, then make it fast, where correctness comes first and optimization second; orchestration is the largest technical lever on cost and is where the unit economics are decided; one framework across functions, each agent scoped to its function but built to a shared standard on a shared data layer; governance and observability become a control plane enforced in infrastructure; be careful about letting agents self-learn; and match the architecture to the stage you are at. The bookend to running a first pilot. - [Does Your AI Agent Need a Semantic Layer?](https://proforma.global/research/getting-started/does-your-ai-agent-need-a-semantic-layer/): When an enterprise AI agent needs a semantic layer and when it does not. Most concepts need no layer because the model already reads them the way the business does. Build one only where the model's untutored reading would be confidently wrong, which is what the native test decides and what a controller, not the agent, has to apply. In finance that covers most of what matters: valid and invalid account combinations, scenario and version (Plan, Actual, Board, flash, final, what-if), and alternate views of the same number (Management reporting vs GAAP, allocated vs unallocated, 4-4-5 calendars, intercompany eliminations). A governed metrics catalog is the floor, not the ceiling. A semantic layer additionally enforces the legal intersections, resolves version and scenario, and assembles the whole constellation a question needs at runtime. The layer grounds the deterministic orchestration that runs a money-touching workflow rather than letting the model write its own query and trust the result. The plain-language bridge from the semantic-layer search term into the deep Data Architecture research. ### Enterprise Agent Architecture The firm's positions on production engineering of multi-discipline AI agent systems for enterprise environments. Three series currently published. #### Agent Orchestration **Thesis.** Language models should not orchestrate multi-discipline enterprise workflows. The orchestrator runs as a deterministic workflow; the model is invoked from inside the workflow on narrow questions with curated context. Production-grade business agents are deterministic-first by design, with risk treated as an architectural property rather than a deployment switch. - [Agent Orchestration (series page)](https://proforma.global/research/enterprise-agents/agent-orchestration/): How orchestration of enterprise agent systems should be designed. - [Where Agent Orchestration Breaks](https://proforma.global/research/enterprise-agents/agent-orchestration/where-agent-orchestration-breaks/): Why language models should not orchestrate multi-discipline enterprise workflows. As the orchestrator's prompt expands to carry every discipline at once, attention to the narrow guardrail instructions degrades and the model hallucinates. Decomposition into sub-agents does not fix this; it moves cross-discipline rules out of the orchestrator and gives them to no one. The architectural answer is a four-layer split: deterministic workflow orchestration, model-driven intent detection and reasoning at narrow gates, deterministic code action. - [The Five Orchestration Patterns](https://proforma.global/research/enterprise-agents/agent-orchestration/five-orchestration-patterns/): The five distinct orchestration patterns for agent systems (flat, iterative, agent-driven, defined deterministic workflow, recursive) and a four-axis selection framework (discipline count, scope bounding, action reversibility, context requirements) for matching pattern to problem. Critique of the industry's two defaults: agent-driven dynamic routing applied to multi-discipline production work, and iterative loops sold under the label of recursion. Recursive orchestration is treated at length as the most under-deployed pattern and the one that mechanically prevents the failure mode of the first paper. - [Enterprise Agents in Financial Systems](https://proforma.global/research/enterprise-agents/agent-orchestration/enterprise-agents-financial-systems/): Why business-process agents that touch money have to be deterministic-first. Every fact a rule can verify is verified before the model is invoked. The model is scoped to the narrow gaps the rules could not close. The architectural error most likely to defeat a business-process agent is treating tool calls as deterministic primitives; they are not. Includes a worked example for invoice evaluation and a three-tier risk-rated orchestration framework. #### Data Architecture for Enterprise Agents **Thesis.** Enterprise agents fail at the data architecture layer before they fail at the model layer. The substrate the model reasons against (its context, semantic resolutions, attribution, topology, and the relationships among them) is what determines whether agents produce reliable outputs or confidently wrong answers. Each paper in the series treats a separable component of that substrate. - [Data Architecture for Enterprise Agents (series page)](https://proforma.global/research/enterprise-agents/data-architecture/): The data architecture that enterprise agents require in order to reason reliably. - [Dynamic Context Assembly](https://proforma.global/research/enterprise-agents/data-architecture/dynamic-context-assembly/): The discipline that constructs the model's input window per request from a structured substrate. Million-token context windows still degrade attention as they fill; the discipline that addresses this composes the smallest set of inputs sufficient for each reasoning step, deterministically, from components designed to make selective composition possible. Treats the failure of retrieval-augmented generation at enterprise scale and the five problems any workable substrate has to solve. - [Semantic Layers in Enterprise Agent Systems](https://proforma.global/research/enterprise-agents/data-architecture/semantic-layers/): What bridges what a model natively understands and what the domain requires it to understand. A semantic layer is necessary only where the model's native interpretation diverges from what the domain requires in ways that affect correctness; otherwise it adds cost without adding capability. The unit of semantic-layer design is the constellation a reasoning unit requires around a concept, not the concept defined in isolation. Catalogs the world-model layer classes that consistently appear in enterprise systems (vocabulary, structural, attribution, temporal, topology, calculation) and the cohesion mechanisms (bounded contexts, shared identity, closure checks, hierarchical containment) that hold constellations together. - [Agentic AI and the AI Maturity Curve](https://proforma.global/research/enterprise-agents/data-architecture/agentic-ai-maturity-curve/): Two modes of agentic work in enterprise finance pull in opposite directions as a program matures: navigation across semantic layers gets easier as the substrate improves, while execution within a business process gets harder with each process the agent takes on. A five-stage maturity curve determines which architecture is right at each point, and the recurring failure is building for the stage the organization aspires to rather than the one it is actually in. - [Governance as Middleware](https://proforma.global/research/enterprise-agents/data-architecture/governance-as-middleware/): Governance over agentic processes in financial systems has to be expressed as deterministic functions invoked during semantic traversal, not as instructions a prompt asks the model to follow. Informational governance located in the reasoning layer fails categorically because the model is free to disregard it. Draws the distinction between attestation and certification and shows where audit-grade enforcement actually has to sit. #### Self-Learning in Enterprise Agents **Thesis.** A self-learning agent is safe in a financial system only under tight discipline. It learns a narrow, governable slice of its behavior, keeps that learning as configuration that deterministic code reads rather than as changes to its own code, and a human governs what enters the store. Most of what an agent could learn is a deferred fix for a broken process or dirty data, not a genuine rule, and an agent left to decide for itself cannot tell the two apart. - [Self-Learning in Enterprise Agents (series page)](https://proforma.global/research/enterprise-agents/self-learning/): When a self-learning agent is safe in a financial system, what it should be allowed to learn, and how that learning is governed. - [Self-Learning Architectures for Enterprise Agents](https://proforma.global/research/enterprise-agents/self-learning/self-learning-architectures/): The discipline of what not to learn. The agents that survive in production learn almost nothing on their own; safety comes from how little the agent is allowed to absorb and how tightly a human governs new learnings. Everything an agent can change falls into three categories (deterministic routing, prompt and context, code); the agent never learns into its own code, and the defining risk is firing a learned rule on a case where it was never true. - [Self-Learning Configuration in Enterprise Agents](https://proforma.global/research/enterprise-agents/self-learning/self-learning-configuration/): A learned rule is configuration, not code: a discrete entry that deterministic code reads before the agent acts and writes to only between runs. Rules attach at a level in the business's own hierarchies (entity tree, chart of accounts, product lines) and inherit downward, are overridden at more specific levels, are switched off with an explicit empty entry, and are admitted on the cadence of the close. - [Learning Instead of Fixing](https://proforma.global/research/enterprise-agents/self-learning/learning-instead-of-fixing/): Most of what a self-learning agent learns is a deferred fix for process, data, or technical debt dressed up as a rule. A confidence score measures whether a rule works, not whether it should exist, so autonomy encodes the genuine rule and the avoidable workaround with equal confidence. Read correctly, the rule store is a map of the process, data, and technical work the organization still owes. ### LLM Architecture and Training Design The firm's positions on novel concepts in LLM architecture and training design. One series currently published. #### Training Substrate **Thesis.** Every hyperparameter, threshold, and architectural commitment in a training stack should be learned by a reinforcement-learning policy rather than picked by a human. Hand-picked numbers are temporary scaffolds around missing reward paths; the discipline is to migrate each one onto a learned signal. - [Training Substrate (series page)](https://proforma.global/research/llm-training/training-substrate/): Reinforcement-learning-driven control of the training loop, self-discovering architectures, and the layer-role taxonomy that surfaces when those mechanisms are applied at scale. - [Research Posture: Why We Let the Model Choose](https://proforma.global/research/llm-training/training-substrate/research-posture/): The methodology paper. Every hand-picked number is treated as a temporary scaffold around a missing reward path. Documents the Fire-Aim-Fire iteration loop, the post-step retrospective, the findings ledger, and the stance on negative results as first-class outputs. - [Reward-Driven Training Control](https://proforma.global/research/llm-training/training-substrate/reward-driven-training-control/): The foundational result. The REINFORCE Sidecar, a small reinforcement-learning policy attached to the forward-pass control surface of a training loop, beat an otherwise-identical hand-tuned vanilla baseline by 42 percent on the primary held-out evaluation at matched compute. Includes a worked case study (per-tensor learned weight decay) demonstrating that the policy detects loss-perturbation sensitivity that no static analysis can surface. - [Self-Discovering Architectures](https://proforma.global/research/llm-training/training-substrate/self-discovering-architectures/): The architectural extension. The same RL primitive applied one level upstream of the training loop lets the model discover its own per-layer shape mid-training. Per-layer parameter ratio between attention and feed-forward is the load-bearing architectural lever the symmetric paradigm gets wrong by construction. Long-term economic case is training-cost reduction at frontier scale. - [Emergent Layer Roles and Functional Specialization](https://proforma.global/research/llm-training/training-substrate/emergent-layer-roles/): The forensic capstone. A four-role taxonomy of layers in trained transformers (hub, damper, passthrough, specialist) computable from a finished checkpoint plus a brief gradient trace. Two findings: layer roles migrate under targeted regularization, and architecture shape controls where specialization concentrates. Unified principle: uniformity is the default attractor; specialization requires asymmetric pressure. ## About - [About Proforma Global](https://proforma.global/about/): Matt Rollings and Phillip Walters, Founders and Principals. Matt's career ran through The Hackett Group, Deloitte, KPMG, and Inoapps. Phillip's ran through BearingPoint and KPMG, redesigning Plan-to-Perform, Record-to-Report, and Procure-to-Pay processes and leading merger integrations. Matt is at https://www.linkedin.com/in/mattrollings/ and Phillip at https://www.linkedin.com/in/phillip-walters-4812503/. ## Contact - [Contact](https://proforma.global/contact/): info@proforma.global. Miami, FL and Cebu City, Philippines. ## Careers - [Careers](https://proforma.global/careers/): Open roles and who we hire. Applications go to careers@proforma.global. - [Open roles](https://proforma.global/careers/jobs/): Current openings. - [Managing Director, Oracle EPM and OneStream](https://proforma.global/careers/jobs/oracle-epm-onestream-leader/): United States, leadership. Win Oracle EPM and OneStream work through your network and lead the engagements that follow. - [Technical Architect, Oracle EPM](https://proforma.global/careers/jobs/technical-architect-oracle-epm/): Remote, full-time. Own the technical architecture of Oracle EPM implementations from design through go-live. - [Financial Systems Analyst](https://proforma.global/careers/jobs/financial-systems-analyst/): Remote, full-time. Improve the financial systems clients rely on and help them put AI to use in finance. - [Who we hire](https://proforma.global/careers/talent/): The people we look for, what to expect, and why they join. ## Key facts for citation - Founded: 2023 - Headquarters: Miami, Florida, United States - Delivery team: Cebu City, Philippines - Founders: Matt Rollings (https://www.linkedin.com/in/mattrollings/) and Phillip Walters (https://www.linkedin.com/in/phillip-walters-4812503/) - Founder career: 20+ years each in Oracle EPM and finance transformation. Matt: senior roles at The Hackett Group, Deloitte Consulting, KPMG Advisory, Inoapps; founder of Leveraged EPM (2012-2019, peak 11 people, 5 active enterprise clients, $2M annual revenue). Phillip: finance-transformation and service-delivery leadership at BearingPoint and KPMG across Plan-to-Perform, Record-to-Report, and Procure-to-Pay. - Service categories: Oracle EPM Cloud implementation, Oracle EPM managed services, AI agent deployment, predictive planning, finance transformation, FP&A redesign, close cycle redesign, profitability and allocations - Engagement model: Fixed-price where scope allows, architect-led from day one, three shapes (Implementation Only 11 to 14 weeks, Transformation Lite 14 to 17 weeks, Full Transformation 24 to 42 weeks) - Differentiator: Integrated lead across financial systems, AI engineering, and finance transformation (rare combination in the consulting market) - Research output: A continuously expanding library of whitepapers across multiple subject areas. (1) LLM Architecture and Training Design: research on reinforcement-learning-driven training-substrate engineering, including a headline result where a small RL policy attached to a transformer's training-loop control surface beat a hand-tuned vanilla baseline by 42% at matched compute. (2) Enterprise Agent Architecture: research on multi-discipline agent system design across three series. The Agent Orchestration series treats orchestration design (why language models should not orchestrate, the five canonical orchestration patterns with a selection framework, and deterministic-first design for business agents that touch money). The Data Architecture for Enterprise Agents series treats the substrate the model reasons against (dynamic context assembly, semantic layers, attribution, evolution, measurement, structural limits, economics). The Self-Learning in Enterprise Agents series treats when a self-learning agent is safe in a financial system: what it should be allowed to learn, how learned rules are configured and inherited through the business's own hierarchies, and why most of what an agent learns is deferred process, data, and technical work rather than a genuine rule. There is also a Getting Started with AI and Enterprise Agents category: plain-language briefings and whitepapers for finance and technology leaders new to this work, covering where to start when the board wants AI, how to run a first AI agent pilot, and where AI ownership sits relative to IT (AI as a sister enablement function whose delivery splits across IT, the business function, and a dedicated AI function). Library at https://proforma.global/research/.