Raju Malhotra is Chief Product & Technology Officer at Certinia where he leads product strategy and AI innovation for professional services.
getty
A study conducted by Service Performance Insight titled "The Impact of AI on Professional Services" found that nearly 90% of current AI initiatives in professional services sit at "level 1" or "level 2" maturity—use cases that assist individuals or add modest automation inside existing workflows. Only about 10% of firms have reached "level 3," where agentic systems orchestrate work across core platforms and reshape operations at scale. (As a disclosure, Certinia licensed the research to make it available to the public at no cost but had no influence on the methodology, data collection or conclusions.)
That gap has a structural explanation. Most firms are deploying a more capable autocomplete feature, and the architecture behind that approach and true agentic AI are fundamentally different.
The first kind of tool is a “scribe." It handles natural language tasks well: meeting notes, document generation and executive summaries. But it has no awareness of the commercial, financial and resourcing structures that govern how a services business actually runs.
An operator, on the other hand, is an agentic system that reasons within the rigid, deterministic constraints of a services organization where margins depend on accurate revenue recognition, contract compliance and precise deployment of talent. An operator executes action, validates its outputs and handles the downstream consequences across the operational architecture.
The structural reason most AI initiatives stall at level 1 or 2 is the absence of a rules-enforced logic layer: the set of hard business rules, validated data models and financial constraints that the AI cannot override. This “deterministic scaffolding” is what prevents an AI system from approximating its way through a balance sheet.
Probabilistic AI operates on likelihood, which is powerful for creative and generative tasks. But in project margin analysis or revenue recognition, that same probabilistic nature becomes a compliance liability. A model can hallucinate, drift mathematically and introduce errors into workflows where anything less than 100% precision is nonnegotiable.
An operator requires grounding in a validated constraint architecture—a metadata-driven ontology that functions as the authoritative source of truth for the business. That layer encodes:
• Project Ontology: How hours, costs and revenue connect across different project types
• Resource Semantics: The meaningful distinction between a senior consultant in New York and a specialist in London, informed by millions of staffing outcomes
• Business Rules: Revenue recognition standards, multi-currency accounting logic and project delivery frameworks that must be enforced, not estimated
General-purpose LLMs trained on open-web data are not inherently grounded in these operational and financial constraints.
Whether evaluating a new system or auditing what's already deployed, the question is the same: Is it genuinely operating at the level you need? Pressure-test any system across these three dimensions.
A true operator is built on domain-specific metadata and ontology. It understands how milestones, revenue recognition and utilization relate, treating "project," "resource" and "margin" as structurally connected concepts. Its training is grounded in real operational history, not generic internet data.
Ask your vendor to explain the services data model and how it was constructed. This should ideally name the volume and type of services data the model was trained on and describe how that training connects to the constraint architecture. If the answer describes a general-purpose LLM with a thin layer of services terminology applied on top, the foundation is that of a scribe.
Operators reason within validated constraints. An LLM may generate a recommendation, but a financial logic layer must validate it before anything is written back into systems of record. Every output should carry a traceable audit trail of the data and business rules that produced it.
Ask the vendor to show how a complex margin scenario is calculated, and what happens when the LLM produces an incorrect result. Look to understand the validation layer, the audit trail and how errors are caught before they reach the system of record.
A real operator acts across systems simultaneously. It can modify a project plan, adjust staffing and update billing schedules in response to a scope change while respecting contractual, financial and resource constraints at the same time.
Ask vendors to demonstrate how the AI handles an in-flight change order across billing, resources and margin impact in a single scenario, end to end, with no human handoff between systems.
The most revealing demonstrations tend to show how the system behaves when constraints conflict or conditions change mid-process. Pay attention not only to whether the AI can execute across systems, but also to how it handles exceptions, policy violations and downstream operational impacts in real time.
Most firms have the right instinct: Start small, prove value and expand. The difference is knowing where to really begin.
Start with your data model. The validated constraint architecture can only be as reliable as the underlying data is coherent. Disconnected project records, inconsistent resource hierarchies and manual billing work-arounds have to be resolved before an operator can function. Audit those gaps first.
Next, identify workflows in your business where approximation is nonnegotiable: revenue recognition close, high-complexity staffing decisions and multi-entity billing reconciliation. These are workflows where a scribe creates real risk and the ROI on operator-level architecture is clearest. Build the validated constraint layer around those workflows before expanding.
Finally, establish the audit trail as a nonnegotiable requirement in every AI output that touches a system of record. The audit trail serves as both a compliance control and a learning mechanism, surfacing where the probabilistic layer is drifting and where the constraint architecture needs tightening.
AI won't transform professional services by making firms better at generating text, but by making the firm itself smarter where every project decision, staffing call and billing action is informed by the full weight of the organization's operational history. That transformation comes from firms that had the discipline to build the foundation before chasing the frontier.
Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?
此内容由惯性聚合(RSS阅读器)自动聚合整理，仅供阅读参考。 原文来自 — 版权归原作者所有。