AI is compressing the data-ingestion phase of professional-services work, allowing teams to review entire data sets instead of prioritizing a small portion.
2
Structured outputs, retrieval-augmented generation, and document extraction can turn days of manual work into minutes, but each workflow needs validation and business context.
3
Enterprise AI adoption depends on close partnerships with users, regular demos, and a focus on business outcomes rather than the latest model or agent framework.
Summary
Kevin Madura and Mo Bhasin describe how AlixPartners is building and deploying an internal generative AI platform across professional-services work. They divide engagements into data preparation, analysis, and client recommendations. AI is already reducing the human effort in the first phase, while allowing teams to apply their methods across an entire corpus instead of reviewing only the most obvious documents. The talk covers three deployed patterns: categorizing large volumes of data with structured outputs and web queries, searching internal material and external databases through retrieval-augmented generation, and extracting structured fields from documents such as credit agreements. The speakers are direct about the work required around these systems. Accuracy depends on business partners, taxonomies, validation, scaffolding, and ways for users to review model confidence. Mo also argues that adoption requires recurring demos and attention to ROI, rather than chasing agents or new models.
AI is likely to reshape knowledge work more slowly than software engineering
Madura uses a chart from METR that measures how long a task an LLM can complete with at least a 50% success rate. He says the sharp improvement in software tasks comes partly from their being verifiable. Knowledge work is messier, so he does not expect the same steep takeoff. The implication is still large disruption for professional services, but the path will be less predictable because real-world work contains judgment and incomplete information.
Professional-services firms may combine senior expertise with AI-scaled execution
Madura describes two common service models. In the junior-led model, senior people direct a large group of juniors who carry out the work. In the senior-led model used by AlixPartners, experienced professionals stay involved in the daily work and deliver results with less leverage. He expects a hybrid in which the knowledge of people with 15 or 20 years in a domain is replicated and made available through AI. He connects this idea to Dwaresh Patel's concept of an AI-first firm.
AI compresses the data-preparation phase of an engagement
Madura divides a typical engagement into gathering and organizing data, analyzing it with a playbook, and producing the recommendation or deliverable the client needs. AI is already reducing the human share of the first phase. He gives document review as an example: reviewing 5,000 contracts at 30 minutes each creates a time and cost limit, which forces teams to prioritize. AI can examine the full corpus, including all vendor contracts or the entire customer base, so the same analysis can be applied to information that humans would have skipped.
Enterprise productivity requires more than individual employee efficiency
Bhasin frames the talk around a gap between AI investment and measured results. He cites claims from Deloitte, the National Bureau of Economic Research, BCG, and S&P Global about planned adoption, limited effects on earnings or recorded hours, difficulty scaling value, and abandoned initiatives. His explanation is that employee productivity and enterprise productivity differ. The useful deployments they discuss are designed to improve the firm's work across engagements, rather than simply helping an individual employee write or search faster.
Structured outputs make large-scale categorization easier to deploy
Bhasin uses vendor and accounts-payable data as an example. A company might need to classify United Airlines under travel or route an IT support ticket to hardware. Earlier approaches required text cleaning, feature preparation, and classifiers such as support vector machines or naive Bayes. With structured outputs, the model can assign items to a taxonomy such as NAICS codes. When the model lacks knowledge about a company, a tool call can run a web query and append information before classification. AlixPartners used this approach to categorize 10,000 vendors with 95% accuracy, doing in minutes work that had taken days.
Classification accuracy depends on business partners and carefully defined taxonomies
Bhasin says the speed and accuracy gains from categorization did not happen without human work. Business partners worked closely with the AI team to reach the desired accuracy, and that collaboration helped turn skeptics into people requesting new applications. The taxonomy carries business context into the workflow. He also warns that model outputs are stochastic rather than deterministic. Getting each step accurate and controlled makes it possible to connect those steps into larger agentic workflows.
RAG can connect enterprise documents to licensed information sources
Bhasin describes an enterprise RAG application that handles hundreds of gigabytes across PowerPoint files, documents, Excel files, and CSVs. It helps consultants answer questions about internal material when they have to get up to speed quickly. The team also embedded third-party API specifications so an LLM could call information sources that had previously been accessible only through a web interface and manual handoffs. He cautions that a prompt such as 'reason across all documents' does not describe what a basic RAG system can do. The team builds these capabilities step by step.
Document extraction needs schemas, validation, and visible confidence signals
Madura shows how a long credit agreement can be converted into fields such as contract parties, maturity date, and senior lenders. The workflow combines a document, a schema, an LLM, and validation around the result. The schema determines the business value because it defines what the system extracts and why. For user trust, AlixPartners uses log probabilities from the OpenAI API and aligns them with structured-output values. It calculates a geometric mean over the relevant tokens as a rough confidence signal, which lets users decide where human review is needed.
Scaling enterprise AI requires recurring demos and an ROI filter
Bhasin says technical implementation is only part of enterprise adoption. His team prototypes in Streamlit and builds the production interface in React, then demonstrates new work to the firm once a month. Those demos help people see what the team can build and support continued investment. He also says the team must resist the pull of every new agent framework, model, or MCP release. Their measure is ROI, earned through individual bug fixes and improvements. Partnerships across the organization remain part of the work.
"The business value really is in the schema of what you're actually extracting and why you're extracting that information."13:28
Who should watch
You are building internal AI products and need examples of workflows that have reached real users inside a professional-services firm.
Your team is evaluating document review, classification, RAG, or extraction and needs practical guidance on schemas, validation, and human review.
You are trying to get business stakeholders to adopt AI and want a concrete approach based on demos, partnerships, and ROI.
