The ROI of AI: Practical Frameworks for Enterprise Value

Opening the conference, Section CEO Greg Shove argued that organizations cannot rely on executive cheerleading or static legal policies to drive adoption. Instead, leadership teams must author an operational AI manifesto defining expectations, token budgets, and cultural norms around AI use. He highlighted that despite widespread tooling, only about 10 percent of enterprise workforces are currently AI fit, meaning most knowledge workers lack the skills to extract meaningful economic value from advanced models.

To fund transformation without inflating corporate overhead, Shove proposed holding back one-quarter of standard workforce backfills over a twelve-month period. In a 1,000-person company experiencing a typical 12 percent annual attrition rate, pausing backfills on roughly 30 positions reallocates approximately $3 million toward inference credits, agent development, and change management upskilling.

Shove warned that a majority of existing enterprise agents are already financially bankrupt because they consume background inference while running unmonitored or generating ignored outputs. While augmenting workers with large language models yields an estimated 5 to 15 percent productivity lift, measuring this individual lift directly is challenging, meaning executives must audit existing agents immediately to terminate non-performing automations.

John Hancock CIO Karthic Sakthivel discussed parent company Manulife's public commitment to realize $1 billion CAD in business value from artificial intelligence by the end of 2027. Sakthivel emphasized that AI investments must originate strictly from defined business problems rather than technological fascination, cautioning against treating AI like a sandwich looking for a picnic.

As a concrete case study, Sakthivel described the deployment of Quick Quote, an internal generative AI tool designed to assist insurance underwriting. Historically, delivering a preliminary, non-binding life insurance qualification assessment took human underwriters at least 24 hours. Quick Quote condensed the initial analysis down to 15 minutes, processing thousands of requests monthly and freeing experienced underwriters to focus on complex, high-margin risk assessments.

Addressing measurement limitations, Sakthivel acknowledged that self-reported time savings represent an unreliable metric because employees frequently understate hours saved to avoid receiving heavier workloads. Furthermore, he noted that out of 100 pilot use cases evaluated across an organization, 99 may fail to produce immediate value, requiring leaders to isolate Goldilocks use cases capable of demonstrating returns within six months.

Section founder Scott Galloway addressed the cognitive dissonance confronting corporate executives who hear frontier lab founders warn of existential extinction risks while simultaneously pitching multi-billion-dollar initial public offerings. Galloway argued that clear regulatory guardrails—analogous to standards enforced by the FAA, FDA, and EPA—are essential features of capitalism that enable industries to scale safely rather than impediments to innovation.

Galloway observed a structural bifurcation in the global AI market, comparing US frontier labs to high-margin German luxury automakers and open-weight models to efficient, high-volume Asian manufacturers. He noted that China heavily regulates domestic AI applications to prevent consumer harm, which correlates with 75 percent of Chinese citizens feeling optimistic about AI compared to just 25 percent in the United States, creating an opening for open-weight models to dominate enterprise market share.

For Fortune 500 business leaders, Galloway offered the counterpoint that managing existential technological threats lies far outside an executive's fiduciary mandate. Corporate leaders should avoid posturing on national policy and instead focus on mundane operational execution: distributing site licenses, leading by personal example on keyboard, eliminating workplace friction, and rewarding workers who apply AI to accelerate business goals.

Elsevier CTO Jill Luber detailed how the scientific publisher achieved a 6 percent revenue lift in the first half of 2026 by integrating generative AI into its core academic platforms. With over 20 million full-text research articles in its repository, Elsevier designed tools such as Claims Radar, which allows researchers to highlight scientific assertions in a paper and instantly identify supporting, contradicting, or related findings across global literature.

To circumvent the massive token expenses and latency associated with running raw large language models across extensive scientific text corpora, Elsevier leveraged a decade of prior machine learning investments. The company pre-computes domain knowledge graphs, taxonomies, and ontologies, effectively absorbing the computational token cost once internally so that end-user queries execute rapidly against structured relationships.

Internally, Luber described using models to modernize legacy codebases, citing an intricate chemistry database SQL query that historically took 32 seconds to execute and had resisted human remediation due to code complexity. An engineering harness paired with an LLM re-engineered the query in two days, slashing execution latency to four seconds. However, Luber stressed that human peer review and editorial judgment remain indispensable, positioning AI strictly as a power tool rather than an autonomous decision-maker.

Anthony Kayapa, Global CTO and CIO at SS&C Technologies, warned against using pull requests (PRs) or closed support tickets as measures of engineering productivity. He noted that developers utilizing coding models frequently generate hundreds of single-line changes, whitespace adjustments, or cosmetic comment edits that artificially inflate volume metrics while delivering zero economic value.

To manage costs and govern interactions across an engineering workforce of nearly 9,000 personnel, SS&C built a proprietary Gateway HQ utility. The platform tracks prompts, token consumption, and code commits, enforcing automated guardrails that block unverified code submissions and prevent employees from using expensive frontier models for trivial tasks. Requests are dynamically routed to local on-premises models first, escalating to commercial foundation models only when deeper reasoning is necessary.

Kayapa outlined the limits of AI-assisted software acceleration, explaining that while modern web frameworks and reporting tools can achieve 30 to 40 percent productivity increases and defect reductions, complex legacy architectures such as COBOL mainframes yield modest 10 to 15 percent gains. Furthermore, SS&C imposes strict monthly token budgets per developer, requiring engineers to justify budget overruns against customer-sponsored revenue milestones.

Michael Dominic, Head of AI at Section, presented an executive framework for managing autonomous agent deployments across individual, functional, and cross-enterprise tiers. He asserted that any decision to embrace AI transformation inherently expands an organization's operational risk surface area, requiring the chief executive to determine whether the strategic risk of inaction outweighs the tactical risk of implementation.

To evaluate agent deployments safely, Dominic introduced a framework balancing operational reversibility against blast radius. Low-blast-radius, highly reversible workflows—such as internal drafting or personal research—should be democratized with minimal oversight. Conversely, workflows with high blast radius and low reversibility, such as automated transaction processing or external client communications, require mandatory human-in-the-loop validation.

Dominic outlined rigorous maintenance hygiene: every agent must launch with an automated 30-day expiration date to prevent ghost agents from quietly draining compute budgets in the background. Teams must inspect agents weekly for runtime errors, monthly for output quality, and quarterly for hard business value. In terms of expenditure, Section recommends budgeting the equivalent of two fully loaded employees per 50 workers for inference, which at Section totals approximately $250,000 to $300,000 annually.

Ted Paris, SVP of Analytics, Intelligence and AI at TD Bank, discussed the financial institution's progress toward achieving a CEO-mandated $1 billion annual value creation target divided between revenue acceleration and cost reduction. Operating within a conservative, highly regulated banking environment, TD established centralized standards covering model risk, data privacy, and cybersecurity before sanctioning production builds.

Paris described TD's strategy of identifying scalable enterprise rails rather than scattered point solutions. TD initially deployed a natural language knowledge management platform within its customer contact centers to allow service representatives to resolve customer inquiries immediately without placing callers on hold. Once validated, those same technical and operational rails were extended across retail store networks and commercial relationship management teams.

To ensure accountability, TD mandates a three-in-a-box governance structure for all major initiatives: a dedicated business lead, an AI subject matter expert, and a technology CIO lead. Paris stressed that the business lead maintains ultimate accountability, requiring them to integrate projected AI savings or revenue directly into their departmental financial plans. He advised organizations to target no-regret opportunities while redesigning end-to-end outcomes to counter nimble, AI-native competitors.

Michelle Quan, COO of Runway, shared how the 250-person creative AI company expanded its marketing spend and digital footprint 100-fold in 2026. Recognizing that traditional creative pipelines could not support the necessary ad volume across fragmented international markets, Runway built an internal ad agent in one week to automate short-form asset generation and localization.

Rather than burdening its senior creative studio with repetitive production, Runway spun up an agile ad tiger team fluent in the agent platform. The team produced up to 1,500 targeted, 10-to-15-second localized ad variants weekly across diverse regions including South Korea, Brazil, Vietnam, and Mexico, analyzing performance data multiple times daily to optimize placement and customer acquisition costs.

Addressing the risk of low-quality automated content, often described as AI slop, Quan emphasized that rapid production must not lower creative standards. Runway pre-embeds approved brand kits, color palettes, logos, and character consistency parameters directly into its tool harnesses. She noted that high-volume generation must remain grounded in clear storytelling and audience relevance, warning that generating volume for volume's sake is a waste of corporate resources.

Eric Safard, Head of AI at Booz Allen Hamilton, outlined strategies for managing tokenomics across approximately 250 active enterprise AI initiatives. Rather than viewing token utilization purely as a cost burden, Safard noted that rising token consumption among skilled senior engineers frequently correlates with a 5x to 8x productivity lift, whereas junior staff often struggle along a steeper learning curve.

Safard proposed a 10-20-70 architectural distribution for enterprise model deployments. Roughly 10 percent of use cases—primarily critical applications like cybersecurity vulnerability remediation—require frontier models with premium pricing power. Another 20 percent rely on mid-tier commercial models for convenience and generalized reasoning. The remaining 70 percent of enterprise workloads are best handled by open-weight models (such as Nvidia Nemotron) fine-tuned on corporate data and deployed in private cloud or sovereign on-premises environments.

Booz Allen recommends an inverse development lifecycle: teams should validate use cases quickly using closed commercial APIs to establish return on investment before undertaking the engineering required to fine-tune open-weight models. Safard cautioned against hunting for overly intricate problems, advising companies to target simple, high-frequency tasks—such as automated compliance tracking and timecard verifications—which reliably generate the strongest immediate returns.

She emphasized that because AI touches diverse business functions, attempting to force all initiatives into a single financial equation leads to frustration and underinvestment.

Measurable ROI applies to discrete workflows with verified historical baselines across time, cost, or quality. In an inbound sales development workflow, automating lead research and personalized messaging reduced human outreach labor from 10 hours to one hour per week per SDR. For a four-person team, net monthly costs dropped from $4,500 to $1,700 (factoring in $200 weekly inference), while booked meeting rates rose from 5 to 7 percent, unlocking an estimated $1.2 million in monthly sales pipeline.

For broader deployments, Directional ROI tracks functional metrics over time (such as Section's engineering team increasing monthly pull requests from 2.5 to 9.4 for an inference expense equal to 0.75 of an engineer's salary), while Assumed ROI models workforce capacity gains. However, Witten cautioned that increasing internal output capacity creates zero economic return unless the business can successfully monetize that additional output, urging executives to match their evidentiary expectations to the maturity of their transformation.

Bhushan Sethi, Lead Client Partner at PwC, argued that corporate leaders must reconnect AI metrics to the classic Balanced Scorecard framework spanning financial, customer, employee, and risk dimensions. Measuring input activities such as token volumes or employee training counts is meaningless unless directly linked to business outcomes like customer retention, pipeline conversion, or operational margin.

Sethi observed that implementations routinely stumble when companies attempt to automate flawed, non-standard processes or encounter unvetted data quality issues. Effective transformation requires mapping workflows down to granular task levels (L3 through L5) to identify precisely which tasks can be handled by autonomous agents and how the remaining responsibilities redefine human job roles.

Addressing workforce dynamics, Sethi confirmed that entry-level hiring is contracting across knowledge-intensive sectors because AI handles routine tasks like reconciliations, presentation drafting, and initial research. To cultivate future leadership, organizations must intentionally train junior staff in uniquely human competencies—such as client storytelling, relationship building, and hallucination detection—while modeling long-term cost sustainability against potential compute inflation, wage shifts, and future AI labor taxation.

Wayfair CTO Fiona Tan described the e-commerce company's strategy of deploying generative tools alongside N8N, a deterministic workflow automation platform, to its 5,000 to 6,000 corporate knowledge workers. By providing non-technical personnel across legal, finance, and supply chain with deterministic platforms, employees could codify standard operating procedures and selectively plug in large language models only where contextual evaluation was needed.

In technical operations, Tan highlighted Wayfair's catalog enrichment initiative across 40 million products. Engineering initially used advanced frontier models to validate automated product data extraction and image discrepancy detection. Once the workflow proved effective, the team shifted execution to specialized models running offline in batch mode during off-peak computing hours, cutting operational costs by 95 percent.

To mitigate runaway token expenditures, Wayfair enforces monthly usage caps across different employee personas, turning off expensive frontier models for general corporate staff and using budget alerts as diagnostic triggers to fix inefficient looping workflows. Crucially, Tan noted that while employees may build personal productivity workflows in sandboxes, any code interacting with production databases must remain strictly governed and maintained by dedicated engineering teams on PagerDuty.