Where AI Value Actually Came From in 2025 (Strategic Moves 4-6)

From Just Curious, a decision-support platform for leaders making high-stakes AI decisions.

Introduction

In yesterday’s issue of Part II, through 11 case studies, we examined how teams created AI value by redesigning workflows: moving decisions upstream, removing human middleware, and building systems that finished work end to end.

Today’s issue completes Part II.

The remaining six case studies focus on a different—but equally decisive—set of constraints. In these organizations, the core workflows were already clear. What determined success next was whether AI systems could earn trust, achieve adoption, and reach production fast enough to matter—inside regulated environments, skeptical teams, legacy infrastructure, and real economic pressure.

Across these cases, value came from a second set of concrete shifts:

• choosing deterministic systems over generative ones where correctness mattered more than flexibility,

• earning adoption by eliminating work operators actively disliked rather than enforcing change,

• and compressing time-to-value by shipping narrow, production-ready systems quickly.

In practice, that looked like:

• a PE-backed preschool platform expanding EBITDA by optimizing staffing hour by hour using deterministic optimization rather than generative AI,

• a Midwest manufacturer replacing fragile Excel-based reporting with a deterministic data model to regain real-time job visibility,

• and a legal research platform reaching market in three months by prioritizing evaluation and first-answer correctness over broad generative capability.

Across the remaining cases, value compounded because AI was use selectively, and was paired with disciplined execution. It was’t about “more AI.”

The six case studies in this issue include:

• How a PE-Backed Preschool Platform Expanded EBITDA by Optimizing Staffing Hour by Hour

• How a Midwest Manufacturer Replaced 30-Tab Excel Reporting with Real-Time Job Visibility

• How a Legal Research Platform Reached Market in 3 Months Without Sacrificing Accuracy

• How a Wealth Management Firm Raised Process Adoption from 25% to 80% by Removing Admin Work

• How a Multi-Site Manufacturer Gained Real-Time Inventory Visibility Across Three Continents

• How a Legal Technology Company Shipped an AI Product in Months Using a SWAT Engineering Model
  

Taken together, these cases show what execution looks like after the initial redesign, when AI systems must survive contact with real constraints, not idealized architectures.

Executive Summary

The Second Half of the Pattern: Trust, Adoption, and Speed

Across the final six case studies in Part II, AI value hinged less on what was built and more on how it was deployed into real operating environments.

These teams were confronting harder questions:

• Where must systems be deterministic to be trusted?

• How does adoption actually happen inside skeptical organizations?

• And how quickly must systems reach production before momentum is lost?

Three additional operational moves determined the outcome.

1) Deterministic systems beat generative ones in high-stakes domains

In finance, compliance, healthcare, and core operations, teams consistently avoided probabilistic reasoning for decision-critical logic. Instead, they paired deterministic pipelines, traditional machine learning, or optimization models with AI only where uncertainty belonged: data ingestion, synthesis, or language translation.

Deterministic systems were easier to audit, easier to explain, and faster to deploy. In several cases, explicitly not using generative AI was the decision that made the system viable at all.

2) Adoption followed relief, not mandates

Across these cases, usage spread because systems removed work operators actively disliked.

When AI eliminated repetitive documentation, manual reconciliation, or administrative drag, adoption followed immediately. Only after systems proved useful did organizations earn the right to automate more complex workflows.

Adoption, in practice, was a lagging indicator of relief.

3) Speed-to-value mattered more than architectural elegance

The final pattern was about tempo.

Winning teams aggressively descoped, integrated with legacy systems, and shipped narrow solutions quickly, often in weeks or months, not years. They rejected multi-year ERP replacements, speculative abstractions, and over-designed platforms in favor of systems that worked now.

In many cases, hard-coded logic and imperfect integrations were not technical debt. They were strategic accelerants.

What these cases collectively show

Taken together, the full set of 17 case studies demonstrates that AI advantage is no longer theoretical. And not evenly distributed.

Organizations that redesign workflows, apply AI selectively, earn trust through determinism, and move quickly to production are already compounding operational gains.

Those that do not often stall, not because the technology (or AI) failed, but because execution did.

In Part III, we’ll turn to that failure mode directly: why initiatives that are sound in principle repeatedly break down in practice, and how organizational structure—not model capability—quietly caps returns.

Move 4: Favor Deterministic Systems in High-Stakes Domains

How a Midwest Manufacturer Replaced 30-Tab Excel Reporting with Real-Time Job Profitability, with Michael Cohen, Stalliant

Starting Conditions (Status Quo)

A private equity-backed custom manufacturer in the Midwest struggled with operational blindness. The newly installed COO spent nearly 50% of his workweek manually wrangling data in Excel, relying on VLOOKUPs across 20 to 30 spreadsheet tabs, just to generate baseline operating reports. This manual latency forced critical business decisions to be made on stale information.. The lack of real-time visibility directly threatened profitability, as the firm’s ability to capture a 5–15% margin on any given job depended entirely on accurate, up-to-the-minute quoting of labor and materials. As Michael Cohen described, the COO “spent so much time, almost 50% of his time just wrangling information together,” maintaining “20, 30 tabs inside Excel just to make the sort of decisions that he should be making daily as COO.”

The Intervention

The team rejected the idea of a “magic wand” generative AI solution for the financial data. Instead, they focused on foundational data engineering to build a reliable information engine. They deployed Power Query and Azure Fabric to create a data pipeline that connected directly to QuickBooks. The intervention intentionally prioritized deterministic data modeling over probabilistic LLMs to ensure 100% accuracy in financial reporting, reserving generative AI for a later phase involving unstructured inputs like bills of materials.

The Redesign

The workflow shifted from manual CSV extraction to an automated data model. Previously, the COO functioned as a high-paid intermediary, stitching together disconnected systems by hand. The redesign established a live data layer where reporting updated automatically. This eliminated the need for manual reconciliation, allowing the operations team to view job profitability in real time and focus on strategic decision-making rather than spreadsheet maintenance. As Cohen explained, “we started with Power Query… with the data model… to connect to QuickBooks and move that information out of QuickBooks so that it could be usable,” shifting the COO from “a clerical role where he’s wrangling all this information” to “a bona fide head of operations” with a real-time view of the business.

The After State

The transformation immediately restored real-time job profitability visibility and returned approximately 50% of the COO’s time. Between reclaimed executive time and reduced manual Excel work across the staff, the firm projected nearly $500,000 in annual savings. Beyond efficiency, the firm began capturing the 5–15% margin per job that was previously at risk, simply by having the accurate data required to quote jobs correctly. As Cohen noted, reclaiming “50% of the time for head of operations” alone represented “$100,000 a year,” with “another three to $400,000 a year from the employees trying to figure out all of the Excel processes.”

What They Learned (and Would Do Differently)

Private equity sponsors are structurally allergic to “black box” AI in finance due to prior exposure to inaccurate reporting. The team learned that using mature, deterministic technologies for the core financial model is required to build trust. Generative AI should only be introduced once the underlying data infrastructure is proven accurate.

Who This Is Relevant For

Lower middle-market manufacturers, construction firms, and distributors where complex job costing is currently managed via spreadsheet workarounds rather than functional ERPs.

Operator Takeaways

• Audit executive time; if leadership spends half their week on data entry, it indicates a systems failure, not a staffing problem.

• Prioritize deterministic data models over generative AI for financial reporting to ensure accuracy and investor trust.

• Quantify the cost of stale information; calculating the margin lost on inaccurate quotes often justifies the infrastructure investment immediately.

How a Legal Research Platform Launched an AI Assistant in 3 Months Without Sacrificing Accuracy, with Csongor Barabasi, Bonsai Labs

Starting Conditions (Status Quo)

A five-year-old legal technology company faced existential pressure from a wave of AI-native legal research competitors. While the firm possessed a successful legacy product and a valuable proprietary dataset, it lacked AI-powered features. Internal attempts to modernize were too slow due to constraints in attracting top-tier AI engineering talent. The critical constraint was user trust. The team identified “first answer correctness” as the primary KPI: a single incorrect citation or summary would permanently undermine adoption. As Csongor Barabasi explained, “the whole legal AI boom happened,” and the client “was struggling to attract AI talent fast enough… to keep up with this demand,” while leadership set an aggressive goal “to launch a legal research assistant… and do this in three months.”

The Intervention

The company deployed an external AI engineering “SWAT team” to compress a multi-year roadmap into months. They explicitly avoided a generic chat wrapper, instead building a custom legal research assistant grounded in the firm’s proprietary data. To address strict industry compliance and data privacy requirements, the team architected the solution within a Microsoft Azure environment, serving OpenAI models securely rather than sending data to public APIs.

The Redesign

The redesign prioritized evaluation before generation. Before writing the application code, the team collaborated with in-house lawyers to build a rigorous “evaluation dataset.” This allowed engineers to quantitatively test precision and recall against a golden standard rather than relying on qualitative feedback. The system indexed millions of documents into a Retrieval-Augmented Generation (RAG) pipeline, creating a closed loop where answers were strictly derived from the firm’s verified legal archives rather than the model’s general training data. As Barabasi emphasized, “one of the key KPIs for a legal AI product is the first answer correctness,” because “if the first answer is even slightly wrong, then you break trust and the product is never used again.”

The After State

The product reached beta in three months and general availability in six. This speed allowed the incumbent to neutralize the competitive threat of new entrants. The new AI product generated seven-figure net new ARR within six months. Furthermore, the successful deployment validated the company’s technical roadmap, helping unlock a new round of funding. As Barabasi noted, “we got to beta launch within three months,” had “paying customers” by that point, and “by those six months, they have already hit seven figure ARR only from this product as a net new ARR.”

What They Learned (and Would Do Differently)

In expert-user domains, you cannot iterate in public with low accuracy. The team learned that establishing an evaluation dataset before development is non-negotiable when user tolerance for hallucination is near zero.

Who This Is Relevant For

B2B SaaS companies, legal and compliance firms, and information services providers with high-value proprietary data and skeptical, expert user bases.

Operator Takeaways

• Build an evaluation dataset before building the product to measure “first answer correctness” objectively.

• Deploy LLMs through secure enterprise environments (e.g., Azure) to satisfy compliance requirements without building infrastructure from scratch.

• Embed subject matter experts directly into the engineering team to validate outputs during the build phase, not post-hoc.

How a PE-Backed Preschool Platform Increased Labor Productivity 50% by Optimizing Staffing Hour by Hour, with Amrit Saxena, SaxeCap

Starting Conditions (Status Quo)

A private equity-backed preschool platform operating over 100 sites faced a data visibility crisis. Having grown through inorganic M&A, the organization relied on disparate information systems that did not communicate. Leadership could not identify overstaffed locations without hours of manual data extraction and Excel analysis. As Saxena describes, “When you ask their leadership… ‘which of your sites is over staff today?’ They would respond and say, ‘look, to figure that out, we need to go and pull data from a bunch of different systems, then do analysis in Excel, take us a few hours…’” Because labor is the single significant variable cost in this industry, the inability to tightly manage student-teacher ratios kept margins middling despite the scale of the operation.

The Intervention

The team explicitly chose not to use Generative AI. Instead, they deployed canonical operations research and machine learning techniques. They first integrated fragmented data sources into a single repository, then built a machine learning model trained on historical attendance data to predict student presence down to the hour. Crucially, this model had to account for specific state regulations where staffing ratios are determined by the age of the youngest student in the room, not just total headcount.

The Redesign

The workflow shifted from static scheduling to dynamic, algorithmic recommendations. The system now runs an optimization model that prescribes specific actions to site principals throughout the day—such as moving three toddlers to a different classroom or shifting a teacher's start time by an hour. As Saxena explains: “We built a little machine learning model trained on all the historical attendance data… to go predict for every hour of every day, how many students of each age level will you have in each classroom… And basically what the system does is it dynamically figures out these recommendations all day, every day… and pings the site leader… and asks them to make the changes.” The principal acts as the human-in-the-loop, accepting or rejecting these recommendations. Rejections provide direct feedback to the model, teaching it which logistical moves are impractical at specific sites.

The After State

The intervention transformed the unit economics of the business. Labor productivity increased by over 50 percent, and EBITDA margins expanded by more than one-third. As Saxena summarized the impact, “In this specific case, labor productivity increased over 50%. EBITDA margins increased by over a third, completely transformative.” The organization moved from reactive, manual staffing adjustments to a proactive system that maximizes capacity utilization while adhering to regulatory mandates.

What They Learned (and Would Do Differently)

Generative AI is often the wrong tool for logic-heavy optimization problems. For logic-heavy, regulatory-constrained environments involving numerical optimization, traditional machine learning and operations research often outperform LLMs. They also learned that capturing the nuance of regulations (e.g., age-based ratios) was critical to making the model useful in practice.

Who This Is Relevant For

Multi-site healthcare facilities, logistics networks, and regulated services businesses where labor is the primary variable cost.

Operator Takeaways

• Don’t force Generative AI onto problems better solved by optimization and operations research.

• Centralize data from disparate ERPs first; you cannot optimize what you cannot see.

• Build feedback loops so frontline managers can reject recommendations and retrain the model on ground truth.

Move 5: Earn Adoption by Removing Pain

How a Wealth Management Firm Raised Advisor Adoption from 25% to 80% by Removing Admin Work, with Jordan Gurrieri, BlueLabel

Starting Conditions (Status Quo)

A family office specializing in retirement planning marketed a proprietary onboarding framework called the “Plan Path.” However, internal execution was broken: only 25 percent of advisors actually completed the full process. The workflow was heavy on manual documentation, requiring advisors to juggle multiple tools, perform duplicate data entry, and rewrite technical financial reports into client-friendly language. This operational friction had a direct cost: advisors who followed the full process closed 40 percent more business, making low adoption a primary revenue bottleneck. As Jordan Gurrieri explained, “only 25% of advisors actually would follow through on all of the steps to deliver in that plan,” because “creating each plan was a heavy lift,” involving “a lot of documentation… a lot of manual work, inputting data, juggling multiple tools.”

The Intervention

The team used a short sprint to identify high-friction points rather than attempting a full platform replacement. They mapped the advisor workflow and targeted specific pain points like duplicate entry and “lost client language.” They launched a lightweight pilot focused on two specific tasks: using AI to parse PDF documents (such as Social Security reports) to map data fields automatically, and generating the “What I Want” section of the client plan.

The Redesign

The workflow shifted from manual transcription to AI-assisted ingestion. Instead of advisors manually keying data from client documents, the AI parsed files and populated the necessary fields. For the qualitative “What I Want” section, the AI drafted narratives using the client’s own terminology, which advisors previously struggled to capture consistently. The rollout strategy relied on “transfer and transformation,” converting senior advisors—initially the biggest skeptics—into internal champions who led peer training. As Gurrieri noted, “in PDF reports, we could leverage AI to do parsing of the reports and mapping of the fields from customers… like their social security reports into the fields required for the path plan.”

The After State

Advisor adoption surged from 25 percent to over 80 percent. The efficiency gains returned approximately one day of work per week to each advisor. Crucially, higher adherence to the “Plan Path” improved the firm’s close rate by five percentage points, generating millions in incremental assets under management. As Gurrieri observed, the firm “went from about 25% advisor adoption to upwards of 80%,” saved “about a day’s a week… worth of work for each advisor,” and improved close rates “by five points,” translating into “millions of dollars of assets under management growth.”

What They Learned (and Would Do Differently)

Small, concrete wins drive organic adoption better than mandates. By saving advisors 30 minutes per plan on a specific, hated task, the firm created internal demand for more AI features rather than forcing compliance.

Who This Is Relevant For

Wealth management firms, family offices, and high-touch professional services where sales processes rely on heavy documentation and customization.

Operator Takeaways

• Target “quick wins” that remove specific administrative pain points (like PDF parsing) to earn the right to automate larger workflows.

• Convert skeptical senior practitioners into champions early; their buy-in validates the system for the rest of the organization.

• Measure success by adoption and conversion lift, not just time saved.

Move 6: Compress Time-to-Value Ruthlessly

How a Contract Manufacturer Cut Quote Turnaround from Days to Seconds, with Alex Gruebele, Tau9 Labs

Starting Conditions (Status Quo)

An injection molding manufacturer identified a market gap between low-volume 3D printing and mass-production molding, but couldn’t capitalize on it. Their sales process relied on manual email exchanges to handle Requests for Quotation (RFQs), creating significant latency and putting them at a disadvantage against larger competitors offering instant pricing. The delay between interest and pricing was a direct threat to conversion, as being the first to quote often doubles the chance of winning the deal. As Alex Gruebele described, “in the world of like ordering parts from a contract manufacturer, you generally have to like email somebody to like get a request for quotation,” a process that is “slow and people lose a lot of business… because they don’t respond fast enough.”

The Intervention

The team chose to build a custom e-commerce and quoting platform rather than force-fitting an off-the-shelf SaaS tool that couldn’t model their pricing logic. They utilized AI-enabled coding tools to accelerate the entire software development lifecycle—from design to testing—compressing the build time to just four months. The core technical intervention was building a custom 3D modeling engine that could ingest customer files, analyze them for manufacturability, and apply specific pricing logic automatically.

The Redesign

The workflow shifted from human-mediated negotiation to automated self-service. Previously, sales engineers had to manually calculate costs and draft responses. In the new system, customers upload 3D models directly to a web portal. The system instantly processes the geometry and pricing variables, presenting a final cost that enables immediate checkout via credit card. This removed the human bottleneck entirely from the initial transaction layer, where speed determined conversion. As Gruebele explained, the team focused on “building out a sort [of] custom 3D modeling… engine that can understand a 3D model and do limited… manufacturability analysis and inform the pricing algorithm as well.”

The After State

Quote turnaround dropped from hours or days to seconds. The platform unlocked high-margin, mid-volume work that depended on immediate responsiveness. By automating the intake and pricing of complex parts, the firm positioned itself to win bids simply by being the first to respond with a transactable price. As Gruebele noted, the team worked “to literally get it down to seven clicks between… part file upload to… credit card.”

What They Learned (and Would Do Differently)

Speed is the ultimate feature. User testing revealed that even the difference between a ten-second quote and a one-minute quote caused drop-off.

The team also learned to stop over-engineering for future flexibility. With AI reducing the cost of code, hard-coding current pricing logic proved more effective than building complex abstractions for hypothetical future needs.

Who This Is Relevant For

Contract manufacturers, custom machinery shops, and industrial service providers where complex manual estimation slows down sales velocity.

Operator Takeaways

• Audit your “time-to-quote” metric; if it’s measured in hours, you’re losing bids to competitors who are simply faster.

• Resist scalable abstractions early; hard-code current pricing logic and refactor later as requirements change.

• Prioritize custom interfaces when off-the-shelf software forces compromises in your core manufacturing workflow.

How a Multi-Site Manufacturer Unlocked 5–10× Larger Projects with Real-Time Inventory Visibility, with Turgut Jabbarli, VestedInYou

Starting Conditions (Status Quo)

A mid-sized manufacturing company with operations across the U.S., Canada, and Europe was growing 20–30 percent annually but hit a hard ceiling. Despite strong demand, the owner had zero real-time visibility into inventory levels or project health, relying on fragmented reports from disparate teams to understand inventory levels and project status. This lack of a central nervous system made leadership risk-averse; they refused to bid on larger, more complex projects because a single operational slip-up could be catastrophic. As Turgut Jabbarli described, the lack of visibility had already produced costly surprises: “they thought the inventory was in the project side. It wasn’t, it was going to be delayed,” forcing the owner to “airship the whole material which cost him extra 100K” and “completely burned his margin,” pushing the project into a loss.

The Intervention

The firm rejected a traditional multi-year, multi-million dollar ERP implementation, which would have delayed value and increased operational risk in a mid-market context. Instead, they opted to build a custom, modular AI-first ERP. The team conducted a month-long deep dive into stakeholder requirements to map actual data flows. They built a new solution using open-source technology and Python, leveraging AI to compress the build timeline to three to four months. Crucially, they decided not to rip and replace the existing accounting software (QuickBooks), choosing instead to integrate with it to minimize disruption.

The Redesign

The workflow shifted from reactive data gathering to proactive alerts. Previously, the owner spent hours daily synthesizing information from different regions. The new system centralized operations using an agent-driven alert layer to flag specific issues—such as inventory shortages or project delays—directly to staff. The system proactively tells the team what needs attention, allowing the owner to view capacity and inventory in minutes rather than hours. 

As Jabbarli explained, “no one has to go into Excel sheet to type in some numbers,” because data now flows automatically through workflows, with only core approvals requiring management input.

The After State

Operational clarity allowed the owner to assess inventory and capacity in minutes instead of hours. With a reliable system in place, the company began bidding on and winning projects 5 to 10 times larger than their previous average. The firm is now projected to accelerate growth from 20 percent to 50–60 percent because the infrastructure can finally support the volume. As Jabbarli noted, the expectation is that they will “continue their growth and also start… taking on bigger projects,” with those projects “5 to 10x bigger size because now they can support that.”

What They Learned (and Would Do Differently)

You cannot layer AI on top of broken data flows. The team learned that rebuilding the underlying data architecture was a prerequisite to any ERP or AI layer; without it, the system would have failed long-term.

Who This Is Relevant For

Mid-market manufacturing, construction, and logistics companies managing multi-site operations with spreadsheets or legacy software.

Operator Takeaways

• Do not rip and replace core accounting systems immediately; integrate your AI layer on top to maintain financial continuity.

• Audit data architecture first; if your data flows are inconsistent, pause the AI build until the foundation is rebuilt.

• Demand modular builds that deliver production-ready tools in 3–4 months, rejecting the traditional 12-month ERP timeline.

Conclusion: From Evidence to Constraint

Part II documented where AI value actually appeared inside operating organizations.

Across seventeen case studies—spanning industries, company sizes, and technical approaches—the same operational moves recurred. These were not abstract strategies or vendor narratives. They were concrete changes to how work entered systems, where decisions were made, and how responsibility was distributed between humans and software.

What these cases establish is not that AI is generically transformative, but that value becomes visible only after workflows are redesigned with discipline. In nearly every successful deployment, the decisive factor was not model sophistication, but a willingness to intervene earlier in the workflow, remove unnecessary human mediation, and narrow scope until systems could reliably complete work end to end.

Importantly, these outcomes emerged under real constraints. Most organizations operated with imperfect data, legacy systems, regulatory oversight, and limited tolerance for risk. Progress did not come from eliminating those constraints, but from working within them, treating AI not as a layer to be added, but as a forcing function for operational clarity.

The consistency of these patterns exposes something else:

If the moves that created value are this repeatable, then the persistence of stalled or failed AI initiatives cannot be explained primarily by technical uncertainty. The gap is not a lack of tools, models, or examples. It is the difficulty of executing these moves inside existing organizational structures.

In practice, the same interventions that unlocked value also challenged entrenched incentives, ownership boundaries, and governance norms. Moving decisions upstream redistributed control. Removing human middleware altered roles. Designing systems that completed work reduced discretion. Compressing time-to-value required abandoning familiar planning rituals.

These pressures rarely appear as explicit resistance. Instead, they surface as delay, scope dilution, architectural compromise, and “temporary” workarounds that quietly reintroduce the very frictions AI was meant to remove.

Part III examines these failure modes directly: why initiatives that are sound in principle repeatedly stall in practice, and why the obstacles are structural, predictable, and often self-inflicted.

If Part II shows what becomes possible once workflows are redesigned honestly, Part III explores why most organizations struggle to make—and sustain—those changes.

Stu

About Just Curious

Just Curious is a decision-support platform for leaders making high-stakes AI decisions.

We work with private equity firms and middle-market operators who are past experimentation and need clarity on what actually works, before time, budget, or momentum lock them into the wrong path.

We’ve built a curated network of applied AI experts; operators, builders, and technical strategists who have deployed AI systems in production inside real businesses. When a team is evaluating an AI initiative, we run the need through that network to surface concrete approaches, realistic scopes, trade-offs, and execution paths.

Teams submit a short description of their problem. We anonymize it, gather multiple expert perspectives (including budget ranges and timelines), and return them side-by-side. Like a lightweight, upfront Mini-RFP. No vendor pitches. No obligation.

→ Submit a brief

Listen to the conversations

Many of the insights in this review come from long-form conversations with operators, builders, and AI leaders published on the Just Curious podcast. Full interviews are available on Spotify and Apple Podcasts.

→ Listen here