Why AI Is Finally Starting to Work (And Why Most Teams Still Miss the Point)

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

Introduction

AI is starting to work, but not exactly in the way most expected.

Over the past year, we spoke with more than 400 practitioners and conducted 100+ in-depth interviews with Just Curious Applied AI experts—operators, technical founders, builders, and AI strategists—as well as private-equity operating partners and enterprise leaders responsible for moving AI from pilots into production inside real businesses.

Many succeeded. Some stalled. Almost all learned the same lessons faster than they expected.

What became clear across these conversations is that AI value rarely hinges on model quality or tool selection. The real inflection points show up much earlier: in how work is actually done, how data is prepared for decisions, and whether organizations are willing to redesign workflows rather than automate what already exists.

This is Part I of the Just Curious 2025 Year in Review, drawn from those conversations. Rather than surveying tools or predicting future capabilities, this series synthesizes what practitioners learned in practice as AI moved from experimentation to expectation across middle-market companies, private-equity portfolios, and large enterprises.

Part I focuses on diagnosis. We focus on four patterns that appeared repeatedly—often independently—across very different organizations:

• Why many AI initiatives break before the model ever matters

• Why “having data” is not the same as being ready to operationalize it

• Why teams that start with tools tend to stall, while those that start with workflows compound value

• Why private-equity constraints don’t hinder AI efforts, but clarify which ones are operationally sound

The remaining sections of this Year in Review move from diagnosis to application:

• Part II examines real-world case studies and operator playbooks—not vendor narratives—showing how teams redesigned workflows, deployed AI or automation, and measured outcomes in practice.

• Part III focuses on failure modes: internal resistance, architectural missteps, and quiet forms of self-sabotage that repeatedly caused otherwise promising initiatives to stall.

• Part IV concludes with a practical operator’s guide—a set of recommendations and non-negotiables distilled from these conversations—outlining what leaders must do before committing capital, tools, or organizational credibility to anything labeled “AI.”

This is not meant to be a trend report or a tool guide. It’s a diagnostic of how AI actually can create value once the novelty fades, and why some organizations are already pulling ahead.

Btw, we’ll return to applied expert profiles and individual case studies in a few weeks. 

Here, we start with the patterns that determine whether AI becomes a marginal productivity boost or a durable operating advantage.

Executive Summary: Four Diagnostic Patterns

Across interviews, four diagnostic patterns surfaced repeatedly, regardless of industry, company size, or level of AI maturity. Together, they explain why many AI initiatives stall quietly, while a smaller number compound into a durable operating advantage.

Organizational Perception vs. Operational Reality

Leadership teams often misunderstand how work actually gets done. Process knowledge is anecdotal rather than observed, masking process debt and “shadow workflows” that live in Excel, inboxes, and informal coordination. The implication is simple: you cannot automate what you haven’t actually seen. Teams that skipped direct observation consistently automated assumptions, and paid for it later.

Data Readiness as the Primary Gating Factor

In practice, an organization’s AI strategy is its data strategy. Many teams start from “below zero”: fragmented, inconsistently defined, or poorly tagged data that supports reporting but collapses when AI is asked to make decisions in motion. Dashboards create a false sense of readiness. Without decision-grade data, momentum fades quietly, through narrowed scope, stalled pilots, and rising cost.

Workflow Redesign Before Tool Selection

Teams that began with tools struggled to convert capability into impact. The highest ROI emerged when organizations redesigned—or eliminated—workflows first, treating AI as a change initiative rather than a software deployment. Sequencing, not sophistication, determined outcomes. Where work wasn’t redesigned, AI mostly automated waste.

Private Equity as a Constraint-Driven Multiplier

Private equity doesn’t change the objective of AI adoption; it tightens the constraints. Compressed timelines, limited tolerance for experimentation, and immediate pressure for financial impact force initiatives to be narrower, more operational, and more disciplined. Those constraints don’t hinder AI but expose which initiatives were never operationally sound to begin with. When aligned, even modest workflow improvements can compound into portfolio-level value.

Taken together, these patterns reflect what experienced practitioners independently converged on—often correcting the same misconceptions and warning against the same failure modes—despite operating in very different organizational contexts. 

They also explain why AI success is increasingly uneven: not because the technology differs, but because organizations do.

Organizational Perception vs. Operational Reality

Operator takeaway: If you automate what’s documented instead of what’s actually happening, you’ll automate the wrong thing.

The most common AI failure mode in 2025 had nothing to do with models. It was a blind spot.

Leadership teams often did not understand how work actually moved through the business. Process knowledge was built from check-ins, documentation, and system diagrams that looked coherent on paper, but masked reality.

Felix Rosner of Marble described how leadership teams typically understand their own operations:

“Most of the understanding is really based on them having catch-ups or interviews with their team leads… it’s anecdotally driven. It’s not really data-based.”

That blind spot turns bottlenecks into guesswork. What looks constrained from the top is often a workaround, an exception, or a manual handoff that only becomes visible when you sit with the people doing the work. 

AI initiatives surface this mismatch quickly: teams start with confidence—owners identified, workflows documented—then slow down and grind to a halt as automation meets reality.

Just Curious applied AI experts consistently described this as process debt: unofficial workflows built to survive brittle systems. In many organizations, the real process lived outside the system of record—in Excel, inboxes, and ad-hoc coordination—because the “official” systems didn’t actually fit the work.

In practice, this gap often only became visible once teams observed the work directly.

In one PE-backed services organization, leadership couldn’t explain why a 50-person team was at capacity despite stable demand. Instrumenting the actual workflow revealed that a large share of time was spent on manual data entry—copying information from emails into legacy systems—work that never appeared in process documentation. Automating the “official” workflow would have missed the real bottleneck entirely.

Michael Cohen of Stalliant offered another example:

“We’ve seen the ERP system that finance looks at and just exports CSVs from and builds a model in Excel as opposed to actually using the system in the first place.”

These shadow processes are rarely visible to leadership until automation forces them into the open. When AI is layered on top of a misunderstood workflow, failure is predictable; brittle at the edges, confusing in the exceptions, and quietly bypassed by the people who know how the work really gets done.

These failures rarely look technical. They surface instead as slow, quiet adoption drift. For the most part, operators don’t reject tools because they fear technology. They reject tools that don’t match how they actually work and solve an actual problem. Much of that resistance is grounded in tacit judgment: the non-obvious pattern recognition operators use to handle exceptions the system was never designed to see.

This is why tribal knowledge is both the bottleneck and the opportunity. Jordan Gurrieri of BlueLabel noted:

“Areas where a workflow relies a lot on tribal knowledge is also a great place to think about applying AI… How do we capture that tribal knowledge and make it available to everybody?”

Teams that made progress didn’t bulldoze this tacit judgment and force it into rigid rules or premature automation. They observed how operators actually handled exceptions. Some used simple heuristics—like asking what tasks they would train a brilliant intern to do first—to surface where real leverage lived.

The cost of getting this wrong isn’t abstract. Turgut Jabbarli of VestedInYou described what happens when teams automate assumptions rather than reality:

“They thought the inventory was tracked in the project system. It wasn’t. A single automation built on that assumption cost them roughly $100K.”

Different methods, same conclusion: you can’t automate a process you haven’t actually seen. The teams that moved forward invested early in observing work as it happened—then redesigned around reality, not documentation.

Stop / Start / Watch

The patterns above translate into a small set of practical posture shifts: what to stop doing, what to start doing, and what to watch for as AI moves from experimentation into operations.

Start

• Start with observation, not diagrams. Shadow operators and trace workflows end-to-end.

• Start mapping exceptions. The edge cases usually are the process.

• Start treating tribal knowledge as an asset. Decide what should be captured and what should remain human.

Stop

• Stop assuming SOPs reflect reality. Documentation is often aspirational.

• Stop diagnosing bottlenecks from management interviews alone.

• Stop automating workflows you haven’t pressure-tested in the wild.

Watch

• Shadow workflows (CSV exports, manual reconciliations, copy-paste chains).

• Quiet rejection: tools that exist but are socially bypassed.

Data Readiness as the Primary Gating Factor

Operator takeaway: If data can’t support decisions in motion, AI will stall.

In 2025, teams learned the same lesson repeatedly: AI strategy is data strategy, whether leaders plan for it or not. While executive teams were eager to deploy generative and agentic AI, many initiatives stalled because the underlying data environment couldn’t support real decision-making inside live workflows.

The point came up repeatedly. As Aditya Nair noted:

“Your AI strategy hinges on your data strategy. The first thing I would look at is whether you have a data strategy in place, and how much of the data is siloed.”

What made this especially difficult was that many organizations believed they were already data-mature. Dashboards worked. KPIs were reviewed. Board decks shipped on time. But once AI systems were asked to operate inside workflows—at the edges, in real time, without human interpretation—those assumptions broke down.

Practitioners described this as starting from “below zero.” Data existed, but it was fragmented, inconsistently defined, improperly tagged, or trapped in formats AI systems couldn’t reliably consume. 

This became obvious when teams tried to apply AI to seemingly straightforward finance use cases. A team attempting to use AI to automate expense allocations across a global business discovered that the underlying financial data didn’t agree with itself. Departments and cost categories were defined differently across dozens of subsidiaries, making automation impossible without first standardizing the chart of accounts. The initiative stalled because the underlying data couldn’t support consistent decisions, regardless of model quality.

Aanikh Kler of Lazer Technologies put it succinctly:

“Not every company was prepared to build properly tagged, filtered, and clean data pipelines for what AI needs nowadays to function at a very high level.”

This is why data problems rarely kill projects upfront. Instead, momentum fades. Scope narrows. Pilots look promising but fail to scale because the data architecture—and the skills required to maintain it—were never production-ready. Be mindful of this!

Foundation first or value first?

Practitioners disagreed not on whether data mattered, but when to fix it.

One camp argued that without a solid BI layer and governance model, AI would simply automate noise, amplifying errors faster than humans could catch them. Others warned that waiting six to twelve months for a full data transformation often guaranteed stagnation.

Csongor Barabasi of Bonsai Labs captured the tension:

“Every company thinks they have to do a large data transformation… but you can have side-load and standalone products which already generate more value.”

The real dividing line was context, not ideology. Teams under tight timelines, especially in private-equity-backed environments, often couldn’t afford perfection. Instead, they proved value narrowly: one workflow, one decision, sometimes using manually curated data to justify deeper investment later.

Across both camps, there was agreement on one point: AI does not forgive dishonest data.

In practice, this meant scoping data work to the decision at hand. A full organizational data overhaul was rarely required upfront. What mattered was whether the data supporting a specific operational or financial decision was accurate, accessible, and reliable enough to act on.

Institutional knowledge is the hardest data problem

The most stubborn data challenge wasn’t technical. It was economic.

In many organizations, the most valuable information wasn’t stored cleanly in systems. It lived in fragmented definitions, inconsistent tags, and undocumented assumptions embedded in day-to-day operations. When AI systems were layered on top of that reality, costs rose quickly.

Alex Brownstein of 3C Ventures described the downstream impact:

“If you have an AI built on poorly organized or poorly managed data, it will not serve you super well, and it can be more expensive to deal with the data than almost anything else.”

As a result, successful teams stopped cleaning data for its own sake. Instead, they worked backward from specific decisions, documenting what actually mattered in practice and fixing only the data required to make those decisions reliable.

Why this matters for leaders

Leaders overpay for AI when they underestimate the role of data in both accuracy and cost control. Poor data increases hallucinations, erodes trust, and quietly kills adoption. It also inflates spend: more retries, more guardrails, more human review.

At the same time, as software becomes cheaper to build, proprietary data tied to real workflows becomes one of the few remaining defensible moats. Organizations that connect trustworthy data to decisions unlock speed, not just insight. Those that don’t remain stuck reconciling numbers across shadow spreadsheets while AI potential stays theoretical.

The lesson practitioners learned the hard way was consistent: AI doesn’t demand perfect data, but it does demand honest data, tied to real decisions.

Start / Stop /  Watch

Start

• Start treating data readiness as a strategic constraint, not an IT hygiene task.

• Start working backward from decisions, not datasets.

• Start proving value narrowly, side-loaded or manually curated data is acceptable early.

Stop

• Stop assuming dashboards equal readiness. Reporting data ≠ operational data.

• Stop waiting for perfect foundations before testing use cases in time-constrained environments.

• Stop underestimating the cost of bad data because it compounds.

Watch

• “Below zero” starting points: data exists but is fragmented, inconsistent, or inaccessible.

• Hidden cost creep from retries, guardrails, and human review.

Workflow Redesign Before Tool Selection

Operator takeaway: If you don’t redesign the workflow first, AI will only automate waste.

Across interviews, one shift surfaced repeatedly: teams that created real value with AI did not begin by choosing technology. They began by rethinking how work should happen.

In successful projects, AI was rarely the headline. It was an ingredient—sometimes powerful, sometimes modest—inside a broader redesign of how tasks moved through the organization. Teams that led with tools, by contrast, struggled to translate capability into impact.

The pattern showed up again and again: copilots in search of workflows, agents deployed without ownership, impressive demos that never made it into day-to-day operations. The issue wasn’t capability, but sequencing. As Chris Taylor of Fractional AI observed:

“They’re thinking tools first instead of thinking use case first. The more you can think use case first and then infrastructure to support it, I think that tends to lead to better outcomes.”

In these cases, AI wasn’t misapplied. It was prematurely applied. Without clarity on what outcome mattered, teams defaulted to what the tool could do, not what the business needed.

By contrast, teams that saw meaningful ROI followed a different sequence. They started by identifying where work slowed down, where errors accumulated, and which decisions actually moved outcomes. Only after that did technology enter the conversation.

In practice, this often meant changing who—or what—did the work, rather than accelerating an existing handoff.

In one commerce operation, a core workflow relied on an outsourced team manually matching shopping lists to product SKUs, a process that took nearly a full day to complete. Rather than giving that team an AI co-pilot to work faster, the workflow was redesigned so the system produced the mapping directly, with humans only reviewing exceptions. The largest gains came from changing the structure of the work, not from applying a better tool.

This reframing often led to simpler solutions than expected. In some cases, AI played a small role. In others, it eliminated entire steps. Edmundo Ortega of Machine+Partners described one such moment:

“We realize that, my gosh, you don’t even need this process anymore. Like you can just jump over the whole thing.”

The difference had little to do with technical sophistication. It came down to how teams approached the problem in the first place. They treated AI as a change initiative, not a software rollout. Jordan Gurrieri of BlueLabel emphasized why that distinction mattered:

“Don’t treat AI initiatives like traditional software initiatives. We need a mindset shift. People and process are going to be the greater hurdles than the technical hurdles.”

Several practitioners warned that novelty-driven adoption actively works against this mindset. Arman Hezarkhani of Tenex captured the risk memorably:

“If somebody came to you and said, ‘Hey, I have an AI-powered hamburger. Do you want to eat it?’ You’d be like, what the fuck does that even mean?”

Justin Massa of Remix Partners described the downstream effect this creates inside organizations:

“I observe most people either try, fail, and tune out, or try, find something interesting, and immediately think the next step is to buy a tool.”

The teams that broke this cycle reframed the problem. Rather than asking what to buy, they focused on changing how work was performed. As Massa put it:

“In order to capture the full ROI, everything should start as a technique.”

Why this matters for leaders

AI creates leverage only when it is applied to a workflow worth fixing. Leaders who begin with tools risk automating waste, accumulating software sprawl, and burning trust with operators. Those who begin with outcomes often find that the right technical approach becomes obvious—and materially cheaper—once the work itself is redesigned.

Start / Stop / Watch

Start

• Start with outcomes, not platforms. Define what should change in throughput, cost, or quality before discussing technology.

• Start asking whether a workflow should exist at all. AI often creates the most value by eliminating steps, not accelerating them.

• Start treating AI initiatives as change efforts, with explicit ownership and adoption plans.

Stop

• Stop leading with tools, copilots, or agents. This almost guarantees misalignment between capability and need.

• Stop assuming novelty equals value. “AI-powered” is not a business outcome.

• Stop forcing users into new interfaces when existing workflows can be augmented instead.

Watch

• Impressive demos with no owner. These rarely survive contact with daily operations.
  Pricing and value-capture mismatches, especially when automation dramatically compresses time-to-output.

Private Equity as a Constraint-Driven Multiplier

Operator takeaway: In portfolio environments, AI only compounds when ownership, scope, and incentives are aligned from day one.

In private-equity environments, the objective of AI stays the same, but the operating constraints tighten quickly.

In PE-backed companies, incentives are already aligned around value creation. What differs is the operating reality: compressed timelines, low tolerance for non-impactful experimentation, and immediate pressure to translate effort into financial results. Under those conditions, only a narrow set of AI initiatives can survive long enough to prove value.

In portfolio environments, this often showed up as a focus on decision throughput rather than efficiency in isolation. A private credit firm needed to evaluate significantly more deals without adding headcount or slowing diligence. Rather than pursuing a broad AI program, they focused narrowly on automating the preparation of decision materials by connecting fragmented internal systems. The constraint—fixed team size under time pressure—forced a surgical application that materially increased throughput.

Across interviews, Just Curious applied AI experts pointed to a consistent but understated pattern: AI initiatives unfold differently in private-equity-backed companies than in founder-led startups or large enterprises, largely due to differences in incentives, timelines, and operating constraints.

In theory, PE environments should be ideal for AI-driven value creation. Targets are clear. Operational leverage matters. In practice, many initiatives stall for reasons that have little to do with the technology itself. Operating context, especially time pressure—ends up defining what success looks like.

In private-equity environments, AI initiatives are judged on speed as much as sophistication. As Amrit Saxena of SaxeCap noted, the bar is early, demonstrable impact:

“We like to come into a business and deliver demonstrable EBITDA uplift within, call it two months of getting started… our first and foremost objective is creating enterprise value as quickly as possible with technology.”

Where AI worked in portfolio environments, ownership was explicit, the scope was narrow, and initiatives were framed as operational bets, tied directly to margin, throughput, or cycle time, not as platforms to be rolled out broadly.

Crucially, value in these contexts wasn’t always defined as cost reduction. In several cases, AI was used to increase throughput and velocity, allowing teams to do more, faster, without sacrificing quality. Brandon Gell of Every described one such KPI shift:

“For this PE firm specifically their KPI was if we do one deal a month, how can we go to doing two or three deals… with the same level of quality. We’re not really trying to decrease your cost as a business—we’re trying to say this is going to give you the ability to go capture the next 2× of revenue.”

This reframing mattered. It allowed AI initiatives to be evaluated not just on efficiency, but on their ability to expand capacity under fixed organizational constraints.

The tension becomes sharper in multi-company portfolios. Sponsors want repeatable leverage; operating teams need solutions that respect local realities. The most effective approaches standardized how problems were identified and tested, not the solutions themselves.

Where portfolio-level leverage did emerge, it followed a consistent pattern: narrow scope, fast proof, and repeatability. Jay Singh of Casper Studios pointed to modular execution as the key enabler:

“The mistake we’ve seen not work is trying to build everything at once. What works is building modularly—focus on one discrete workflow, get it into production quickly, build trust in that system, and then compound from there rather than swinging for the moonshot.”

Timing also played a critical role. Ownership transitions created rare windows where change was both expected and enforceable. Teams that used early post-close periods to reset workflows, decision ownership, and metrics were far more likely to see AI initiatives survive long enough to compound.

Why this matters for leaders

In PE contexts, the biggest risk is not choosing the wrong model. It’s wasting time on initiatives that don’t move the value-creation needle.

Discipline matters more than novelty and clarity matters more than ambition.

AI compounds in portfolios only when it is treated as a series of tightly scoped operational experiments, not as a promised transformation.

Start / Stop / Watch

Start

• Start framing AI initiatives as operational bets, each tied to a specific value lever (margin, throughput, cycle time).

• Start narrowing scope aggressively: one workflow, one metric, one accountable owner.

• Start thinking portfolio-wide about problem identification, not solution standardization.

Stop

• Stop launching broad “AI transformations.” These dilute ownership and delay results.

• Stop assuming portcos have AI-ready DNA. Most have strong operations, not applied AI depth.

• Stop over-indexing on vendors without embedded operational understanding.

Watch

• Political drag when AI initiatives threaten informal power centers.

• False standardization: one-size-fits-all solutions that fit no portfolio company well.

Conclusion: From Diagnosis to Application

Once teams narrowed data to specific decisions and workflows, the bottleneck shifted again. At that point, information wasn’t the bottleneck. Workflow design was.

Across interviews, this marked a clear inflection point. The organizations that moved forward did not stop at understanding their processes or cleaning their data. They made deliberate choices about what work should exist at all, where human judgment truly mattered, and which constraints were worth redesigning rather than optimizing.

This is where consensus gave way to divergence.

Some teams translated these insights into durable operational gains. Others encountered new forms of friction—political, architectural, and cultural—that proved harder to surface than any data issue. Tooling and raw technical talent were rarely the deciding factors. It came down to sequencing, ownership, and a willingness to treat AI not as a capability to be layered on, but as a forcing function for organizational change.

The remaining sections of this series move from diagnosis to application:

• Part II examines real-world case studies and operator playbooks, showing how teams redesigned workflows, deployed AI or automation, and measured outcomes in practice.
  

• Part III focuses on failure modes: the internal resistance, architectural missteps, and quiet self-sabotage that repeatedly caused otherwise promising initiatives to stall.
  

• Part IV concludes with a practical operator’s guide—a set of recommendations and non-negotiables distilled from these conversations—outlining what leaders must do before committing capital, tools, or organizational credibility to anything labeled “AI.”

The goal of this series is not to predict the future of AI, but to clarify what has already become true, and what leaders must internalize if they want AI to move from experimentation to repeatable, defensible value creation.

Looking forward! 

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