What AI will not fix, and what it will reveal

The Quiet Disappointment Underneath the Hype

The first time I watched an "AI transformation" project fail without anyone calling it a failure, I was sitting in a leadership review. The pilot worked. The pilot metrics looked good. The pilot did not survive contact with the rest of the organization. Months later, the slide deck still said "in production" but no one had touched it since the first few weeks.

That experience is a common one across the AI projects I see. The technology is not the problem. The organization underneath the technology is the problem. And nobody is having the right conversation about it, because the right conversation is uncomfortable: it is not about which model to license, it is about what we have been avoiding for a decade and finally have no excuse to keep avoiding.

This essay is about that conversation.

Something is changing in the rooms where AI decisions get made.

Two years ago, "AI strategy" was a sentence that ended with a budget. Today it ends with a question: "Where did the productivity actually go?"

The data has caught up with the slide decks. MIT NANDA's 2025 report on generative AI in business (a preliminary study of 300+ public deployments, 50+ interviews and a survey of 153 leaders, not peer reviewed) reported the most quoted statistic of the past year: 95% of enterprise generative AI pilots delivered no measurable P&L impact. Not 50%. Not 70%. Ninety-five.

In my experience, where AI pilots stall the cause is usually organizational rather than technical: strategy, ownership, governance, and change management. The technology worked. The organization couldn't absorb it.

Three other findings sit alongside this:

• 55% of the business leaders who made employees redundant because of AI say they made wrong decisions (Orgvue, 2025 survey of 1,163 senior leaders). In a February 2026 Careerminds survey of 600 HR professionals, 35.6% of companies that cut roles had rehired more than half of them, about a third reported losing critical skills, and only 21.4% said AI fully replaced roles without operational problems.
• After AI adoption, email time rose 104%, chat and messaging rose 145%, the average focused session shortened by 9%, and multitasking rose 12% (ActivTrak, 2026, 443 million hours of activity from 163,638 employees). An eight-month UC Berkeley Haas study at a roughly 200-person tech company, published in Harvard Business Review in February 2026, found employees took on broader tasks and extended their workdays after adopting AI.
• Gartner predicts organizations will abandon 60% of AI projects unsupported by AI-ready data through 2026, and 63% of the data management leaders it surveyed said their organization lacks, or is unsure it has, the right data practices for AI. (Gartner, February 2025)

The pattern is clean. The AI itself is doing its job. The organizations underneath the AI are not.

This is the part of the cycle where promises get audited. Forrester's 2026 predictions said AI would be "trading its tiara for a hard hat" as enterprises focus on demonstrable ROI over vendor promises. Executives are no longer asking whether AI is interesting. They are asking where the revenue, the cost reduction, or the headcount efficiency actually shows up - and increasingly, the answer is nowhere obvious yet.

This essay starts from a single observation: AI doesn't solve the structural problems. It exposes them faster. Every weakness in your data, your decisions, your design, your culture is now being magnified at machine speed. The hype cycle is settling. The structural problems are getting sharper, not softer. And the next three to five years will be defined less by who adopted AI fastest, and more by who fixed the foundation underneath it.

The Permanent Problems - What AI Won't Fix

Some problems sit so deep in how organizations actually work that no amount of model capability addresses them. They are not engineering problems. They are not data problems. They are problems of structure, judgment, and human behavior - and AI mostly amplifies them.

The map below distinguishes the problems AI exposes (that were always there) from the problems AI creates (that didn't exist before), and within each, which are permanent and which will evolve. The four quadrants give you the operational lens for which work pays off where.

Six of the permanent (top-left quadrant) problems are worth naming.

1. Bad data foundations

The bottleneck for every AI initiative is not the model. It is the data the model sees. When organizations attempt to scale AI, they encounter fragmented data ecosystems, inconsistent business definitions, siloed systems, incomplete metadata, and unclear lineage. The 60% abandonment rate Gartner projects for AI projects without AI-ready data is not a model problem - it is a data foundation problem dressed in newer clothes. AI doesn't fix bad data. It produces faster, more confident outputs from bad data. Section 06 of this essay treats the data foundation in depth.

2. Organizational seams

Conway's Law (1968) is older than the modern computer industry and remains undefeated: organizations design systems that mirror their communication structures. In the AI era, this works in both directions and accelerates. AI agents deployed across siloed teams reinforce the silos. They create pockets of local efficiency within boundaries that remain as rigid as ever. Each department's agents optimize for local efficiency rather than global coherence - exactly the way the humans already do, but faster, with less friction, and less visibility into the broader misalignment they are encoding.

A finance agent optimizing approvals will not coordinate with a procurement agent optimizing throughput. A risk agent optimizing fraud detection will not negotiate with a product agent optimizing onboarding conversion. The system gets more efficient at producing the same cross-functional warfare that lived in the org chart before.

3. Decision quality under pressure

AI is excellent at producing options. It is poor at choosing between them when the cost of being wrong is asymmetric. The hardest decisions in a growing company - which customer segment to abandon, when to fire a senior employee, whether to take the round at a lower valuation - are not decisions a model improves. They are decisions where judgment, context, relationships, and timing matter more than analysis. AI gives faster bad decisions if leaders aren't clear. A leadership team that lacked decision-making discipline before AI will lack it more visibly after, because the volume of decisions accelerates and the ones that get rushed are exactly the ones that mattered.

4. Risk blindness

This one is core to my practice, so I'll be precise. AI optimizes the visible. Risk lives in the invisible. LLMs and agents are extraordinarily good at the question you asked, the metric you defined, the workflow you scoped. They are extraordinarily poor at surfacing the question you didn't think to ask, the metric that quietly drifted, the workflow that depends on an undocumented person who is about to leave.

The gap between what the system was designed to surface and what actually matters doesn't shrink as models improve. It widens, because the optimization happens faster while the human capacity to spot what was missed stays roughly the same. This is the structural reason risk management becomes more valuable in the AI era, not less.

5. Trust and accountability gaps

When an AI agent makes a decision, who owns the outcome? The engineer who built the prompt? The product manager who scoped the workflow? The executive who signed the policy? The vendor whose model was used? The customer who interacted with it?

Most organizations have not answered this question, and they will only realize the answer matters after the first incident that requires it. Regulated industries (banking, healthcare, insurance) are starting to install accountability frameworks because they have to. The rest of the economy hasn't yet, and the lag will be expensive. AI doesn't fix accountability. It makes it more diffuse, and the cost of diffuse accountability is paid in slow, expensive incidents.

6. Cultural alignment

No model replaces shared understanding. The reason a high-performing team executes faster than a low-performing team is not better tools or better KPIs - it is the implicit agreement on what matters, what good looks like, and what gets prioritized when something has to give. That alignment is built through repeated decisions, public conflict, modeling by leaders, and time. AI cannot produce it, and AI deployed on top of a misaligned culture mostly accelerates the misalignment.

The companies who try to use AI as a substitute for cultural work are the ones who will discover, by 2028 or so, that they automated their dysfunction.

The New Problems AI Is Creating

If the permanent problems are the ones AI fails to solve, the new problems are the ones AI is actively producing. Most companies are not measuring these yet. I expect they will be soon, because the costs are starting to show.

1. The configuration tax

Treated in depth in section 04. The short version: the savings AI promised get eaten by the cost of configuring AI to behave. Prompt engineering, eval harnesses, RAG tuning, observability tooling, vendor management, compliance review. The new BI overhead, with different names.

2. AI brain fry and multitasking inflation

This is the most under-discussed cost. In March 2026, BCG researchers writing in Harvard Business Review (Julie Bedard and colleagues, a survey of nearly 1,500 full-time workers) described "AI brain fry" - mental fatigue that sets in when interacting with AI exceeds a person's cognitive capacity, particularly when overseeing several AI tools at once.

The mechanism: once a worker assigns a task to a chatbot, they switch to another project while waiting for output. They check it. They iterate. They switch again. The workflow takes on a new rhythm that itself creates cognitive load. Task volume rises. Multitasking rises. Focused work falls. The early productivity gain can turn into fatigue, turnover and quality problems later.

The BCG study linked the strain to overseeing several AI tools at once, and ActivTrak's data shows multitasking up 12% and focused sessions 9% shorter after AI adoption. There are limits on multitasking and attention that no model removes.

3. The hollowed-out career ladder

SignalFire's 2025 talent report found that new graduates made up 7% of hires at the 15 largest tech companies in 2024, down from 15% before the pandemic, and that those companies cut new-graduate hiring by 25% versus 2023. A Harvard working paper (Lichtinger and Hosseini, 2025), using resume and job-posting data, found that junior employment fell relative to senior employment at firms that adopted generative AI, driven by slower hiring. Interest rates and other factors also weigh on junior hiring, so the share attributable to AI alone is not settled.

The compounding effect arrives later. This is my projection, not a measured result. The cohort currently learning with AI (2024-2026) becomes mid-level engineers in 2027-2029 and seniors in 2029-2032. If that cohort has skill gaps because AI did the foundational work for them, the industry hits a senior engineer shortage just as it needs those engineers most. You can't poach seniors away from other companies because every company has the same pipeline problem. A 2025 Microsoft Research and Carnegie Mellon survey of 319 knowledge workers points to a related risk: higher confidence in AI was associated with less critical thinking effort (self-reported).

The bottom of the career ladder is thinning. In my estimate the bill arrives around 2029, and it isn't possible to backfill quickly.

4. Vendor lock-in 2.0

The previous generation of vendor lock-in was about contracts, integrations, and switching costs. The new generation is about depending on models you cannot audit, made by vendors who can change pricing, behavior, or availability without notice.

If your customer support flow depends on a specific model from a specific vendor, and that vendor deprecates the version you built against, or doubles the price, or shifts the safety profile - you have no leverage. The companies who treat this as a procurement question instead of an architectural question will pay an asymmetric cost in 2027-2028.

5. The new technical debt

In my experience, agentic AI deployments tend to accumulate undocumented prompt sprawl, undocumented retrieval pipelines, undocumented eval logic, and undocumented orchestration. By the time anyone wants to migrate or audit, it will be the same situation that legacy code created in 2005 - except generated faster, by more people, with fewer comments. The new technical debt looks different from the old one but produces the same outcome: nobody understands the system, nobody wants to touch it, and the cost of changing anything grows faster than the cost of leaving it broken.

This is the under-priced bill of the current era. [Original to this essay - I have not seen this framed sharply enough in current coverage, but the pattern is visible in the AI projects I see.]

6. Decision velocity vs. decision quality gap

Faster decisions are not always better decisions. AI compresses the time between "we should decide something" and "a decision has been made" - and that compression often skips the most valuable part of the process, which is the friction that surfaces the disagreement. [Original to this essay.] Most companies treat decision speed as a virtue. In high-stakes contexts, decision speed is a leading indicator of decision quality degradation. The leaders who survive this era will have built deliberate friction back into the process for the decisions that matter.

The Configuration Tax

The most underestimated cost in enterprise AI is the cost of making AI tools actually work in production. I call this the configuration tax. [The framing is original to this essay; the cost pattern is widely reported but rarely named as a single tax.]

The components, as I see them:

• Prompt engineering and versioning - the rapidly maturing discipline of writing, testing, and managing prompts at production scale. The tooling category alone (Langfuse, PromptLayer, Helicone, and others) is now an enterprise spend line.
• Eval harnesses - the test infrastructure for AI outputs. Most teams ship without one and discover quality regressions in customer-facing incidents.
• RAG pipeline tuning - chunking strategy, embedding choice, retrieval thresholds, reranking. Each of these is its own subspecialty now.
• Agent orchestration platforms - the new middleware layer. CrewAI, LangGraph, Autogen, internal frameworks. Often three or four coexisting in the same company because different teams picked differently.
• Observability and monitoring - tracking what the agents actually did, why, with what inputs, at what cost. This category barely existed two years ago and is now table stakes.
• Compliance and policy review - what is the agent allowed to do, what data can it see, how do we prove it complied. Regulated industries are likely to spend more here.
• Integration engineering - connecting agents to existing systems, APIs, data sources, identity, audit trails.
• Vendor management - juggling multiple model providers, pricing volatility, contract terms, fallback behavior, region-specific availability.

By the time an enterprise has deployed AI tools at scale, it has often rebuilt the same operational overhead it was trying to eliminate. The savings AI promised get consumed by the cost of configuring AI to behave. Gartner predicts that over 40% of agentic AI projects will be canceled by the end of 2027 due to escalating costs, unclear business value or inadequate risk controls.

This is the diminishing-return paradox of AI tooling: every additional capability requires additional configuration to use safely; the configuration overhead compounds; eventually the marginal new capability is worth less than the marginal new configuration cost. [Original to this essay.]

The fix is not "stop configuring." The fix is to recognize the configuration tax as a permanent cost line, design for it intentionally, and ruthlessly prune any AI capability that doesn't earn its place. The companies that do this may look slower than the ones that don't, for a quarter or two.

The Next Migration Nobody Wants to Talk About

Here is the part of the conversation that mostly doesn't happen at conferences.

Most companies have not finished the previous migration.

The numbers, drawn from 2025 reporting:

• Only 9% of enterprises say their digital transformation has let them fully retire or replace all legacy applications (Pega and Savanta, survey of 500+ IT decision makers, October 2025).
• The average enterprise in that survey reported losing more than $370 million a year to technical debt. It is a vendor-commissioned survey, so read the figure as directional.
• The U.S. federal government spends more than $100 billion on IT each year, with agencies typically reporting about 80% on operations and maintenance of existing systems (U.S. GAO).
• McKinsey has reported that only about 10% of cloud transformations capture their full value. In my experience, many stall in a hybrid state, pay for both architectures, and rarely consolidate.

Translation: for most enterprises, legacy and cloud migration are unfinished. The hybrid-cloud-with-legacy-underneath architecture is the actual production reality for most enterprises, and it has been for years.

Now those same enterprises are being asked to migrate to an AI-native architecture - on top of cloud migration, on top of legacy modernization. The next migration is stacked on two unfinished migrations underneath it. The compounding cost is enormous, and almost nobody is acknowledging it.

If we want to be honest about what AI-native architecture requires, it includes: clean data pipelines (which require fixing legacy data systems first), event-driven infrastructure (which requires fixing batch-oriented legacy logic), real-time observability (which requires modernizing logging and monitoring), identity-aware access control across services (which requires resolving identity sprawl from cloud migration), and policy-as-code governance (which requires writing the policies the old systems handled implicitly).

Most companies cannot afford to do all three migrations simultaneously, and most cannot afford to leave any of them unfinished either. [Original framing - I have not seen this stacked-migration problem articulated as a single named pattern, but the pattern is visible in the projects I see.]

The honest answer most leadership teams haven't given themselves yet: you may not be able to migrate to AI-native architecture without first resolving the legacy and cloud migrations underneath it. The shortcut path - "we'll just put an LLM in front of the legacy mess" - works for a quarter and then starts producing the cost overruns Gartner is warning about.

The companies that will get through this with a coherent architecture are the ones that name the three migrations explicitly, sequence them, accept the painful trade-offs, and refuse to pretend the bottom layers are done when they're not.

The Data Foundation, Now Doubly Critical

I wrote about the data foundation extensively in The KPI Trap. Most of what was true for KPI architecture is now doubly true for AI architecture, because AI multiplies the cost of bad data instead of merely tolerating it.

The six prerequisites for an AI-capable data foundation:

1. Data modeling that matches the business. Schemas that capture events at the right granularity, entity relationships that reflect how customers actually move through the product. If "active user" is defined three different ways in three tables, no AI built on that data can mean what you think.
2. Instrumentation. Events captured reliably, at the right time, with monitoring for client-side failures, server drops, and schema-change breaks. Most data quality problems are collection problems, not analysis problems - and AI surfaces them at higher volume.
3. Single source of truth. Or an explicit, written map of which system is authoritative for which metric. Without this, every AI-generated report has plausible deniability.
4. Metric definitions written down and shared. A data dictionary that any human or agent can reference. Most metric disputes are definition disputes wearing a measurement costume - and now they're definition disputes that AI agents inherit and propagate.
5. Data ownership and quality monitoring. Clear owners for each metric's data quality, with modern observability tools (Monte Carlo, Bigeye, Anomalo) or simple freshness checks so pipeline failures don't surface in AI-generated outputs.
6. Governance. A light intake process for new data sources and AI uses, plus a regular review where unused or contradictory data gets retired.

Without this foundation, AI delivers inaccurate insights at machine speed. Garbage in, garbage out - but now at the throughput of an LLM.

The companies that win the next five years will not be the ones with the most AI tools. They will be the ones with the cleanest data, the clearest definitions, the most defensible governance, and the discipline to not deploy AI on top of a foundation that can't carry it.

The 2030 Manager - What They'll Actually Need

Harvard Business Review's February 2026 piece "To Thrive in the AI Era, Companies Need Agent Managers" (Suraj Srinivasan and Vivienne Wei) introduced a role that did not formally exist three years ago: the agent manager. The agent manager's job is to orchestrate how AI agents learn, collaborate, perform, and work safely alongside humans - how agents are configured, what data they see, how they hand off to each other and to humans, and how they're governed for safety, compliance, and business alignment.

My read, building on the HBR piece, is that the manager role is shifting from task assignment and status collection (largely automatable) to four capabilities that don't automate well:

1. Sharper problem framing. AI can answer questions; it cannot decide which questions matter. The manager who frames the right question saves a hundred hours of model output downstream.
2. Stronger judgment under uncertainty. When the data is ambiguous, when the model is hedging, when two functions disagree - the manager has to call it. This is the work that doesn't get easier when AI gets better.
3. Orchestrating hybrid human-agent teams. Designing workflows where agents do first-pass work, humans handle nuance and relationships, and the handoffs are explicit and well-governed. As agents multiply, a manager can end up with more AI agents than human direct reports, which changes how work is planned and reviewed.
4. Bridging the executive-manager AI perception gap. Executives often overestimate AI capability based on demos. Managers often underestimate it based on early integration pain. Someone has to translate between them in both directions.

What this means in practice:

• Fewer managers, but the ones who remain need a sharper toolkit. The middle layer of "manager who exists to track status" disappears. The manager who exists to design decisions, frame problems, and orchestrate hybrid teams becomes more valuable, not less.
• The training gap is large and unaddressed. Forrester's AIQ research found that only 23% of organizations offer training in prompt engineering. The companies who treat manager development as the AI strategy will outperform the companies who treat tool procurement as the AI strategy.
• A new role is emerging: the decision quality auditor. [Original to this essay.] Inside the manager function, someone has to be responsible for asking "are we deciding this well, or just quickly?" - the human counterpart to AI's speed. This role doesn't exist on org charts yet. It will, by 2028 in some form, because the cost of low-friction wrong decisions made at AI speed is going to force it.

The 2030 manager looks less like a controller and more like an architect. Less time on assignment, more time on design. Less time watching dashboards (the system watches itself now), more time deciding which questions are worth answering at all.

The Human Problem - When Everyone Multitasks

Behind the productivity-paradox headlines is a quieter, harder problem: the texture of work is changing in ways that exhaust people.

The 2026 ActivTrak analysis found that after AI adoption, email time rose 104%, multitasking rose 12% and the average focused session shortened by 9%. The UC Berkeley Haas study and the BCG brain-fry study point the same way, and the BCG study linked the strain to overseeing several AI tools at once. The number of tools is itself a variable.

Three patterns I'm watching closely:

Role inflation

[Original to this essay - this is my term for a pattern I see across most AI-adopting companies.]

One person is now expected to do the work of three, because AI "augments" them. The product manager runs research, drafts spec, writes test cases, and ships internal communications - all "augmented." The designer prototypes, codes, and runs user research. The engineer also does QA, writes docs, and manages stakeholder updates.

What looks like a productivity gain in the first quarter is a sustained increase in cognitive load that the employee absorbs personally. Reports of burnout tend to show up in the dashboard late. By the time they do, the person is often already half out the door.

The companies who treat AI augmentation as a free productivity multiplier are running a hidden compensation reduction without acknowledging it. The hours go up, the cognitive load goes up, the comp stays flat. The market eventually corrects this. The companies that anticipate the correction will retain the people they actually need.

The skill atrophy problem

Already covered in section 03. Juniors not learning fundamentals, mid-careers depending on AI for work they used to do themselves, seniors who were trained pre-AI becoming disproportionately valuable and disproportionately rare. The pipeline issue resolves itself only if companies deliberately invest in structured under-augmentation for development purposes - giving juniors and mid-careers projects where they have to do the work without AI, specifically to build the underlying skill. Most companies will not do this. The ones that do will own the senior talent in 2029-2032.

The quietly reversing layoffs

A quieter 2026 trend: in a February 2026 Careerminds survey of 600 HR professionals, 35.6% of companies that made AI-led layoffs had rehired more than half of the roles they eliminated, and 52.1% had rehired within six months. About a third reported losing critical skills and expertise, and more than half said AI required more human insight than expected. Orgvue's 2025 survey found that 55% of the leaders who made AI-related redundancies admit they made wrong decisions, and Forrester's 2026 predictions expect half of AI-attributed layoffs to be quietly reversed.

The lesson is not "AI doesn't work." The lesson is that the cost of getting AI adoption wrong is now visible in the data. These are self-reported surveys, not a controlled comparison, but they suggest that moving cautiously has been the more defensible choice.

Why Risk Management Gets Sharper, Not Softer

This is the part of the essay where I declare the bias I've been working from the whole time. I am an operator who became a consultant. Today I work as an interim and fractional product and program manager for tech companies. I help organizations manage what is about to break before it becomes expensive.

The AI era does not reduce the demand for this work. It amplifies it.

Three reasons:

1. AI optimizes the visible. Risk lives in the invisible.

I made this point in section 02 and want to extend it here. LLMs and agents are designed to optimize the metric you defined, the workflow you scoped, the question you asked. They are not designed - and cannot be designed in any general sense - to surface the question you didn't think to ask, the leak you didn't know was there, the second-order effect that emerges from the interaction of two systems that were "working fine" in isolation.

The gap between the optimized surface and the unobserved depth widens as AI gets better at optimizing the surface. Risk lives in that gap. Someone has to be looking at the things the system was not designed to surface - and that someone is not the AI, by structural definition.

2. The speed of failure is now machine speed.

Pre-AI, a bad decision propagated through an organization at human speed. People talked to each other, raised concerns, slowed each other down, sometimes caught the problem before it spread. AI-mediated workflows don't have that natural friction. A bad assumption encoded in a prompt or a policy can propagate across a very large number of decisions in days.

This makes early risk detection disproportionately more valuable than late risk remediation. The 10x rule I've written about elsewhere (The 10x Rule of Organizational Change) may be larger still in AI-mediated systems (my view, not a measured figure). The cost of catching a leak in the design phase versus the production phase has always been large. In AI-mediated environments, it is overwhelming.

3. The skills that surface invisible risk don't get automated.

Leak detection is a pattern-recognition skill built from operator experience - the kind of experience that comes from seeing the same shapes repeat across companies and knowing what questions to ask in week one to surface them in week three instead of month nine. AI can support this work (and I use AI extensively in my own engagements). It cannot replace it, because the pattern recognition required runs across domains AI was not trained to integrate: technical architecture, regulatory exposure, team dynamics, financial pressure, executive incentive structure, and the specific personalities involved.

The structural reason the operator-led risk practice scales in the AI era is that the work itself is the kind of work AI is structurally bad at: integrative, judgment-heavy, dependent on cross-domain pattern recognition, and grounded in lived experience rather than indexed knowledge.

The Infrastructure Trap - Where Companies Fall Short

From my own work and from the public failures I'm watching, three failure modes show up consistently. [The framing of these as "three foundation failures" is original to this essay; the underlying patterns are well documented.]

Failure 1: Configuring AI before mapping flows

The most common failure pattern I see: a team buys an AI tool, deploys it on top of existing workflows, and discovers six weeks later that the AI is automating a process that was already broken. The workflow had three undocumented exception paths that humans were quietly handling. The AI doesn't handle them. The exceptions become incidents. The incidents become rollbacks.

The fix is unglamorous: map the current flow before you automate it. Where does information actually move? Who actually decides? What are the exception paths and who handles them today? What is the implicit knowledge in the heads of the people doing the work? AI deployments succeed when the underlying flow is understood and fail when it isn't.

Failure 2: Buying AI instead of building the foundation

A close cousin of failure 1: the budget went to tools instead of plumbing. The team has Copilot, Claude, ChatGPT Teams, Perplexity Enterprise, Cursor, Glean, Notion AI - and the data underneath all of them is still fragmented, undefined, and ungoverned. The tools work brilliantly in demos and poorly in production, because production requires foundation work that nobody allocated budget for.

The honest sequencing: foundation first, tools second. Most companies do the opposite, because tool adoption is visible and foundation work is invisible. The companies that do this in the right order will look slower for a year and faster forever after that.

Failure 3: Designing for the demo, not the failure mode

Most enterprise AI deployments are designed for the happy path - the flow that works, the user who asks reasonable questions, the data that is clean. They are not designed for what happens when the model is wrong, when the user asks something hostile, when the data is missing, when the upstream service is degraded. The failure modes are where the real cost lives, and most procurement processes don't evaluate them.

Risk-adjusted AI architecture is the discipline of designing for the failure modes first and the happy path second. This sounds expensive. It is, briefly. The companies who skip this step pay much more later, in incidents, regulator attention, and customer trust.

What Comes Next - 7 Predictions for 2027-2030

Predictions are dangerous and worth making. Here are seven I'll defend. These are original to this essay - they are informed bets, not researched facts. Read them as the consultant's framework for what to plan around, not as forecasts you can take to the bank.

Prediction 1: AI-native companies emerge - but they won't be most companies.

A small cohort of companies will rebuild from the ground up around AI-native architecture, data-as-product, and hybrid human-agent teams. They will be fast, lean, and impressive. They will also be fewer than the headlines suggest, because rebuilding from scratch requires conditions (greenfield, capital, talent, leadership willing to accept short-term cost) that most companies don't have. The majority will run a hybrid model for the rest of the decade, and the hybrid will be the actual modal reality.

Prediction 2: Manager headcount drops; manager quality requirements rise.

The 1:8 manager-to-IC ratio common in tech becomes 1:15 or 1:20 as routine management is automated. Total manager headcount drops. But the standards for the remaining managers rise sharply - framework thinking, decision design, hybrid team orchestration, the ability to bridge executive expectations and ground-floor reality. The companies that cut managers without raising the bar for the ones who remain will discover that they cut the wrong people.

Prediction 3: Data foundation becomes the #1 hiring priority by 2028.

The role currently called "data engineer" or "analytics engineer" or "platform engineer" will become the most-recruited role in the next two years, surpassing AI/ML engineering. Companies will realize that the bottleneck was never the model and was always the data underneath it. Compensation for senior data foundation talent will rise sharply because supply is constrained and demand has shifted.

Prediction 4: The "fractional everything" model expands.

As manager headcount drops and the skill requirements rise, more companies will buy senior talent on a fractional basis rather than full-time. Fractional CTOs, fractional COOs, fractional risk officers, fractional data leaders. The full-time C-suite of 2020 becomes a leaner core with fractional expertise around it. This is not a threat to operators-turned-consultants; it is a thesis I'm betting my practice on.

Prediction 5: Senior engineering shortage hits hard in 2029-2032.

This is the most confident prediction because it is mostly arithmetic. The pipeline through juniors and mid-careers is hollowing out now. The companies who didn't backfill that pipeline with deliberate development programs will be competing for an inadequate supply of seniors. Compensation will spike. The companies that survive will be the ones who invested in development programs no model can replace.

Prediction 6: "Decision quality" becomes a measurable function.

Some company - probably in 2027 or 2028 - will publish a credible methodology for measuring decision quality independent of outcome. Once one company does it well, others will copy. By 2030, "decision quality" will be a standing function inside larger companies, alongside finance and people operations. This is the role I called decision quality auditor in section 07, and it is the natural complement to AI's speed advantage.

Prediction 7: Risk management eats the org chart.

This is the most self-interested prediction in the list and I want to be transparent about that. Risk management - in the broad sense I use it - becomes a horizontal function that touches every department, not a vertical function that lives in legal and compliance. The CRO becomes a more powerful role than the CFO in some industries. Risk literacy becomes a required executive skill. The companies that build this in early will look paranoid for a quarter and resilient for a decade.

The Work That Doesn't Disappear

If there is one argument this essay makes, it is this: the structural work doesn't go away when AI arrives. It gets sharper, more expensive, and harder to skip.

The data foundation work that companies were avoiding in 2022 is now blocking their AI deployments in 2026. The cross-functional alignment work that looked optional in 2020 is now amplifying every silo's local optimization into a global misalignment. The decision-quality work that nobody was measuring in 2018 is now the bottleneck on every leadership team trying to keep up with the velocity of AI-mediated business.

The work is the same work. The cost of skipping it is higher.

For the companies I'm building with right now, the message is simple:

• The model is not the bottleneck. The organization underneath the model is the bottleneck. Always was. More visibly now.
• The foundation is not optional. Data, decisions, design, culture. Skipping any of them now produces incidents faster than it used to.
• The configuration tax is permanent. Plan for it, manage it, prune it. It will not go away.
• The next migration is stacked. Legacy underneath cloud underneath AI. Name them. Sequence them. Don't pretend the bottom layers are done when they're not.
• Risk management is the discipline that doesn't get easier. AI amplifies what you can see. Someone has to be responsible for what you can't.

The AI hype cycle is settling. The structural problems are getting sharper. The companies that fix the foundation underneath the AI - not the ones that adopt the AI fastest - will own the next decade.

The framework is unchanged. The cost of skipping it is higher. The shape of the work is the same shape it has always been: see what is invisible, manage the risk before it becomes expensive, build systems that bend without breaking. AI is the accelerant. It is not the answer.

The Economics - What Skipping the Foundation Now Costs

Most of this essay has been structural. The economic argument is worth making separately, because most boards reading this will want to know: what does delaying foundation work cost us in dollar terms?

The cost of building AI on a broken foundation

• Stalled pilots. MIT NANDA found that 95% of enterprise generative AI pilots delivered no measurable P&L impact. Every pilot that stalls is tooling spend and engineering time that returns nothing, before counting opportunity cost.
• The configuration tax. An AI ops stack (prompt management, evaluation, observability, agent orchestration, integration engineering) is a recurring cost in tooling and engineering time. Gartner attributes its forecast of agentic project cancellations partly to escalating costs. If the underlying data foundation cannot support the agents, that ongoing spend produces little.
• The reversed layoff cost. In the Careerminds survey, as reported by trade press, about 31% of organizations said rehiring cost more than the layoffs had saved, and another 42% said the two roughly cancelled out. Replacing an employee is commonly estimated at 0.5x to 2x annual salary (Gallup).
• Skill atrophy debt. The current 2024-2026 cohort of juniors learning with AI hits mid-level in 2027-2029. If their development was outsourced to the model rather than built in fundamentals, the senior shortage in 2029-2032 will be both expensive (compensation premium) and structural (you can't poach from competitors who have the same problem). Companies that under-invest in junior development now pay the bill in a window they cannot accelerate out of.

What foundation work returns

• AI projects that actually ship. Gartner's forecast is specifically about AI projects without AI-ready data, which makes the foundation the first thing to fix. The investment is mostly front-loaded.
• Configuration tax stays bounded. A clean foundation means the ops layer is smaller, simpler, and survives model upgrades without rework. The companies who treat AI infrastructure as a permanent line item save more by keeping that line item small than by chasing the latest tooling.
• Talent retention. Senior operators stay at companies whose AI strategy is grounded in real work, not slide decks. Retention is an additional benefit.
• Strategic optionality. A company that has done the foundation work can choose its AI vendors, switch them, or self-host. A company that hasn't done the foundation work is locked in to whichever provider it integrated first, with switching costs that compound every quarter.

The size of the investment and the timing of the payback depend on the organization, and I have no verified benchmark for either. The sequencing argument does not depend on them: fixing the foundation first avoids paying for stalled pilots twice.

Where to Start - A 90-Day Foundation-First Roadmap

The mistake most companies make is starting AI work with model selection. The companies whose AI actually ships start with data and decision architecture. Here is the sequence I recommend.

Days 1-30: Audit the foundation, not the AI

1. Inventory your data architecture honestly. Which entities exist as canonical records? Where are the duplicates? Which dimensions have shifted definitions in the last 18 months without versioning? If your data foundation cannot answer "who is our customer, definitively," your AI agents cannot either.
2. Map your existing measurement stack against the three layers (Outcome / Execution / Foundation, from The KPI Trap). Find the layer you are skipping. It is almost always Foundation.
3. List every current AI initiative. For each one, ask: does it depend on data the foundation can support today? If no, this initiative is on the wrong side of the sequence and should pause until the foundation underneath it is ready.
4. Document the change-detection infrastructure. If your engineering team can't answer "what shipped in the last 7 days, what was the linked ticket, what was the impact" in 30 seconds, the AI agents you build won't be able to either.

Deliverable at day 30: a single-page diagnostic of where the foundation is, where the AI initiatives sit relative to it, and which initiatives need to be paused or re-sequenced.

Days 31-60: Fix the bottom layers

1. Pick the two most blocking foundation gaps from the audit. Usually one is data (canonical entity, source of truth, definitions) and one is decision architecture (who owns what, how decisions get documented).
2. Build the change-log infrastructure. Engineering changes, ticket closures, deploys - automatically captured into a single source of truth that downstream AI agents (and humans) can query. The Stage 0 pattern from Ship KPIs Like Features applies directly.
3. Pause the AI work that depends on the broken layers. This is the most political moment. Hold the line. The pilots that resume after the foundation is fixed will be the ones that actually ship.
4. Start the configuration tax discipline. Hard caps on AI tooling spend. Quarterly review of which AI tools are actually delivering. Vendor consolidation where redundant.

Deliverable at day 60: a working change-log feeding at least one downstream system, two foundation gaps materially closed, and an honest list of AI initiatives that need to be re-sequenced.

Days 61-90: Resume AI with the foundation underneath

1. Restart the top one or two AI initiatives with the corrected foundation. Verify that the data, change-log, and decision architecture are ready before each restart.
2. Adopt the agent-driven implementation pattern from the Ship KPIs companion piece. Six agents (one infrastructure, five lifecycle). The framework now stops being a slide and starts being operational.
3. Establish the cadence. Quarterly foundation review. Weekly portfolio review. The discipline survives leadership turnover only if it is calendared, documented, and explicitly assigned to a Governance Owner.
4. Communicate the framework to the leadership team. Not as theory. As the operating discipline for how this company will adopt AI going forward.

Deliverable at day 90: AI initiatives that resume on a foundation that supports them, a measurement architecture that catches problems early, and a leadership team that understands why the sequence matters.

Why this is the right sequence

The alternatives I have seen tried - "ship AI first, fix foundation later," "do both in parallel," "let the AI tooling force the foundation work" - tend to share one failure mode. The AI initiatives consume the bandwidth that the foundation work needed. The foundation never gets fixed. The AI initiatives slowly degrade. The team blames the model when the model was never the issue.

Foundation-first is the slower path, and in my experience the one that finishes.

This Is the Work I Do

I work as an interim and fractional product and program manager for tech companies. Work in this area often begins the same way: a leadership team that has invested in AI tooling, is not seeing the return, and cannot diagnose where the friction lives.

The pattern I expect to find is consistent. The data foundation is incomplete. The change-detection infrastructure does not exist. The decision architecture is informal. The AI tooling is fine. The organization underneath the AI tooling is not ready to absorb it.

What you get is a written assessment of the foundation, a sequence for AI initiatives, and a measurement design intended to catch problems earlier than a board meeting nine months later.

If this essay surfaced uncomfortable questions about your own AI strategy - especially the gap between the pilot metrics and the production reality - those are exactly the questions worth a conversation. Here is how to start.

Related Reading

• When your measurement stack becomes the risk - the article on why measurement stacks become the risk they were supposed to monitor. The data foundation argument runs in both directions.
• The 10x Rule of Organizational Change - why every risk caught late costs 10x more than the same risk caught early. The economics behind why risk management gets sharper in the AI era.
• The Enterprise AI Transformation Operator Playbook - the practical companion piece. How to actually run an AI transformation that produces results.
• The Operator-Consultant Method - the four-phase methodology behind every engagement.

Sources

• MIT NANDA, The GenAI Divide: State of AI in Business 2025, July 2025. Preliminary report, not peer reviewed (300+ public deployments, 50+ interviews, survey of 153 leaders). 95% of organizations saw no measurable P&L impact from generative AI pilots.
• Gartner, "Lack of AI-Ready Data Puts AI Projects at Risk", 26 February 2025. 60% of AI projects unsupported by AI-ready data abandoned through 2026 (prediction); 63% of 248 data management leaders lack or are unsure of the right practices (Q3 2024 survey).
• Gartner, "Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027", 25 June 2025.
• Srinivasan, S. and Wei, V., "To Thrive in the AI Era, Companies Need Agent Managers", Harvard Business Review, 12 February 2026.
• Ranganathan, A. and Ye, X. M., UC Berkeley Haas, AI work-intensification study, published in Harvard Business Review, February 2026. Eight months, roughly 200 employees at a US technology company.
• Bedard, J. et al., "When Using AI Leads to 'Brain Fry'", Harvard Business Review (BCG), 5 March 2026.
• ActivTrak Productivity Lab, 2026 State of the Workplace, 11 March 2026. 443 million hours, 163,638 employees. Vendor research, directional.
• Orgvue, "55% of businesses admit wrong decisions in making employees redundant when bringing AI into the workforce", 2025. Survey of 1,163 senior leaders (vendor survey, directional).
• Careerminds survey of 600 HR professionals, February 2026, as reported by Human Resources Director. Vendor survey, directional.
• Forrester, "Predictions 2026: AI Moves From Hype To Hard Hat Work", and Forrester's 2026 future-of-work predictions on AI-attributed layoffs being quietly reversed.
• Forrester, "Your Employees Aren't Ready For AI - And It's A Problem" (AIQ research); the 23% prompt-engineering training figure as reported by ITBrief.
• SignalFire, State of Tech Talent Report 2025. New graduates were 7% of hires at Big Tech (top 15 by market cap) in 2024; new-graduate hiring down 25% on 2023.
• Hosseini Maasoum, S. M. and Lichtinger, G., "Generative AI as Seniority-Biased Technological Change", SSRN working paper, 2025.
• Lee, H.-P. et al., "The Impact of Generative AI on Critical Thinking", Microsoft Research and Carnegie Mellon University, CHI 2025. Survey of 319 knowledge workers; self-reported.
• Pegasystems, "Average Global Enterprise Wastes More Than $370 Million Every Year Through Technical Debt", 14 October 2025. Survey by Savanta of 500+ IT decision makers (vendor-commissioned, directional).
• U.S. Government Accountability Office, GAO-25-107795. Federal IT spending above $100 billion a year, typically about 80% on operations and maintenance.
• McKinsey & Company, "Unlocking cloud value: Achieving operational excellence through SRE". About 10 percent of cloud transformations achieve full value.
• Conway, M. E., "How Do Committees Invent?", Datamation, 1968.
• Sections marked "Original to this essay", the seven predictions, and the 90-day roadmap are my own analysis and are not drawn from a study.