.svg)
.jpg)
Written by:
Nick Polachek - EVP, Head of Strategy and AI Transformation, Launch Consulting Group
Jason Talley - Head of Delivery, Launch Consulting Group
Jess Lampe - Global Lead Technologist, Launch Consulting Group
Buy the technology, train the users, and capture productivity. That rule worked for ERP, for cloud, for collaboration software. It is not working for AI.
Two years into the largest technology investment cycle in enterprise history, most companies cannot point to the money. Adoption is everywhere. Returns are not. The tools install, the licenses ship, the pilots demo well, and the P&L does not move.
The spend is real. The usage is real. The return is missing.
The uncomfortable part: the playbook that worked for every prior technology is part of why AI is not paying off.
Three things have changed in the last twelve months.
Spending outgrew patience. Worldwide AI spending will reach $2.5 trillion in 2026.1 At that scale, “we are still early” no longer answers “what did we get.”
The failure rate got measured. Roughly 95 percent of generative AI pilots produce no measurable P&L impact, against thirty to forty billion dollars spent.2 About six percent of companies attribute more than five percent of EBIT to AI.3 The distance between those two numbers is the story.
The market started canceling. Gartner expects more than forty percent of agentic AI projects to be canceled by the end of 2027, on cost, unclear value, and weak controls.4 The reckoning is not a forecast. It has started.
These are facts about outcomes, not the technology. The question is why most companies cannot convert AI into value, and what the few who do are doing differently.
The usual explanations do not hold. Not model quality; everyone has the same models. Not underinvestment; the laggards spend more. The failure is older than the technology. Companies run AI as a technology program when it is a business strategy decision.
Most have the pieces: a business strategy, an AI strategy, and a technology roadmap. They are connected backwards. Technology leads; the AI strategy is a list of tools, and neither trace to what the business is trying to win. So, the spend never reaches the P&L. The CEOs pulling ahead have taken direct ownership, because the choices are strategic, not technical.5
Exhibit 1: Business strategy sets the direction; AI and technology serve it. Most companies run it upside down. Source: Launch analysis.
This also reconciles the contradictory headlines. Studies that measure enterprise EBIT find near-universal failure.6 The study that measures productivity among firms already tracking returns finds 74 percent positive.7 The vendors see record revenue. All are right in their own frame. AI creates local value today; the six percent have connected it to the P&L. The rest pay for value they cannot bank.
AI is a business strategy decision, not a technology initiative. The ROI gap closes only when business strategy, AI strategy, and technology strategy point the same direction, built in that order: the business sets the outcomes, the AI strategy names where AI changes how the company competes, and the technology of platform, data, and software is built to serve that, not to lead it.
Aligning business strategy, AI strategy, and technology strategy is not the framework itself. It is the prerequisite. Once those three are pointing in the same direction, organizations still need to pass six execution gates before AI investment becomes measurable business value.
The companies that convert AI spend into outcomes do the same things in the same order. Six disciplines, each a gate. Clear all six and the investment reaches the P&L. Skip one and it joins the 95 percent.
Exhibit 2: The ROI gap, framed as six sequential disciplines. Source: Launch analysis.
These are not stages to admire from a distance. They are decisions a leadership team makes, or fails to make, in the next two quarters.
The fastest way to waste an AI budget is to spend it on everything.
The 95 percent run portfolios. The six percent run a shortlist. Spread thin, no use case gets the redesign, data, and adoption it needs to reach production. Breadth feels like progress. It is the opposite.
Triage on three axes. Value: is the outcome worth winning. Feasibility: can the data and systems support it now, not after an eighteen-month program. Fit: is this an AI problem at all, or a process, incentive, or data problem in disguise.
Exhibit 3: Triage by value and feasibility, after filtering for genuine AI fit. Source: Launch analysis.
What the disciplined few do
They concentrate. Two or three problems that are high value, feasible, and genuinely AI get the full weight of the organization. The rest are parked or killed, on purpose. They favor specialized use cases over broad rollouts, because tightly scoped work returns more reliably. Saying no is the discipline.
Why concentration beats coverage
One use case in production beats ten stalled pilots. Concentrated investment is what gets one across the line.
Redesign the work around agents. Never automate a broken process.
The companies pulling ahead stopped buying tools and started rebuilding work. They redesign the workflow around what an agent can do. The test: remove the AI, and the workflow no longer makes sense, because it was built for the agent.
Never automate a broken process. Point an agent at a slow, manual, exception-ridden workflow and you scale the dysfunction and add running cost. Fix the process first.
Where the value comes from: about ten percent from algorithms, twenty percent from data and technology, seventy percent from people and process.8 Most firms spend the inverse.
Exhibit 4: Roughly 70 percent of AI value comes from people and process; most enterprises spend the inverse. Value split per BCG; spend split illustrative.
What the disciplined few do
They invert the spend and treat the workflow, not the tool, as the unit of design: which steps go away, which go to the agent, which a person still owns. The result looks strange to anyone who ran the old process, because it was built for the agent. The deeper playbook is in our companion paper, AI as a Value Project.
Why redesign is the work
The same model returns nothing on an unchanged process and real money on a redesigned one. The seventy percent is not overhead. It is the project.
Adoption is the work, not what happens after it.
Most treat adoption as a downstream step: launch, send the training link, move on. Then usage is shallow, people route around the new process, and shadow tools fill the gap. That is value leaking from a sound use case.
Design the human-agent model deliberately. The Director sets intent and constraints. The Verifier owns and approves the output. The Transformer feeds results back so the next cycle starts smarter. The agent executes in between. People move up to directing, verifying, and improving.
Exhibit 5: People direct, verify, and improve; the agent owns the execution between the checks. Source: Launch analysis.
What the disciplined few do
They fund adoption inside the use case. Incentives make the new way the easy way. Training is built on the redesigned workflow, not the tool. Roles are named. They watch real usage, not seat counts.
Why adoption is where value lands
A workflow no one runs is a slide, not a return. The impact shows up only when the people change how they work.
Plan the platform. Fund the platform. Do not wait for it.
Every enterprise needs a real platform under its AI: cloud, modernized data and software, a secure agent platform. It deserves a plan and a budget. The mistake is sequence. Put the platform first, fund an eighteen-month program, and the budget is gone before the business sees anything.
The excuse hides in a reasonable sentence: our data is not ready. It is the most expensive line in enterprise AI. You do not need all your data clean. You need the slice one use case depends on clean enough to trust. The firms waiting to fix everything are still waiting when competitors ship.
Exhibit 6: Plan and fund the platform, but sequence it behind the use cases so it never blocks the first ones. Source: Launch analysis.
What the disciplined few do
Use cases first, platform as you scale, software always. The first use cases ship on just enough foundation; each hardens it for the next. Modernization still happens, planned and funded, but pulled forward by real demand, not built on speculation.
Why the sequence protects the return
A platform built ahead of the use cases bets they will arrive. One built behind them is paid for by value already in hand.
Built in from day one, it speeds you up. Bolted on later, it stops you.
Most treat governance as a gate to clear before launch. That is why it reads as a tax. It is an operating model: owners, policies, and controls that travel with the work. The fastest movers design it in from the first sprint, so speed and trust compound instead of trading off.
Bolt it on later and the cost is real. Agents drift, cost creeps, quality degrades unwatched, and one incident stalls everything. More often the program simply freezes, because no one can vouch for what the system does. As agents take on more autonomous work, partial coverage stops being partial protection.
Exhibit 7: Bolted-on governance stalls the program; built-in governance releases the next round of funding. Source: Launch analysis.
What the disciplined few do
Guardrails, observability, and cost controls from the start. Ownership clear, outputs auditable, spend attributable to the agent and use case, so a misconfiguration shows up in hours. The full structure is in our companion paper, The Six Domains of AI Governance.
Why governance unlocks the next dollar
Funding follows confidence. The discipline that lets finance approve the next phase is the one that keeps cost and quality from drifting. Governance is not the drag. Its absence is.
Ship frequently. Measure to the P&L. Reinvest the proof.
The clearest difference between the six percent and everyone else is measurement. Most never instrument for it. They launch on a productivity promise, track usage and hours saved, and find at review time that none of it rolls up to a number a CFO will defend. Projects die for lack of evidence, not lack of impact.
The fix is financial, not technical. Baseline the current cost before you build: labor at loaded rates, cycle time, error rate, rework. Set three to five auditable KPIs. State the expected gain. Net it against the fully loaded cost, not just the license. Ship in short cycles, measure against the baseline, and reinvest the proven savings.
Frequency beats the big bang. A multi-year program defers proof to the moment patience runs out. Small, governed increments produce evidence early and compound it.
Exhibit 8: Frequent shipping banks value early and compounds it; the big-bang program defers all proof. Launch illustrative example.
What the disciplined few do
Measurement is a precondition of funding, not a closing report. No baseline, no build. Short, outcome-based renewal cycles over multi-year commitments. And they separate signal from vanity: seats and prompts are inputs; the P&L is the return.
Why the measurement is the moat
Vendor revenue and desk-level productivity are both real. Neither reaches the financials until someone baselines, ships, and books the result. That last step is the difference between using AI and profiting from it.
This is the business’s problem to own
The gap is not a technology problem to delegate. It is an operating problem the business owns. Five shifts follow.
· Own the outcome, not the rollout
Put a business executive, not technology, on the hook foreach investment’s result, with a P&L line. If no one will sign for theoutcome, it is not ready to fund.
· Fund the seventy percent, not the ten
Reweight the budget from licenses and platforms toward theredesign that produces the return. No large line for process redesign means thebudget is set up to fail.
· Make adoption a line item
Budget training, incentives, and role redesign inside theuse case. Track usage, not seats. A capability no one uses is a cost.
· Plan the platform, sequence it behind the usecases
Keep the modernization plan and budget. Refuse to let itgate the first use cases. Never accept “the data is not ready” as a reason toship nothing.
· Put a number on it first
Require a cost baseline and three to five KPIs to fund anyuse case. The disciplines are ordered; skipping to scale is how a fundedprogram ends with nothing. The retrofit is two quarters of work, minimum.
Five moves, none requiring a new platform or a budget cycle.
1. Pick one. Cut the portfolio into a single use case that is high value, feasible now, and genuinely an AI problem. If you cannot name it, that is the first problem to solve.
2. Name a business owner. An executive, not a technology lead, who owns the P&L outcome. If no one takes it, the use case is not ready.
3. Baseline the cost before you build. Current-state labor, cycle time, error rate, and rework, in dollars. If you cannot measure it manually, you cannot prove AI improved it.
4. Redesign the workflow and design adoption in. Rebuild the process around the agent and name who directs and verifies it. If removing the AI leaves the old process intact, you are automating the wrong thing.
5. Govern and measure from sprint one. Stand up cost attribution, output checks, and three to five KPIs before the first agent goes live. Fund the platform in parallel, but do not let it gate this use case.
The diagnostic is this week’s deliverable. The program is the year’s.
Nick Polachek is EVP, Head of Strategy and AI Transformation at Launch Consulting Group, where he advises executives reshaping their organizations around AI.
Jason Talley is Head of Delivery at Launch Consulting Group, where he leads the teams that take enterprise AI from pilot to production.
Jess Lampe is Global Lead Technologist at Launch Consulting Group, where he guides the architecture and engineering that bring enterprise AI to production at scale.
