Most companies ship AI and move on. But deployed AI is not a project — it's a decaying asset. This report measures how fast its value falls, why, and the operating model that makes it compound instead.
The industry has settled its verdict on the first wave of enterprise AI: it mostly failed. MIT's landmark 2025 study put a number on it — 95% of generative-AI pilots produced no measurable business impact.1 McKinsey found that while 78% of organizations now use AI somewhere, only 6% capture enterprise-level returns.2 Gartner expects more than 40% of agentic projects to be cancelled by 2027.3
Every one of those numbers describes a starting failure — the pilot that never crossed into production. That story is now well told, and it hides the more expensive one.
The systems that do reach production are quietly losing value from the day they launch. A model that answered 88% of tickets in month one answers 71% by month five as language, edge cases, and upstream data drift. An agent praised in its demo accretes silent failure modes no one is watching. A forecasting tool slides out of the workflow it was built for the moment the team reorganizes around it. None of this shows up as a cancelled project. It shows up as a number that was real once and isn't anymore.
We call the governing quantity the Adoption Half-Life: the time it takes a deployed AI system to lose half the value it delivered at launch, absent deliberate intervention. It is the single most useful — and least measured — number in enterprise AI. This report defines it, explains the physics behind it, gives you five archetypes to locate yourself, four chapters written for the four people who own the problem, and the operating model that turns decay into compounding.
The question that separated winners from losers in 2025 was not "can we build it?" By 2026 it's "can we keep it working — and growing — after everyone stops watching?"
— The thesis of this report
AI value doesn't fall randomly. It leaks through four specific gaps between a working model and durable business impact. Name them, and you can instrument them.
The distance between a system that works in the demo and one that works at the operator's desk. MIT traced most pilot failure not to model quality but to brittle workflows and misalignment with day-to-day operations.1 A tool the workflow has to bend around is abandoned the first busy week.
Models, data, and the world move apart after launch. An agent's accuracy erodes as upstream schemas change and base models are updated beneath it. Without a persistent audit trail, drift goes unnoticed until it causes a visible failure.3 This is the purest form of half-life: silent, compounding, invisible on any launch dashboard.
When the build team leaves and no one owns the last mile, the system enters an ownerless middle: too live to ignore, too unowned to improve. Gartner attributes cancellations to management and governance, not model capability.3 The half-life of an unowned system is the shortest of all.
As autonomy rises, unaudited agents accrue a new liability class — call it agentic debt. Escalating cost, unclear value, and inadequate risk controls are Gartner's top three cancellation drivers.3 Every ungoverned agent shortens the half-life of the whole portfolio, not just itself.
A number is only useful if you can compute it. Here is how Applore instruments half-life across live engagements — and how you can start on Monday.
Definition. Pick the one metric the system exists to move — resolved tickets, cycle-time saved, conversion, decisions automated. Measure it at launch (V₀). Track it monthly with no improvement work applied. The Adoption Half-Life (t½) is the number of months at which the value delivered falls to V₀ ⁄ 2.
Across Applore's enterprise engagements the pattern is consistent, and it runs opposite to the market. Systems built with three disciplines — an accountable adoption owner, a workflow redesigned around the operator, and a partnership that stays past handover — don't merely resist decay. Their value compounds. Every named engagement below remains in active use at the 12-month mark, with measurable impact that grew rather than faded. In half-life terms, these systems have a negative half-life: they are worth more a year in than they were at launch.
Drawn from Applore's live, named enterprise engagements (delivery-team reported). The contrast with the market benchmarks above is the entire point.
Delivered, published engagement outcomes across Applore's portfolio — the raw material the half-life model is calibrated against.
Buying from specialized partners succeeds about 67% of the time. Internal-only builds succeed roughly one-third as often.
— MIT, The GenAI Divide, 2025 1 · the clearest argument yet for a specialist operating partner
Every organization sits at one of five stages, distinguished not by how much AI they've built, but by the half-life of what they run. Find yours.
Impressive demos, nothing in production. Value is entirely anticipatory. The organization mistakes motion for adoption.
Dozens of pilots, none owned past launch. This is where 95% of the market sits — and where the Desk and Ownership gaps do their damage.
A few real systems in production, each a silo. Value is real but fragile — no shared evals, no drift monitoring, no portfolio view.
Shared infrastructure — evals, observability, governance. Decay is slowed and monitored. Most of McKinsey's returns concentrate here.
Systems improve in production. Feedback, adoption, and data flywheels mean value grows post-launch. The 6% live here.2
The gap between Stage 02 and Stage 05 is not a technology gap — every archetype has access to the same models. It is an operating gap. McKinsey's high performers are 3.6× more likely to pursue transformational change and 55% rework their workflows around AI rather than bolting it on.2 Movement up this ladder is the entire game.
Applore's work with JK Tyre & Industries — a manufacturer operating across 105 countries — began with a single application and compounded into four production systems, each still in active use and still expanding at the 12-month mark.
It is the Compounder archetype made concrete: value that grows after launch because someone owns each system, the workflow was rebuilt around the operator, and the partnership stayed engaged well past handover. The opposite of a pilot declared "done" at go-live.
The decay is one phenomenon, but the lever you hold depends on your chair. Read your own — then read the one above and below you.
You can list every AI system in the enterprise. You cannot say which are gaining value and which are quietly rotting — because success was declared at launch and never revisited. Meanwhile shadow AI multiplies the ungoverned surface, and the board wants an ROI story you can't yet substantiate.
Counting deployments as wins. A portfolio of 40 live systems with a 3-month half-life is worth less than 6 systems that compound — but your dashboard shows 40.
Institute a quarterly half-life review across the portfolio. Fund the operating model, not just the build. Govern agent sprawl before it becomes agentic debt you can't price.
Every system you ship without an evaluation harness, an audit trail, and drift monitoring is a system whose half-life you've chosen not to see. The build-vs-buy decision is really a half-life decision: MIT found internal-only builds succeed a third as often as specialist partnerships,1 largely because internal teams under-invest in exactly the post-launch machinery that fights decay.
Treating evals and observability as "phase two." Phase two never gets funded, and the first drift failure arrives as an incident, not an alert.
Make the eval harness and drift monitor part of the definition of done. Buy the plumbing, build the differentiation. Measure how long drift lives before you catch it — then drive it down.
Adoption is not a launch event; it's a maintained state. The Desk Gap closes only when someone on your team owns the workflow fit, the retraining loop, and the daily reality of the operators using the thing. The single highest-leverage org move of 2026 is naming that owner — the role most enterprises simply don't have.
Declaring victory at "it's in production." Production is where the half-life clock starts, not where the work ends.
Assign a named adoption owner per system. Instrument real usage from day one. Run a monthly workflow-fit review with the actual operators, and feed it back into the roadmap.
For an early or scaling company, capital efficiency is the strategy. A system with a short half-life is a recurring re-build cost disguised as a launch. The winning pattern is rarely "hire a big AI team early" — it's to partner to ship the first compounding systems fast, transfer the capability, and put the moat where your data flywheel actually is.
Building AI infrastructure that isn't your differentiation, then spending your best engineers maintaining its decay instead of shipping product.
Buy the commodity, build the flywheel. Choose partners who stay accountable to adoption, not just delivery. Make "value retained per dollar" a line your board sees.
Five practices separate the Compounders from everyone else. They are not technology choices. They are operating disciplines — and they are learnable.
Start from the operator's real workflow and constraints, not a showcase scenario. Ship the version that survives a busy Tuesday. This is where the Desk Gap closes and where most half-life is won or lost.
Ship the usage telemetry with the system, never after. If you can't see whether the value curve is rising or falling this month, you are flying blind on your most expensive assets.
Every live system gets one accountable human for its last mile — workflow fit, retraining, drift response. The Ownership Gap is closed by an org-chart decision, not a model upgrade.
Treat every agent as a liability that must be evaluated, audited, and paid down. Eval harnesses and audit trails are the interest payments that keep agentic debt from compounding into a cancelled program.
The decay begins exactly when the launch team leaves. Contract and staff for the adoption phase deliberately — the months after go-live are where compounding is manufactured, or forfeited.
Where the half-life lens says the market goes next.
"Adoption owner" becomes a real title. The role that closes the Ownership Gap moves from improvised to org-chart, the way "SRE" did a decade ago.
Agentic debt gets a balance sheet. As Gartner's 40% cancellations land, boards start demanding a register of live agents, their evals, and their risk exposure.3
Buy-vs-build swings decisively toward specialist partners for non-differentiating systems, tracking MIT's 67%-vs-one-third success gap.1
Evals become procurement table stakes. "How will we detect drift?" moves from a phase-two afterthought to a question asked before the contract is signed.
The high-performer gap widens, not narrows. Compounders pull away as their flywheels turn; the 6% becomes a more valuable — and more exclusive — club.2
Workflow redesign beats model choice. With frontier models commoditizing, the 55% who rework workflows capture the returns the model-shoppers don't.2
Half-life enters the vocabulary. "What's the half-life on that?" becomes a normal question in AI steering reviews — because the alternative is discovering decay as an incident.
Five questions. An instant read on your archetype and where your value is leaking.
Answer honestly about your most important live AI system. This is directional, not an audit — but it points precisely at which gap to close first.
This report combines two evidence bases. Third-party research — cited throughout — establishes the market backdrop: MIT's GenAI Divide (300+ deployments, 150 executive interviews, 350 employees surveyed),1 McKinsey's State of AI 2025,2 and Gartner's 2025 agentic-AI poll of 3,400+ organizations.3 Applore's proprietary data — the half-life readings — is drawn from our live enterprise engagements and is the report's original contribution.
Applore runs a half-life diagnostic across your live AI systems — where value is leaking, which gap to close first, and what a compounding operating model would return.
Request a half-life diagnostic Read the insights