40%+ of Agentic AI Projects: Dead by End of 2027

• Cost of Context killing enterprise AI - and creating the next great M&A opportunity
• If you recognize your company in the description below, the window to act - as buyer or as seller - is open now

Palo Alto, August 2026 - It's a CIO's nightmare. Your CEO is demanding AI across the enterprise, all 370 applications of it. So that's what you're implementing. Then one day an agent misreads a column called rev_7_cust_ob, approves a discount it should have escalated, and the mistake surfaces in a customer's renewal. Versions of this are playing out everywhere. That's why - you might have noticed - "context" was the hottest topic at every conference this year.

At WCP, we advise on M&A and capital raises across enterprise software and AI, which gives us the vantage point to see which companies acquirers pursue, what they pay, and why. An unmistakable theme has recently emerged among strategic acquirers and financial sponsors - value in enterprise AI is migrating toward the layer that makes context targeted, accessible, governed, and durable.

In mid-2026, acquirers are looking for semantic and retrieval infrastructure, the integrators and managed service firms that build and run it, and the software platforms adaptive enough to expose their context cleanly. Those companies are commanding higher multiples. If you run one of the companies this migration favors - for example, a SaaS platform holding an ocean of operational data, an MSP or systems integrator, or a builder of the context layer itself - this piece is addressed to you.

The Bill Comes Due
According to MIT (The GenAI Divide: State of AI in Business 2025), "Despite $30-40 billion in enterprise investment into GenAI…95% of organizations are getting zero return" and "just 5% of integrated AI pilots are extracting millions in value, while the vast majority remain stuck with no measurable P&L impact."

According to Gartner (June 25, 2025), "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." Meanwhile, DoiT/Sapio Research (Feb 2026, n=500 finance leaders) found that "79% experienced cost overruns in the past 12 months."

At WCP, our belief is that the cancellations and the cost overruns are part of the same story: the Cost of Context. The Cost of Context is the full operational, technical, and economic burden of gathering, permissioning, interpreting, and reconciling enterprise information so a machine can act on it. The Cost of Context is high, and it is not worth paying in every case. For example, it may clear the bar for high-value professional work (e.g. - closing the quarter accurately and on time), but it utterly fails for low value work (e.g. - surveying employees on which cafeteria coffee they prefer).

Which is exactly why bringing the Cost of Context down matters: every dollar taken out of the Cost of Context expands the set of tasks AI can profitably touch. The Cost of Context is an unpriced line item in enterprise AI today, and it is where the next several years of value creation, and M&A, will concentrate. Bringing Cost of Context down is how AI adoption turns into return on investment, and the companies that bring it down are the ones already commanding premiums from acquirers.

When those agentic projects are canceled, the demand behind them will not evaporate: the budgets, use cases, and board mandates will migrate to whoever can deliver working AI. That migration - thousands of stranded projects, teams, and point solutions finding new homes at once - is the key driver behind what we believe is an upcoming historic consolidation.

An Architecture Built for Humans
The average enterprise runs several hundred SaaS applications - 371 on average, and 473 at large enterprises to be specific (Productiv) - and most of these are unmanaged by a central IT function. Customer data sits in a CRM, financials in an ERP, telemetry in a warehouse, conversations in Slack. Each application brings its own authentication, permissions, schema, and business semantics. Each is another context boundary.

Humans cross those boundaries astonishingly well - which is precisely why the Cost of Context remains mostly invisible. Harvard Business Review researchers found that mid- and back-office workers switch between applications roughly 1,200(!) times a day, a toggling tax that consumes roughly 9% of the workday, about five weeks a year per employee. Human judgment quietly became the integration layer holding the fragmented enterprise together, and because that layer never appeared on an invoice, no one priced it.

"Back-office workers switch applications roughly 1,200 times a day." That consumes about five weeks a year.

AI strips that hidden layer away. An agent carries no institutional memory; it cannot intuit which field is stale or which informal process actually governs a decision. It needs context made explicit, accessible, governed, and machine-readable - and it pays, in tokens and reasoning quality, for every ounce beyond what the task requires. An architecture built around human tolerance for fragmentation turns out to be actively hostile to machine execution. The market is still mispricing this asymmetry: enterprises that spent two decades accumulating applications were, without knowing it, accumulating context debt, and the companies that can retire that debt are the ones this cycle of M&A is being built around.

From Data Consolidation to Context Consolidation
At many enterprises, the instinct is to pour everything into one lake and point the AI at it. Cognizant, in one of the sharper sessions at Databricks in Spring 2026, explained why that only gets you halfway: consolidating data into a lake makes it ready for analytics. Making it ready for AI takes more. A human analyst can work with a table named cust_mstr_v3 because they know, or can ask, what it means; an AI agent needs a semantic layer - a description of what each asset means, how entities relate, and which definitions are authoritative. Above that sits a retrieval and routing function that narrows any business question to the few assets that actually matter, because no agent can afford to crawl everything on every request.

Not all context is of equal value. The knowledge that arrives with the model itself is available to every competitor - and to every customer and vendor - so it confers no advantage. The context that lets a business compete is the enterprise's own, and it arrives in two streams: the systems of record, which are right essentially 100% of the time but delayed and incomplete, and the operational exhaust - CRM entries, email, Slack, text messages - which is timely and highly relevant to the task at hand, but messy and unauthoritative.

How those two streams are merged for use by AI is one of the crucial questions in enterprise software, and the answer is evolving in real time. Every enterprise will settle it differently, according to its competitive dynamics, privacy obligations, and proprietary processes. The consequence: the context layer cannot be bought as a product. It has to be composed, tuned, and maintained enterprise by enterprise.

Cognizant's other point may matter even more: that translation layer between enterprise data and the agents that use it is a living thing. You cannot hire a services firm, say "build me a semantic layer," and check the box. A week later someone prefixes a table with a double underbar that everyone quietly reads as "deprecated" - a convention people absorb over coffee, and an agent has to be told. Schemas drift, definitions evolve, business questions change; the semantic layer has to be maintained continuously.

Regulated industries add a further requirement. If an AI system approved a decision on the first of the month, an auditor, regulator, or litigant may later ask exactly what data, definitions, permissions, and rules the system saw at that moment. In financial services, healthcare, and other supervised sectors, point-in-time context reconstruction is the price of admission for putting AI into production.

Semantic layers, retrieval, governed permissions, temporal records: together they turn modernization from a project with an end date into a continuous management function. Project revenue ends; context ownership recurs. That permanence is what makes these firms valuable, and it is why acquirers have started underwriting them as infrastructure rather than services.

Why MSPs Become Strategic - And Are Hot Acquisition Targets in 2026
The same MIT study (The GenAI Divide: State of AI in Business 2025) finds that AI solutions delivered through external partners reached full deployment roughly twice as often as internally built ones, about 67% of the time versus a third (with the caveat that the study did not necessarily differentiate organizational capabilities from implementation approach). Enterprises are learning to buy the integration discipline that makes AI work rather than build it, and the capabilities required map almost exactly onto what managed service and IT services firms already do: systems integration, identity and access management, data engineering, governance, and ongoing operational support. The MSP sector did not pivot to AI. AI arrived at the MSP sector's doorstep.

Multiple private equity managers we spoke with at M&A West are underwriting exactly this. The winners will be long-term operators of AI-ready environments, holding the semantic layer, permission model, and context record on the customer's behalf, year after year. We call it recurring context ownership: a provider embedded in a customer's systems, definitions, permissions, and decision history is far harder to displace than one selling commodity support, and far more strategic to a buyer.

"Recurring revenue earned the last cycle's multiples; recurring context may earn the next one's."

Where Consolidation Happens in the SaaS Market
For years, SaaS economics rested on standardization: one product, many customers, configuration at the edges. That model produced scale and beautiful margins, along with rigid workflows, proprietary data models, and application silos. Core systems of record are far too embedded to be displaced outright. The danger for an incumbent is subtler, and in our coverage work we watch it forming: a closed, rigid application becomes an obstacle to AI-driven work.. Each silo is another tax the enterprise's AI must pay, and buyers are beginning to ask which of their vendors levy it.

The same position that creates the danger creates the opportunity. SaaS vendors already sit inside the enterprise, holding trusted relationships and critical operational data, a position that is genuinely hard to replicate. The winners will use AI to expose their context cleanly, personalize workflows to each customer's logic, orchestrate across adjacent systems, and adapt continuously. This is the sorting question among acquirers: not whether a vendor has AI features, but whether its architecture can surrender context to someone else's agent. The vendors that can will be consolidators. The vendors that cannot will be consolidated.

Forcing Function for the Biggest Consolidation Wave Since Cloud Economics Began
Companies that make context relevant, efficient, and inexpensive, and keep it that way, convert a one-time AI project into a permanent operating advantage. That asset commands premiums - the semantic and retrieval infrastructure, the identity and governance systems, the context-owning MSPs, and the adaptive software platforms. Companies that fail at this will be marginalized or will fail - therefore our belief that more than 40% of agentic AI projects will not see 2028. Every canceled project releases budget, talent, and an unmet board mandate that lands on whoever can deliver context economically. A 40% death rate is not the end of the enterprise AI cycle; it is the forcing function for the largest consolidation wave enterprise software has seen since the move to the cloud.

You can already see this thesis on a conference floor. At the spring 2026 summits, the exhibitor booths were filled with venture-funded companies building context - semantic layers for specific domains and processes - and data providers lining up to feed them. Agentic AI operates in the semantic layer: its winners will be the next wave of successful software companies, and the data companies that feed it - hundreds of vertical data businesses emerging in financial services, healthcare, and elsewhere - will win alongside them. By our observation, that is where the majority of AI venture capital outside the data-center buildout is going.

If you run a SaaS platform sitting on an ocean of operational data - a system of record for customers, money, people, or an industry vertical - you already hold the context everyone else's agents need. The opportunity of the next two years is to expose that context cleanly and consolidate: the path Salesforce, ServiceNow, Snowflake, and Databricks are already buying their way along. The alternative is to be priced as the silo an acquirer has to work around.

If you run an MSP, systems integrator, or data-services firm, you are sitting on the most under-priced asset in this market: recurring context ownership. The semantic layer, the permission model, and the point-in-time context record you operate on a customer's behalf are infrastructure, even if your income statement still says services. Accenture and Cognizant are assembling AI-readiness practices by acquisition; sponsors such as Thoma Bravo and Vista have built platforms on far less durable revenue. Firms priced on services multiples today will be underwritten as infrastructure before this window closes.

If you build the context layer itself - semantic layers, retrieval and routing, catalogs and governance: the likes of dbt Labs, Cube, AtScale, Collibra, Alation, and Glean - you are the scarce asset this cycle is organizing around, and the premiums are already visible. The same goes for the data providers that feed the layer: if your company owns a relevant, rights-clean data stream in a vertical, you are part of this thesis whether you think of yourself as an AI company or not.

And if you build agentic AI applications, the 40% figure is a map of your market: it describes your customers' failed internal builds and your undercapitalized competitors. The survivors will be the ones whose economics of context work. For many of the rest, the right move will be to transact into a platform while scarcity value is high.

The software industry has seen this pattern before. In every prior platform shift, the durable value settled not in the layer that drew the headlines but in the layer everything else was forced to depend on. Cloud had its equivalent; so did mobile. In enterprise AI, that layer is context. The shakeout now beginning will strand almost half of today's agentic projects, and the capital and demand they release will fund the wave of M&A that builds the context layer for everyone else. Most of the companies that own meaningful context are still valued as the services firms or point solutions they used to be, and that gap will not survive contact with the next two years of deal flow. If you recognized your company in one of the paragraphs above, that window is open now.

Sources & notes
MIT NANDA, "The GenAI Divide: State of AI in Business 2025" (pilot success and build-vs-buy figures); Harvard Business Review research on application switching ("the toggling tax"); Productiv and Okta application-sprawl data; Gartner commentary on AI-ready data and its forecast on agentic-project cancellations; a 2026 survey of enterprise finance leaders on AI cost overruns; FinOps Foundation and industry reporting on token and inference costs. Qualitative observations are drawn from the author's conversations at M&A West and the 2026 Snowflake and Databricks summits, including a Cognizant session on enterprise AI readiness at Databricks. This piece reflects the author's views, is provided for general informational purposes, and is not investment advice or a recommendation with respect to any security.

Authored by Russell Tillitt, Mukesh Ahuja, Alex Bonilla