Not because it’s a secret, but because nobody’s counting. The agent sprawl problem is shadow IT’s sequel, and it’s already bigger.
Ask the revenue org, not the IT org
When agent sprawl gets discussed, it is usually framed as an infrastructure story. That framing is why most executive teams underestimate it.
Ask your CRO how many agents are touching the pipeline right now. Ask your CMO how many are writing, sending, scoring or routing something on the company’s behalf this week. In most organizations, the answer is a range, not a number, and the range is wide.
That is not a technology problem. It is a commercial control problem, and it sits inside the functions that carry the number.
Sprawl starts where budget autonomy is highest
There is a straightforward reason agents proliferate fastest in go-to-market teams. Those teams have discretionary budget, monthly targets, and the shortest distance between deciding to try something and having it running.
The spending pattern reflects it. Roughly half of generative AI budgets are directed at sales and marketing, even though back-office automation frequently returns more. Meanwhile, around 90% of employees use personal AI tools for work while only about 40% of firms hold official subscriptions.
This is the same dynamic that produced SaaS sprawl a decade ago, with one difference that matters commercially. A duplicate SaaS licence wastes money quietly. A duplicate agent takes action, and two agents optimizing the same funnel with different assumptions will produce two different answers about the same customer.
Support is where agents are thickest
At the point of customer contact, AI agents are increasingly answering product questions, resolving support requests, and assisting sales teams before a human ever joins the conversation.
McKinsey’s State of AI 2025 found that organizations are rapidly experimenting with AI agents, with 62% already piloting or using them, and that marketing, sales, and customer-facing functions are among the areas reporting the strongest revenue gains from AI adoption. However, most organizations are still in the early stages of scaling AI across the enterprise, and only 39% report measurable EBIT impact at the company level
The commercial opportunity is clear, but so is the integration challenge. Most AI capabilities have been introduced independently inside CRM systems, support platforms, analytics tools, and developer workflows by different teams and at different times. Gartner predicts that over 40% of agentic AI projects will be canceled by the end of 2027 because of escalating costs, unclear business value, and inadequate risk controls, highlighting that the challenge is no longer access to AI but connecting it into a coherent operating model that delivers measurable business outcomes.
You are also acquiring agents without deciding to
The second source of sprawl is not bottom-up at all. It arrives through procurement, bundled into software you already bought.
Gartner expects roughly 40% of enterprise applications to embed task-specific agents by the end of 2026, up from under 5% a year earlier. Your CRM, your sequencing tool, your support desk, and your analytics platform are each shipping agents into your workflow as product updates, not as purchase decisions.
Nobody signed a contract for those agents. Nobody scoped their permissions or decided what they may do unsupervised. They inherited whatever access the parent application already had, which in a revenue stack usually means the customer record.
The practical consequence is that an accurate agent count cannot be produced from a purchase ledger. Procurement knows what you bought. It does not know what arrived inside it.
Three costs that show up on the P&L
Duplicated build and duplicated spend. Salesforce’s 2026 Connectivity Benchmark, summarised by IBM, found around half of enterprise agents operating in silos rather than as part of a coordinated system. When agents are built independently, the same data pipeline gets constructed several times and each version carries its own compute cost and its own maintenance burden.
One customer, several versions of the truth. The same research found 27% of the APIs connecting these agents running without governance or audit trail. When a prospect gets a sequence from one agent, a conflicting offer from another, and a support answer that contradicts both, the damage is to conversion and trust rather than to a security posture.
Attribution that quietly stops working. Revenue operations depend on knowing what touched a deal and in what order. Once an unknown number of agents are enriching, scoring, routing and messaging without registration, pipeline attribution becomes an estimate. Forecast confidence degrades before anyone can explain why.
The ROI paradox nobody wants to own
The market data on returns looks contradictory until you account for visibility. A large majority of organizations report measurable economic impact from agents, yet far fewer report significant, attributable ROI from agents specifically.
Both findings can be true at once. Value is being created and simply cannot be traced, because the estate creating it was never inventoried.
That gap is the likeliest explanation for the most sobering forecast in this category. Gartner expects more than 40% of agentic AI projects to be cancelled by the end of 2027, citing escalating costs, unclear business value and inadequate risk controls. Projects rarely die because the technology failed. They die when nobody can produce a defensible answer to what it cost and what it returned.
Compliance stopped being a back-office cost
For commercial leaders, the most consequential shift is that AI governance has moved into the sales cycle. Enterprise buyers now ask about agent inventory, data handling and human oversight during security review, and increasingly write AI-specific terms into master agreements.
An organization that cannot answer those questions does not lose a compliance argument. It loses the deal, or spends six weeks in a security questionnaire it could have cleared in one.
The regulatory calendar reinforces this rather than relieving it. Under the EU AI Act, high-risk obligations for Annex III systems have been deferred to 2 December 2027 following the Digital Omnibus package adopted in June 2026. A deferral is not a repeal, prohibited practices and general-purpose model obligations are already in force, and the Act reaches any organization whose systems touch users in the EU regardless of where it is headquartered.
Marketing leaders should note one date in particular. Transparency obligations covering artificially generated content land in December 2026, which is a content and disclosure requirement sitting squarely in the CMO’s remit rather than the CISO’s. Penalties for high-risk non-compliance run to €15 million or 3% of global turnover.
Better regulation is a commercial asset, not a tax
It is worth resisting the reflex to treat this as friction. The frameworks converging on this problem, including ISO/IEC 42001 and the NIST AI Risk Management Framework, all begin at the same place, establishing what you operate and who owns it.
That artefact does double duty. The inventory a regulator wants is the same inventory a CFO needs to allocate agent spend, the same one a CRO needs to trust attribution, and the same one a buyer’s procurement team needs to clear you.
Governance built only to satisfy an auditor is a cost. The same work, framed as operational visibility, is the thing that lets you scale agents without losing the ability to explain what they did. Regulation deserves to stay at the core of this conversation precisely because it is currently the only force compelling companies to build something they need anyway.
Five things a commercial leader can do this quarter
Count what exists, including what came bundled. Audit the agents your revenue stack shipped you, not only the ones your team built. Treat an unregistered agent in a customer-facing workflow as a defect.
Put a named owner and a cost center against every agent. Not a team, a person, with the computer and licence cost attributed to their budget. Ownership without cost attribution does not survive a planning cycle.
Decide who speaks to the customer. Establish which agent is authoritative for pricing, offers and support answers, so a prospect receives one version of the truth rather than three.
Instrument agent activity into your attribution model. If an agent touched a deal, revenue operations should be able to see it. Otherwise your forecast is measuring a system you cannot observe.
Set disclosure standards now. Define how and where you tell customers they are interacting with, or reading output from, an AI agent. December 2026 makes this a legal requirement in the EU; buyer expectations are already making it a commercial one.
The list is the differentiator
The enterprises that come out of this cycle ahead will not be the ones that deployed the most agents. They will be the ones that can produce a current, accurate list, with an owner and a cost beside every line, and hand it to a regulator, a CFO or a prospect’s security team without a three-week scramble.
That is an unglamorous ambition for a technology this significant. It is also the one almost nobody has met yet, which is exactly why it is worth being early to.
