The Agents You Cannot Name

Deloitte's 2026 State of AI in the Enterprise report, published in April and drawing on 3,235 IT and business leaders across 24 countries, found that only 21% of organisations have a mature governance model in place for agentic AI. 74% plan to adopt it within two years. The gap between those two figures is not a compliance footnote. It is the operating environment you are being asked to sign contracts inside.

I sat in a procurement review three weeks ago where the finance lead held up an invoice for $127,000 in API spend across OpenAI, Anthropic, and Amazon Bedrock for the previous month. Nobody in the room could name more than six of the agents that produced the bill. The platform team knew there were agents in customer support, sales automation, and developer tooling. Marketing had something running in campaign generation. One of the regional subsidiaries had stood up a procurement agent without telling anybody. The rest was inference cost attached to systems whose owners had left the company, experiments somebody forgot to turn off, or shadow deployments that had migrated from a Slack bot into production without an intake ticket.

This is not an edge case. 72% of organisations experienced an unexpected AI cost spike or surprise bill during the past twelve months, according to FinOps practitioner data published in August 2026. 98% now manage AI spend, up from 31% two years ago. Managing it and understanding it are different verbs. When Gartner said in April 2026 that by 2028 an average global Fortune 500 enterprise could have more than 150,000 AI agents in use, up from fewer than 15 in 2025, the headline sounded like progress. The operational reality is that most enterprises cannot audit the 40 they are running today.

The failure mode here is not technical. Many organisations now face what engineering teams privately call the POC graveyard: hundreds of pilots, dozens of demos, very few systems operating reliably at enterprise scale. Governance failed before the agents shipped. The intake process approved the proof of concept. The steering committee nodded at the slide deck. Nobody specified who owns the production deployment, what it is allowed to call, how much it can spend per month, or what happens when it fails at 03:00 on a Saturday.

A simple chat request in 2023 averaged 800 input tokens. An agentic step in 2026 routinely passes 50,000 input tokens. Input pricing appears identical on the rate card at $2.50 per million, but the cost per completed task is 60 times higher. An agentic workflow, where an autonomous AI agent reasons iteratively, breaks down a task, calls tools, verifies outputs, and self-corrects, may trigger ten to twenty LLM calls to complete a single user-initiated task. Token prices fell. Uber's engineering organisation ran out of its entire annual AI coding budget two-thirds of the way through the year in April 2026. Two months later, Microsoft pulled back most internal developer access to Claude Code over the same problem. Both got caught by the same maths: token prices are falling, but the bill keeps climbing.

You approved the agent because the ROI model showed headcount savings. The model assumed one call per task. The agent makes 14. It also stuffs the context window with retrieved documents you did not price in, retries on errors you did not account for, and runs continuously rather than on-demand because nobody drew a line between "monitor this queue" and "poll this queue every 90 seconds forever". The FinOps team sees the spend. They cannot attribute it to a feature, a team, or a customer. When the teams making spend-generating decisions don't feel the budgetary consequences, governance collapses.

The governance gap compounds when agents call tools. Agent sprawl happens when teams create AI agents across functions without centralised visibility, ownership, governance, cost controls, or auditability. Agents can reason, call tools, access data, trigger workflows, and act on behalf of users. That makes their risk behavioural, not just operational. A chatbot that hallucinates a wrong answer is a quality problem. An agent that misinterprets an instruction, calls a write API, and corrupts a production database is an incident. 35% of organisations admit they could not shut down a rogue AI agent if one emerged, and 36% have no formal plan for deploying AI agents at all.

Most enterprises do not have an agent registry. An AI registry catalogues every AI system in the organisation. Each entry records system name, purpose, risk classification, model provider, training data sources, deployment date, human oversight requirements, and compliance status. The registry is the single source of truth for AI governance. Without one, you have no list of what is running, who approved it, or what it is allowed to do. You have a Bedrock bill, a support ticket from a customer whose order got cancelled by something they did not recognise, and an engineering lead who thinks the agent they shipped last quarter is still in pilot.

Shadow AI makes the inventory problem worse. Akamai's Enterprise AI Usage Risk Report 2026, released in the first week of August, reveals how decentralised shadow AI, highly active AI power users, and silent browser extensions are exposing critical corporate assets to entirely new classes of cyber risk. Netskope's Cloud and Threat Report for 2026 found that 47% of generative AI users access tools through personal accounts, and the average enterprise experiences 223 data policy violations per month tied to AI usage. With 78% to 89% adoption across all enterprise departments in 2026, shadow AI use is standard operating procedure. Employees who cannot use AI through official channels use it through personal devices and accounts. The primary effect of a ban without a sanctioned alternative is that usage becomes less visible to security teams, increasing rather than decreasing risk exposure.

Prompt sprawl is the shadow IT of the agent era. High-performing prompts are frequently scattered across personal notes, chat histories, code repositories, and team messaging channels. Without a centralised system, teams waste countless hours rewriting successful prompts, fixing inconsistent outputs, and duplicating effort across departments. A March 2026 report by US-based AI firm Braincuber Technologies found a single untracked prompt change can degrade user experience without anyone noticing. One documented case saw a customer satisfaction score drop from 87% to 61% over just eleven days, because the team had no change logs in place. The agent works differently this week because somebody edited the system prompt in production to fix a corner case. Nobody logged the change. The evaluation pipeline that should have caught the regression does not exist.

An IBM Institute for Business Value study from 2025, surveying 1,300 senior AI decision-makers, found that those who reported their companies ignored the issue of technical debt saw returns on projects drop by 18% to 29%, with timelines expanding by as much as 22%. AI debt is the new technical debt, and it compounds faster. There are seven sources of AI debt: pilot sprawl, undefined success criteria, ungoverned data foundations, tool and subscription sprawl, shadow integrations, skipped change management, and vendor lock-in by default.

The buyer's problem is that you are being asked to fund the next wave of agents before anybody can account for the last one. The board wants ROI. The CFO wants unit economics. The CISO wants an audit trail. You want to know which agents are running, what they cost, and who can turn them off. The answer you get is a Grafana dashboard showing aggregate token consumption and a Jira epic titled "Agent Governance Framework" with 11 subtasks, none of them assigned.

The contracts you are signing now will determine whether you govern this or pay for it indefinitely. If the vendor cannot give you per-agent cost attribution, you cannot allocate spend to a business unit. If the platform does not log every tool call the agent makes, you cannot audit what it did when something breaks. If nobody owns the production deployment, the agent becomes technical debt the moment it ships. You approved a pilot. You are financing a liability.

Most EU AI Act provisions become binding on 2 August 2026, and only 12% of enterprises have mature AI governance processes in place, according to HFS Research and Infosys, even as agentic AI deployment moves into production at scale across most large organisations. The margin for running agents without a formal governance structure has closed. The agents you cannot name are the ones that will fail an audit, breach a data policy, or cost you three months of budget in a weekend because nobody set a spend limit.

If you cannot list every agent running in production, you do not have an AI strategy. You have shadow IT with API access and a monthly bill you cannot explain.

Tarry Singh is the founder and CEO of Real AI (realai.eu), an enterprise AI advisory and deployment firm working with global enterprises on production agent systems, model risk, and AI sovereignty strategy. He also leads Earthscan (earthscan.io) for Energy AI, and is a founding contributor to the EU-funded HCAIM and PANORAIMA programmes for responsible AI education across European universities. He writes at tarrysingh.com.