AI Agents: Separating capability —
from marketing language
The shift from “AI that chats” to “AI that acts” is underway. Here is how to evaluate if an acquisition target has genuine agentic capabilities or just a well-packaged conversational interface.
Start here: the plain-language definition
If a Large Language Model (LLM) is an incredibly well-read analyst sitting at a desk, an AI Agent is that same analyst given a corporate credit card, access to your CRM, and permission to send emails on your behalf.
Standard AI answers questions. An AI agent executes multi-step tasks. When given a high-level goal (e.g., “Resolve this customer’s refund request”), an agentic system will independently break the goal down into steps, access external databases to verify the purchase, calculate the refund amount, trigger the payment API, and draft the apology email. It acts autonomously.
“A company boasting about its ‘AI agents’ without strict API governance is not advertising a competitive moat. It is advertising an unquantified operational liability.”
How it works — without the engineering lecture
To understand the commercial viability of an agentic system during due diligence, you must break it down into its three functional components:
When a target company claims they have built an "AI Agent," diligence must immediately focus on The Hands.
If the system cannot independently execute a task via APIs, it is not an agent. It is a chatbot with a clever marketing team.
The four questions your board should be asking
Agents introduce execution risks that standard generative AI does not. If an AI writes a bad summary, a human corrects it. If an AI agent executes a bad database deletion, the damage is immediate. Here is how to interrogate the risk.
Is there a "Human in the Loop"?
True autonomy is dangerous in high-stakes environments. Does the system execute tasks completely on its own, or does it prepare the workflow and pause for human authorization before taking an irreversible action (like moving money or sending a contract)?
What are the API guardrails?
Agents interact with software via APIs. If the agent is compromised or hallucinates, what prevents it from deleting records or exposing client data? Access control and read/write permissions for AI systems must be audited just as strictly as human access.
Who assumes the liability?
If a B2B SaaS company provides an AI agent that incorrectly modifies a client's core data, who is financially responsible? Software licenses must be heavily reviewed to understand indemnification clauses surrounding autonomous AI actions.
What is the fallback mechanism?
Agents fail when edge cases arise or third-party tools go offline. What happens when the agent gets stuck? If the system continuously loops or crashes silently without gracefully handing the issue to a human operator, the technical debt is severe.
What this means for valuation
The presence of genuine agentic architecture alters both the ceiling and the floor of a company’s valuation.
Upside: True agents decouple revenue growth from headcount. Traditional software makes humans more efficient; agentic software executes the work entirely. Companies that can demonstrably prove their agents handle end-to-end tasks (e.g., resolving 40% of Level 1 IT tickets without human touch) command premium multiples because their gross margins will scale exponentially, not linearly.
Downside risk: The cost of agentic failure is high. If an agentic workflow is fragile—requiring constant engineering intervention to maintain API connections and prompt chains—the maintenance costs will quietly erode EBITDA. Furthermore, “fake agents” (complex scripts masquerading as autonomous AI) will be uncovered during rigorous technical due diligence, often derailing deal confidence.
The one signal that separates leaders from laggards
During technical due diligence, stop looking at “User Engagement” metrics for AI features. The only metric that matters for an AI Agent is the Task Completion Rate (TCR).
If a company claims to have an agentic product, ask for the data on how many times the agent successfully resolved a multi-step user intent from start to finish without a human taking over. If they cannot provide that data, they are selling a concept, not a capability. Price the asset accordingly.
