Your company is ready for agentic AI when it can write down what the agent should do, with which data, how far it can go on its own, and how you'll know it made a mistake. If any of those answers depends on "Jane knows how that works," the next investment isn't an agent. It's closing that gap.
The terms are worth separating, because the market blurs them. An assistant answers when someone asks. An AI-powered workflow follows fixed steps defined in code. An agent decides its own steps: it reads the context, picks which tool to use, acts and checks the result. Anthropic's guide Building effective agents (December 2024) draws the same line and offers advice that belongs in every board meeting: find the simplest solution possible and only add complexity when it's needed. The same piece notes that an agent's autonomy comes with higher costs and the risk of compounding errors.
Where the market is
According to McKinsey's The state of AI in 2025, 62% of respondents say their organizations are at least experimenting with AI agents, and 23% say they're scaling an agentic system somewhere in the enterprise. Within any single business function, no more than 10% are scaling agents.
Gartner is blunter. In June 2025 it predicted that over 40% of agentic AI projects will be canceled by the end of 2027 because of escalating costs, unclear business value or inadequate risk controls. The same release warned about "agent washing," vendors rebranding chatbots and RPA as agents, and estimated that only about 130 of the thousands of self-described agentic AI vendors actually offer the capability.
In short: many companies have started, few have scaled, and a good share of projects will stop. The five signs below help you tell which side your company is on.
1. The process is written down, exceptions included
An agent needs a process someone can describe, including what to do when something falls outside the norm. The common cases are easy. What makes or breaks an agent is how it handles exceptions: the order with no invoice number, the customer asking for a discount outside policy, the contract with a clause nobody has seen before.
This week's test: ask the process owner to list the 10 most frequent exceptions and what's done in each. If the list doesn't come, or each person on the team gives a different answer, the process still lives in people's heads.
2. Data has an owner and a known quality level
An agent that reads bad data takes the wrong action, and takes it fast. In February 2025, Gartner predicted that through 2026 organizations will abandon 60% of AI projects unsupported by AI-ready data. In the same survey, 63% of organizations said they either don't have, or aren't sure they have, the right data management practices for AI.
This week's test: for each system the agent will read from, ask who owns the data and what its known error rate is. "I think it's fine" isn't an answer.
3. A simple agent has already been measured
Companies that jump straight to multiple agents talking to each other discover context, cost and reliability problems late and expensively. One agent with a narrow task, measured for 60 to 90 days, teaches more than a large pilot nobody can evaluate.
This week's test: is there any agent or AI automation in production whose cost per case and error rate you know? If not, that's your first project.
4. Decisions follow criteria, not seniority
If today's rule is "whoever has been here longest decides," the agent has nothing to follow. Criteria can be an approval matrix, a table of limits, a written policy. What they can't be is gut feel.
This week's test: pick 20 recent decisions in the process and ask two different people to explain them. If the reasoning diverges a lot, criteria are missing. Making those rules explicit usually improves the process even if the agent never gets built.
5. There's appetite for governance, not just speed
An agent scales mistakes as well as it scales good work. Governance here is concrete: limits on what the agent can do without approval, a log of every action, someone responsible for reviewing what went wrong, and a way to switch it off fast. The NIST AI Risk Management Framework, published in January 2023, organizes this into four functions (govern, map, measure and manage) and is a solid starting point if you'd rather not invent your own structure.
This week's test: ask who answers for it if the agent sends the wrong email to 500 customers. If nobody knows, governance doesn't exist yet.
Hypothetical example: email orders at a distributor
The scenario below is a hypothetical example built to show how to apply the signs. It is not a client.
A distributor receives purchase orders by email, as PDFs and in the message body. Three people read them, check stock and pricing, and enter them in the ERP. Leadership wants an agent to do all of it on its own.
Applying the signs:
- Process: the common cases are clear, but exceptions (substitute products, customers with blocked credit) depend on one person. Partial.
- Data: the product catalog has duplicate codes and ERP inventory updates once a day. Missing.
- Simple agent measured: none. Missing.
- Criteria: the discount approval matrix is written down. Present.
- Governance: nobody has defined what the agent can enter without approval. Missing.
One and a half signs out of five. Building the full agent now would amplify the duplicate catalog. The shorter path is different: an agent that only reads the email and pre-fills the order for human approval, while the catalog gets cleaned and exceptions get documented. In 90 days the company has a measured agent (sign 3), written exceptions (sign 1) and data that can support the next step.
Readiness checklist
Answer yes or no:
- The 10 most frequent process exceptions are documented.
- Every system the agent will read from has a named owner.
- The error rate of that data is known, even roughly.
- There's an agent or AI automation in production with a measured cost per case.
- Process decisions follow written approval limits or policies.
- What the agent can do without human approval is defined.
- Every agent action is logged and auditable.
- Someone is responsible for reviewing errors, and there's a fast way to switch the agent off.
- The project's success criterion has a number and a deadline.
Seven or more yeses: you can design an agent with real autonomy. Four to six: start with an assisted agent that suggests while a human approves. Fewer than four: the investment with the best return right now is in process and data.
Next step
If your company passed the checklist, the next step is to test one use case with a fixed timeline and a success criterion agreed up front, which is how the Cyborg Forge works. If you came in below, Advisor As a Service helps prioritize which gaps to close first. Either way, our methodology shows how each stage is measured.