Three Questions That Expose Fake Agents * The A-P-M Test

Gartner counts roughly 130 vendors delivering genuine agentic AI [1]. The market contains thousands of claimants. That ratio reveals a $7.92 billion industry [2] where the majority of products labeled AI Agent fail the most basic test of agency. You are funding counterfeits.

The A-P-M Test is a three-question diagnostic that separates real AI Agents from rebranded chatbots in under 60 seconds. I built it to give you a pass/fail instrument: Autonomy (can it act without you?), Proactivity (can it initiate before you?), Memory (can it learn from you?). You will walk away from this article with a weapon you can deploy in your next vendor pitch, your next board meeting, and your next procurement decision.

Why Do Most AI Agent Vendors Fail the Test? * The agent-washing epidemic

Agent-washing is the practice of rebranding a basic automation tool as an AI Agent without delivering true agency. I coined the term in AI Agents: They Act, You Orchestrate because the agentic AI market needs a clinical name for its dominant pathology: rule-based chatbots and scripted workflows marketed as autonomous agents.

The evidence is damning. In 2025, every 2010-era chatbot was rebranded as an autonomous agent [3]. Vendors rebrand everything from scripted chatbots to simple RPA flows as AI Agents despite these tools relying on fixed rules and zero learning [4]. Writer.com's research confirms the mechanism: most mismatched expectations come from Level 1 assistive tools marketed as Level 3 or Level 4 systems [5]. The vendor sells you a macro. The slide deck calls it an agent.

The cost of this fraud compounds. Gartner predicts 40% of agentic AI projects will be canceled by end of 2027 due to escalating costs, unclear business value, and inadequate risk controls [6]. Meanwhile, 88% of executives are increasing their agentic AI budgets [7]. That is not a trust gap. That is a procurement failure operating at industrial scale.

Complex evaluation frameworks from the analyst firms have failed to stop it. Gartner publishes multi-criteria matrices. Forrester offers capability assessments. Deloitte distinguishes between chatbots and agents across multiple axes [8]. These frameworks are thorough. They are also unusable in a 30-minute vendor demo. You need a weapon you can deploy in real time.

Autonomy: Can It Act Without You?

Autonomy is the first hallmark and the minimum threshold for agency: an AI Agent executes a complex, multi-step task from start to finish without requiring your input at each stage. If you must manually guide it through every step, you have a convenient macro, not an agent.

The test is binary: does it complete a workflow, or does it complete a single step? A food delivery app automates payment but still requires you to choose, order, and confirm. That is a tool you wield. An autonomous AI Agent, given the intent arrange the Q3 offsite, executes the entire workflow: selecting the venue, checking calendars, and sending invites. Zero taps from you.

Deploy this question in every vendor demo. Ask the vendor to show an end-to-end workflow without human intervention. If the demo requires a human click between stages, the answer is no. Cognition's Devin, one of the most advanced autonomous software agents on the market, still resolves only 14% of GitHub issues autonomously [8]. That 14% is real autonomy. The other 86% is assisted tooling. Most vendors cannot match even that honest ratio.

Proactivity: Can It Initiate Before You?

Proactivity is the hallmark that separates a servant from a strategist: a proactive AI Agent identifies an opportunity or a threat and initiates action before you ask. A reactive tool waits for your command. A January 2025 arXiv paper defines proactivity as behaviors initiated by the AI Agent without user prompts [9]. Most vendors claim this capability. Almost none deliver it.

The Event Horizon is the threshold where a tool stops waiting for commands and starts initiating action before you do. A system that reports a server outage after it happens is a relic. A system that alerts you to degrading performance metrics before the outage cascades is proactive.

Tonkean launched what it calls the first proactive agents for enterprise operations in September 2025, agents that monitor, reason, and suggest actions without user prompts [10]. That is the standard to test against. In your next vendor demo, watch the sequence: does the agent only speak when spoken to? If every action requires your prompt, you are looking at a reactive tool with a proactive label. That is agent-washing, and you are the mark.

Memory: Can It Learn From You?

Memory is the hallmark that transforms a one-shot tool into a long-term collaborator. An AI Agent with memory evolves. A tool without it resets to zero with every interaction. A chatbot is a digital amnesiac; every conversation is the first. An AI Agent without a past cannot be trusted with your future.

The test is simple: tell the agent something once. If you have to repeat it, the agent has no memory. Does it get smarter with every interaction, or does it start over?

The industry is waking up to this. Leena AI built a three-tier agentic memory architecture: user-specific, organizational, and domain memory that enables agents to learn from corrections and compound value over time [11]. AWS launched AgentCore with long-term memory designed for behavioral consistency and adaptation [12]. Dust.tt declared agent memory the foundational capability for building an ambient AI operating system [13]. These companies understand that memory is the compounding asset. Without it, every dollar you spend on an agent delivers linear, not exponential, returns.

Memory also reveals your lock-in risk. Where the memory lives determines who controls the value. If the vendor owns the memory layer, they own the compounding intelligence you paid to build. That is a strategic dependency you must architect around, not stumble into.

The Trinity Is Indivisible * Partial credit does not exist

The A-P-M Test treats the three hallmarks of agency as constitutional requirements, not a pick-two menu. A system missing any one hallmark fails the test entirely. Partial agency is a contradiction in terms.

Autonomy without memory is a dangerous, unguided missile, executing tasks with no recall of past failures. Proactivity without memory is a ghost, initiating actions untethered from your history or preferences. Memory without autonomy and proactivity is a database, storing context that no system acts on.

I designed the A-P-M Test as a pass/fail instrument, not a scoring rubric. Three yeses, or the product is not an agent. This eliminates the partial credit thinking that lets agent-washing survive. It gives you a devastating one-line summary for the board deck: "Of the 12 vendors we evaluated, three passed the A-P-M Test." That sentence is worth more than every multi-criteria matrix an analyst firm has published, because it forces a binary decision.

The Gartner prediction that 40% of agentic AI projects fail maps directly to this indivisibility. Those projects are built on systems that pass one or two hallmarks but not all three. They look agentic in a demo. They collapse in production.

The Verdict You Were Avoiding

The article so far has built the case against vendors. Here is the reframe you will not enjoy: you are complicit.

Every failed pilot you greenlit was a pilot you could have killed in the first meeting by asking three questions. The 40% cancellation rate is not a vendor problem. It is a procurement problem. You approved the budget. You skipped the diagnostic. You accepted a vendor's slide deck as evidence of agency instead of demanding a live, end-to-end, unprompted workflow with persistent memory recall.

The A-P-M Test doubles as a self-evaluation tool. Every time you skip it, you subsidize the counterfeit market that makes real AI Agents harder to find and harder to fund.

The next vendor pitch is already on your calendar. The next budget request is already in your inbox. You now own the three questions that separate a $7.92 billion market from a $7.92 billion fraud. The 40% who cancel their agentic AI projects by 2027 will be the ones who never asked.

The A-P-M Test is one framework from AI Agents: They Act, You Orchestrate by Peter van Hees. The book maps 18 chapters across the full architecture of the Agent-First Era, from the Perceive-Reason-Act Cycle that powers true agency to the Autonomy Spectrum that determines where your agents sit on the continuum between tool and collaborator. If agent-washing and the diagnostic gap resonated, the book gives you the complete blueprint: the Delegation Ladder, the Friction Tax quantified, and the Human Premium Stack that separates the roles agents cannot touch. Get your copy:

🇺🇸 Amazon.com
🇬🇧 Amazon.co.uk
🇫🇷 Amazon.fr
🇩🇪 Amazon.de
🇳🇱 Amazon.nl
🇧🇪 Amazon.com.be

References

[1] Beam.ai, "40 Percent of Agentic AI Projects Will Fail," Beam.ai Agentic Insights, Jan 2026. https://beam.ai/agentic-insights/40-percent-agentic-ai-projects-will-fail-heres-how-to-be-in-the-60

[2] Precedence Research via Arcade.dev, "Agentic Framework Adoption Trends," Nov 2025. https://blog.arcade.dev/agentic-framework-adoption-trends

[3] Xtract.io, "Your 2025 AI Wrapped Reality Check," LinkedIn, Dec 2025. https://www.linkedin.com/pulse/your-2025-ai-wrapped-reality-check-road-2026-xtract-io-hl6ac

[4] Outreach.io, "Agent Washing: Why AI Projects Fail," Oct 2025. https://outreach.io/resources/blog/agent-washing-ai-projects-fail-guide

[5] Writer.com, "Agent Washing," Writer Blog, Aug 2025. https://writer.com/blog/agent-washing/

[6] Gartner, "Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027," Gartner Press Release, June 25, 2025. https://www.gartner.com/en/newsroom/press-releases/2025-06-25-gartner-predicts-over-40-percent-of-agentic-ai-projects-will-be-canceled-by-end-of-2027

[7] PwC, "AI Agent Survey," PwC US, 2025. https://www.pwc.com/us/en/tech-effect/ai-analytics/ai-agent-survey.html

[8] Deloitte, "Autonomous Generative AI Agents Still Under Development," Deloitte TMT Predictions, Nov 2024. https://www.deloitte.com/us/en/insights/industry/technology/technology-media-and-telecom-predictions/2025/autonomous-generative-ai-agents-still-under-development.html

[9] arXiv, "Proactive AI Agent Behaviors," arXiv preprint, Jan 2025.

[10] Tonkean, "Tonkean Announces Proactive AI Agents," Tonkean Blog, Sept 2025. https://www.tonkean.com/blog/tonkean-announces-proactive-ai-agents--first-proactive-agents-for-enterprise-g-a-teams

[11] Leena AI, "Leena AI Agentic Memory," Leena AI Blog, Oct 2025. https://blog.leena.ai/leena-ai-agentic-memory-enterprise-operations-transformation-2025/

[12] AWS, "Building Smarter AI Agents: AgentCore Long-Term Memory Deep Dive," AWS Machine Learning Blog, Oct 2025. https://aws.amazon.com/blogs/machine-learning/building-smarter-ai-agents-agentcore-long-term-memory-deep-dive/

[13] Dust.tt, "Agent Memory: Building Persistence into AI Collaboration," Dust.tt Blog, Aug 2025. https://dust.tt/blog/agent-memory-building-persistence-into-ai-collaboration