Gartner predicts that more than 40% of agentic AI projects will be cancelled by the end of 2027 — killed not by bad models, but by escalating costs, unclear business value, and inadequate risk controls. Gartner has a name for one root cause: "agent washing," vendors rebranding chatbots and RPA as autonomous agents. In logistics, where an agent will quote real money and message real customers, buying the washed version is an expensive mistake.

The demo won't tell you which one you're looking at. Demos are staged. These 12 questions will — ask them before the demo, in writing, and watch which vendors answer cleanly and which get vague.

Why the questions matter more than the demo

A demo shows you the happy path: a clean RFQ, a perfect response, a smiling sales engineer. Production is the unhappy path — a garbled email thread, a missing lane, a customer asking for something outside policy. The difference between a real agent platform and a rebranded chatbot shows up in exactly four places: governance, integrations, audit trails, and pricing. That's what these questions probe, and it's the same evaluation lens we apply in the 2026 freight broker AI platform comparison.

Governance: what is it allowed to do?

1. What actions can the agent take autonomously, and where is that authority defined? A real answer names specific authority bands: "quotes up to $X auto-send, above that they route to a rep." A red-flag answer is "it's fully autonomous" or "you can configure it however you want" with no structure.

2. How do I set escalation rules, and can I see them? Escalation rules are the product. If defining "when the agent stops and hands to a human" requires professional services, the governance layer doesn't exist. This is the discipline behind AI agent governance in logistics — guardrails you can read, not promises you can't.

3. What happens when the agent is wrong? Who finds out, and how? Every agent is wrong sometimes. The question is whether errors surface through monitoring and guardrails, or through an angry customer email. Listen for confidence thresholds, sampling/review workflows, and rollback. Silence here is a dealbreaker.

4. Can I turn down its autonomy without turning it off? You want a dial, not a switch: draft-only mode for the first month, auto-send within limits later. Vendors with one autonomy level are selling a demo, not a deployment path.

Integrations: does it work in your stack or beside it?

5. Which systems does it read and write — TMS, email, SMS, WhatsApp, telematics? "Integrates with everything" means nothing. Ask for the list of live, production integrations with named systems (your TMS by name), and ask whether the agent writes back — updating the load record — or just reads and messages.

6. Is this a native agent platform or a bolt-on to something else? A chatbot skinned over a TMS module behaves very differently from an agent built to act across systems. The TMS add-on vs native AI agent distinction matters most when a workflow spans three systems — bolt-ons tend to stop at their own border.

7. What does implementation actually involve, and who does the mapping? Get weeks, not adjectives. Ask how many customer deployments used their standard connectors versus custom engineering, and what happens when your TMS version upgrades.

8. What does the agent do when a system is down or data is missing? Production means dirty data. A real agent detects the gap and escalates; a fragile one hallucinates a plausible answer. This single question separates production systems from proofs of concept.

Audit trails: can you defend what it did?

9. Show me the log of a single action, end to end. For any quote sent or message answered, you should be able to see: what came in, what data the agent pulled, what rule it applied, what it sent, and when. If the audit trail is an afterthought, your compliance conversation is too.

10. Can I export the agent's decision history for a customer dispute or an internal review? In freight, disputes are routine. "The agent said the rate included fuel" needs to be answerable from a record, not a support ticket.

Pricing: where's the trap?

11. How is it priced — per seat, per message, per action, per load? And what happens at 3x volume? Per-seat pricing on a product whose purpose is replacing seat work is a contradiction. Usage pricing that explodes at scale is the "escalating costs" line in Gartner's cancellation forecast. Model your volume at 12 months, not at pilot.

12. What's the total first-year cost — license, implementation, integrations, overages? Ask for a number, in writing, before the demo. Vendors who won't quote until after the "discovery process" are telling you the answer.

How to score the answers

| Signal | Green flag | Red flag | |---|---|---| | Authority limits | Named bands, customer-configurable | "Fully autonomous" or "trust the model" | | Escalation | Self-serve rules UI, shown live | Requires vendor services to change | | Error handling | Confidence thresholds, review queues, rollback | "Our accuracy is very high" | | Integrations | Named production connectors, writes back | "Open API" as the whole answer | | Audit trail | Per-action log, exportable | Available "on request" | | Pricing | Modeled at your 12-month volume, in writing | Deferred until after the demo |

Two or more red flags in the governance column — walk. That's how projects end up in Gartner's 40%.

FAQ

Should we run a pilot before asking these questions?

Ask first, pilot second. A pilot with a vendor that fails the governance questions wastes the same months as a bad full deployment — pilots are where "agent washed" products look their best, because nothing real is at stake.

What's the single most important question?

Number 3: what happens when the agent is wrong. Every vendor has rehearsed answers about capability. Almost none have rehearsed honest answers about failure — which is exactly why it reveals the most.

Do these questions apply to copilots as well as agents?

Yes, with lower stakes. A copilot that drafts for human review can fail softer. But anything that sends quotes or customer messages without a human click needs full marks on governance, audit, and error handling.

The bottom line

With 40% of agentic AI projects headed for cancellation, the difference between the 40% and the 60% is decided before the contract is signed — in whether you probed governance, integrations, audit trails, and pricing, or watched a polished demo instead. Twelve questions, in writing, before the demo. The vendors worth buying from will thank you for them.

Debales.ai builds governed AI agents for freight — authority bands, escalation rules, and full audit trails are the product, not an add-on. Book a demo and bring these 12 questions — or see the platform first.

---

Sources: Gartner press release, "Over 40% of Agentic AI Projects Will Be Canceled by End of 2027" (June 2025); Gartner January 2025 poll of 3,412 attendees; Gartner 2028 agentic AI forecasts.