Why a score is not a journey
If your assistant ordered the wrong replacement part or cancelled the wrong reservation, asking which word triggered the mistake tells you almost nothing. You need to see where its steps diverged.
When an apprentice baker burns a loaf of sourdough, handing you the ingredient list does not explain what went wrong. The flour was fresh. The salt was measured. The water was pure. The mistake happened somewhere between kneading the dough, waiting for the rise, and setting the oven timer.
Until recently, explaining an AI system looked a lot like reading that ingredient list. If an automated model declined a loan application or flagged an email as spam, engineers highlighted the specific words or numbers that tipped the score. If the word "urgent" carried heavy weight, that was your explanation.
That approach worked well enough when software made single, one-shot predictions. Today, AI assistants do not stop at predicting a score. They act.1
From single guesses to ten-step errands
An assistant does not just answer a question. It carries out a multi-step errand across the web.
Picture asking an assistant to reschedule a medical appointment, rebook a train ticket, and notify your employer. To finish that errand, the software takes a dozen connected steps. It logs into a clinic portal. It reads an open calendar slot. It checks train schedules on a booking site. It fills out a confirmation form. Finally, it sends a brief note to your supervisor.
Suppose you wake up to discover you were booked on a train leaving at midnight instead of noon. Asking which words carried the highest weight tells you nothing. Did the assistant misread the clinic calendar? Did the rail booking site update seat availability midway through? Or did the assistant forget your morning meeting halfway through the task?
Where assistants lose the thread
Recent research confirms what everyday intuition suggests. Traditional explanation tools completely fail when applied to multi-step AI agents.1
When researchers analyzed hundreds of automated errands, they discovered that failures rarely stem from a single bad word. Instead, assistants fail because they lose track of what changed along the way. In fact, studies found that state confusion was nearly three times more common in failed runs than in successful ones, cutting task completion rates in half.1
When an assistant loses its place, an ingredient list cannot help you. You do not need to know which words seemed important in the abstract. You need to know what the software saw at step three, what decision it took at step four, and why it changed direction at step five.
Comparing a static score with an action journey
Three things an action journey must show
To make automated assistants truly accountable, every multi-step errand needs an auditable journey record. We think that record must answer three simple questions.
1. What did the assistant observe at each step? When software opens a page or queries an online service, it sees a momentary snapshot of facts. A journey record captures that snapshot. If a rail site updated seat prices midway through, you can see whether the assistant acted on current facts or outdated data.
2. What rule guided each move? Every time an assistant clicks, reserves, or sends, it should link that move to a clear instruction. If it booked an evening departure, did your note permit evening travel, or did the software overstep? Without a rule attached to each step, you cannot tell an honest mistake from an unauthorised action.
3. Where can a human step in? A journey record should not be a sealed black box delivered only after disaster strikes. You should be able to pause the journey, inspect the steps taken so far, and revoke active permissions with a single tap.
Why receipts matter when agents deal with agents
Today, an assistant mostly interacts with websites built for human eyes. Tomorrow, assistants will coordinate directly with other assistants across machine channels, a shift we explore in Why the internet needs a second road.
Your calendar assistant will negotiate with a clinic scheduling assistant. That clinic assistant will coordinate with a pharmacy inventory agent. In a chain of automated handoffs, no human being sits watching every second.
If something breaks, who was responsible? Without signed, step-by-step receipts, every system can blame the next. The clinic claims your assistant requested midnight; your assistant claims the clinic confirmed noon. When every participant is software, the signed journey log is the only credible witness.
What this means
When an assistant takes ten steps on your behalf, you need a checkable record of each step. You can spot mistakes where they happened rather than guessing from a final score.
What you should ask today
You should not need a degree in machine learning to trust an automated assistant. Before trusting software with your errands, ask a straightforward question: When this makes a mistake, will you show me the steps it took, or just tell me to trust its score?
If a provider cannot show you the journey, you are flying blind. Keeping a checkable trail of every step is simply how we ensure that as assistants grow more capable, they remain honest, accountable, and firmly under your care.
1. Sindhuja Chaduvula, Jessee Ho, Kina Kim, Aravind Narayanan, Ahmed Y. Radwan, Mahshid Alinoori, Muskan Garg, Dhanesh Ramachandram, and Shaina Raza, From Features to Actions: Explainability in Traditional and Agentic AI Systems, arXiv:2602.06841 (February 2026).
