---
title: "Product update: the Logbook, and a confidence threshold you control"
description: "Every agent action as one log line: agent, location, output, confidence, status. Flag a line to feed the agent's learning; set the autonomy threshold yourself."
date: 2026-08-31
author: "The BPA team"
category: "Product updates"
tags: ["product update", "logbook", "confidence threshold", "audit log"]
canonical: https://www.businessprofileagent.com/blog/product-update-logbook
html: https://www.businessprofileagent.com/blog/product-update-logbook
publisher: Business Profile Agent
---
# Product update: the Logbook, and a confidence threshold you control

Every action an agent takes is now one line in a log-file style listing: who did what, where, with which confidence, and what happened to it. A flag on any line sends feedback into the agent's next learning round. And the threshold above which agents act alone is a slider, not a setting we own.

**In short:** The Logbook records every agent action in Business Profile Agent as a single line with timestamp, agent, action, location, output, confidence score and outcome. Any line can be flagged with a reason; the feedback lowers the agent's confidence on similar cases and is reviewed in its next learning round. The confidence threshold above which agents act alone in Autopilot is set per client, with a default of 80%. Logs are retained for at least six months.

Since 29 August, the bottom strip of the workspace has a tab that was not there before: **Logbook**. It opens a listing that looks deliberately like a log file, because that is what it is. One line per agent action. No charts. The newest line at the top, the eight agents colour-coded, and a flag at the end of every row.

We built it because the question every pilot customer asked in the second week was the same: "What did it do while I was not looking?" A dashboard answers "how much". A log answers "what, exactly" — and lets you disagree with a specific line.

## Why a log and not another dashboard

Google's May 2025 paper on secure agents lists three principles, and the third is that "agent actions and planning must be observable" — logging that "enables trust, debugging, auditing". Singapore's January 2026 governance framework for agentic AI says the same in more words: actions must be "traceable and controllable", with logs covering tool calls, access history, reasoning and workflow steps. The EU AI Act's Article 12 requires high-risk systems to "technically allow for the automatic recording of events (logs) over the lifetime of the system". Our agents are not high-risk under the Act. We adopted the wording anyway; it is a good specification.

### What you can do with a line

Read the output as it was sent. See the confidence the agent had and the threshold that applied. See whether it was published, held for approval, or handed to a person. Flag it. Open the location. Restore the snapshot taken before the action. Everything a manager asks after "what did it do?" is one click from the line.

## What one line contains

| Field | Example | Why it is there |
|---|---|---|
| Timestamp | 09:41:07 | Order of events; the API's pacing is visible. |
| Agent | Review Agent | Eight agents, each with its own colour and its own learning round. |
| Action · location | Replied to a 2★ review · Kestrel · Basel | What and where, in the words a manager would use. |
| Output | The reply text, as sent | Never a summary — the customer saw this exact text. |
| Confidence | 76% | Scored by the proof-reading agent on a second model. |
| Status | held for approval · < 80% | Derived from the gear and the threshold at that moment. |
| Flag | ⚑ | Send feedback; the line turns to "flagged · feedback sent". |

## Flag → feedback → learning

MIT's 2025 study of enterprise generative-AI projects put the failure rate at 95% and named the cause: "Most GenAI systems do not retain feedback, adapt to context, or improve over time." The flag is our answer to that sentence. Flagging a line asks one question — what went wrong: wrong tone, factual error, wrong language, too long, should not have replied — and takes an optional note in plain words ("we never promise a callback; point people to the booking link"). The feedback is logged with who and when, lowers that agent's confidence on similar cases immediately, and is reviewed in the agent's next learning round. The output itself stays until you edit it; the log does not rewrite history.

*What the threshold does*

### "Agents act alone from 80%" — yours to move

- **In gear 3, Autopilot**, an action above the threshold publishes and logs; below it, the draft is held for a person with the reason it scored low.

- **In gear 2, Drafting**, the threshold changes nothing about who publishes — a person does — but the score still shows, so you can see what Autopilot would have done.

- **Per client, per action type.** A pharmacy chain can run positive-review replies at 75% and hours corrections at 95%.

**8** — agents logged

Review, FAQ, Proof-reading, Validation, Watch, Content, SEO·GEO·LLM, Policy Guard.

**80%** — default threshold

A slider in the Logbook header.

**6 mo** — minimum retention

Logs are ours; imported Google data is refreshed and never kept beyond 30 days.

## What the customer sees

From 2 August 2026, Article 50 of the EU AI Act applies: people must be told when they interact with an AI system, and AI-generated text published to inform the public must be disclosed unless it "has undergone a process of human review or editorial control" under someone's editorial responsibility. The Commission's guidance says ordinary marketing copy is generally out of scope, and that superficial checks do not count as review. Our default gear is Drafting — a person reads and approves — and Google publishes a reply as the business, without the responder's name. The Logbook is the record that the review happened, by whom, and when.

The customer research points the same way. Zendesk's 2026 survey found 95% of consumers expect an explanation for AI-made decisions; a Gartner survey found 64% would prefer companies did not use AI in customer service at all, mostly for fear of not reaching a person and of wrong answers. Both are arguments for a log that a person reads, a threshold a person sets, and a stop button a person owns.

## What is next

Export of the Logbook as CSV per client and period, for the people who keep their own records. An alert when the share of held drafts crosses a level you choose — the earliest signal that a location's reviews have changed character. And, on request from two pilots, a read-only Logbook link for the client's own compliance team.

## Questions we get on this

**What does the Logbook record?**  
Every action taken by any of the eight agents in Business Profile Agent, as one line: timestamp, agent, action, brand and location, the exact output sent, the confidence score, the resulting status (published, held for approval, awaiting approval, logged, or handed to a person) and whether the line was flagged. The listing updates live as agents work.

**What is the confidence threshold and how do I set it?**  
The confidence threshold is the score above which an agent in Autopilot (gear 3) may publish without a person. It is a slider in the Logbook header, default 80%, and can be set per client and per action type. Below the threshold the draft is held for approval with the reason it scored low. In Drafting (gear 2) the score is shown but a person always publishes.

**What happens when I flag an agent action?**  
You choose a reason — wrong tone, factual error, wrong language, too long, or should not have replied — and can add a note in plain words. The feedback is logged with who and when, immediately lowers that agent's confidence on similar cases, and is reviewed in the agent's next learning round. The published output stays as it is until you edit it.

**Does the customer see that an AI drafted the reply?**  
Google publishes review replies as coming from the business and does not show the responder's name. In the default Drafting gear a person reviews and approves every reply, which the EU AI Act's Article 50 treats as human review under editorial responsibility. The Logbook records who approved each reply and when.

**How long are Logbook entries kept?**  
At least six months, mirroring the retention period the EU AI Act sets for high-risk system logs. The Logbook contains Business Profile Agent's own records of its actions; imported Google Business Profile data is refreshed regularly and is not kept for more than 30 days, in line with Google's API policy.

## Sources

1. [Google — An introduction to Google's approach for secure AI agents (May 2025)](https://storage.googleapis.com/gweb-research2023-media/pubtools/1018686.pdf)
2. [IMDA Singapore — Model AI Governance Framework for Agentic AI (Jan 2026)](https://www.imda.gov.sg/-/media/imda/files/about/emerging-tech-and-research/artificial-intelligence/mgf-for-agentic-ai.pdf)
3. [EU AI Act — Article 12, Record-keeping](https://artificialintelligenceact.eu/article/12/)
4. [EU AI Act — Article 50, Transparency obligations](https://artificialintelligenceact.eu/article/50/)
5. [European Commission — FAQ on transparency obligations under Article 50](https://digital-strategy.ec.europa.eu/en/faqs/transparency-obligations-under-article-50-ai-act)
6. [MIT NANDA — The GenAI Divide: State of AI in Business 2025](https://mlq.ai/media/quarterly_decks/v0.1_State_of_AI_in_Business_2025_Report.pdf)
7. [Zendesk — CX Trends 2026](https://www.zendesk.com/newsroom/press-releases/contextual-intelligence-becomes-the-new-standard-for-exceptional-customer-experience-in-2026/)
8. [Gartner — 64% of customers would prefer companies didn't use AI for customer service (2024)](https://www.gartner.com/en/newsroom/press-releases/2024-07-09-gartner-survey-finds-64-percent-of-customers-would-prefer-that-companies-didnt-use-ai-for-customer-service)
9. [Google Help — Reply to reviews](https://support.google.com/business/answer/3474050)
10. [Google — Business Profile APIs policies (30-day data retention)](https://developers.google.com/my-business/content/policies)

---
Business Profile Agent — Google Business Profiles at scale, with agents that only go as far as you let them. https://www.businessprofileagent.com