BergLabs Field Notes · Monthly
Field Notes from the operating layer.
Each issue sets one AI failure against one that worked and names the design choice between them. Every claim is sourced, and every issue is published here as readable text as well as a LinkedIn carousel.
- 01Case in pointA failure in the news
- 02Why it happensThe mechanism
- 03The fixOne design choice
- 04Proof pointOne that worked
- 05Term explainedOne AI term
- 06Fact-checked2+ named sources
- Case in point
One AI failure from the last three months.
- Proof point
One deployment that got it right.
- Term explained
One AI term, defined in plain words.
Issues
The latest issues.
- Issue 02 · October 2026
Voice at the front door.
BergLabs Field Notes15 slidesThe phone line is the cheapest place to start enterprise AI and the most expensive place to get it wrong. One AI receptionist that locked a patient out, one police line that cleared its queue, and the routing rule that separates them.
Read Issue 02 - Issue 01 · September 2026
The human in the loop.
BergLabs Field Notes15 slidesOne AI failure that made the news for the wrong reasons, one that did not, and the single design choice that explains both: a supervision layer between the model’s output and the customer’s outcome.
Read Issue 01 on LinkedIn
Human in the Loop
The video series behind the notes.
Short walkthroughs of how the work is made, checked and kept on the record.
- Video 01
From a robot take to a trainable feed
Why robot policies stall between capture and training, and how a managed queue closes the gap.
Plays from LinkedIn, which may set its own cookies. Watch on LinkedIn
- Video 02
Who made this label, and who checked it?
One BergFlow project from setup to batch upload, the annotator queue, reviewer decisions and QC sampling.
Plays from LinkedIn, which may set its own cookies. Watch on LinkedIn
Articles
Long-form notes between issues.
Every article from the BergLabs editorial archive, newest first. Filter by the work it describes or search by workflow.
Showing 44 of 44 articles
- Managed operations8 September 2026
Field-service diagnostic copilots with cited procedures
A diagnostic copilot should identify the equipment version, retrieve the applicable manual and propose the next approved test with its source. Technicians keep repair authority.
BergLabs EditorialRead the note - Commerce8 September 2026
Passenger commerce fulfilment with changing order states
One shared order state joins route eligibility, cutoffs, stock, vendor acceptance and payment, and sends conflicting states to an operator.
BergLabs EditorialRead the note - Finance8 September 2026
Payment reconciliation automation with human review
Deterministic rules handle clear matches; finance reviewers own the ambiguous exceptions, with the settlement and ledger evidence kept behind every decision.
BergLabs EditorialRead the note - Data and annotation8 September 2026
AI-data workforce acquisition for verified capacity
Applicant volume becomes useful only after identity, language, task skill and availability checks show a worker can meet a real production requirement.
BergLabs EditorialRead the note - Managed operations8 September 2026
Workforce dashboards that measure delivery responsibly
Measure task coverage, quality and service risk with the smallest set of work events needed. Desktop activity is a poor substitute for useful output.
BergLabs EditorialRead the note - Data and annotation8 September 2026
Contributor data quality and approved payout workflows
Quality acceptance and payment release are separate decisions: the first checks the contribution, the second checks the payee and the amount.
BergLabs EditorialRead the note - Sales and growth8 September 2026
Evidence-based B2B account scoring with clear reasons
Rank sales attention on signals that keep a source, an observation date and a reason to act, so a seller can inspect the trigger before outreach.
BergLabs EditorialRead the note - Sales and growth8 September 2026
AI search visibility: connect citations to qualified action
Measure source correctness, page readiness and qualified action alongside brand presence, so an AI answer leads somewhere useful.
BergLabs EditorialRead the note - Data and annotation8 September 2026
Multimodal AI data production with traceable review
One traceable record links each licensed asset to its task, instructions, prelabel, reviewer decision and accepted output, with BergFlow as the reference platform.
BergLabs EditorialRead the note - Finance8 September 2026
Collections voice workflows and promise-to-pay controls
A voice workflow can transcribe an authorised call and prepare follow-up; a reviewer confirms amounts and dates and handles hardship and disputes.
BergLabs EditorialRead the note - Legal and compliance8 September 2026
Distributor onboarding with complete approval evidence
Automation assembles an evidence-complete case; an authorised compliance reviewer owns activation and the exceptions.
BergLabs EditorialRead the note - Managed operations8 September 2026
Embedded AI operations pods with accountable ownership
One backlog, a named business owner and a regular review of exceptions, releases and outcomes, measured on service performance.
BergLabs EditorialRead the note - Data and annotation8 September 2026
Specialist vetting and capacity matching for live work
A current capability record, with expiry dates on domain skill, language, security training and availability, matched to a real assignment.
BergLabs EditorialRead the note - Workflow automation8 September 2026
Enterprise AI agent management with bounded authority
Explicit task boundaries, scoped tool permissions and checkpoints, with a stop for approval before consequential actions.
BergLabs EditorialRead the note - Data and annotation8 September 2026
Hybrid AI-human annotation with calibrated routing
Confidence thresholds validated per task and risk class, independent sampling of high-confidence items, and trained review for ambiguous labels.
BergLabs EditorialRead the note - Commerce8 September 2026
AI-assisted seller catalogue quality control
Extract candidate attributes, flag category errors and send ambiguous listings to a reviewer, with reported results labelled as first-party.
BergLabs EditorialRead the note - Commerce8 September 2026
Ecommerce search relevance and catalogue-quality operations
Connect buyer intent, query-product judgements and catalogue corrections to a stable evaluation set.
BergLabs EditorialRead the note - Legal and compliance8 September 2026
Distributor marketing-claims monitoring with review
Capture dated page evidence, compare it with an approved claim library and route possible violations to compliance review.
BergLabs EditorialRead the note - Workflow automation8 September 2026
ETL modernisation with data contracts and dual runs
Inventory old jobs, define versioned data contracts, then reconcile old and new outputs in a dual run before the data owner approves promotion.
BergLabs EditorialRead the note - Sales and growth8 September 2026
Multichannel speed-to-lead with CRM ownership controls
Resolve identity, consent and ownership before responding, so forms, chat and events do not trigger duplicate outreach.
BergLabs EditorialRead the note - CX8 September 2026
Technical-support escalation with reproducible evidence
Automation assembles the version, redacted evidence and a reproducible problem statement; a support engineer verifies severity and scope.
BergLabs EditorialRead the note - CX8 September 2026
AI customer operations with verified resolution
Keep identity, context and case state across channels, verify the issue was resolved, and hand off disputes and consequential changes.
BergLabs EditorialRead the note - Workflow automation8 September 2026
AI workflow evaluation with end-to-end failure tests
Release decisions need scenario-level evidence that a full run reaches the right state when tools, policies and handoffs fail.
BergLabs EditorialRead the note - Commerce8 September 2026
SKU competitive benchmarking with comparable evidence
Match the exact variant, pack size, seller, market and availability, then have an analyst verify material conclusions against captured evidence.
BergLabs EditorialRead the note - Trust and safety8 September 2026
Content moderation with policy evidence and appeals
A versioned policy, preserved context and an appealable decision, with models triaging and trained reviewers handling context.
BergLabs EditorialRead the note - Finance8 September 2026
Accounts-payable automation with payment authority preserved
Reconcile invoice, purchase-order and receipt evidence before preparing a payable; finance keeps authority over exceptions and payment release.
BergLabs EditorialRead the note - Workflow automation8 September 2026
Document intake automation with a controlled exception queue
Turn an authorised email, upload or scan into a validated case with field-level provenance, and route conflicts to review.
BergLabs EditorialRead the note - Trust and safety8 September 2026
Image, video and landing-page quality in one review loop
Evaluate the media and the destination together, in one case record, with contextual review for consequential decisions.
BergLabs EditorialRead the note - Trust and safety8 September 2026
Marketplace moderation that preserves legitimate supply
Combine content, seller context and policy evidence to prioritise harm while keeping legitimate inventory live.
BergLabs EditorialRead the note - Commerce8 September 2026
AI catalogue generation with verified product facts
Listing copy generated from verified product facts, with human review for unsupported claims and variant conflicts.
BergLabs EditorialRead the note - Commerce8 September 2026
Seller onboarding for verified listing readiness
Onboarding should end in a verified readiness state, with an authorised reviewer approving activation.
BergLabs EditorialRead the note - Commerce8 September 2026
AI property image selection for accurate booking pages
Rank booking photos while reviewers check room coverage and whether each image reflects what a guest can book.
BergLabs EditorialRead the note - Sales and growth4 September 2026
Can AI search find your brand? An SEO, AEO and GEO audit
Follow the buyer from the question to the cited source, the landing page and a qualified next step.
BergLabs EditorialRead the note - Research21 July 2026
When do more AI voices hurt?
A controlled study of multi-LLM deliberation: a strong best-single selection baseline can beat more complex protocols at lower cost.
BergLabs EditorialRead the note - Data and annotation14 February 2026
Why expert data operations is becoming a core B2B capability
The bottleneck in AI has moved from compute and algorithms to data, and to the teams that produce it well.
BergLabs EditorialRead the note - Data and annotation14 February 2026
Annotation quality at scale: building reliable QA across millions of labels
How sampling, calibration and review keep label quality stable as volume grows.
BergLabs EditorialRead the note - Finance14 February 2026
AI payment reconciliation for PropTech and BFSI
Where AI helps reconciliation in PropTech and BFSI, and where finance review stays in the loop.
BergLabs EditorialRead the note - Research13 February 2026
From data annotation to enterprise intelligence: the BergLabs journey
How managed data work grew into an operating layer for enterprise workflows.
BergLabs EditorialRead the note - Trust and safety13 February 2026
Building trust and safety at scale: AI-powered content moderation
A guide to combining model triage with trained human review in moderation programmes.
BergLabs EditorialRead the note - Data and annotation13 February 2026
Data labelling in 2026: trends, tools and human-AI collaboration
Where labelling work is heading, and how people and models split it.
BergLabs EditorialRead the note - Commerce13 February 2026
E-commerce search, rebuilt around meaning
From keyword matching to semantic understanding, and the relevance data that makes it work.
BergLabs EditorialRead the note - Robotics13 February 2026
Manipulation is not pick-and-place: why robotics needs better training data
Why dexterous manipulation needs richer, better-reviewed training data than pick-and-place tasks.
BergLabs EditorialRead the note - Data and annotation13 February 2026
RLHF, explained: a practical guide
A practical guide to reinforcement learning from human feedback and the review work behind it.
BergLabs EditorialRead the note - Data and annotation13 February 2026
Why AI projects stall, and how to fix your data pipeline
The data-pipeline causes behind stalled AI projects, and the fixes that move them forward.
BergLabs EditorialRead the note
No article matches that search.
About the series
What Field Notes is.
A monthly issue on enterprise workflow automation and AI transformation in 2025 and 2026.
Get each issue by email: write to partnerships@berglabs.ai with “Field Notes” in the subject line.
Each issue also runs on LinkedIn as a carousel, alongside our Human in the Loop video series.
01Public cases and our own work
Some issues analyse recent public workflows against named sources; others turn what we have seen in BergLabs delivery work into a note. Single-source figures are labelled.
02One mechanism, explained
The failure and the success are chosen because one design choice separates them. The issue explains that choice, then shows where it runs in BergAuto and BergFlow.
03Published as text
The carousel goes out on LinkedIn; the full issue is published here as text, with every source line, so it can be read, searched and cited.
Questions
Questions about Field Notes
What it is, how it is checked and how to get it.
Field Notes
What is BergLabs Field Notes?
A monthly issue on enterprise workflow automation and AI transformation. Each one sets an AI failure against one that worked, drawn from public cases or from BergLabs delivery work, and explains the design choice between them. Issues run on LinkedIn and as text on this page.
How are stories chosen and checked?
From two places: recent public cases, checked against two or more independent named sources where possible, and BergLabs delivery work, described without naming clients. Figures that rest on one source are labelled.
Turn the lesson into a pilot.
Every issue ends in the same place: one workflow, scored, reviewed and measured.