Insights
Blog
Field notes on document intelligence, voice agents and building enterprise AI that survives an audit.
60 posts
Character Accuracy Is a Vanity Metric. Measure These Four Instead.
A vendor quoting 99% accuracy is telling you almost nothing useful. Here are the four numbers that actually predict whether a document AI deployment will save money.
What an Agentic Document Pipeline Actually Is
Agentic" has become a marketing word. Here is the concrete architecture behind it for document processing — what each component does, and why one big model call is not the same thing.
What Straight-Through Processing Rate Really Tells You
STP is the single most quoted number in document automation and the easiest to inflate. Here is how to read it honestly, and why the STP-versus-error curve matters more than any single figure.
The State of Handwriting Recognition in 2026
Handwriting was the wall that stopped a generation of digitization projects. Vision language models moved it — but not uniformly, and knowing where the remaining hard cases are decides whether a project succeeds.
Build or Buy Document AI: An Honest Decision Framework
Foundation model APIs made the first 70% of a document pipeline easy to build, which is exactly why so many in-house projects stall. Here is where the real work is, and how to decide.
The Latency Budget of a Voice Agent
Sub-second response is the difference between a conversation and an interrogation. Here is where the milliseconds actually go and which ones you can get back.
Prompt Injection Through Documents Is a Real Attack Surface
If your pipeline reads documents supplied by outside parties and acts on what it finds, those documents are untrusted input reaching a model that follows instructions.
Designing a Review Queue People Can Actually Work
The review interface determines the unit economics of document automation more than model accuracy does. Here is what separates a queue that takes seconds per exception from one that takes minutes.
Build the Evaluation Harness Before the Pipeline
Without a scoring harness, every prompt change is a guess and every model upgrade is a gamble. Here is how to build one for document extraction, and what to put in it.
ACORD Forms Are Standardised. Your Inbound ACORDs Are Not.
A standard form should be the easy case for document AI. In practice ACORD submissions arrive in a dozen states of completeness and fidelity — here is how to handle the reality.
Where Document AI Sits in the Emerging AI Regulation Landscape
Most document extraction is low-risk under the frameworks now taking shape. The obligations attach to what you do with the output, which is an argument for a specific architecture.
The Security Questions to Ask an AI Vendor That Standard Questionnaires Miss
Your existing vendor assessment covers encryption and access control. It probably does not cover model providers, training use, or what happens to your data inside an inference call.
Ambient Documentation and Dictation Solve Different Problems
They are often evaluated as competing products. They have different failure modes, different oversight requirements, and different specialties where each is clearly right.
Scan Once, Extract Many Times: Capture Standards That Do Not Age
Capture settings chosen for today's extraction pipeline will constrain every future one. Here is how to scan so the images outlive the software that reads them.
Bank Statement Analysis Is a Classification Problem, Not an Extraction Problem
Pulling transactions off a statement is the easy part. Deciding what each one means — income, transfer, business, recurring obligation — is where the accuracy actually matters.
Bills of Lading Are Legal Instruments, Not Just Data
Extracting fields from a transport document is straightforward. Understanding which document is the original, what it entitles the holder to, and what its clauses mean is not.
Extracting Clinical Data Without Creating a HIPAA Problem
The engineering decisions that determine whether a clinical document pipeline is compliant are made early, in architecture rather than in policy. Here are the ones that matter.
Closing Document QC Is a Consistency Problem
Most closing defects are not judgement calls. They are dates out of order, figures that disagree between documents, and signatures that are missing — all of which a machine can check on every file.
Commercial Invoices and Packing Lists: The Line-Item Problem
Trade documents look structured and are not. The header fields extract easily; the line items are where the value is and where the difficulty is.
Submission Intake Is the Real Bottleneck in Commercial Underwriting
Underwriters spend a large share of their time assembling submissions rather than underwriting them. Automating intake raises quote capacity without touching a single pricing decision.
Confidence Scores Are Not Probabilities (Until You Make Them One)
The number a model attaches to an extraction is rarely the probability that it is correct. Here is how to calibrate confidence so your review thresholds mean something.
Why Customs Filing Takes Hours and What Actually Removes Them
The declaration form is not the bottleneck. Reconciling a document set that disagrees with itself is — and that is the part worth automating.
Data Residency When Your Model Provider Is Somewhere Else
A vendor hosting in your region may still send your documents through an inference endpoint in another. Here is how to establish where processing actually happens.
The Real Total Cost of a Document AI Deployment
Per-page pricing is the smallest line in the budget. Here is the full cost model, including the four categories that consistently get left out of business cases.
Classification and Splitting: The Step Everyone Skips
Most document AI failures happen before extraction, when a 40-page bundle is treated as one document or routed to the wrong parser. Here is how to build the step that decides everything downstream.
Automating FNOL Without Losing the Plot
First notice of loss is the highest-leverage automation target in claims and the easiest to get wrong. Here is what to automate, what to leave alone, and how to measure it.
Freight Operations Run on Exceptions. Automate the Detection, Not the Resolution.
Most logistics document work is comparing what was planned against what happened. The comparison automates well; deciding what to do about the difference does not.
Claims Attachments: The Unstructured Half of Healthcare Payments
Healthcare claims are largely electronic. The documentation that supports them is not — and that mismatch is where a great deal of payer and provider effort disappears.
Tariff Classification Is a Legal Determination, Not a Text-Matching Task
Automated HS code assignment is tempting and frequently wrong in ways that surface years later. Here is what to automate and where to keep a person.
Income Documents Are Where Lending Automation Gets Difficult
Pay stubs, tax returns and profit-and-loss statements resist automation for a specific reason: extracting the numbers is easy, and deciding which numbers count is not.
Model Governance for Insurers Deploying Document AI
Regulators are asking insurers to explain, monitor and document the AI systems in their operations. Here is what that means concretely for a document processing deployment.
Claims Document Triage: Sorting the Pile Before Anyone Reads It
A claim file accumulates dozens of documents from a dozen sources. Getting each one to the right place at the right time is a bigger win than extracting every field perfectly.
KYC Document Verification Beyond Reading the Fields
Extracting a name and date of birth from an identity document is the easy half. Establishing that the document is genuine, current and belongs to the person presenting it is the other half.
A Practical Playbook for Land Record Digitization Programmes
Most land record programmes fail on sequencing and scope rather than on technology. Here is the order of operations that keeps them on schedule.
Your Reviewers Are Generating Training Data. Are You Keeping It?
Every correction a human makes is a labelled example produced by work you already paid for. Most pipelines throw them away. Here is how to capture and use them.
Medical Record Review: Find and Cite, Do Not Summarise
Hundreds of pages of records, one clinical question, and a reviewer with twenty minutes. The right architecture retrieves and cites evidence rather than generating a narrative.
Indexing a Loan File at Packet Speed
A mortgage file is 300 to 800 pages of forty document types in no reliable order. Getting the stacking order right is a prerequisite for everything downstream — and it is automatable.
Validation Patterns That Catch What the Model Missed
The cheapest accuracy improvements in document AI are not model upgrades. They are rules — arithmetic, cross-references, and reference-data lookups that turn a guess into a verified value.
Multilingual and Multi-Script Documents in One Pipeline
Real archives and real enterprises produce documents with two scripts on one page, transliterated names, and language that switches mid-sentence. Here is how to build for that rather than around it.
Why Pathology Dictation Is a Harder Speech Problem Than It Looks
General-purpose transcription fails on pathology for specific, fixable reasons — vocabulary, structure, and the fact that a single misheard word changes a diagnosis.
When to Trust a PDF's Text Layer and When to Look at the Pixels
A PDF that contains selectable text seems like a solved problem. Often it is — and the cases where it is not will silently corrupt your data.
Automated Redaction Is Asymmetric, and That Changes the Design
Over-redacting is correctable. Under-redacting is not. Every design decision in a redaction pipeline should follow from that asymmetry.
Prior Authorization: Automating the Packet, Not the Decision
Prior auth is slow because assembling and reviewing clinical evidence is manual. Automating that assembly cuts days out of the cycle without touching medical necessity determination.
Private Deployment Options for Document AI, and What Each Costs You
Nothing leaves our environment" is achievable. It is not free, and the trade-offs are worth understanding before it becomes a hard requirement in a procurement document.
Treat Prompts Like Code, Because They Are
A prompt edit is a production change with non-local effects and no compiler to catch mistakes. Here is the release discipline that makes changing them safe.
Clearing a Public Records Backlog Without Adding Headcount
Records requests, permit applications and case files pile up for the same structural reason. Document automation helps, but only if applied to the step that is actually the constraint.
The Document Work That Starts After the Loan Closes
Servicing generates decades of inbound correspondence, and most automation attention stops at closing. Here is what the post-close document workflow looks like and where it pays.
What SOC 2 Does and Does Not Tell You About an AI Vendor
The report is a genuine signal about operational discipline. It says almost nothing about whether the AI works, and buyers routinely read more into it than it contains.
Finding Subrogation Opportunities Nobody Has Time to Look For
Recovery potential is usually visible somewhere in the claim file. The problem is that reading every file is not economic — which is exactly the kind of problem document AI is suited to.
Why Tables Are Still the Hardest Part of Document AI
Models that read prose reliably still fall apart on merged cells, multi-page rows, and columns whose meaning is defined three pages earlier. Here is why, and what actually helps.
How to Evaluate a Voice Agent Properly
Word error rate tells you almost nothing about whether a voice agent works. Here is the evaluation stack that predicts how it will perform with real callers.
Turn-Taking Is What Makes a Voice Agent Feel Human
Interruption handling, backchannels and knowing when someone has finished speaking do more for perceived quality than any improvement in what the agent actually says.
Appointment Scheduling Is the Best First Voice Agent Deployment
Bounded scope, clear success criteria, immediate measurable value, and a low cost of failure. Here is how to build one that works and what makes it harder than it looks.
Disclosure, Consent and Recording for Voice AI Deployments
The compliance requirements around automated calling, recording and AI disclosure are jurisdiction-specific and moving. Here is the framework to design against.
What Hotels Actually Need From a Voice Agent
Front desks miss calls because staff are with guests. The fix is not a smarter chatbot — it is an agent that handles the small number of calls that make up most of the volume.
When Not to Use an LLM Agent
The failure mode of the current moment is using a language model for work that a regular expression, a lookup table, or a form field would do better. Here is how to tell the difference.
Land Records, Reimagined
Land record digitization has stalled for two decades because OCR cannot read hand-written regional scripts and manual data entry does not scale. Vision agents change the math. Meet Axar — a vision-first agent for land records, built for governments, banks, courts, and the organisations downstream of paper registers.
Why Vision Language Models Are Replacing Traditional OCR in Enterprise
Traditional OCR breaks on handwritten text, mixed layouts, and context-dependent fields. Vision Language Models see documents the way humans do — and that changes everything for enterprise document processing.
The Case for Human-in-the-Loop AI in Regulated Industries
Fully automated AI is a liability in regulated industries. Human-in-the-loop AI delivers the volume benefits of automation with the auditability and trust that compliance demands.
From Handwritten Records to Structured Data: Lessons from Digitizing Government Archives
Millions of handwritten government records sit in paper archives worldwide. Here is what we learned building a VLM-powered pipeline to digitize them at scale — and what it means for any document-heavy enterprise.
No posts match that search.