Field notes
Case studies and working notes from the automations we ship: what the numbers actually were, what broke, and what we would not automate again.
Brokers now let AI agents like Claude into live accounts. What the agent can really do, which safeguards matter, and how to test one in stages.
We ran TypeSafe's Jev on 1,370 labeled examples. Above 0.9 confidence it was right 91-94% of the time. Below that it was overconfident. What to automate.
Vibe coding prototypes a workflow in an afternoon. Running it on real invoices or customer data needs engineering discipline it wasn't built with.
Your first AI agent pilot should prove you can govern one, not automate it. How to pick the process, and what to build before scaling to higher-stakes work.
Small businesses lose $15-25K on a first AI vendor that doesn't work out, and 38% can't even export their data. The exit questions to ask before you sign.
You cannot specify an AI system in advance. Ship a walking skeleton, run it in shadow mode, then human-in-the-loop, and measure three numbers — automation risk, catch rate, review rate — before letting it run on its own.
Estimate back-office automation ROI with three numbers: hours per month on the workflow, loaded hourly cost, and the package price — Spinner from €2,500, Bottlenose from €5,000. Measured reference: ~58 hours/month given back.
A serious AI audit leaves you four things you can act on without the vendor: a process map, a gap analysis, an ROI estimate with its assumptions written down, and a build plan. If it never says no, it is a sales document.
Zendesk and Sierra just declared per-seat pricing dead. Before you sign an outcome-based contract, here's what "resolution" quietly means — and doesn't.
Gartner: customers are 3x more likely to ask ChatGPT than a company chatbot. What that means for whether your support AI is worth building at all.
98% of companies have deployed AI somewhere in customer service, but only 15% connect it across systems to resolve issues end-to-end. What closes that gap.
Gartner: half of companies that cut support staff for AI will rehire by 2027. What Klarna's reversal shows, and what to measure before you cut headcount.
The internal IT helpdesk is an easier automation target than customer support — known requesters, repeating request types, systems you own — and a riskier one, because the agent can act. Where the line between answering and acting belongs.
A well-built AI support system resolves 50-70% of tickets on the email and ticket channel — not the 90% in the demo. What decides where you land, how to estimate your own ceiling from 200 real tickets, and why deflection is not resolution.
AI ticket triage makes four decisions on every request — what it is, who owns it, how urgent it is, whether it can be answered now — and writes them into your existing helpdesk. The architecture, the escalation rules, the metrics.
You can automate email-to-ticket workflows inside your own cloud account: N8N plus a model you control reads, classifies and turns emails into structured tickets. In production: 85% of tickets auto-created, ~58 hours/month given back.
Every guide compares outsourcing AP to a BPO against buying automation software. Both skip the option that fixes what each one leaves out.
Vendor invoice management, AP automation, invoice processing: same problem, different names. The question that predicts fit isn't the label.
Off-the-shelf AP tools stop at the 10-20% of odd invoices. Building in-house costs $140K+ and a team to run it. The option between the two.
New CFO survey: 45% still review AI-processed invoices by hand. Extraction isn't reconciliation — the four layers a genuinely touchless AP flow needs.
60 mid-market CFOs say fraud controls delay good payments — firms with recurring friction lose 1.92% of revenue, 6x more than peers.
Insurers are adding AI-specific exclusions in 2026. What voids coverage when an agent approves a bad payment, and the control that keeps you covered.
Six recurring leaks in accounts payable: duplicate payments, price and quantity drift, unapplied credit notes, lost early-payment discounts, supplier bank-detail fraud, contract-price drift. And how to size your own leak first.
Start with accounts payable document flows — high volume, rule-based, checkable within a day — and leave forecasting for last. A ranked shortlist for mid-size finance teams and the four questions that rank any candidate.
Three-way matching — invoice against purchase order against goods receipt — is one of the few finance workflows automation can take end to end. Why matching fails on normalisation rather than OCR, and the pipeline that works.
Automating invoice processing end to end takes four layers: OCR, extraction, validation, and reconciliation against orders and delivery notes. Built in production: 10,000+ invoices, 100% automation, real fraud found.