Tech Reads
Practical reads on quality management, ERP projects and AI operations: the subjects we build products and run projects in.
Plenty of quality teams run a certified QMS on spreadsheets, a shared drive and one analyst's statistics file. Here is what that gap costs, and what closing it looks like.
ISO 9001 audit readiness fails at the evidence gap. Here is what auditors ask for clause by clause, and how to prove the work continuously.
Lean Six Sigma tooling is siloed: the charter in a doc, the stats in a separate package, the tracker in a spreadsheet, the CAPA in email. Here is what DMAIC in one place looks like.
Native quality dashboards read live QMS data, drill into the source records, and sit beside SQL, Python and R, with no export step in between.
Document control is where most ISO 9001 systems quietly fail clause 7.5. What controlled documented information actually requires, and why a shared drive cannot deliver it.
AI in quality management works when analysis lives beside your records, not when a chatbot is bolted onto a database and called intelligence.
Five ways enterprise AI projects fail, and how to prevent each: the confidence collapse, the scope creep spiral, the reversibility trap, the adoption cliff and the vendor dependency trap.
The ERP integration patterns that hold up at scale, whatever the platform, and the one pattern that corrupts every system it touches.
Many AI system failures happen at the handoff. An item approved in 20 seconds on a Friday afternoon was not reviewed. What works: right-sized context, delivery timing and reversibility windows.
Payment history and news sentiment miss most vendor problems. Three behavioral signals to track instead: RFQ response time variance, invoice discrepancy rate and communication changes.
The Odoo 17 changes that matter in production: spreadsheets on live ERP data, the website rewrite, what the AI features are worth, and why 17 is the first version that looks ready for multi-entity operations in MENA.
Function calling sounds cleaner until the model calls a function it shouldn't have. Structured output sounds simpler until a nested array comes back malformed. When each breaks, and which to use.
A three-week integration becomes a ten-week project, with seven weeks spent on the debt around it. Technical debt is a cost accounting problem that nobody itemizes.
A $2.4M capex gets approved without committee review because a threshold hasn't been updated since 2019. Approval chains fail from threshold decay, delegation gaps and workarounds nobody talks about.
How an ERP go-live slips three months. Nothing technical causes it. It starts with UAT designed to find zero issues and a data migration tested on 5% of the real data.
Infrastructure metrics tell you the pipes are working. They say nothing about whether the outputs are correct. Correction rate does.
An AI workflow update goes out on a Tuesday night and breaks the invoice approval queue for 8 hours. The prompt and the model are fine. A 200-token system prompt addition pushed one document type over the context limit, and nothing detected it. The patterns that prevent this.
Four people, ten days, every quarter: 160 person-days a year on compliance reporting. The data lives in five systems, the reconciliation logic lives in spreadsheets, and the audit trail gets rebuilt every time. That is a data architecture problem, not a staffing one.
A manufacturer on Odoo asks why scrap costs are missing from the reports. Nothing is broken. Scrap was never being recorded, and three months of waste cost never reached the books.
A test suite can pass at 100% while the model gives wrong answers in production. Standard QA tests structural correctness. For AI systems, structural correctness and behavioral correctness are separate problems.
Build or buy enterprise AI? Budget is the wrong first question. What matters is whether the capability is core to how the company competes.
Three years after go-live, the API decisions that hurt most are the ones made in an afternoon without realizing they were decisions. Each one is expensive to fix across a live system.
Automate a procurement workflow well and processing time falls sharply. Automate one step too many, and the system picks a supplier with a 3-week lead time for an urgent order because the price was right.
A 12-step procurement workflow runs cleanly through step 6. At step 7 it starts approving vendors it rejected at step 3. The context window is full, and the agent can no longer see its own earlier decisions.
Fourteen screens to process a standard supplier delivery. The features are all there. The software was designed by people who never had to use it at 3pm on a Friday with three trucks backed up.
Some automation projects should be stopped mid-build: the technology works and the economics do not. A framework for deciding when to walk away.
An AI vendor plans to treat a 2009 SOAP API as a read-only data source. The project dies when everyone realizes that posting approved invoices back to the ERP goes through the same API, and nobody designed for that.
A 12-step procurement workflow got to step 9, with the vendor selected, the PO drafted and the approval routed, and then the orchestrator crashed. The LLM was fine. The prompts were fine. Three hours of work vanished because nobody designed what "resume from step 9" meant.
Why a pipeline that looks cheap per call gets expensive in production, and the batching, caching and routing changes that bring the bill back down.
Vendors promise 80% reductions in AP processing time. The claim is technically accurate and practically misleading. It applies to the invoices that were already closest to automated.
A system can have logs, dashboards and approval workflows and still be unable to answer an auditor. Logging and auditability are different things.
Fine-tuning can cost tens of thousands of dollars before it beats a good prompt. The cost math, the cases where it wins, and three gates to clear first.
Eleven questions to put to an ERP vendor before you sign: license tiers, total cost, upgrades, data export, support terms, localization and year three.
Plenty of enterprise model choices are made on a demo and a leaderboard rank. Real LLM evaluation needs your own documents, your failure modes and a few tests that usually get skipped.
Some companies stay on an old ERP version for years because every upgrade would break their custom code. Here is how that happens, and where to draw the line between configuration and customization.
Most RAG tutorials get you to a working demo in an afternoon. The jump to a system that reliably serves enterprise users at scale is where most implementations quietly fail.
ERP data migrations fail on data quality problems that look like technical ones. The data is the problem, and fixing it takes longer than building the new system.
Ask one LLM to run a complex enterprise workflow end to end and it will lose context somewhere. Here is why we split the work across narrow agents, each with one job.
Extraction is the easy part of a document-to-ERP pipeline. The projects stall on ERP permissions, schema mapping and the three-way match.
One hallucinated invoice total can end up in a payment run. These are the seven layers we design into financial automation, from structured output to drift detection.
A cheaper ERP license saves money in year one. If the API is poor, integration workarounds can eat that saving and more. The module selection can be fine while the API is not.
Where Odoo and Business Central each win, where each one will frustrate you, and the upgrade and licensing traps neither vendor documents properly.
Automation ROI models tend to count headcount savings that never happen. The real numbers come from cycle time, error reduction and growth capacity. Here is the honest math.
Normal code review checks correctness. AI systems also need review of prompt injection surfaces, failure mode coverage, output validation and observability.
A worked scenario: an AI agent starts routing invoices to the wrong approval tier. What failed, what a review should have caught, and the checks that prevent it.
Every AI vendor looks good in a demo. The questions that matter are about what happens when the model is wrong, what the audit trail looks like, and whether you can talk to a real client.
Vendors promise 10x ROI. An honest model counts implementation, the parallel-run months, maintenance, correction time, and only the headcount you really reduce.
Enterprises tend to deploy AI first and discover the compliance questions later. Four questions on data residency, auditability, human override and regulatory change to answer before go-live.
MENA enterprises run younger ERP stacks, have multi-entity complexity that forces good data discipline, and operate in a regulatory environment that rewards auditability. These are structural AI advantages.
In-context, vector and structured fact memory each fail in a different way at scale. Here is the failure map for all three.
Nine areas to test before an Odoo 16 to 17 upgrade, from custom JavaScript and QWeb reports to Python dependencies, OAuth clients and the customer portal.
The 80% reduction claims for invoice automation can be real, but only after solving problems the demo never shows. Five of them, in order of how badly they hurt.
A prompt injection attack can run in a document processing agent for days before anyone notices. Your existing security framework was not built for this. Here is the AI attack surface.
AI spend runs through IT but the ROI lives in operations. Most budgets miss three of the five real cost categories. Here is how to structure the conversation so you fund the right things.
An agent with 6 tools and an agent with 43 have different problems: tool selection errors, compounding latency and a growing permission surface. What to expect and how to fix it.
Every ERP ships with hundreds of standard reports, and management teams use a handful. The rest of the value is trapped in Excel workarounds that nobody fixes, because nobody owns the reporting gap.
A forecasting model with every input anyone can think of does worse than one with a dozen well-chosen inputs. More data makes the model worse. Here is what predicts cash flow over 13 weeks.
AI systems fail at the data layer far more often than at the model layer. Four pipeline failure patterns cover most of what goes wrong.
AI governance frameworks are sized for Fortune 500 companies. Here is what a 200 to 500 person company needs: a system register, three risk tiers and an incident response document.
A prompt change can improve average accuracy and still cause a regression on tail cases that runs for weeks before anyone catches it. Here is the discipline that prevents this.
Most data problems in multi-company Odoo setups trace to the same four mistakes: intercompany rules, shared records without ownership, user access, and tax setup not validated per entity.
We would decline a resume screening project and build an onboarding automation. The line between HR AI that saves time and HR AI that creates discrimination liability is clear once you know where to look.
An AI deployment can work perfectly and still fail on adoption, with the team processing by hand as before. Five resistance patterns, and what works against each.
Adding a second model adds latency, cost and failure modes. Sometimes it clearly pays for itself, and sometimes a better-prompted single model does the job. Here is how to tell.
Standard ERP field service modules cover the core workflow and only part of what the business needs. The gaps are predictable: scheduling optimization, mobile UX and contract billing.
MCP standardizes AI tool connectivity and cuts the development work per integration. Per-tool authorization, input validation and audit logging are still yours to build.