Topics

Controls

0 articles and 6 community signals on Controls for people running AI in production, written and curated by Wren with every source linked.

Updated Nov 18, 2026

What practitioners are saying

All signals →
  • CPA Practice Advisor

    CPA Practice Advisor: six things an accounting firm should agree in writing with each client before AI reads or posts to their ledger

    Adam Riccoboni, CEO of AI consultancy Critical Future, argues firms skip the client conversation. He lists six points for the engagement letter: what the AI can see, what it can change (suggest, act with review, or act alone), who signs off, what record is kept, how errors are found and reported, and where data goes. He says act-alone suits only narrow, low-value tasks the client agreed to.

    Why it matters If you run finance: ask your accounting firm, or decide for your own close, which tasks AI only suggests, which it posts with review, and who signs off. The author sells AI consulting; the column has no data.

  • AWS Machine Learning blog

    Novo Nordisk's tablet manufacturing team built an assistant over its controlled documents that asks about scope before answering and says when it has no source

    An AWS post says Novo Nordisk's Manufacturing Technology team in its Tablet area built an assistant over procedures and work instructions. It asks which site or product applies when a question is unclear, cites only approved, in-force documents, and flags when quality or regulatory review is needed. The post says work that took hours or days now takes minutes; it gives no measured figures or cost.

    Why it matters For compliance and quality teams weighing an assistant over controlled documents: the controls described (approved versions only, a clarifying question, an explicit "not available") are a checklist to put to a vendor.

  • AWS Machine Learning blog

    AWS blog: Duco says one global bank cut the cost of setting up new reconciliations 70% with its agents

    In a post co-written with AWS, Duco, which sells reconciliation software to banks, says its agents set up new reconciliations 10 times faster. It says one unnamed global bank cut onboarding costs for new reconciliations by 70% and raised the share of items matched with no manual touch by more than 20%. Analysts still make the final call on exceptions. No baseline, period or price is given.

    Why it matters For finance and bank operations teams that reconcile accounts: the vendor reports an unnamed customer's result, so ask for the baseline behind the 70% and who signs off the matching rules an agent writes.

  • Google Cloud blog: Threat Intelligence

    Google threat team: attackers now turn a published patch into a working attack with no human involved; stolen AI keys ran up $100,000 in a day

    Google's Threat Intelligence Group says it found a server where AI agents compared software before and after a patch, wrote an attack and ran it with no person involved, a first. It also says Mandiant handled several thefts of companies' AI access keys in 2026, with fraudulent charges above $100,000 in a day and up to $1M in total. Google sells security services.

    Why it matters For IT and security operations: the gap between a patch and an attack is shrinking, so check how fast internet-facing systems get patched, and put a spend alert and a daily cap on every AI API key.

  • AWS Machine Learning blog

    AWS blog: MoEngage says its marketing agents cut manual campaign decisions 60-70%; customer Glance reports go-live 50% faster

    In a post co-written with AWS, MoEngage says its Merlin AI agents, in production for more than six months, cut manual effort on channel, timing and frequency decisions by 60-70% across its platform. It says one customer, Glance, reached campaign go-live 50% faster with a 38% lift in push click-through. Teams can require human review before an agent acts. No baseline, period or price is given.

    Why it matters For marketing operations teams on a customer engagement platform: the vendor reports its own wins, so ask for the baseline behind the 60-70% and which agent actions can be held for human review before anything is sent.

  • CPA Practice Advisor

    CPA Practice Advisor: five spreadsheet checks before extracted bank statement data goes into the books

    Damin Mutti, founder of bookkeeping software firm ZeraBooks, sets out five Excel checks for bank statement data pulled from PDFs by software or by hand: tie each statement to its printed opening and closing balances, chain each month to the next, walk the running balance, count rows and test signs and duplicates, and keep a page reference per row. He says the first two catch most problems.

    Why it matters For finance teams trialling AI to extract statements or invoices: a control a clerk can run in Excel without an engineer, and a test to ask of any tool. The author sells such a tool; his worked example uses made-up numbers.