Blog
Notes from production
Our engineering teams on systems in production: what held, what we changed and why.
The CIO's Test for AI Governance
Strategy5 min
An AI policy says what agents should do. Their accounts and tools decide what they can do. A three-question test for every policy line shows a CIO which parts of AI governance hold, and which exist only on paper.
Read the article- The Map Is the ArchitectureThe process map drawn in the first week, every handoff between a person, a system and a document, is usually filed and forgotten. We build from it: each handoff becomes a written agreement, each step is marked as reading, checking or deciding, and the running system reads the map. That is what lets a company add a market in weeks, in any industry, and it fixes where AI belongs.Subash Natarajan5 min
- Where ERP Agents StopThe AI agents now shipping inside Microsoft's Dynamics 365 and Business Central are useful and deliberately limited. They prepare and recommend inside one business system; entering transactions into the books, approving them and anything that crosses systems stay with you. How to use them, and what is still worth building.Kris Payne5 min
- Who Owns the System When the Builders LeaveMost handovers transfer documents. Ownership means the client's team can change the system safely without calling the people who built it. The quality of a handover is set by the smallest change the team is afraid to make.Alexandru Bene5 min
- Capacity, Not HeadcountA steering committee asked how many full-time posts an order desk automation would save. The better question was how much demand the desk turned away. Why cutting people caps growth, why knowledge and judgement are worth more as answers get cheap, and how that changes what we automate, measure and decline.Subash Natarajan5 min
- Design for Close WeekFinance automation is sized on the average month and judged at month-end, when volume, deadlines and scrutiny peak together. How the month-end calendar should set capacity, scheduling, the rules for late documents, change freezes and the first month run side by side with the old way.Kris Payne5 min
- The Decision Card: Designing Human Approval for AI in FinanceFifty-five approvals in four minutes is a signature, not a control. How we design the card a treasurer sees when an AI system stops for a person, which cases may be approved together, and how we measure whether review is happening.Alexandru Bene5 min
- Model Retirement: Your Lock-In Lives in Your Test SetEvery AI model a workflow depends on will be retired by its provider, and the replacement will be better on average and different in the cases that matter. How tied you are depends on whether you own the real cases that can test a replacement. How we fix the version, test a replacement on what it gets newly wrong, and write the way out into the handover.Subash Natarajan5 min
- How Good Systems Earn Their Users' TrustThree weeks after go-live, a senior warehouse lead was re-checking every delivery the system had confirmed against a printed list. Getting a team to use a new system means getting their trust to match what the system deserves, and the number that moves it is how fast the system changes when an expert corrects it.Alexandru Bene5 min
- Price the ExceptionsAutomation rate is the number on the steering committee slide. The cost, the risk and most of the skill sit in the cases that are left over, and growth adds new kinds of exception faster than it adds volume. The table of exception costs we keep for every workflow, and the payback rule for new automation.Kris Payne5 min
- Where an AI Agent Should Keep Its StateAn AI agent works with three kinds of information: the company's main systems, the case file of the work, and what the model has in view for the current step. Most failures we are asked to review come from treating the third as one of the first two, and from plans that go out of date while the agent works.Subash Natarajan5 min
- Tool Design as a Control for AI Agents in ERP SystemsThe instructions say what an AI agent should do. Its tools, the actions it is allowed to take, decide what it can do. We build those actions as narrow business steps with checks, not as open access to a system, and publish them as a list an IT lead and an auditor can read.Alexandru Bene5 min
- Regulated AI Is an Operating ProblemWhere data may be processed, what each step may see, how changes are approved and what evidence exists afterwards decide whether AI may run in a regulated operation. They change the design, not only the paperwork, and two of them pull in opposite directions.Subash Natarajan5 min
- Autonomy Is Earned One Decision Class at a TimeSponsors ask when the approvals can be switched off. Autonomy is not a setting on an AI agent. It works like a credit limit on each type of decision: granted on evidence, enforced by what the system lets the agent do, and cut back by rule.Subash Natarajan5 min
- A Posting Protocol for AI in the General LedgerMatching bank lines gets the attention. Writing entries into the accounts decides whether every one of them can still be explained. A fixed key per entry, a check on the accounting period, one front door, and the rule that automation never corrects its own entries.Kris Payne5 min
- Waiting, Touch and Rework: Measuring What AI Changes in OperationsHours saved is usually the least reliable number in an AI business case, because it counts waiting as work. How we build a value statement a CFO can defend a year later: waiting, hands-on work and rework measured apart, a step-by-step rollout as the comparison, and no revenue claim without a measured link.Alexandru Bene5 min
- Data Contracts for ERP ExportsAn AI model will read a changed data file without complaint, which is exactly why it must never be the thing that checks data coming in from another system. The written agreement we check every delivery against: shape, meaning, completeness, timing and owner, with a failed delivery held back whole instead of repaired, and an agreement that later becomes the brief for replacing the system.Subash Natarajan5 min
- A Confidence Score Is Not a Probability of Being RightThe number a model reports next to its answer is not the chance that it is correct, and the case types it is most reliable on are where its worst errors hide. We replaced confidence cut-offs with a permission per case type, earned on reviewed cases and withdrawn automatically.David Grayhurst5 min
- Intelligence Is Getting Cheap. Accountability Is Not.The price of an AI answer is falling faster than any company can plan around. The cost of standing behind that answer is not falling at all. Where lasting value in company AI will sit, and what that means for the business case, the design and the vendor contract.Anubhai Mehta5 min
- Organise AI Around Workflow OwnersThe AI centre of excellence was built around the scarce skill of 2023: knowing how models work. That skill is becoming cheap. What stays scarce is an owner accountable for a workflow's results and engineers who can connect AI to the systems it runs on. How to reorganise before the pilot list gets any longer.Anubhai Mehta5 min
