The clipboard came back
A new system went live on schedule and worked while the engineers were still on site. Three weeks later the shift log was back to paper, and nobody had switched anything off.
The questionnaire that showed up after the system was built
A customer's security team asks where the data goes, who approved the model, and what happens the day it disagrees. Those are architecture questions. Answering them after the system is built means admitting none of it was decided when it mattered.
This week in AI
Anthropic started trusting its AI to help defend power grids and water systems. In the same week, Wikipedia's operator said it caught AI agents it believes OpenAI built quietly testing whether Wikipedia's own tools could be used to reach other websites, unsupervised. A new large model from Mistral and a new fast-answer tool from OpenAI rounded out a week about how much independence an AI agent should be given.
What good means, before anything is built
A vendor's demo is scored against their own best examples. What matters is how an approach does on the cases your own people argue about — and that has to be settled before you choose anything, not after.
Fewer fields, kept honest
A form with more required fields looks more careful. The version that actually gets filled in is the one with fewer.
Sponsorship at the top, problems from below
A budget with no named problem buys a demo. A problem with no budget stays a memo nobody funds. We will not start a programme until both halves are in the room.
Stuck with last year's model
A better, cheaper model ships somewhere every few months. Moving to it should take an afternoon. For most systems it takes a quarter, because the model was never kept separate from everything built on top of it.
The audit that took three people and a week to answer
A customer's security team asks for every time your AI touched their contract in the past year — who it showed, what it drafted, who signed off. The answer should be one export. Here it took three people a week, because the record was never kept in one place to begin with.
The access that moved slower than she did
A sales rep changes teams on a Monday. Her mail and her CRM role catch up by Tuesday — but the agent that has been drafting her calls keeps reading the old team's pricing sheet for two more weeks, because no system, including that one, was ever told exactly what she was allowed to see.
The clause and the thread that negotiated it
A customer asks why a clause reads the way it does. The answer already exists — the clause in the contract, the reasoning in an email thread months old — just not in the same place. One retrieval index is what turns that into one answer instead of a two-day search.
This week in AI
Four frontier AI labs cut their prices this week, two of them twice over. Amazon and Google both quietly added a hard stop so a cloud AI agent cannot spend past a number you set while nobody is watching. And a second company shipped a free rival to the typed-answer model we explained last week.
The crack the designer never saw
A technician logs why a bracket cracked. Three buildings away, a design engineer draws a similar one with no way of knowing. VEDA, Nirdesh and VCAD are three separate products, and they are only worth building this way if the verdict from one reaches the other two before the part is designed again.
Tiered by consequence
A drafting assistant and a system that decides whether a part passes get the same monthly review, run off the same checklist. The board has forty minutes either way, so the harmless pilot waits six weeks and the one that actually matters gets rushed through with it.
The log nobody can produce
A customer's security team asks for every action an AI agent took on their data last quarter, and who approved each one. The record exists, but only in three systems that do not agree with each other — which is the same as not having one.
The contract that went in for a summary
A renewal call is an hour away and a forty-page supplier contract needs summarising fast. Pasting it into a public chatbot takes thirty seconds, and nobody had ever told this person not to. The pricing inside belongs to a customer who never agreed to let it leave the building.
The list that took two days to build
A customer's security team asks a simple question: which models are running, on what data, approved by whom. Getting an answer takes two days of asking around, because the answer was never written down anywhere. A registry turns that search party into a query.
The deal that stayed in negotiation
A pipeline stage is supposed to say where a deal actually stands. Most of the time it says only when someone last had a spare minute to click it. A maintenance agent proposes what the record says changed — a person still decides whether it did.
This week in AI
An Indian AI lab shipped a genuinely useful speech model this week, then said it would also rent out other companies' models from the same servers — two different claims about sovereignty, easily confused. Firefox began routing its AI assistant through a privacy proxy instead of a server that remembers you. And the cheapest way to make an AI answer showed up not in a smarter model, but in a shorter one.
The question you already asked them
A rep who has to rebuild an account from memory before every call ends up asking the customer something they already answered. A research agent's job is not to sound like it knows the account. It is to know it, with a source under every line.
Filed correctly, and still wrong
A revised purchase order can be saved in exactly the right folder and still be wrong — the number nobody set beside the one it replaced. An intake agent's job is not to file paperwork faster. It is to notice when what just arrived disagrees with what the business already believes.
The call that leaves the script
An outbound calling agent can work every number on a list without tiring. What it must never do is guess its way past the sentence nobody approved it to say — and that turns out to be the whole design problem.
Small enough to change on a Tuesday
Every case study on an AI conference stage belongs to a company with a platform team and a budget line for pilots that fail. We built the same thesis for the business that has neither: one box, one record, and an afternoon to build on it.
This week in AI
Two frontier labs cut prices on the same day, a new kind of AI that answers in a checkbox instead of a paragraph got tested by outsiders within a week of launching, and the tools for running a model on hardware you own kept getting better. None of it changes what we build. All of it says which way the ground is moving.
The system that shut the line down itself
A vendor evaluation turns tense when one system stops the line on its own, no confirmation needed, and the salesman calls it full autonomy. We build the opposite: the one decision that changes anything in the physical world always waits for a person to make it, and to own it.
The vendor could not say which country
A shopfloor AI demo goes well until someone asks where the answer actually came from, and the vendor cannot name the country, let alone the building. We treat that question as the whole point: sovereignty is not a setting added later, it is the condition that decides whether any of the rest is even possible.
Nobody was going to fill in the form
A plant sets out to capture what its longest-serving operator knows before he retires. Months of scheduled interviews later, the folder holds four pages. We stopped asking people to write down what they know, and started writing it down while they did their actual jobs.
Five hires and no use case yet
A company decides to get serious about AI, hires a team, and watches a backlog grow faster than anything ships. We help build the opposite: a handful of senior specialists who ship one use case, then make the second one cheaper.
The record nobody wrote down
A vendor's system keeps running long after the vendor is gone, but the reasoning behind it rarely does, so a small change turns into a support ticket. We build the exit into the engagement: runbooks, the written record of every decision, and a team that does not need to call us.
Forty ideas, none of them first
A workshop produces a long list scored for impact and feasibility, and six months later nothing has shipped because every idea had a defender and none had an owner. We start engagements the other way: two or three named problems, sized, with a written list of what AI is the wrong tool for.
The document that outranked the current one
A shared drive full of drawings and service notes hands back the document that matches your words best, not the one that is still correct. We treat that as a relationships problem, not a search problem, so every answer points back to the exact document, and revision, it came from.
The interview nobody has time for
A role goes out and three hundred résumés come back. A recruiter reads the first forty properly and skims the rest, so the shortlist ends up reflecting who wrote the best document, not who would do the best job. We give the interview to every applicant instead, and leave the decision exactly where it always was.
Two kinds of sure
A predictive score cannot say whether it is certain because the physics say so, or certain because it has seen this pattern before. We build the two kinds of certainty as separate models, on purpose, so a person can tell them apart before deciding.
The score you never checked again
A system was measured once, at the vendor bake-off, and the number went straight into the sign-off deck. Nobody has run that test since. We treat evaluation as something you keep doing, not something you did.
The assistant with no job description
A demo assistant is asked to do five different jobs at once, and when one of them goes wrong nobody can say whose job it was. We build agent ecosystems the other way: a named agent per job, each with a scope, a budget and a person who signs off.
The boundary with exactly one gate
A customer's security team asks where the data actually goes, and the honest diagram has a box labelled the cloud. We draw the boundary first, with exactly one gate, before we choose a single tool.
Where the load actually travels, before the part is made
A bracket comes back cracked at the corner nobody reinforced, while the rest of it is heavier than it needed to be. VCAD shows where the load actually goes the moment the part is drawn, on every part, not just the ones that seemed to matter.
The carbon number a European buyer asks for at quotation time
A quote goes out on price and lead time. Increasingly, a European buyer asks for one more number before signing: the carbon in the part itself. VCAD reads that number off the same drawing that checks whether the part will hold.
Describe the bracket, get the bracket
A mounting bracket isn't hard to design. It's just a day and a half of clicking. VCAD gets a design engineer to a real, editable first draft, so the day goes to the part that's actually in question.
The screen that stopped taking taps
A touchscreen only reads a bare, dry fingertip. A shop floor rarely has one spare. Nirdesh was built for the hand that is already busy.
What the OEM never found out
Most complaints about a manual never reach the people who wrote it. Nirdesh catches the disagreement the moment it happens, not weeks later in an inbox.
The man with both hands dirty
The manual has the answer on page 212. The person who needs it right now cannot hold a tablet, take off a glove, or stop to scroll.
The technician's no
A ranked list is not what makes VEDA useful over time. What a technician does when the top answer is wrong is.
Small, on purpose
A giant model in a datacentre knows a little about everything. A small one on your box knows a lot about the machines you actually own — and it doesn't go quiet when the network does.
The fault nobody can place
Three technicians, three theories, two hours gone. The machine had already said what was wrong — just not anywhere anyone could read it.
Late for a reason in another system
The project tool says the job is on track. It is three weeks late, and the reason is that one man was on leave — a fact the company held all along, in the system nobody thinks of as a delivery system.
The drawing that arrived as a photograph
A dimensioned drawing photographed at an angle. A spec sheet snapped at a trade stand. A whiteboard. All of it real business, none of it in any system — because intake is the gap every tool starts working after.
Eleven deals, four of them dead
Everyone in the room knows the pipeline number is wrong, so nobody trusts it, so nobody updates it. The problem is not discipline. It is that the CRM asked for the wrong thing.
A draft that cites its source
The danger is not that an assistant gets something wrong. It is that it gets it wrong fluently, in a register that signals it has checked. A draft you cannot check is worse than no draft.
The folder called final_v3
A customer asks why the price went up. The answer is in an email from fourteen months ago, about a drawing revision, in a folder nobody can identify — and the person who knew has left.
Six vendors, a slice each
Every one of your systems now has an AI assistant, and each is good inside its own slice. None can answer a question that crosses two — which is every question that matters.
Nothing published yet.
Nothing under that filter yet.