Thanks for having us in your inbox again. We are building Aimable in the open: real work done with AI, without sensitive data leaving the building.

When we started Aimable, the goal was simple to say and hard to do: let people use AI on the data they actually care about, the sensitive kind that never belongs in a public chatbot. Working with our first customers taught us something that reshaped how we build. Keeping data safe is what makes AI allowed inside a serious organisation. But getting real work done faster is why anyone reaches for AI in the first place. Safety is the floor. Productivity is the reason. So we went looking for concrete pieces of work where AI clearly lifts productivity and the sensitive-data problem is real at the same time.

The investing world was an obvious place to look, because it is one we know. Between us we have built companies, made angel investments, and worked alongside VC funds, PE funds and investment teams. We know first-hand how much work it is to compare opportunities and turn a pile of documents into a memo you can actually decide on.

So we were glad when our first AFM-regulated customer came to us with exactly that ask: help us with our screening memos, our investment memos, our contracts. We are proud of this customer and proud of the work, and when we can say more about who they are, we will. For now: building it with them, in the open, has taught us more about productive and safe AI than building in isolation ever could.

Aimable is not a one-trick tool. The same platform also powers a knowledge-base assistant that lets support teams answer from policy instead of memory, which you can watch in our HR2day walkthrough. But rather than skim ten use cases, this issue goes deep on one. The question we hear most from readers is simple: how do I actually get more productive with AI, without being careless with data I am responsible for? That is the question we want this newsletter to keep answering, and this issue answers it through a single use case, the investment memo, seen from a few angles.

This one is also an invitation. If you work in investing, VC, M&A or private equity, and screening memos, investment memos or contracts are eating your week, we would like to talk. Reply to this email, or send any of us a message on LinkedIn, and we will gladly walk you through how Aimable works and where it could help.

In this issue:

1. A day of deal screening, done in minutes, with nothing leaving the building.

2. 🎙 On the podcast: what surprised one investment platform about their AI memo.

3. This week from Ludger: the filter that learns what counts as sensitive.

4. On our radar.

A day of deal screening, done in minutes, with nothing leaving the building.

Here is what that looks like in practice. In a video we posted this week, Arjé Cahn, our CPO, takes the documents a real deal arrives in, PDFs, spreadsheets, a term sheet, and gets a clean summary back from one of the most capable models available in a couple of minutes. The work that used to eat an afternoon, done before the coffee goes cold.

The part worth watching is what the model never received. Aimable has already marked the personal data it found across the documents: the names of the leadership team, the company, a column of email addresses, the ticker. Arjé corrects it where it is wrong, adds a company name to the dictionary, and confirms. Only then does the instruction, “summarise this deal,” go out to the model. The summary that comes back is genuinely good. And the sensitive parts in it are highlighted, so you can see for yourself, line by line, that none of those names or numbers ever left your environment.

Read that as a productivity win first. What you can finally do is put the best model in the world to work on documents you were never allowed to go near with it. The filtering is what makes it allowed. The highlighting is what makes it trustworthy: you do not have to take our word for it, you can check.

The work in the video runs on skills we built on top of the Aimable platform, with the safety built in rather than bolted on, and we can tailor those skills to the way a specific team actually works.

→ Arjé’s LinkedIn post and the full video on our site.

🎙 On the podcast: what surprised one investment platform

The video above shows the what. The podcast is the why. In our new episode, Arjé and Ian sit down with the story behind that investment memo. The customer is a Dutch, AFM-regulated investment platform. Their screening memo used to take a full day of manual work, pulling databases, slide decks and financial documents into a templated summary. Now it takes about fifteen minutes, and it comes out as a Word document in their own template, with an Excel sheet underneath.

The surprise was where the win actually landed.

“Each screening memo used to take a full day of manual work. We built them something that does it in fifteen minutes. But the productivity win turned out to be somewhere we didn’t expect.”

The conversation gets into finished work versus chat, why an AI model cannot be trusted to do the arithmetic but a small skill wrapping a calculator around it can, and the unglamorous truth that some of the highest-value automations look like Ian’s accountant chasing down a single missing invoice. Worth a listen if you have ever wondered how to make AI hand you finished work instead of talking about the work.

→ Listen to the episode on Trusted Intelligence.

This week from Ludger: the filter that learns what counts as sensitive

The video and the podcast both rest on one quiet thing: a filter that decides, document by document, what to hold back before anything reaches an outside model. Ludger Visser, our founding AI engineer, has spent months on it. The interesting part is what happens over time: filtering sensitive data keeps getting easier and faster, because the filter keeps learning what counts as sensitive for a given customer.

What counts is rarely obvious, and the off-the-shelf options make it worse, because they are almost all American models trained on English. Take the term “CAO”, Dutch shorthand for a collective labour agreement, a word that turns up constantly in HR documents. To an American model it looks like a three-letter company name, so it gets stripped out even though it is not sensitive at all. Context decides the rest. A founder’s name sitting next to their shareholding has to stay in the building; the name of a billion-euro public company mentioned in passing usually does not. A target company headquartered in a city of a million is hard to identify from that alone; the same business described as the largest employer in a village of 200 is not. Sensitivity depends on the sentence, the document, and the customer, so the filter learns on two levels. When a user marks something the filter got wrong, it stops making that mistake for them within seconds. And when a mistake turns out to be general rather than specific to one customer, the fix rolls out to everyone.

And we do this, in the plainest sense, in the Netherlands. The filter runs on our own infrastructure here, or on the customer’s own servers, before a single name reaches a model in the US.

On our radar

A few things we did not cover in depth but are watching.

• The EU AI Act deadlines are slipping. Arjé flagged a piece by Arnoud Engelfriet at ICTRecht on the Omnibus package pushing several AI Act deadlines back. Read it for a clear-headed Dutch legal take on what actually moved. One caveat worth holding onto: the GDPR has not moved at all. The moment you upload a file with someone’s personal data into an external AI, you are sharing that data with another party, and that is a GDPR question whatever happens to the AI Act timeline. The deadline that slipped was never the thing protecting you.

• The token tikker. Pim Verschoor, our commercial lead, passed around Anton Leicht’s essay “Cut Off.” The comfortable assumption that access to the best models keeps getting cheaper and more abundant may not hold. If your AI plans quietly assume infinite cheap tokens, that is worth a second look.

That is the week. The thread through all of it: the teams getting real work done on sensitive data are not the ones who waited. They are the ones who picked the right piece of work, pointed good AI at it, and could see exactly what it did and did not touch.

Trusted Intelligence is published by the Aimable team, from what we build, the conversations we have, and the things that make us think.

Subscribe at trustedintelligence.substack.com. Forward it to a colleague trying to figure out AI for their organisation. Catch up on earlier issues.

For most of the past three months, our engineering team has been doing something quite invisible. A project we have been calling “Patagonia”, internally. A near-complete rewrite of the layer underneath everything we ship. New data model, new auth, new API surface, new lines drawn between what is platform and what is workspace. The kind of work that makes a product look exactly the same on the outside while every assumption underneath gets replaced.

This week, we’re super excited to announce that the first two things that ride on top of it landed in production at the same time. Alongside them, two pieces went up on our Substack that we never sent by email. We are doing that now. Four items. Two you can use. Two we want you to read.

In this issue:

1. You can now build Aimable into your own product.

2. Your AI agents play by the same rules as your people.

3. Ian on the Aimable Lab.

4. And Arjé on what actually breaks when you run agents.

Thanks for reading!

You can now build Aimable into your own product

The Aimable Platform API is live. One endpoint. Your developers stop calling OpenAI or Anthropic directly and start calling Aimable instead. What they get back is what they always wanted: a clean answer from the model of your choice. What happens in between is the part that used to keep your CISO up at night. Personal data filtered out on the way in and restored on the way back. Your rules enforced. Your model routing applied. Every call written into the same logbook your auditor already knows how to read.

Most enterprise AI tools treat governance as a feature. A toggle in the admin panel. “Enable PII redaction.” The trouble with governance as a feature is that developers route around it. Not because they are reckless, but because features add friction and people building things optimise for speed. So they call the model directly. They skip the toggle. The audit log gets a hole in it, and nobody notices until the auditor does.

The Aimable Platform API takes a different bet. Governance is not a feature on top of the model. It is the layer the model is called through. Every chat in the Workbench, every n8n agent, every Python script, every Office add-in: same gateway, same rules, same logbook. Bart Evers, our CTO, put it more simply on the day it went live: “This is the first time we sell to two audiences with the same product.” The team building, and the organisation that needs to know what the team built is safe.

Send your developers to: https://docs.aimable.ai/

Your AI agents play by the same rules as your people

The second thing that shipped this week is a full agent governance layer on top of the Platform API. Your automated agents, the ones running in n8n, in OpenClaw, in custom Python, or in a partner tool, can now operate as first-class users of your Aimable Space. They authenticate. They get scoped permissions. They appear in the audit log next to your human users. They follow the same policies, with the same data protection in front of every model call.

This is the part of the manifesto we have been waiting longest to make concrete. Humans and agents play by the same rules. A nice line in the brand book. A different thing to actually wire up. For most of the last year, the agent story was a slide. This week it is a feature.

What it changes in practice: when someone in your organisation, or at one of your vendors, builds a workflow that calls AI, you do not have to choose between letting it through ungoverned or blocking it. You give it an agent identity inside the relevant Space. It gets the same ground rules as the people working there. One rulebook. No shadow AI. No parallel governance regime for the automated half of your work.

Take a look in our Lab

A couple of weeks ago, Ian Zein, our CEO, launched a new page on our site: Aimable Lab. It is where we put the things we are building before they land in the core product. Not roadmap slides. Working surfaces, each with its own page and an animated demo of how the thing actually behaves.

Five things live there today. A hosted agent that runs shifts inside your Space. Notes grounded line by line in the meeting transcript they came from. Skills, a packaged recipe of prompt, tools, policies, and output schema that people and agents call the same way. Office add-ins that bring Aimable into Word, Excel, and PowerPoint. And a canvas where you highlight a section of a document and ask for a change in place.

The running thread across all five: yours, not theirs. These surfaces plug into the Space and policies your team already has, instead of spinning up a parallel vendor world.

→ Read Ian’s full post: Take a look in our Lab

The agent works on day one. The trouble starts in week three.

Last week our commercial lead, Pim, could not find a meeting note. The internal agent that usually drops them into our Slack had been silent for five days. Nobody noticed it was broken. Pim assumed the agent had skipped his meeting.

Arjé, our CPO, went looking. What had happened was not an AI problem. A key the agent depended on had expired. The agent went into a retry loop. The retry loop wrote logs. The logs filled the disk. The host went down, and a few other things went with it.

Three failures like that, in three different agents, including one at home (Rosie, the family agent that lives in our WhatsApp). None of them dramatic. None of them about the model. All of them about the things that quietly drift in the weeks after the agent goes live: keys expiring, model versions shifting under you, dependencies you forgot the agent had, log files filling disks. Field notes from three of Arjé’s quietly failing agents, and where he is landing on what an AI platform is actually for.

→ Read Arjé’s full post: The agent works on day one. The trouble starts in week three.

Trusted Intelligence is published every other Friday by the Aimable team. From links we share, conversations we have, and things that make us think.

If you enjoyed it, please forward it to a colleague who is trying to figure out AI for their organisation.

Next issue: a deeper look at one Skill we built with a finance customer, plus what is moving in Europe. Thanks for reading.

In this issue:

1. Bart built Aimable into Excel, Word, and PowerPoint. Here is what that looks like.

2. “We use Microsoft Copilot, so it’s probably fine.” A new default kicks in on 17 April. Here is what to do before then.

3. 🎙 Ian & Arjé on the podcast: meet Aimee, Rosie, and Zosia.

4. Private equity is paying extra for companies that barely use AI. Really.

5. This week from Ludger: when your laptop becomes a datacenter.

Aimable, now in Excel, Word, and PowerPoint

In our sales Slack on Monday, our CEO Ian Zein wrote a one-liner about something Bart had just demoed:

So we are.

For the past couple of weeks, Bart has been quietly building Aimable as a set of add-ins for Microsoft Office: Excel, Word, and PowerPoint. You open a spreadsheet, a document, or a deck, and from a side panel you can ask Aimable anything grounded in your organisation’s own verified documents, collections, and knowledge sources. The answer comes back written straight into the cell, the paragraph, or the slide you were working on, with source citations, under the same rules as the rest of the platform.

Bart put the shift in one line: “With Aimable’s Office add-ins, you do not have to ask an AI to generate an Excel file dozens of times. You pull the AI straight into the .xlsx, .docx, or .pptx you are already working in.” That inversion is the whole thing. The familiar workflow today is: go to a chat tool, ask it to generate a spreadsheet, copy the result back into your real document, check whether it is right. Aimable flips it. The document stays where it is, and the AI shows up next to it, grounded in your own sources, playing by your rules.

Yes, Copilot has AI inside Office too. The difference is what the AI is allowed to see, which sources it is grounded in, and where the data actually goes. With Aimable, your team works with its own verified knowledge, in the tools they already open every morning, without a new pile of questions for the auditor at the end of the quarter.

We are testing the add-ins in a pilot with a new customer right now, with the whole company invited to use them. No polished video yet, but a short clip of them in action below. When Bart first demoed it, our internal agent Aimee (see podcast!) captured the reason it works: “As soon as people see they can safely use sensitive documents, it sells itself.”

If your team has been quietly wishing AI would do this kind of work inside the tools they already live in, hit reply. We would rather have ten conversations about this now than ten LinkedIn posts about it six months from now.

“We use Microsoft Copilot, so it’s probably fine.”

Since we are already on the subject of Office this week, it is worth flagging a Microsoft documentation page that started making the rounds in our Slack on Monday. By the time you read this, you have one week to do something about it.

The sentence we hear in almost every sales conversation goes like this:

“We use Microsoft Copilot, so it’s probably fine.”

People assume that because Copilot is Microsoft-branded and came bundled with their enterprise agreement, the AI part must already be safe, compliant, and thought through by someone above their pay grade. It usually isn’t, or at least not in the way they think.

The page is about a Microsoft 365 Copilot feature called Flex Routing. Tweakers, the largest Dutch tech publication, wrote about it under a headline that translates to “Microsoft changes Copilot for 365: data processing can suddenly happen outside the EU.” Both stories point at the same calendar entry. On Friday 17 April, Microsoft is enabling a setting by default that lets your live Copilot prompts be processed in the United States, Canada, or Australia during periods of peak demand. Microsoft has placed a notice about it in the admin Message Center (reference MC1269223). Almost nobody seems to have read it.

In plain language: when your team gives Copilot a prompt, the model has to process it somewhere. Microsoft calls this “inferencing.” For European customers, that processing always happened inside the EU Data Boundary. As of next Friday, during busy hours, your prompts may be sent to a server outside the EU instead. Your stored data still sits inside Europe and encryption stays on. But the moment your colleague clicks “Summarise this contract” in Word during a busy morning, the actual processing of that prompt may happen on a server in Texas.

In our Monday meeting, our commercial lead Pim flagged the tactical point: the comfort people draw from their Microsoft enterprise agreement is doing a lot of quiet work it is not really qualified for. Bart made the engineering point that where the inferencing happens is exactly where the legal exposure sits, and almost no Dutch organisation has thought about peak-demand routing as a data residency event.

We are not saying block Copilot. Copilot is fine for plenty of work. We are saying:

That is the new question for every IT lead in the country to ask this week. And it is also the moment where an add-in like Bart’s starts earning its keep.

→ https://learn.microsoft.com/en-us/microsoft-365/copilot/copilot-flex-routing
→ https://tweakers.net/nieuws/246504/microsoft-wijzigt-copilot-voor-365-dataverwerking-kan-ineens-buiten-eu-gebeuren.html

🎙 This week on the podcast: meet Aimee, Rosie, and Zosia

Ian and Arjé recorded an episode about something strange. Each of them has quietly built a personal AI agent this past month, and neither is letting it run on a work laptop.

Ian’s is called Zosia, after a character in the Apple TV show Pluribus. She lives on a small Mac Mini, has her own phone number and her own calendar, and every morning at nine she sends Ian a WhatsApp with a rundown of his day. A few people in his family group chat briefly assumed Ian had hired a real-life personal assistant.

Arjé’s is called Rosie, and she runs the chores rota for his kids over WhatsApp. Things got interesting when the children started negotiating with her, and invented the nicknames they now insist she uses. The oldest is addressed as “the Legendary Cool Dude.” The middle one is “the Untouchable Alpha Lord of Chill.” Rosie complies. The kids are effectively programming an AI by talking to it, and nobody in the house finds that strange.

The work agent is Aimee. She reads our Slack, our meeting transcripts, and the articles we share, and she surfaced half of this newsletter before we wrote a word of it. It is also her quote that anchors the first item above. One well-configured agent is already doing the work of maybe three people on our team. That is the productivity story nobody is telling loudly enough yet. Ian puts it plainly on the recording:

→ https://trustedintelligence.substack.com/p/our-ais-run-the-household-and-the

Private equity is paying extra for companies that barely use AI

Arjé flagged a an article on FD.nl this week that is worth a second read. Dutch PE firms are actively seeking out “blue-collar” companies, businesses with physical operations and limited AI exposure, because they see them as safer bets. Translation: professional investors have quietly started pricing AI risk into what your company is worth.

In our management meeting it turned into one of those honest conversations where you realise the framing has shifted overnight. For months we have been talking about AI risk as a compliance story. This is the week it started looking like a valuation story. The helpful reading: organisations already using AI on their own documents, with a clear logbook of who did what, now have something most of their peers do not. A story that holds up in a due-diligence room.

→ https://fd.nl/bedrijfsleven/1591386/blauwe-boorden-populair-op-overnamemarkt-want-die-hebben-weinig-last-van-ai

This week from Ludger: when your laptop becomes a datacenter

Ludger Visser, our founding engineer, has been buried in a family of AI models Google DeepMind released this week. The range goes from a tiny 2 billion parameter version that fits on a Raspberry Pi to a 27 billion parameter version that reasons well on a normal laptop. Ludger’s one-line summary in our Slack: “This is the most efficient model you can run locally right now.”

The longer version, which he walked the team through on Monday, is the one worth sitting with. Local models do not replace the cloud. They unlock a second destination for the work you do not want leaving the building. A marketing brief can happily go to a frontier cloud model. A conversation about an HR file should probably stay on a local Gemma 4 model, on a machine that never touches the open internet. Same working environment, different destinations, chosen for you per question.

Ludger keeps saying “one model is never enough” in our meetings. After this week, the cheaper-to-believe version of that sentence is finally sitting on a laptop you can lift with one hand.

→ https://deepmind.google/models/gemma/gemma-4/

On our radar

A few things we did not cover in depth but are watching:

• The cofounder and CEO of Linear, Karri Saarinen, argues that unreliable AI is a design problem, not a model problem. “The model is never the problem” has been the opening line of our own point-of-view document since September. Reassuring to see someone building a very different product land on the same idea.

• The Dutch IT lawyer Arnoud Engelfriet asked a deceptively simple question this week: are you even allowed to let an AI assistant sit in on your meetings? The legal answer is less obvious than most people assume.

Trusted Intelligence is published every other Friday by the Aimable team: Ian Zein, Arjé Cahn, Bart Evers, Ludger Visser, and Pim Verschoor. From links we share, conversations we have, and things that make us think.

If someone forwarded this to you, sign up at https://trustedintelligence.substack.com to get it directly. If you enjoyed it, forward it to a colleague who is trying to figure out AI for their organisation.

Thanks for reading!

We’re Ian Zein, Bart Evers, Arjé Cahn, Ludger Visser, and Pim Verschoor. Since September 2025 we’ve been building Aimable, the AI platform that helps organisations finally use AI on their sensitive data. Safely, with answers that are grounded in their own sources. We believe AI should make people more productive, not more anxious, and that the safety to do so should be built into the platform, not bolted on afterwards.

You’re receiving this because we’ve crossed paths at some point in the past couple of months and we thought you might find it interesting to hear what we’re up to. We hope you do. If not, there’s an unsubscribe link at the bottom, no hard feelings!

We want to start sharing what we see, what we think, and what we’re learning about AI along the way. Not as experts who have it all figured out, but as founders who are deep in it every day, talking to companies, building product, and trying to make sense of a world that moves faster than anyone’s governance can keep up with.

Let’s get into it.

What’s Happening in AI

Something interesting happened this week. Two Dutch software companies, Moneybird and AFAS, sent completely opposite signals about AI. One opened the door wide. The other slammed it shut. A security incident in the open-source world quietly proved both of them right. And in between all of that, Arjé and Ian sat down to record the very first episode of our podcast, Beyond the Prompt, about something deceptively simple: why does AI still give most people garbage?

Here’s what we noticed.

In this issue:

1. Someone we’d like you to meet: our amazing customer HR2day shares their story.

2. Two Dutch software vendors. Opposite reactions. Same problem.

3. A vibe-coded pull request just compromised an AI routing tool. Now what?

4. OpenAI wants to store your files. Think about that for a second.

5. 🎙 Ian & Arjé Beyond the Prompt: why AI gives you garbage (and the two-document fix)

Someone we’d like you to meet

We talk to a lot of organisations that are trying to figure out how to use AI without putting their sensitive data at risk. Most of those conversations happen behind closed doors. But every now and then, someone is willing to share their story publicly, and that’s worth highlighting.

Marco Boerlage runs HR2day, one of the Netherlands’ leading HR software companies. His team handles sensitive employee data for hundreds of organisations. What stuck with us when we first met Marco is how personally he takes this. For him, data privacy wasn’t a compliance exercise or something to delegate to legal. He wanted his people to have great tools. And he wanted them to sleep well at night knowing nothing leaks.

Marco joined us for a webinar where he talks about that journey. No jargon, no compliance speak, just a practical conversation about what it actually takes to make AI work when the data is sensitive. If you’re leading a team that deals with confidential information and you’ve been wondering where to start, this is worth 15 minutes of your time:

You can watch the full webinar online on our website: https://aimable.ai/en/webinar-recording.

The Moneybird-AFAS paradox

Moneybird, the accounting platform half of Dutch freelancers and small businesses use, just shipped a command-line interface that lets AI agents talk directly to your financial data. Invoices, contacts, ledger entries, all accessible to automated workflows.

The same week, AFAS decided to restrict AI integrations. Their reasoning: the governance can’t keep up with the pace of adoption. Marcel ten Hoopen wrote about it on LinkedIn, and his take stuck with us: “AFAS is right about the risk. But stopping isn’t a strategy.”

We talked about this at length during our Monday meeting. Arjé knows the Moneybird founders, and Ian brought up the AFAS post. It turned into one of those conversations where you realise you’re looking at the same coin from two sides. We genuinely believe Moneybird is making the right call. The market is heading this way. Your bookkeeper’s software is becoming agent-ready whether you planned for it or not.

But we also get why AFAS hit the brakes. Honestly, if we weren’t building what we’re building, we’d probably feel the same unease. We talk to companies every week that feel exactly that tension. They see the opportunity. They feel the urgency. And they have no idea how to govern what’s coming. One founder told us last week, quite bluntly: “I know my team is using ChatGPT. I just don’t want to know what they’re putting into it.” That kind of wilful blindness is more common than anyone admits.

What we keep coming back to: the answer isn’t choosing between Moneybird’s optimism and AFAS’s caution. It’s building the layer in between. That’s what we work on every day, and weeks like this remind us why we started in the first place.

→ Moneybird CLI · AFAS discussion

Who watches the watchers?

LiteLLM, an open-source tool that a lot of companies use to route their AI requests, got compromised this week. Not through a sophisticated hack. Through a pull request that appears to have been vibe-coded. AI-generated code that looked clean but wasn’t.

NVIDIA’s Jim Fan framed it as: you need boring, predictable, well-tested traditional software to safeguard the AI systems everyone is racing to build. He calls it “software 1.0 controlling software 3.0.”

Ian shared this in our Slack with the note: “This is exactly how I see my safety gates story.” And then Bart pushed back a little, which is what Bart does. Are we software 1.0 or 3.0? And the honest answer, one we had to sit with for a minute, is: both. We use AI where AI excels: understanding context, working with language, connecting knowledge. But the parts that enforce the rules? That protect your data? That log what happened and prove it to your auditor? Those are deliberately, almost stubbornly, old-fashioned. Deterministic. Predictable. And after this week, we feel pretty good about that choice.

It’s not glamorous. Nobody starts a company dreaming of building boring software. But the moment someone’s vibe-coded pull request compromises the tool that routes your AI traffic, you understand why the guardrails need to be the least exciting part of the stack.

If you’re running a business: do you know how the tools your team uses to “manage” AI were actually built? It’s worth asking.

→ Jim Fan on LinkedIn

OpenAI’s quiet land grab

OpenAI launched ChatGPT Library last week, a feature that lets users upload and permanently store personal files inside ChatGPT. Documents, spreadsheets, reference material, all living in OpenAI’s cloud.

On the surface, it’s a convenience feature. Underneath, it’s a platform play. Once your documents live inside ChatGPT, leaving becomes expensive. And you might not even notice it happening.

We’ll be honest: this one worries us. Not because the feature is bad, it’s actually quite useful. But because we see the pattern play out in real time with almost every company we talk to. Last week, in a sales conversation, a prospect told us: “We use Microsoft, so it’s probably fine.” We hear versions of that sentence weekly. And the answer is always the same: fine, until your auditor asks where your company data actually lives. Then it turns out it’s partially in your systems, partially in a consumer AI tool your team adopted organically, and now partially stored permanently on OpenAI’s servers.

It starts with one person uploading a document. Then a team. Then it’s just how things work around here.

We don’t think blocking ChatGPT is the answer. We’ve tried that conversation with companies and it doesn’t work. People find workarounds. What does work is giving your team something they actually prefer to use. Something where they can bring their own documents, get answers with sources they can check, and where you, not OpenAI, decide where the data lives. That’s what we’re trying to build. We don’t have it perfect yet, but we’re getting closer every week.

→ BleepingComputer

🎙 This week on Beyond the Prompt

We launched a podcast. It feels a little scary, to be honest.

In the first episode, Ian and Arjé talk about something that came up at an event last week: why do people still apologize for using AI? Ian admitted he used to frantically remove em-dashes from AI-generated text so nobody would know. Arje caught himself doing the same. At some point we both stopped, because we realised the embarrassment was the problem, not the AI.

But the real insight in this episode is something simpler. The reason AI gives most people garbage output is that it doesn’t know them. It doesn’t know your company, your tone, your standards. We found a fix: two short documents, one describing how we communicate and one describing what we stand for, that we feed to AI with every task. Two to three pages total. Ian now generates his LinkedIn posts entirely with AI, some reaching 35,000+ impressions, without changing a word.

The trick: you don’t write those documents alone in a room. You have a conversation with your team, record it, and let AI draft from the transcript. Two rounds and it’s solid. We share exactly how in the episode.

On our radar

A few things we didn’t cover in depth but are watching closely:

• Running LLMs on your iPhone. On-device AI inference is becoming workable. The day your team can run AI without data leaving their phone changes a lot of security assumptions. We want to dig into this properly next week.

• AI navigator for ISO 27001. Someone built a fine-tuned RAG system specifically for compliance frameworks. It confirms something we hear constantly: generic AI gives generic compliance answers. Not good enough when the stakes are regulatory.

What’s Happening in AI is curated weekly by the Aimable team, from links we share, conversations we have, and things that make us think.

Subscribe on Substack to get it in your inbox.

Thanks for reading!