Your AI colleagues never sleep. Have you hired your first one? | Software Talks AI
October 2026 · Issue #5 Monthly SaaS executive newsletter

Your AI colleagues never sleep. Have you hired your first one?

29% of office workers still aren't using AI. 10% aren't even allowed.

This month

Welcome to another edition of Software Talks AI, Monterro's monthly read on AI for B2B SaaS leaders. Always-on AI colleagues, a new kind of AI model that everyone's talking about, and Microsoft's Copilot overhaul are some of this month's topics.

But first: the gap.

Every year, the Swedish Internet Foundation releases a report called "The Swedes and the Internet." This annual report details the internet habits of Swedes, some of the most digitized people in the world. The report itself is fresh off the presses, having been published on September 28th, although the data that the report is based on is from early January.

Despite Sweden being highly digitized and one of Europe's leading AI hubs, the study revealed that only 62% have used an AI platform such as ChatGPT. Among office workers, 71% had used AI. Almost 10% of people in the workforce weren't allowed to use it, and more than 30% weren't sure if they were allowed to. Only about 20% of people use AI daily.

It's still a massive gap compared to what we're seeing among frontrunners, not least in the Monterro sphere. In the Monterro portfolio, agentic AI systems are working autonomously through complex workflows, delivering entire tasks end-to-end. Now, many are establishing AI colleagues.

Kaisa's autonomous AI agent has now answered many thousands of missed calls for car dealers in four countries, working around the clock.

Umbraco's public MCP servers let any developer bring their own AI colleague to build websites in Umbraco's CMS.

Planima has AI agents writing code around the clock on their own computers.

Pythagoras cut the time it takes to create project documents from four hours to ten minutes using AI, while Valona Intelligence saves upwards of five hours per report.

Back in August, we wrote that the gap between the most AI-capable firms and the median firm had tripled in five months. The latest data shows the same thing from the other side. While some companies are setting up AI colleagues that work around the clock, a large share of the workforce does not yet know whether it is allowed to open ChatGPT at work. That gap is very visible.

The good news for anyone behind: the tools have never been cheaper or better, and this month they got cheaper still.

And: we just did a live webinar introducing latecomers to the power of Claude Cowork. The webinar is aimed at those who don't yet have much experience with Claude Cowork. Check out the recording for practical, hands-on advice for utilizing powerful AI.

Watch the recording →

Top stories

What moved in AI this month, and what it means for B2B software leaders.

Story 01

AI colleagues move into Slack and Teams

AI agents working in Slack and Teams have existed for a long time, but the idea of having long-running, always-available agents as colleagues has become increasingly popular in the last few months as both the technology and frontrunners are maturing rapidly.

There are plenty of platforms offering these kinds of agents. Devin and Viktor may be everyday names, but they're also the names of AI companies offering such agents working around the clock in Slack, as do Claude Tag, Abundly, and Dust.

Just now, in late September, Microsoft announced numerous new offerings in the always-on space, including Microsoft Copilot Autopilot and Agents 365. OpenAI also just announced its own alternative, Dots, which its CFO embarrassingly referred to on air as Muse, Meta's alternative for this type of AI.

Point is: there are a lot of players offering always-on AI colleagues. These colleagues can work in Slack or Microsoft Teams, post in chats, be triggered by events such as an incoming ticket appearing in HubSpot, or run on a schedule. Multiple people can message the AI colleague, even talking to it together in a thread on Slack. The colleague has a named owner, but it runs on its own usage limits, and can be contacted by many people.

An early challenge that companies run into is permissions. While your personal AI assistant mirrors whatever permissions you have in HubSpot, Salesforce, or Microsoft 365, an independent AI colleague might need its own set of permissions. In some scenarios, the colleague might even need its own service account. Carefully think about what permissions the colleague needs, and start narrow.

Over autumn and winter, we can expect to see AI colleagues become more and more commonplace.

Read the full story →
Story 02

Jev draws attention away from LLMs

On September 15th, a San Francisco-based startup called TypeSafe AI released Jev, and it became the fastest-adopted model in the history of Vercel's AI gateway: nearly 13% of paid teams used it within 24 hours.

Unlike a large language model (such as Claude), Jev does not write text. You give it a piece of information (an email, a support ticket, an agent's planned action) and a set of typed questions, and it returns answers with probabilities. It costs $0.042 per million input tokens, output is free, and an answer comes back in roughly a tenth of a second.

The output might sound narrow, as it "only" returns probabilities, but it's actually incredibly versatile. A large share of AI spend in software products goes to small decisions: which queue does this ticket belong in, is this lead worth calling, is this agent action safe, should this request go to a cheap model or an expensive one. Jev makes those decisions close to free. Jev is also significantly less likely to hallucinate than a large language model, making it appealing for financial workflows.

We think the real power of Jev comes from combining it with a large language model (LLM). A model like Claude does the planning, writing, and reasoning. A fast decision model like Jev handles the thousands of quick calls along the way. For instance, if one wants to use Jev to triage customer support tickets, then Jev is the one who says "this ticket should go to second-line support," and the LLM is the one who actually hands the ticket over to that team.

Jev may be hard to use for EU-based SaaS companies today due to compliance, but OpenAI has already announced its own version, and plenty of open-source alternatives are appearing at a rapid rate. We can expect EU-hosted alternatives to appear before long.

Read the full story →
Story 03

Agent platforms keep improving

As always, the mainstream AI platforms keep improving at a steady pace.

Throughout September, Anthropic has made numerous larger updates to its Claude application, including folding Cowork and Chat into one. They also moved Claude Design into the main conversation, while introducing Claude Slides and Claude Docs. OpenAI made the equivalent move in July, when ChatGPT Work merged chat, agent work and Codex into one app.

Just a week ago, Microsoft announced numerous updates to Copilot. Copilot is now one app with three parts: Home (chat plus Microsoft's own Cowork agent), Code (build apps and tools by describing them) and Autopilot (always-on agents). Full versions of Word, Excel and PowerPoint now run inside Copilot, and users can choose between OpenAI and Anthropic models. The pricing has continued to evolve: regular AI use stays on the standard per-user license, while agent work (including Autopilot) and the newest models are billed by usage, which Satya Nadella called "per-seat-plus-consumption." This could end up making Copilot an expensive alternative.

In August we wrote that Microsoft 365 Copilot tends to be around half a year behind the competition. In our view, ChatGPT and Claude are still ahead in many respects, especially in how well the agent works across tools. But Microsoft is clearly speeding up, and for companies that live in Microsoft 365, the distance may be shrinking.

Of course, several of the announcements listed here are not yet available for companies based in the EU, and annoyingly, some features that are available in the EU still don't offer proper EU data residency.

Read the full story →

Key metrics

Numbers worth keeping.

Ecosystem 140,000+

public MCP servers listed online.

Growth 1,088%

Anthropic's revenue climb in 2025.

Concentration 2

customers account for 25% of Anthropic's 2025 revenue.

Our take

AI colleagues are gaining momentum

In April we described the new agent tools as "a junior colleague you onboarded this morning." In August Jacob Bergdahl wrote that frontrunners were treating their AI agents as colleagues. In September the vendors made it literal. Claude Tag, Copilot Autopilot and OpenAI's Dots all give an agent a seat in your team's channels, a memory of what the team is doing, a schedule, and an identity that IT can govern.

What makes an AI colleague different from a chatbot is that it does not wait to be asked. It runs a weekly digest every Monday at 09:00, follows up on a thread that went quiet, triages every new support ticket, or watches a repository and opens a pull request when CI breaks.

Anthropic says 65% of its product team's code is now created by its internal version of Claude Tag.

Inside the Monterro portfolio, the companies furthest ahead are already running this pattern. The frontrunners have even been doing it for years, but it's been scaled up significantly over the last month.

When selecting a vendor, consider: how easy it is to create an AI colleague; what the pricing looks like; how permissions are configured; what platforms the colleague can live in; if export options are good (making sure to avoid vendor lock-in); for Monterro companies, if it can run Enclave inference.

When setting it up: 01 Start with one AI colleague, in one or two channels, doing recurring work that is easy to verify. 02 Set spending caps per channel. 03 Review its audit log often at first.

At first, you may find that you need to help the AI along, guiding at every step. Over time, as the AI colleague improves, it can work increasingly autonomously. For instance, an AI colleague that triages customer support tickets might triage some tickets incorrectly at first. As you correct its mistakes, it will learn how to better fulfil its job, until ultimately, it can work completely autonomously.

Try this

Put it into practice before the next issue.

1

Ask Claude to make a video presentation. You might be surprised at how Claude can create beautiful motion graphics, combined with audio effects and music.

2

Let your AI work every weekend. Whether you're using Claude, ChatGPT, Copilot, or another mainstream AI platform, you should have the ability to schedule your AI to work over the weekend. Take a common workflow of yours and write it down, step by step. Consider how far the AI can get without your input, and schedule it to work until that point every Sunday. Come Monday, it'll already have started your work.

3

Show, don't tell. Have you been making a specific type of report for years now, and want to hand over that duty to an AI? Throw your old reports at it, preferably the best ones. Ask it to extract your style of writing them. This way, instead of you describing the reports to the AI, it will figure out itself what a good report looks like.

Footnote: Before using Claude Cowork, check with your organisation that it complies with your internal policies. Avoid inputting sensitive or personal data into AI tools unless your organisation has given explicit approval.

Infuse AI

Monterro, built for the AI era

Accelerating AI transformation across our portfolio companies with Infuse AI.

“Only a few months after launch, our AI products are already contributing meaningful revenue growth and creating measurable value for customers. By year-end, we expect these products to account for around 50% of new sales.”

Henrik Lenerius, CPO at Kaisa, a Monterro portfolio company

Get Software Talks AI in your inbox

One email a month on AI for B2B SaaS leaders. Written by the Monterro team.