Rewiring the company, not just the toolchain | Software Talks AI
May 2026 · Issue #2 Monthly SaaS executive newsletter

Rewiring the company, not just the toolchain

Are you redesigning around AI, or just adding it to the workflow?

This month

Welcome to another edition of Software Talks AI, Monterro's monthly read on AI for B2B SaaS leaders. Mythos, Cowork-competitors, “Bring your own AI”, and Claude Design are some of this month's topics. Anthropic and its Claude products continue to produce the most headlines, as the company continues to grow its product suite at a rapid pace.

But first: a hot discussion topic this month has been on the difference between AI-assisted and AI-first organizations. The companies pulling ahead right now are not the ones that bolted Copilot onto an existing workflow. They are the ones that dismantled the workflow and rebuilt it around the assumption that AI is the primary builder, with humans providing direction, judgement, and quality gates.

Many teams claim to be AI-first while running the same sprint cycles, the same Jira boards, the same weekly standups, and the same QA sign-offs. They just added AI to the loop, rather than redesigning it. If product management still takes weeks while build time has collapsed to hours, planning becomes the bottleneck. If marketing takes a week to announce what engineering shipped on Tuesday, marketing becomes the bottleneck.

The work for leadership in 2026 is to push AI-native operating models into every function, not just the ones that touch code, and to redefine entire workflows. When implementing AI in the organization, don't just translate a workflow into AI. Look at the workflow's outcome and explore how AI can achieve that function.

Monterro companies are currently reinventing their workflows, which we aim to share in greater detail in upcoming newsletters. Some companies have a goal of having no human-written code by the end of the year, some have increased their output by 6x, and some have reinvented their onboarding and customer success workflows.

Top stories

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

Story 01

Mythos worries banks

Anthropic's new Mythos model is perhaps the biggest AI story of the month. It's an AI model built for defensive cybersecurity research, currently available only to a small set of partners (such as Microsoft, Apple, the Linux Foundation, JPMorganChase, and around 40 other organizations) into a defensive coalition called Project Glasswing. In seven weeks of internal testing, it identified more than 2,000 previously unknown software vulnerabilities, including a 27-year-old bug in OpenBSD and a 17-year-old remote code execution path in FreeBSD. It can chain vulnerabilities together autonomously across operating systems and browsers.

The banking sector is paying close attention. At the IMF spring meeting in Washington, regulators issued specific warnings about Mythos and its successors. Banks run on legacy systems that are hard to patch, and so far Anthropic has not granted access to any banks in Europe. In the US, JP Morgan and others are reportedly exploring stress testing with Mythos-class tools. Coinbase and Binance have both approached Anthropic.

The pattern across coverage from the BBC, the Guardian, and CoinDesk is consistent: this is less about a single model and more about a permanent shift in the economics of vulnerability discovery. The cost of finding a serious zero-day issue has dropped from weeks of expert work to a few thousand dollars of compute. Many software leaders are nervous about Mythos. The hosts of the New York Times podcast “Hard Fork” raised the question on whether we will need to essentially rewrite all software following AI such as Mythos, given their exceptional ability at finding vulnerabilities.

Read the full story →
Story 02

Claude Cowork gains competition

Claude Cowork, currently the most capable mainstream AI agent, has been a wakeup call to Anthropic's competitors. It led to Microsoft quickly scrambling to build Copilot Cowork and to OpenAI cancelling ChatGPT initiatives in order to better focus on their core business. Now, Microsoft and OpenAI are both announcing new agentic features.

Microsoft has greatly improved Copilot's agentic capabilities in Word, Excel, and PowerPoint, with their own internal data showing engagement up 52 to 67 percent and satisfaction up 21 to 65 percent across the apps. Copilot can now take multi-step actions directly in your documents, not just suggest them. At the same time, OpenAI introduced workspace agents in ChatGPT, powered by Codex. These are shared, team-level agents that run in the cloud, integrate with Slack, and can be built without engineering support. A Rippling sales consultant built one end-to-end that replaced five to six hours of weekly account research per rep.

The shape of the market is becoming visible. Claude Cowork, Microsoft 365 Copilot, and ChatGPT workspace agents are converging into a category we call agent harnesses for knowledge workers: a workspace where an AI agent has durable memory, access to your files and apps, and the ability to take action across them. They behave less like assistants and more like a junior colleague you onboarded this morning. For leaders, the question is no longer whether to enable one of these, it is which to standardize on, how to govern data access, and how to teach the organization to delegate effectively.

Read the full story →

Key metrics

Numbers worth keeping.

Source: hai.stanford.edu/ai-index/2026-ai-index-report

Coding ~100%

AI performance on SWE-bench Verified, up from 60% just one year ago, as AI rapidly improves on coding

Talent ‑80%

Drop in AI researchers moving to the US, reversing a decade-long trend

Competition 2.7%

Performance gap between leading US and Chinese frontier models, effectively closing the gap between the two nations

Our take

Your users are now AI agents

There's been two duelling approaches to utilizing large language models (LLMs) in products. On one side is integrating LLMs themselves into a product, and on the other is building endpoints for external LLMs (such as Claude Cowork or ChatGPT Codex) to hook into. In order words: you either add an LLM to your product, or you let users bring their own LLM.

The choice depends on the context, and some software owners have been implementing both approaches, but what we're discovering is that currently the latter approach seem to be the winning path for most software. Allowing external agents to hook into your product's capabilities means that users can effectively create any workflows or user interfaces they want through state-of-the-art AI tooling. You have lower costs, as the cost of the LLM falls on an external product, while you can retain revenue by creating pricing models for accessing your capabilities.

An example of a SaaS that has gone all out on this approach is Salesforce. They've spent the last two years making every capability in the Salesforce platform exposed as an API, MCP, or CLI command. Translated for non-developers: this allows an AI agent to utilize all Salesforce functions without having to navigate a slow user interface.

There are already significantly more agents than humans navigating the internet, and its inefficient for agents to navigate software using UI's made for humans. We now need to rethink our software and consider how we can make it more accessible for AI agents. AI agents are now your users.

At Nordic Software Summit, held in Stockholm om August 20th, we'll be doing a keynote on this very topic. Attendance is free, so please sign up if you're interested in learning more!

Nordic Software Summit →
Try this

Put it into practice before the next issue.

1

Ask your AI to interview you until you've reached a mutual understanding. Example prompt: “I want you to help me create a skill that prioritizes my e-mails. Ask me questions on how I want you to prioritize until we have reached a mutual understanding.”

2

Try Claude Design: Anthropic's new design product, currently in research preview. It takes a rough idea and turns it into an interactive prototype, a branded slide deck, or a marketing landing page. It can read your design system from your codebase and apply it automatically, and you can import directly from Figma or upload your branding guidelines. As always, check your enterprise's compliance policies before sharing company data.

Read more →
3

OpenAI has released a new image generation model in ChatGPT. It outperforms Gemini 3.1 (Nano Banana 2) in many areas, so try using detailed, complex prompts to get the most out of it.

Read more →

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.

“We measure success in production deployments and organizational adoption, not slide decks. We have already successfully infused AI across many of our portfolio companies. Because we work across so many software companies, we learn fast, spot patterns early, and turn those learnings into best practices that can be used across the portfolio, insights no single company could build alone.”

Nils Hulth, AI Lead at Monterro

Get Software Talks AI in your inbox

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