Photo: Tonyglen14 · CC BY 2.0 · Philadelphia, 2016. Archival photograph; cropped and resized.
November 20, 2026 · Rob Tyrie · Ironstone Advisory
I think the next killer app for AI is the work suite. The important competition may be over who organizes the work. An application can remain essential while losing its place as the starting point for the working day.
I had dinner with friends from the technology business this week. We talked about the money flowing into frontier models, but what interests me is the delay between investment and impact. A market correction can change expectations overnight. It does not instantly erase the engineering and infrastructure already under way. I expect those investments to keep arriving on our desktops as new capabilities.
And those capabilities have ordinary work to do. Get the spreadsheet into shape. Build the presentation around the evidence. Carry the context from one task into the next without making somebody reconstruct it all.
Think about Lotus 1-2-3 and Excel. Graphical interfaces helped change how people worked with spreadsheets. I see a similar opening for AI work suites: make the work more accessible, and the starting point can move. Excel and PowerPoint can remain part of the process.
In the books experiment for my NeoLibrary design, several agents helped identify and enrich records in a shared spreadsheet. We separated proposals from independent verification and gave one agent responsibility for writing to the dataset. One proposed author attribution looked plausible and was wrong. Verification caught it before it reached the sheet. I had initially carried messages between the agents myself, which was an instructive way to discover that being the integration layer is not a particularly scalable occupation.
The chatbot made asking easy. The next wave has to make doing dependable.
That is the connection running through this week's stories. What may an agent read? What may it change? Who can redirect it? What happens when the infrastructure cannot keep up? Somebody has to organize all of that before the finished work deserves our trust.
My practitioner model, Ask → Reason → Create → Build → Orchestrate → Evolve, distinguishes the capabilities involved. Knowing how to ask a clever question and building a repeatable workflow are different accomplishments. Leaders need to know which one they are assessing.
Over the next twelve months, I expect those differences to show up at work. Here is how I would test that expectation:
On comparable assignments, does useful work arrive sooner, with less correction and review effort?
Who gets the next assignment or greater responsibility, and does the quality of their work justify it?
Do rewards follow demonstrated improvement, or just enthusiasm for the tools?
Before calling something a performance problem, ask whether the organization has made competence possible. People need tools, training and time to learn. People who can reason logically, work patiently with data and get software safely into production have valuable skills beyond coding. Teach them the business. Give them responsibility for outcomes and the workflows their colleagues actually need.
Evolve is where my model makes the obligation explicit: what failed, what did you change, and how do you know it improved?
Read this week's stories with that question in mind. Then ask it about the work on your own desk.
The agent has arrived at work. Now it meets the screen, the customer record, the power grid and the people expected to supervise it. Five stories about where useful capability meets a boundary, and a long read on what it takes to make these systems dependable.
Signal announced protections against mobile agents capturing messages on supported systems, following Reuters reporting on early safety problems with Meta’s Muse. Encryption protects the trip. An agent reading the screen is already at the destination. The issue is what an agent can see through the device, not a demonstrated breach of Signal’s encryption.
ChatGPT for Sales brings account research, call prep and follow-up into one conversation. Connected tools such as Attio can save next steps back to the CRM. The CRM doesn’t disappear. The interesting question is whether it’s still the first thing a salesperson opens in the morning.
Gemini Tasks is rolling out for free in Canada. In the U.S., eligible users can also reply to its text updates to steer work already in progress. That texting feature is currently U.S.-only. Google doesn’t need to teach people a new habit to make that useful. They already know how to reply to a text.
PJM proposed speeding up its interconnection process and, separately, setting rules for how large electricity loads behave during grid disturbances. Neither proposal puts power on the grid today. The critical path for AI expansion may run through a utility and a regulator, not a model lab.
The EU Council called for AI literacy frameworks beyond primary and secondary school. It didn’t introduce an adult placement test. The OECD/EU school framework already describes three levels across four areas: engaging with AI, creating with it, managing it and shaping it. The broader DigComp 3.0 uses four levels, based on how complex the task is, how much thinking it requires and how independently someone can do it.
EU Council · OECD/EU school framework · DigComp 3.0
My practitioner model asks what you can make work reliably: Ask → Reason → Create → Build → Orchestrate → Evolve. Making a useful artifact is one thing. Building a workflow that runs twice, deciding what people and agents should each control, and improving it when the evidence says it isn’t working are different accomplishments. I’d assess those capabilities separately. At Evolve, show me what failed, what you changed and how you know it got better.
Bhardwaj and colleagues review 126 studies covering how agents plan, use tools and work together. Much of the evidence still comes from controlled prototypes. Read it for the design patterns, then ask the harder question: what would we need to see before trusting one of these workflows to run unattended in our own organization?
The connection is control. What the agent may read. Where it may act. Whether the infrastructure can support it, and whether somebody can show that the work improved. Follow those questions into your own operation. That is where these stories become useful.
Follow my writing on Substack.