By Janus Boye
The bet on AI is getting bigger. This year, some teams saw substantial improvements. Others are still working to get internal approval. The full cost is still an open question. How ready are we to rely on AI for more of our work in 2027?
I spent this week with our Boye & Co peer groups in San Francisco, New York City, Boston and Montréal. We compared notes on what changed in 2026, from how our roles are evolving to what happens when the systems we depend on fail. Major cloud outages disrupted essential services, and critical software vulnerabilities and security breaches put customers at risk.
These ten observations are my first attempt to make sense of those conversations. Which ring true for you? What are we missing?
1. Context engineering: From prompts to context and memory
If two teams give different answers about who’s eligible for a service, which answer should an AI agent use? In New York, Anu Jagga Narang of AT&T brought us back to the question of who decides which source to trust and how we know where the information came from. Before an agent can act on what an organisation knows, someone has to decide which source to use and who keeps it current. Matthew Garrepy’s New York recap
A skill file can describe how a task should be done, and the Model Context Protocol (MCP) can connect the agent to the information and tools it needs. The harder part is preserving what the organisation has learnt. A policy might tell an agent what the rule is today. Knowing why the rule changed and when an exception was agreed can matter just as much. That knowledge often sits with the people who made the decision.
2. What still works when your digital infrastructure is down?
The Microsoft 365 disruption on 12 and 13 November lasted just under 24 hours, according to Microsoft’s preliminary report. Exchange Online was hit hardest, with effects on parts of Teams, SharePoint and other services that depend on it. When the tools people use to coordinate a response share the same dependency, even working around an outage can become part of the problem. Microsoft’s preliminary incident report
Can staff still handle urgent requests? Can we reach the people who need to know what’s happening? The answers tell us which parts of the organisation can keep working while a supplier restores service, and where we might need another way to operate.
3. Digital sovereignty is really happening
We spent a little more time on digital sovereignty in Canada than we did in our US meetings. It brought me back to Digital sovereignty – some assembly required, the Boye Aarhus 26 presentation by Mathias Bolt Lesniak, Project Ambassador at TYPO3. He picked up a conversation he and Jeffrey “Jam” McGuire, Partner at Open Strategy Partners, had opened in a members’ call earlier in the year. Where a service runs matters. So does who gets to decide how it operates and what happens when the supplier changes the terms. Boye Aarhus 26 session
Jam posed a question that belongs in a procurement conversation. Could we operate and adapt this system ourselves if we needed to? Open source can give us more options, including working with another supplier or bringing some of the work in-house. To exercise those options, we still need people who understand the system and know what it would take to move it. By the time we want to leave a supplier, much of our freedom to act has already been decided. Digital Sovereignty 2026: Some Assembly Required
4. Security is going to get worse before it gets better
This week brought two reminders of how much website security depends on decisions made across teams. The open source PHP project issued updates to address vulnerabilities, including critical ones. Separately, Optimizely confirmed that malicious code had been added to some customer projects through its Web Experimentation service. PHP security updates · Optimizely security advisory
In the PHP case, someone needs to know where it’s running and make sure it gets updated. In the Optimizely case, customers need to know which pages load the supplier’s scripts and who can disable them. Now add agents that can read internal documents and use tools on our behalf. Every new connection adds something for teams to keep track of and a decision about who gets access. That’s why I expect security to get harder before our ways of working catch up.
5. AI search changes (almost) everything
Someone using ChatGPT, Perplexity or Google AI Overviews may get enough of an answer to make a decision without ever visiting the websites it draws on. Sagility has taken that possibility seriously enough to publish a page specifically for AI assistants. It sets out what the company does, defines its offerings and explains how they relate. It made me wonder how easy it would be to find that same clarity on the rest of our own websites. Sagility’s page for AI assistants
Search on our own websites still matters. Someone who already knows we offer a service may use a different name for it than the team that publishes the page. They may find three old versions before they get to the current one. An AI assistant drawing on those same pages can run into the same ambiguity.
That’s where content modelling and taxonomy come in. A content model can give every service a place for the information someone needs to act, including who’s eligible and what to do next. A taxonomy can connect the language visitors use with the terms teams use internally. When digital teams want to scale AI across hundreds of services, those decisions about structure and language need to hold up across the whole site.
6. A rollercoaster year for designers
“Your AI is showing” was the phrase Matthew Garrepy picked up from Lubna Abu-Osba’s contribution in New York. It gets at a tension designers have been working through this year. It’s easier to produce another version, but a polished result can still look like it could have come from anyone. Matthew Garrepy’s New York recap
The interesting question is what designers do with that extra room to explore. A team might use it to test ideas it would previously have set aside for lack of time. It could also end up reviewing more variations without getting any closer to something that speaks to its audience. Knowing which ideas are worth pursuing and why is part of the work, even when making another version takes seconds.
7. Tech companies are preparing for a tougher year
Some tech companies are going into 2027 planning for a possible recession. Some CEOs have already been asked to plan for substantial staff cuts. If customers pull back on spending, companies that keep winning business will have more room to invest. Those already missing their growth targets may come under pressure from their owners to cut jobs quickly and in large numbers.
For some people, that’s already happened. In October, Dropbox announced it was eliminating about 200 roles, roughly 9% of its global workforce, including jobs in engineering, product and design. Earlier in the year, Amazon announced cuts affecting about 16,000 roles. Dropbox’s announcement · Amazon’s announcement
It’s reasonable for people in tech to be worried. A budget plan can be provisional for months while employees are already making decisions about whether to stay, look for another job or change direction. How candid can leaders be about what’s under discussion, and what support can people count on if their job goes?
8. Meta’s MUSE is bigger than we may think
Meta’s Muse can use a browser on someone’s behalf. That means a person can use a website without ever opening it themselves. It’s still early, but Muse already accounted for 59% of the agent traffic HUMAN Security measured across its network in October. That gives us an idea of how quickly the mix of visitors can change. HUMAN Security’s October traffic report
For the people running websites, there are two questions. Can an agent help someone get what they need from the site? And what does that activity cost to serve? Some automated requests may show up in server logs and on the hosting bill while passing under the radar of the usual analytics. Looking at those sources together may change our understanding of who’s using the site and what it costs to serve them.
9. What do we hand over to agents, and who stays responsible?
In New York, Mario Oliver described how agents were both helping him and adding to his workload as a product manager. An agent can prepare more work in less time, while the person responsible spends more of their day reviewing it and sorting out exceptions. What looks like a productivity gain from one part of the organisation can look like a different job from another. Matthew Garrepy’s New York recap
That changes how people learn the job, too. Much of the judgement that goes into reviewing work comes from having done it yourself. If an agent handles the routine cases and people mainly see the exceptions, where does someone new to the role get the experience to make those calls?
There’s also a decision about what the agent can do on its own. Preparing an update for review and applying it to a customer’s record carry different consequences. The team that configures the agent may be different from the team that has to explain a mistake to the customer. The people who answer for the decision need to be able to see what the agent did, pause it and put things right.
10. What gets better when we work faster?
In New York, Amit Patel of Caylent brought the discussion back to outcomes. A team may be able to produce ten times as much, but what gets better for the people the work is meant to serve? Matthew Garrepy’s New York recap
Suppose a team can now publish more service pages in a week. Can someone find the right service, understand whether they’re eligible and take the next step? The publishing team can count the pages. Search data, conversations with users and questions coming into the support team can tell us whether those pages helped.
Those teams also see different parts of the cost. The software bill tells us what we paid to run it. The people checking the output, correcting errors and keeping the source material current can tell us how much work it takes to make it useful. Before we decide what’s worth scaling in 2027, we still have to connect what those people know.
