Convergence Day · 1 Sep 2026

10:45 · Block one

Nearly eight months of change in 45 minutes

20–21 January to 1 September. 224 days — seven months and twelve days.

We last went deep at Tändstickspalatset, with JJ and Christina. Today we pick that thread back up.

Today's AI is the worst you'll ever have.

This is a principle, not a prediction. It does not mean wait for better AI. It means build habits and operating systems that improve as AI improves.

The AI you met in January is already the weakest AI this room has worked with. The question is whether your ways of working have moved too.

The unfinished promise

The room left January highly energized, intending to build an AI colleague for CMT meetings: meeting-aware, context-holding, something close to an executive assistant. The initiative met friction and quietly fizzled. We are not going to guess at the blocker.

Two things are different now. The Microsoft platform is materially more capable than it was in January, and the Convoy gives us a mechanism for removing friction instead of abandoning the idea.

The landscape

224 days. Three working modes. The tools moved; the unit of work changed.

Chat answers. Agents act. Cowork executes longer work. Better prompting still matters, but delegation, evidence and authority now matter more.

  • ~4.2 months

    Estimated doubling time for frontier software-task horizons since late 2024. The direction is strong; the precise rate is uncertain.

    UK AISI, interpreting METR

  • 2–17 weeks → 14 hours

    Epoch's MirrorCode: Opus 4.7 reimplemented a specified 16,000-line tool in 14 hours, passing 2,000 of 2,001 tests, at a cost of $251.

    A precisely specified reimplementation benchmark — not ambiguous strategy work.

    Epoch AI · MirrorCode

  • 48×

    Average machine execution per session: 26 minutes for an agent versus 33 seconds for search across 10,000 matched early-adopter pairs.

    An observational, product-specific field study; exact magnitudes are approximate.

    Perplexity Research + Harvard Business School researchers

Explore it yourself →

The working modes

Chat, Agent or Cowork?

A practical guide to choosing the right operating mode.

The money

$190 billion

Microsoft's planned capital expenditure for calendar 2026. Even at that level, it expected AI capacity to remain constrained through the year.

Microsoft FY26 Q3 earnings

  • From seat to seat + meter.

    Microsoft says its per-user businesses are becoming per-user plus usage. Cowork is usage-billed. The leadership question is no longer only who has a license, but which work creates enough value to justify consumption.

    source

  • SEK +500 million.

    Mycronic raised its 2026 net-sales outlook from SEK 8.75B to SEK 9.25B. Mycronic's own report explicitly ties demand in High Volume and Global Technologies to the AI-infrastructure build-out.

    source

  • $251 can buy 14 hours of agent work.

    It cannot buy the right brief, the right boundary or the right review.

    source

Intelligence got cheaper. Usage got priced. Judgment became the scarce input.

The reality checks

Honesty is a feature. Capability moved faster than reliability, governance and organizational absorption.

  • Longer does not mean generally reliable.

    AISI says the direction is clear but the exact time-horizon rate is uncertain: the longest tasks have the fewest human baselines, and narrow evaluations do not predict every real-world system.

    source

  • The spectacular demo had a perfect kind of problem.

    MirrorCode reimplemented existing software against extensive visible and hidden tests. Epoch explicitly says translation to day-to-day engineering remains an open question.

    source

  • More autonomy means more checkpoints, not fewer.

    In the Perplexity field study, agent sessions paused for user input in 13% of queries versus 0.3% for search. Autonomous work still needs approval and clarification.

    source

  • Availability is not readiness.

    Microsoft features roll out gradually and still depend on licensing, admin enablement, tenant configuration, permissions and governance. Today we use what is actually switched on.

    source

From January to now

  • Context

    thenFiles you pointed at

    nowWork IQ across accessible mail, files, chats, meetings, people and projects

  • Work mode

    thenChat plus first specialist agents

    nowAgent Mode in Office plus Cowork for multi-step work

  • Model choice

    thenMostly one default model

    nowMultiple models, auto-routing and Claude available in Copilot Chat

  • Economics

    thenPer-seat license

    nowSeat plus consumption, credits and usage billing

  • Governance

    thenPromising prototypes

    nowAgent registry, ownership, lifecycle and tool governance

What Microsoft changed

  • Work IQ

    Copilot answers using context across emails, files, chats, meetings, people and projects.

    Why it matters: the executive-assistant idea can draw on connected work context rather than a single pasted artifact.

    official source

  • Specialized and custom agents

    Researcher and Analyst for deep work; Agent Builder for your own agents grounded in work data, SharePoint and connectors.

    Why it matters: a useful helper can now start as a bounded agent prototype rather than a full implementation project.

    official source

  • Copilot Cowork

    Generally available. Multi-step work across Microsoft 365 with approval checkpoints and usage-based billing.

    Why it matters: this is the surface where a design becomes something that actually runs.

    official source

  • Agent governance

    Agent registry and Agent 365-style visibility: lifecycle, ownership and tool governance for the agents in the tenant.

    Why it matters: the leadership question moves from 'can we?' to 'under what authority, with what evidence?'

    official source

Meeting recap, transcripts and Copilot in Teams are the closest thing to January's idea already within reach — where they are enabled for you.

Licensing, admin enablement and tenant configuration all apply. Today we use what is actually switched on.

Which of these bites hardest in CMT right now?

0 people answering · change your mind any time

  • Volatile

    Technology and demand move faster than annual plans.

  • Uncertain

    Available capability depends on license, admin settings and tenant enablement.

  • Complex

    Context lives across meetings, mail, files, business systems and permissions.

  • Ambiguous

    The same AI can be assistant, agent or coworker. The right authority boundary is contextual.

Four principles, one upgrade

  1. 01Invite AI into everything you do
  2. 02Be the human in the loop
  3. 03Treat AI like a human — but remember it is not
  4. 04Today's AI is the worst you'll ever have

All four still hold. The frame moved from co-intelligence — thinking with a helper — toward co-existence: working alongside systems that may operate for longer and sometimes outperform you. The leadership upgrade is from permanent checking to designed authority, evidence and escalation.

Co-existence is Ethan Mollick’s framing.

Ethan Mollick · Co-Existence and the End of Co-Intelligence

Five vectors became six

We used five ways AI creates organizational value. A sixth became impossible to ignore.

  1. 01

    Proactive compliance and risk mitigation

    Spot risk earlier and make evidence part of the workflow.

  2. 02

    Organizational memory and knowledge activation

    Turn meetings, decisions and expertise into usable continuity.

  3. 03

    Decision velocity — and quality

    Compress preparation while making assumptions and evidence visible.

  4. 04

    Cognitive capacity and well-being

    Move repetitive load away from people without moving judgment away.

  5. 05

    Talent development, attraction and retention

    Let people grow through better work, not drown in low-value work.

  6. 06

    Access and inclusion

    When conversation becomes the interface, more people get to act.

The sixth vector matters here: the ability to build and delegate should not belong only to the technically fluent.

The five capabilities of the agent era

Knowledge work is shifting from doing every step yourself to specifying, delegating and judging work done with agents.

Rate yourself, honestly. 1 is not yet, 5 is can teach it. · 0 people rating

  • 01

    AI capability mapping

    Knowing what today's models can and cannot do, refreshed monthly rather than yearly.

    1 not yet · 5 can teach it · no ratings yet (0 ratings)

  • 02

    Context & harness management

    Giving the system the context, tools, permissions, guardrails and evidence it needs.

    1 not yet · 5 can teach it · no ratings yet (0 ratings)

  • 03

    Problem & product prototyping

    Turning a fuzzy need into something testable before lunch.

    1 not yet · 5 can teach it · no ratings yet (0 ratings)

  • 04

    New opportunity identification

    Spotting what is suddenly possible that was not possible last quarter.

    1 not yet · 5 can teach it · no ratings yet (0 ratings)

  • 05

    Rapid new skill acquisition

    Learning at the pace the tools change — the meta-capability.

    1 not yet · 5 can teach it · no ratings yet (0 ratings)

Five shifts we are living with

  • 01

    From answers to ongoing work

    Agents can plan, use tools and keep going. As the work gets longer, the brief matters more than a clever prompt.

  • 02

    From apps to conversation

    More work starts with intent instead of navigation. When conversation becomes the interface, the threshold to begin drops.

  • 03

    From task to outcome

    Delegation moves from ‘do this step’ toward ‘deliver this result, within these boundaries.’

  • 04

    From co-intelligence to co-existence

    Ethan Mollick calls this co-existence: not an occasional helper, but a working relationship you negotiate and re-negotiate as systems improve.

  • 05

    From checking everything to designing evidence

    Human judgment where it matters, not manual approval everywhere. Leaders define authority, required evidence and escalation paths.

The thing being managed is no longer primarily people or agents.It is the flow of agency through the organization.

The old manager loop

  1. assign work
  2. coordinate people
  3. review deliverables
  4. decide

The emerging leader loop

  1. define outcome
  2. design roles & authority
  3. orchestrate capacity
  4. observe evidence
  5. intervene on exceptions
  6. improve the system

Where must CMT build muscle first?

  1. 1Define outcome
  2. 2Design roles & authority
  3. 3Orchestrate capacity
  4. 4Observe evidence
  5. 5Intervene on exceptions
  6. 6Improve the system

Drag to reorder, or use the arrow buttons.

connecting…

The room ranking appears once you submit yours.

Leadership becomes less about controlling execution, and more about designing the conditions under which execution can safely happen without you.

Sources