← back to the archiveCover illustration for “AI adoption needs a default-path test”
POSTday 173·2d ago·Published ·by Andy Padia

AI adoption needs a default-path test

In short: AI adoption is real when the assisted workflow becomes the easiest legitimate path and every exception explains why it did not.

Atlassian's State of Teams 2026 surveyed 12,035 knowledge workers and 173 Fortune 1000 executives. Eighty-nine per cent of the executives said AI increased the speed of work. Only 6% were sure they could point to clear organisation-wide ROI.

That gap will not close with a firmer instruction to use AI. My rule is that adoption starts only when the assisted route becomes the easiest legitimate way to finish a named workflow, and when every exit from that route leaves an exception reason.

Personal speed does not prove the workflow changed

Atlassian collected its survey data in January and February 2026. The report says 80% of work occurs at team level, while many AI programmes still concentrate on individual productivity. It calls the resulting coordination cost a fragmentation tax.

McKinsey found a related split in a different sample. Its 12 March 2025 State of AI survey tested 25 organisational attributes. Workflow redesign had the largest effect on respondents' ability to report EBIT impact from generative AI, yet only 21% of organisations using generative AI said they had fundamentally redesigned at least some workflows.

Both are survey findings, not a causal experiment. They still expose the operating mistake. A person can draft one step faster while approvals, handoffs, duplicate entry and rework remain exactly where they were. The tool changed. The path did not.

The default-path test measures behaviour inside the work

I would pick one recurring workflow and make the assisted route the normal route inside the system where the work already happens. Keep the old route available as an explicit exception, not as an invisible side channel.

Then measure the path itself:

SignalWhat to record
Eligible workCases that should enter the assisted route
Default-path completionCases accepted without leaving that route
ExceptionsThe exact reason a person switched paths
Accepted outputWork that passed the existing owner or quality rule

The exception reasons are the useful part. “Policy forbids this data,” “the model missed a required field,” and “review took longer than doing it manually” point to different fixes. A generic non-adopter label hides all three.

This is editorial judgement, not a Trigent or client deployment I ran. I have not tested the proposed default-path measure against EBIT. It is a more honest adoption test because it records the behaviour at the workflow boundary instead of inferring it from seats, prompts or self-reported confidence.

A mandate can start the trial, then get out of the way

A mandate has one useful job: it makes trying the new route part of the work. I argued the same narrow case when reviewing Klaviyo's L3 fluency requirement. The mandate becomes misleading when compliance remains the headline after the workflow is supposed to have changed.

Give the mandate a short life. Once the default route exists, report eligible cases, accepted completions and exceptions. If people keep leaving for the old path, read the reasons before sending another training reminder. The problem may be permission, missing context, poor output or a broken handoff. None of those is repaired by asking for more prompts.

What's in it for you

  • You can distinguish resistance from a legitimate product, policy or quality failure.
  • You can improve the exact point where people abandon the assisted workflow.
  • You can keep speed evidence without pretending that it already proves enterprise return.

AI adoption is real when the assisted route becomes the default for accepted work and every exception teaches you why it did not.

#ai-adoption#workflow-design#enterprise-ai#measurement#change-management

Sources

  1. atlassian.com/blog/state-of-teams-2026
  2. mckinsey.com/capabilities/quantumblack/our-insights/the-state-of-ai-how-organizations-are-rewiring-to-capture-value
← older drop
Claude Code: Anthropic's agentic coding tool
newer drop →
Speculative decoding: a draft model speeding up a big one

related drops

explore all 557 drops →
← back to the archiveday 175