When AI Only Made the Mistakes Faster

There is a conversation that keeps happening at companies about a year into generative AI. Writing is faster. Summaries turn up. Meeting notes assemble themselves. And yet the time it takes to actually decide anything has not moved. If anything there is more to check now, and the week feels fuller than it did.

This is written for people responsible for AI adoption inside an organization. It is also for anyone who has noticed that their own work sped up while the whole did not. The cause is not individual technique, so it turns fairly quickly into a question about who decides what.

Most organizations are standing on a bridge

We lay out the progression like this: moving information off paper, rebuilding the work itself, starting to use generative AI, beginning to redesign work around AI, and finally having that be ordinary.

Most organizations sit between the third and fourth. Generative AI is in use. The shape of the work has not changed to assume it. That is a bridge, and a lot of people are standing on it.

The fourth stage onward is increasingly labeled "AX." That term does not have a settled definition yet. In Japan it is still under discussion in a government working group, and what has been published is a secretariat draft. We do not write "the ministry defines AX as." What matters operationally is not the label but what happens to people stuck on the bridge.

Five things block the crossing

These are the five we work from.

1. The data is not in order
2. The work is not standardized
3. The systems do not talk to each other
4. Authority and responsibility are vague
5. There is no culture that absorbs change

You will have noticed that none of the five is about AI. Every one is homework left over from before generative AI arrived. Whatever went unfinished in the earlier stages is sitting at the foot of the bridge, waiting.

Which is why the mess is what speeds up

Here is the awkward part. Introduce a fast tool on top of unfinished homework and the good work is not the only thing that accelerates.

Hand unstandardized work to AI and the number of different ways of doing it grows faster than before. Ask for summaries when the underlying data is patchy and thinly-supported summaries circulate faster than before. With nobody assigned to check, those go straight into the next meeting.

The tool is working correctly. It was told to make things faster and it is making them faster. What went wrong is that something which should not have been sped up was in the pile.

Only one of the five moves this week

Getting the data in order, standardizing the work, connecting the systems. Those three take time, money and people. They do not start this week. Culture takes longer still.

That leaves authority and responsibility. This one needs no new budget and can start immediately. It also limits the damage from the other four while they remain unfinished. Standardization can still be a mess, but if someone is designated to check, the errors stop before they leave the room.

This is where organizations get stuck, though. Who decides and who checks is not written on the org chart. Job titles are on there. Which person actually plays the role of saying "this is wrong" is a separate matter.

Read the current relationship, not the title

RELACTIS is the assessment we built for that. Twenty-four questions, about four minutes, and you get seven factors (Creation, Crossover, Autonomy, Advocacy, Systemization, Verification, Engagement) plus which of fifteen types you sit closest to. It costs nothing and needs no signup. Take it alone to see your own shape, or run it across a team for the distribution by role.

Three seats bear on authority and responsibility.

Read a distribution and the empty seats are what you notice first. Nobody checking, while summarization keeps getting rolled out further, is a real combination. That is not negligence. It is a role nobody has been asked to take, and asking usually fills it.

What it will not do, stated plainly: the assessment does not assign the seats. It shows who currently engages in what way, and deciding who to ask remains a human job. If twenty-four questions could settle it, the org chart would already have been enough.

One thing for tomorrow

Bring to mind one task you recently handed to AI. Summarizing, drafting, sorting, whichever. Then ask this:

"If this comes out wrong, who notices?"

If a name comes back, the checking seat for that task is filled. If nothing comes back, you have found the near edge of the fourth blocker. No budget and no meeting required, and you will know this week.