A year ago a lot of people were asking us whether AI would replace their teams. Now they ask better questions. They ask where it actually saves an hour, what it costs to run, and who checks the output. That shift alone tells you how the year went.
The first thing we learned is that the model is rarely the bottleneck. In almost every engagement, the slow part was the data. Records spread across three systems, a spreadsheet that two people maintain differently, a process that lives in one person's head. AI sits on top of all that. If the foundation is messy, the AI is confidently messy, which is worse than a blank page.
So we kept starting where we always start, with a Discovery Sprint. Two to four weeks, ending in a working prototype, not a slide deck. The prototype matters because it turns a vague hope into a concrete thing someone can use on Monday and tell us what is wrong with it. Half of what people imagine they want falls away the moment they touch a real version.
The wins were quieter than the headlines
The features that earned their place were unglamorous. Pulling fields off an invoice or a packing list. Sorting a busy inbox so the urgent thing surfaces first. Drafting a first reply that a human tidies and sends. None of these are clever. All of them give people back time they were spending on work no one enjoys.
We have woven AI into specific operational workflows, automation and exception checks among them, and the lesson there matched everything else. The value was in catching the odd case early and letting people spend their attention where judgement actually mattered.
Where it went wrong was predictable
When a pilot stalled, it usually was not the AI. It was that no one had decided who reviews the output, or the feature solved a problem the team did not really have, or it was bolted onto a process that was broken to begin with. AI does not fix a broken process. It runs it faster.
The other quiet lesson is cost, and not the cost you expect. The bill from the model provider was almost never the issue. The real cost was review time. Every AI output that touches a customer or a number needs a person to glance at it, at least at first. Design that review in and the feature is honest. Pretend it is not needed and the feature quietly breaks trust.
If we had to compress the year into one sentence, it is this. AI works once the data and the processes underneath are right, and getting those right is most of the job. The teams that treated AI as a tool to make a good process sharper got real value. The teams chasing magic spent money and got tired.
We are more optimistic than we were a year ago, but for ordinary reasons. The boring uses keep paying off.
Facing something similar in your business?
Talk it through with our AI guide, or send the team a note. We will tell you straight whether and how we can help.