AI can run the loop. Choosing is still your job.
Dashboards, OKRs and strategy documents were built for slow, expensive information. AI changes the jobs they did, but not evenly.
Many years ago, when I was starting my first company, I made a classic mistake. Our team had an idea for a product, and I asked a well-known entrepreneur I respected what he thought of it. He was enthusiastic. With his encouragement and our own vision, we decided to push forward, and when progress was slow I kept at it, because persistence is a good trait in an entrepreneur. Or so we are told. We never found product-market fit. Much later I asked him whether he had ever used our product. Of course not, he said. He had just assumed someone else would.
Looking back, we were acting on one person’s opinion and our own enthusiasm, and nothing in how we worked would have told us early that we were wrong. Ways to find out already existed. Customer development techniques, such as customer interviews or the then-new Lean Startup, would have tested our idea against real customers, and I’ve used methods like these with good results since. AI wouldn’t have stopped me from trusting the wrong person, but it makes that kind of testing much cheaper: asking many more customers, and trying a few versions of the idea instead of betting everything on one.
Most of the management tools we still use come from that world, where information was slow to gather. So I’ve started asking the same question about each of them: what was this practice actually solving, and do we still have that problem?
My answer starts from what management is for. It has always done three jobs: knowing where we are, choosing what to do and what not to do, and acting and learning from the result. AI changes these jobs very differently. It already helps us know where we are, and where results are easy to measure, it can take over much of the work. It helps least with choosing, and choosing is what makes the rest worth doing.
Our management tools assume information is expensive
Think about how direction gets set in a typical company. A strategy is written every few years. Goals are set and reviewed every quarter, with owners and targets. In between, people meet every week to share status, because there is no other way to know what everyone else is doing. Dashboards follow the same logic: collect the numbers in one place and look at them together on Monday.
Each of these made sense. Gathering information took real work, so you did it rarely and compressed the result into something people could carry around, like a few objectives or a strategy on a page. The rhythm of management followed the cost of information. You couldn’t revisit your choices every week, because you couldn’t afford to find out whether they still held.
That cost has dropped a lot, and I’ve seen it in my own work. Two years ago, I would open a Jupyter notebook several times a week to understand our data. Now I ask the question I have today in plain language, of Claude or of the AI chat in our product analytics tool, and get a reasonable deep dive in minutes, without writing a line of Python or SQL.
Most of what a manager needs to know, from customer conversations to what other teams are working on, can now be current every day. Tools built around a quarterly refresh are out of step with that.
Each tool was doing three jobs at once
When I took these tools apart, I noticed that each one was doing the same jobs in different proportions, bundled together because they were reviewed on the same calendar.
The first job is knowing: where are we, and what is really going on? The second is choosing: where do we focus, and what do we leave out? The third is acting: doing the work and learning from the result.
This is close to what Richard Rumelt calls the kernel of good strategy: a diagnosis, a guiding policy and coherent action. I use plainer words because the pattern holds for management in general. What’s new is that two of these jobs can now run continuously instead of once per planning period.
Here is how the tools split across the jobs:
AI keeps the picture current
Dashboards did one important job well: monitoring, the glance that tells you something has moved that shouldn’t have. AI can take that job over. An alert doesn’t wait for Monday’s meeting, and it can watch many numbers at once, including the slow drift nobody would notice on a chart.
A dashboard also answers only the questions someone chose months ago, and it never told you why a number moved. Working that out was always the manager’s job. AI can now suggest an explanation, often a good one, but deciding whether it names the real problem is still a human judgment.
Get Topi Järvinen’s stories in your inbox
Join Medium for free to get updates from this writer.
The bigger change is that the picture can be shared. If everyone works from the same current view of where we are, and of what’s happening with customers and in the market, I’d expect meetings to need much less time to bring people up to date. It helps even more when the shared picture holds the choices too: what we decided and what we ruled out, with the reasons. Then people can act without waiting for the next meeting and still pull in the same direction.
I think a shared record like this only works if it builds up as part of the work people already do, and if it keeps the decisions rather than every word said. The company I work for, In Parallel, builds one version of this from meetings, team chats and other systems such as project management and CRM. You can also start smaller, with a decision log and a living strategy page that people and their AI tools both read.
Measurable work can run as a loop
AI changes the work itself the most, but the gains vary across tasks and situations. They are largest where the result comes back fast and can be measured.
Say a display campaign is underperforming. The usual response is a meeting where everyone has an opinion about the approach, the message, the target audience or the creative. The alternative is to generate variations automatically and keep the ones that test best, without the debate. Each round tells you what to try next.
The cleanest public example I know is Andrej Karpathy’s autoresearch. The person writes the instructions and chooses the metric. An AI agent edits the training code, runs a five-minute experiment, checks whether the result improved, keeps or discards the change, and repeats, around a hundred times overnight. You wake up to a log of experiments and, with luck, a better model.
The catch is that it works because a five-minute run gives a clear answer. Most management decisions don’t give one that fast. A pricing change or a move into a new market can take months to show a result. AI still helps with that work, but a person has to judge each result, and the automated part stays small. For now, most of the gain is in knowing where we are.
Choosing matters more now
Roger Martin describes strategy as the choice to do some things and not others. Cheaper information doesn’t make that choice easier. If anything it gets harder, because you can now see more options and act on more of them.
Years ago, at a company I worked for before my current job, we discussed a large development sprint to automate our revenue sharing, which was then a manual workflow with manual payments. The person doing the work said he’d be happy to stop, but building the automation would cost more than the time he spent on it, so it should probably wait. The cost made the choice for us, and it turned out well. When we did automate it later, we built a better system, because by then we knew what the work really required. Today building it would be cheap, but I’d still consider doing the steps by hand first, just to learn more.
I’ve come to think this was the lasting value of OKRs: they forced companies and managers to decide, choose and focus. The same is true of strategy work, where the value was in the discussion and the choices it produced, far more than in the SWOT slides or the vision statement.
For the same reason, asking AI to write the whole strategy doesn’t make sense to me. AI can help a lot with understanding the market and planning the work, and it makes a good sparring partner for your own thinking. But a finished strategy from AI skips exactly the part that mattered: people discussing the options and making the choices.
What changes is how tightly we hold the plan. In the old model, changing course halfway through a strategy period felt like giving up. With current information, it’s often simply the right call. You can keep the choices and still change the plan, which is close to how Rumelt describes a guiding policy:
“It is like a signpost, marking the direction forward but not defining the details of the trip.” — Richard Rumelt, Good Strategy Bad Strategy
Autoresearch shows how choices and loops fit together. A person decided what to improve and how to measure it, and the agent did the rest. Turning a choice into a number a loop can chase is part of the human job too. A fast loop aimed at something nobody chose is just busy.
A loop needs a metric you can trust
Running a loop on a bad metric is worse than showing that metric on a dashboard. A weekly meeting was slow, but someone in it could still look at a number and say, “That can’t be right.” An automated loop takes that person out. Goodhart’s law applies: the harder you push on a number, the less it tells you, and a loop pushes much harder than any meeting did. When a metric is safe to put inside a loop is the subject of my next piece.
Spend the time on choosing
So what should a management team do with this? Hand most of the monitoring and reporting to AI, and run loops wherever the result can be measured. Then spend the time this frees on choosing well, and on revisiting your choices when the picture changes rather than when the calendar says so.
Back to my question: do we still have the problem each tool was built to solve? For dashboards, mostly no, because AI can do the monitoring without anyone looking at a chart on Monday. For OKRs and strategy work, yes, and more than before. What they really forced was a choice, and a team that can try almost anything has more to choose between.
As for my first company: with today’s tools, we would have found out much sooner that customers didn’t want what we were building. But deciding what to build, and when to stop, would still have been my job.
I’d like to hear how this looks from where you sit. Reach me at topi@in-parallel.com.
