What “More Productive With AI” Actually Means

I have watched this happen in enough rooms to know the shape of it before it finishes. A leadership team decides it wants to be more productive with AI. Nobody objects, nobody asks what the phrase means, and the decision is recorded as a decision. Six months later the same team is arguing about metrics, and everyone believes the argument is about measurement.

It is not about measurement. It is about a question that was skipped, and every downstream disagreement is a symptom of the skip.

Five destinations wearing the same words

Productivity is a ratio. Something out, divided by something in. The phrase “more productive with AI” leaves both terms unspecified, which is why it accommodates so many private interpretations without any of them surfacing.

The first destination is faster output. The same work, delivered in less time. This is the definition most engineering leaders hold, and it demands investment in cycle time: removing handoffs, shortening review queues, and making sure the constraint is not sitting downstream of the accelerated step. An organization that gets faster at generating work while its approval chain stays the same length has not moved at all. It has built a larger queue.

The second destination is fewer people. The same output at a smaller headcount. This one is rarely said aloud in the meeting, which does not stop it from being present in the meeting. It demands investment in role redesign, in genuine consolidation of responsibilities, and in the difficult work of deciding which capabilities the organization is willing to no longer have. It also produces a specific behavioral response the moment anyone suspects it is the real destination, and that response is not enthusiasm.

The third destination is cheaper delivery. Lower total cost per unit of work. It looks like the second destination and is not. Cost per unit can fall with the same headcount, and it can rise with a smaller one, because the cost that matters includes rework, review, incident response and the coordination overhead of a more complicated toolchain. This destination demands unit economics that most organizations do not currently compute, which is why it is the easiest one to claim and the hardest one to prove.

The fourth destination is higher quality. The same volume with fewer defects, fewer reversals and better decisions. It demands investment in review capacity and in feedback loops that tell you what a decision produced. It is also the destination most directly in tension with the first. Speed and quality can rise together, but only when the organization has built the loop that catches the errors speed introduces, and building that loop costs the very time the first destination is trying to reclaim.

The fifth destination is more experiments. A larger number of attempts at things the organization has never done. This one is different in kind from the other four, because it is not about doing existing work better. It is about doing different work. It demands tolerance for failure rates that would be unacceptable in the other four destinations, and it demands the authority to stop something that is not working without a defense of the original decision.

James March named this last tension in 1991 and it has not softened since. Exploitation, meaning the refinement of what an organization already does, and exploration, meaning the search for what it does not yet do, compete for the same finite resources. Returns from exploitation are closer in time, more certain, and easier to attribute. Returns from exploration are distant, variable, and frequently attributable to nobody. An organization that pursues both without naming which one it is funding will fund exploitation by default and be surprised when nothing new appears.

Why the meeting produces agreement and not alignment

Nothing goes wrong in the meeting. That is the difficult part.

The phrase is not chosen to be ambiguous. It is chosen because it is true. Every one of the five destinations is a real form of productivity, and the sentence is a fair description of all of them. When the CFO hears it, cost per unit is what productivity has always meant in her function, so she hears cost. When the head of engineering hears it, cycle time is what productivity has always meant in his, so he hears speed. Neither of them is interpreting loosely. Each is applying the operational definition their function has used for years, correctly, to a sentence that did not exclude it.

So agreement is genuine and instantaneous. Nobody is being political. Nobody is withholding. Five people sincerely endorse the same sentence and leave with five different pictures, and because the pictures never got compared, the difference is invisible until it shows up in a budget request or a quarterly result.

The clarifying question would fix this. It rarely gets asked, and the reason is structural rather than personal. Asking a senior group to define a term that everyone has just agreed on carries a social cost. It reads as slowing things down, or as not following, or as having missed something obvious. Amy Edmondson’s work on psychological safety in work teams describes exactly this mechanism: the interpersonal risk of appearing ignorant or obstructive suppresses precisely the questions that would surface a shared misunderstanding. The safer the team, the more likely someone says what does that mean. In most rooms, silence is cheaper.

What makes this an alignment problem rather than a communication problem is that communication succeeded. Everyone heard the words. The words were the wrong resolution for the decision being made.

How the ambiguity survives contact with a dashboard

The natural response is to instrument. If nobody is sure what productivity means, measure several things and let the data adjudicate.

This does not work, and the reason it does not work is old. V. F. Ridgway showed in 1956 that quantitative performance measures reliably produce behavior optimized for the measure rather than for the purpose behind it. A metric is not a neutral observation. It is an operational definition of a goal, and the moment it is reported it starts directing effort. Choosing metrics before choosing a destination does not defer the decision. It makes the decision quietly, in favor of whichever number is easiest to produce.

The balanced scorecard is the usual next move, and Robert Kaplan and David Norton were right in 1992 that a single financial measure gives an incomplete picture of performance. But a scorecard distributes attention across dimensions, it does not resolve conflicts between them. Put cycle time, cost per unit, defect rate and experiments launched on the same page and you have made all five destinations visible at once without saying which one wins when two of them disagree. Every team reads the scorecard and optimizes the row it owns. The scorecard becomes a record of the unresolved question rather than an answer to it.

Then the arguments start, and they present themselves as methodological. Whether cycle time is the right unit. Whether the baseline is fair. Whether that quarter should be excluded. These arguments feel technical and are not. They are the original unmade decision, surfacing in the only place it is now permitted to appear. You can tell because the arguments never resolve. Technical disagreements end when someone produces better data. Destination disagreements do not, because no data exists that can tell you which outcome the organization should want.

The AI-specific failure

Everything above predates AI. Organizations have been agreeing on undefined goals for as long as there have been organizations, and the usual consequence is drift: slow, absorbable, visible in hindsight.

AI removes the slowness. Capability distributed to every team simultaneously means every team can now act on its private definition immediately, at volume, without asking anyone. The engineering team pursues cycle time and generates more work than the review function can absorb. The finance team pursues cost per unit and drives out the review capacity the engineering team just made necessary. The product team runs more experiments and produces a decision backlog nobody has the authority to clear. Every one of those teams is executing well against the destination it heard.

When an organization deploys AI into an environment without alignment, the technology amplifies the fragmentation. Every team uses the tools for what they think matters. The outputs are locally optimized and organizationally incoherent.

From Builders Build, Chapter 9

Locally optimized and organizationally incoherent is the exact output signature of five destinations pursued in parallel. Melvin Conway observed in 1968 that organizations produce designs which mirror their own communication structures. An organization whose leaders never reconciled their definitions of productivity will produce an AI capability that mirrors that non reconciliation, in the tools it selects, the workflows it automates and the metrics it reports.

What is new is the speed of the mirroring and the cost of correcting it. Before AI, a team pursuing the wrong destination was limited by its own throughput. Now it is not. The divergence that used to take three years to become expensive takes two quarters.

The conversation that settles it

The fix is a conversation, and it is shorter than the arguments it prevents. It is not a workshop and it does not require a facilitator. It requires the leadership team in a room with three questions and a willingness to accept the answers.

Start by writing the five destinations on a wall and asking each leader to rank them privately before anyone speaks. Private first is not a formality. It is the whole mechanism. Ranked in public, the second person adjusts toward the first, and by the fifth you have manufactured a consensus that describes nothing. Ranked privately and then revealed, you get the actual distribution, and the spread in that distribution is the finding. A leadership team that discovers it holds four different first choices has learned more in ten minutes than it would learn from a quarter of metric debates.

Then ask the question that forces a choice: which destination are we willing to give up this year. A ranking that costs nothing is not a decision. If the answer is that all five matter, the team has not chosen, and the default will govern. The default is always exploitation, because it is nearer, safer and easier to attribute, and the experiments will not happen no matter how many times they are endorsed.

Then write the destination down in a sentence a team lead can apply without calling anyone. Not “we will drive AI enabled productivity improvements.” Something closer to: this year, more productive means the same work delivered in less calendar time, at current headcount, with defect rates held flat. That sentence is testable. It also tells a team lead what to do on a Tuesday when a tool promises cost savings that would slow delivery down, which is the only test of an operating definition that matters.

Expect disagreement when this conversation is done well, and treat the disagreement as the point. A group that ranks five destinations honestly will find real conflict, because the destinations really do conflict. Surfacing it is what makes the choice a choice. A room that reaches this question and produces no disagreement at all has usually not reached the question.

What a settled definition changes downstream

A named destination is not a strategy. It is the thing that makes four other decisions possible, all of which are currently stuck.

The metric stops being an argument. Once the destination is speed at flat quality, the measure follows from it rather than competing with it, and the quarterly debate about which number to report ends because the number is implied by a decision the group already made. Metric fights are goal fights in costume, and settled goals dissolve them.

Tool selection stops being a matter of preference. Different destinations favor different capabilities, and a team that knows it is optimizing for experiment volume will evaluate the same product differently from a team optimizing for unit cost. Without a destination, tool selection defaults to whoever is most persuasive, which is how organizations end up with overlapping licenses and no coherent account of why.

Training gets a target. Teaching people to work faster is a different curriculum from teaching people to try more things. Most AI enablement programs teach neither, because they were built to cover the tool rather than the destination, and the tool is the least transferable part of the skill.

And the tradeoff conversation becomes possible. An organization with a named destination can say out loud what it is not pursuing this year, which is the only way anyone below the leadership team can make consistent decisions in the gaps the strategy did not anticipate. That is the practical definition of alignment: not that everyone agreed, but that a person three levels down can make a call you would have made, without asking you.

None of this guarantees the destination was the right one. Organizations pick wrong, and picking wrong is recoverable. What is not recoverable in any reasonable time is a year of five teams pursuing five destinations while the leadership team believes it chose one. That failure produces no learning, because there is no result to evaluate. There is only a set of locally sensible outcomes and an argument about metrics that will not end.

Sources