Why AI Productivity Is Not Reaching Revenue or Margin
Companies running AI across the front office and the floor report the same thing: more work is getting done. Revenue and margin have not moved. The distance between those two facts is not a technology problem. It is a resource allocation problem, and every old rule about scarcity still applies.
The Pattern Has a Precedent
In 1987 the economist Robert Solow wrote that you could see the computer age everywhere but in the productivity statistics. Companies had been buying computers for more than a decade. Output per hour had barely moved. The lag was not caused by the hardware. Companies dropped computers into processes built around paper and got faster paper.
There is an older case that fits even better. In 1865 William Stanley Jevons studied what happened when steam engines got more efficient at burning coal. Coal use went up, not down. Cheaper energy per unit of work meant far more units of work got attempted, and total consumption rose with it. Economists have called it the Jevons paradox ever since.
Tokens are on the same curve. The cost of producing a unit of knowledge work fell hard and fast, and consumption rose to meet it. What did not arrive automatically was any answer to which units of work were worth producing.
Cheap capacity does not allocate itself.
Resources, Activities, Outcomes
A company should be watching three things together, and always in this order:
- Resources. Capital, labor hours, attention, and now tokens.
- Activities. The work those resources produce. Quotes sent, orders planned, cases closed, tools built, etc.
- Outcomes. Revenue, margin, cash flow. The only three that reach the statements.
You spend scarce resources to drive activities, and those activities convert to outcomes. That gives you two ratios, and both of them have to hold.
Activity divided by resource is efficiency. It tells you how much work a dollar or an hour buys. Outcome divided by activity is conversion. It tells you how much of that work reaches the income statement.
Almost every AI report we get handed covers the first ratio and ignores the second. Usage dashboards, seat counts, hours saved, tickets deflected. None of that is an outcome. Efficiency without conversion is spend.
What Token Abundance Did to the Chain
Outcomes got easier for individuals to reach. A front office worker can automate invoice research. A warehouse lead can build a planning tool in an afternoon. A sales team can stand up a daily report to stay on top of the inbox. Every one of those is real work that used to require a ticket, a queue, and a developer.
They can also build a long list of things they want that will never drive a return, and nothing in the system says which is which.
Time, energy, and attention used to be the constraint that kept almost everyone pointed at company value. Short term, to hit their metrics. Long term, to keep their job. Those limits did the prioritizing for you, quietly, without anyone writing a policy. Now people can build, iterate, develop, and do, and a lot of energy and tokens go into work that nobody ever ranked. Most of them do not know what the priorities are, because nobody told them.
Scarcity used to do the prioritizing for you.
Ideas From the Bottom. Direction From the Top.
You want the ideas coming from the bottom up, nearest to where the work happens, where the needs sit and where the problems develop. That instinct is correct and that half is working. The people closest to a broken process have always known which part of it is broken.
Top-down direction is the half that went missing. Leadership wants AI and is frozen in a research stage, running pilots and evaluating vendors, while the people working for them already have the problem in front of them and can see the fix. So the fix gets built anyway, without a ranking.
Where build directives used to come from people watching revenue growth, margin expansion, and cash flow, they now come from people watching activity. That is the whole problem in one sentence.
Bottom-up ideas still need top-down ranking.
Running a Tool Through Both Ratios
Everyone has had the idea to build something that reads email faster, organizes files, pulls updates before a meeting, or cleans up the notes process. Before any of it gets built, run it through both ratios.
- Activity per resource. The tool has to produce more of a unit of work the company wants more of, per dollar and per hour. Reading email faster is not a unit of work the company sells.
- Outcome per activity. The work it produces has to convert better than the work it replaced. A cleaner notes process still yields the same number of closed deals.
Most personal tooling improves neither at the company level. It moves one person's convenience, which is worth something, and which is also why it never shows up in the numbers. Companies are building everything without checking whether they are producing activity or producing conversion. They are spending resources without checking whether those resources drive the right activity in the first place.
Convenience is not conversion.
Three Things to Settle Before Anything Gets Built
AI work needs directives from people who think in outcomes, and those directives have to be settled in advance. Three of them, in writing, with numbers:
- The activity you are working to increase. Named, countable, and already tied to revenue or cost.
- The resources you are willing to commit. Labor hours, build time, and a token ceiling.
- The conversion you expect from that activity. Stated as a rate, so it can be checked in ninety days.
Almost none of the companies we have worked with had any of the three written down before they started. What they had instead was a good stack of working tools, heavy token usage, and the same business as before. A few came out worse, because forty small tools carry maintenance, and maintenance is labor.
Write the target before you write the tool.
Diligence Before the Build
Before AI touches a workflow, four gates. A no at any one of them stops the build.
The first gate catches the most expensive mistake. Plenty of workflows are not workflows at all, they are workarounds people built because a system would not do something ten years ago. Automate one of those and you make it permanent, and you make it faster, and now nobody will ever go fix the system underneath.
The second gate is where assumption does the most damage. Teams point at the step that feels slow rather than the step that is slow. Measure it before you fund anything against it.
The third gate is the honest one. Some constraints are physical, contractual, or a supplier's problem. AI does not move those, and speeding up the paperwork around them changes nothing downstream.
Past the gates, do the modeling work before the build work. Map the improved process first. Put expected activity levels, outcome conversions, and required resources against it, with their cost. Then compare all of it to what the current process costs, produces, and takes in time, so you are looking at an opportunity cost and not a wish.
Automating a workaround makes the workaround permanent.
Put Tokens on the P&L
From a business standpoint, tokens are a form of labor. They do work, they cost money per unit of work, and their productivity can be measured. So treat them the way you treat labor. Give them their own line item, and build a few productivity metrics on top of it.
The new cost of work delivered is fully burdened labor plus token spend.
Run the nominal figures next to the dollar figures. Outcome per head. Outcome per million tokens. Then activity as a ratio to both labor and tokens, again in spend and in raw counts. The reason for tracking both is straightforward: model prices keep falling, and a falling input price will make every dollar-denominated ratio look better while the underlying work stays exactly as productive as it was.
A ratio that moves only in dollars is a price change.
What Changes When Tokens Sit Next to Labor
The measurement does most of the enforcement on its own. Once token spend has a line item and an owner, build requests start arriving with a conversion target attached, because someone now has a number to defend. Tools that produce activity and nothing else stop getting funded, because the line item makes them visible for the first time. The ranking that scarcity used to handle quietly gets handled deliberately instead.
That is the whole correction. Use a standard business approach on a new resource. Put it on the P&L, measure it in productivity and conversion, and the proper use of AI gets forced from there.
BlackArc Industrial builds AI into workflows for manufacturers, distributors, and industrial services companies. We fix the process before we automate it, and we hold the work to one number: cash flow.
Cash flow is the only scoreboard that settles the argument.