The Outcomes Were the Easy Part: What it Takes to Operate Agentic AI at Scale

Twenty years ago we optimized systems. Ten years ago we started optimizing labor. Today we are optimizing AI. And yet most communications providers are still fighting the same battles they fought through all three: delayed activations, orders that fall out and stall, revenue that leaks, bills that go out wrong, operating costs that keep climbing, and customer journeys stitched together from parts that never met.

The reason is simpler than the technology makes it look. We optimized technology and labor in isolation, each on its own track, and largely ignored the thing that decides whether either one produces value: the coordination of work between people, systems, partners, and now agents. Outcomes are not created by technology. They are created by coordinated execution, and that is the part we keep skipping.

I wrote a while back that most AI transformations are expensive ways to make broken processes run faster, that automating a broken process just gets you to broken faster, and that the work that matters happens before any code, in defining the outcome and mapping the real job. That argument won. Outcome-first is the word of the day now, and I am glad it took. But naming the outcome was the easy part. The hard part is everything that has to happen across all those people, systems, partners, and agents to actually deliver it.

The coordination tax

Here is the thing nobody puts on the slide. Every new system, vendor, workflow, partner, and now every new agent introduces a small coordination cost: one more place to log in, one more handoff, one more reconciliation. Each is minor. Nobody notices at first.

But they compound. Across twenty years of modernization they have added up to what is effectively a tax on the enterprise, a coordination tax, paid every day in a currency most operators never put on a report. It shows up as swivel-chair work between systems that do not talk, as exception queues that grow faster than anyone can clear them, as orders that stall waiting on a handoff, and as a customer who feels every seam. The work itself is usually not hard. The trouble is that it is fragmented, and the cost of stitching it back together by hand, over and over, is both enormous and nearly invisible. That tax, not the technology and not the labor rate, is the largest inefficiency most operators have left to recover.

The response is a discipline, not a deployment

So the real question is not how much AI you can deploy. It is how you take the tax out of the work without losing the people who understand how the work actually happens.

The answer, at least as I have come to practice it, is what I call agentic workforce optimization: the disciplined practice of optimizing a blended workforce of people and agents against an outcome, and tuning it over time. Optimization, not automation. I will come back to that distinction, because everything depends on it.

Start with the outcome, but know it is only the floor

You do not reduce the coordination tax by buying more technology, and you certainly do not reduce it by automating the fragmented process exactly as it stands. You start with the outcome. Tony Ulwick’s outcome-driven innovation gives you a way to define success that is durable and measurable, a job statement that does not change every time a new model drops. Teresa Torres put the warning best: the real risk is not that you fail to build something, it is that you successfully build the wrong thing, and rushing to AI just adds speed to that mistake.

But saying “outcomes” and defining one are not the same act. Most teams now believe outcomes matter, which is progress, but believing it and doing the unglamorous work of writing the job down at a system-agnostic level are different things, and the second is still rare. Outcome-first is necessary. It is also just the floor.

The process that is actually different runs past go-live

Most of the industry still treats an AI initiative as a project. You scope it, build it, launch it, move on. That framing is why so many of these efforts die at the handoff.

The shift from agents to agentic workflows, which Cobus Greyling has described well, is really a shift back to design. An agentic workflow is not a pile of capabilities turned loose. It is a set of purpose-built agents, each with a defined job, clear inputs and outputs, and human checkpoints, so the whole thing stays traceable. As the Nielsen Norman Group has argued for years, trust in a system comes from transparency and control, not from raw capability.

What makes the process genuinely different is that it does not end at launch. Marty Cagan’s case for a product operating model applies here almost word for word. This is not a project with a finish line. It is an operating model: empowered teams measured on outcomes rather than output, owning the result through to production and beyond instead of handing it across a wall the moment the demo works. You start with the people who do the work in the room, map the real work into skills, and group those into agents and workflows. Then you keep going after it ships, because the version you launched is the first one, not the last. Design and run are the same effort, not two departments.

The commercial model has to match the way value is really made

Here is where it gets tense, and honestly so. Because outcomes is the word of the day, some customers are using the moment to push all of the risk onto the partner. The instinct, which I have run into in more than one form, comes down to wanting to pay for the steak only after eating it, and only if they liked it.

It is understandable, but pure outcome-only terms quietly assume the outcome is something the partner makes alone, on someone else’s systems, with someone else’s people, and can be cleanly credited at the end. None of that is true. The outcome is co-produced, in the customer’s own environment, by a blended team. You do not get a great result by judging the plate at the end. You get it by being in the kitchen together, with both sides carrying some of the risk and sharing the reward.

It helps to be clear about what the older models priced. Traditional outsourcing priced labor, a seat doing a task somewhere cheaper. Outcome-based AI tries to price the result. Neither captures where value is actually made, which is the execution between them. A model that holds up prices that: a fair and predictable charge for the work that runs every day, a premium for the wins you can see and that recur, and honest reporting of the value that is shared, the kind you show a customer with pride but do not bill as if one side did it alone. Offshore still has its place for commoditized, high-volume work, but bodies are not where the difference sits anymore, and neither is a results-only bet no one can fairly attribute.

Operating it is the part everyone underestimates

If outcome-first is the floor, then operating the thing is where most of the value and most of the failure live. It is the back half of the work, and the part where you look least like everyone else, because almost no one has built the muscle for it.

Operating an agent fleet is an engineering discipline, not a dashboard. The DORA research from Nicole Forsgren, Jez Humble, and Gene Kim gave software teams four measures that fit agents almost exactly: how often you can safely change things, how long a change takes, how fast you recover when something breaks, and how often a change makes things worse. On top of that sits a newer discipline, what some call agent operations: watching what an agent does across a whole task, evaluating its decisions, and pulling a human in the moment a result drifts. Stafford Beer said it most simply decades ago. The purpose of a system is what it does. You operate against what the agents actually produce, not what the process document claims.

And you do not operate it without your people. The people running operations today are not the cost you strip out on the way to automation. They are a large part of the reason anything works now. Most have spent years performing small daily miracles on software that should have collapsed a decade ago, holding service together with workarounds, judgment, and knowledge that lives in no manual. That instinct, the ability to make an imperfect system deliver, is exactly what an emerging agentic system needs to become usable, and it does not come from the model.

This is the difference between two strategies that look alike on a slide. Most AI plans begin by asking how many people we can remove. Agentic workforce optimization begins by asking how much more our people can create once the routine work is carried elsewhere. The first is a cost strategy. The second is a capacity strategy, and over any horizon that matters the capacity strategy wins, because you cannot cut your way to a better customer experience. Some roles will change and some will go. That is a byproduct of the shift, not its purpose, and the operators who treat it as the purpose are the ones who get it wrong. The better version has the technology partner, the client’s IT group, and the operations people stop behaving like three camps that hand work across a wall and start working as one team, until it is hard to say where one ends and the next begins. From there the model settles into something specific: humans on the loop, supervising and correcting the agents that carry the routine work, and humans in the loop at the decisions that still need judgment. The people do not disappear. Their work moves up.

So be exact about the word. You are not optimizing the people away. You are optimizing the workforce, people and agents together, for the best result the combination can produce against the outcome. Lose that distinction and you are just running the old headcount play with new tools, getting the old results faster. You can automate a broken process and get to broken faster. You can also automate around your people and get to unusable faster. Same mistake, one floor down, and the customer feels it before anyone in a boardroom does.

The value is in what compounds

The reason to do all of this the hard way is that it compounds, and the shortcuts do not.
In the first piece I pointed at the next evolution: moving past workflows to skills, discrete pieces of work that are portable and reusable. Once you have mapped the work into well-defined skills, those become building blocks you can reassemble across products, channels, and systems. A skill that verifies a customer works the same whether a person, a chatbot, or an automated workflow calls it. Do this for a year and the operator is not just running a few agents, it is accumulating a library of capability that makes the next thing faster, and every reused skill is a piece of the coordination tax that does not come back. The workforce compounds the same way: a team that helped build the system and operates it on and in the loop becomes the engine that evolves it as the business changes, rather than a group waiting on the next vendor release.

Under all of it is the only scoreboard that counts. The operator’s customer never sees the agents, the org chart, or the contract. They know whether their service arrived when promised, whether their bill is right, and whether their problem was solved the first time they called. When the inside works, the customer simply experiences a company that works, and that is the asset that compounds longest.

The work after the word of the day

Outcome-first won the argument, and I am glad it did. But outcomes were the easy part. The hard part is the coordination tax, and taking it out of the work without taking out the people. It is building workflows that survive contact with reality, writing commercial models that reward shared success, operating a blended workforce with discipline, and doing all of it shoulder to shoulder with the people who already know how the business runs.

Twenty years ago we optimized systems. Ten years ago, labor. Today, AI. The next decade will belong to the operators who learn to optimize the work between them, and their customers will feel it first, in a service that shows up when promised and a bill that is finally right. Call it agentic workforce optimization if you want a name for it. What it really is, is the unglamorous discipline of helping people and agents do their best work together, against an outcome that matters, and then refusing to stop improving it. That, in the end, is the difference between transformation theater and a company that actually works.