Welcome to the AI Brief from the YPO Technology Network. I'm Stephen Forte. If you caught the episode earlier this week on inference getting cheaper, today is the other half of the story. AI may be getting cheaper to run, but it is not getting simpler to install inside a real company. Today we are going to talk about a strange thing that happened this week. The companies selling automation are building armies of consultants. OpenAI, the company behind ChatGPT, is reportedly building a new deployment company. Anthropic, the AI lab behind Claude, announced a similar push. Reuters then reported that both ventures are looking at acquisitions of services firms, the kind of firms that send engineers and consultants into companies to make technology actually work. That is the headline, but the headline is not the story. The story is what it admits. The most valuable AI companies in the world have discovered that the hard part is not only building the model, it is getting the model into the bloodstream of a real company where the work is messy, political, historical, half-documented, and somehow still running payroll every two weeks. Here's the contract for today's episode. I am going to explain why enterprise AI deployment is becoming a services business, why the best agent implementations start with workflow archaeology, and why the future of AI inside your company will not look like one bot replacing one employee. It will look like rebuilding the operating system of the business one workflow at a time. Let's start with the news. TechCrunch reported that Anthropic is launching an enterprise services venture. Bloomberg reported that OpenAI is building its own version called the Development Company. Reuters added the next piece. These ventures are reportedly in talks to acquire AI services companies. The goal is simple. Bring in the engineers and consultants who can help companies put AI models to work. Pause on that for a second. The companies telling you AI can automate knowledge work are raising capital to buy knowledge workers. That is not hypocrisy. It is useful evidence. It tells you where the bottleneck is. For the last two years, the AI industry has sold the story as if intelligence were the scarce resource. Better model, better answer, better benchmark, better demo. And that was true for a while. The model mattered because the model was weak. But inside companies, the bottleneck is shifting. The question is less, can the AI reason? Increasingly, the question is, does the AI understand how this company actually works? Those are not the same question. Most companies do not run on their org chart. They run on exceptions, handoffs, spreadsheets, workarounds, tribal knowledge. And that one person in finance who knows why the export from system A has to be cleaned before it goes into system B, that person is not in the architecture diagram. That person is the architecture This is where the consulting layer matters Not because consultants are magical we are not A shocking number of us still use airline Wi and call it infrastructure The value is that somebody has to map the actual work before the agent can improve it. In our work bringing agents into companies, the first useful step is almost never what agent should we build. The first useful step is what work is actually happening. And the only way to answer that is to go deep. We interview people across the firm, not just the department head, not just the sponsor, almost everybody who touches the work. We ask what they do, what they receive, what they change, what they approve, what they copy, what they reformat, what they check, what they know is wrong, but still have to do because the system upstream was designed during the Bush administration and nobody wants to touch it. That last category is larger than most CEOs would enjoy admitting. The important thing is that we are not looking for one person to replace. We are looking for work patterns. That distinction matters. A traditional automation conversation often starts with a role. Here is Susan. Susan does seven tasks across four domains. Can we build an agent that does Susan's job? Usually that is the wrong question because Susan's job may contain one task that belongs with sales operations, one that belongs with finance, one that exists only because a vendor portal is terrible, one that is really compliance review, one that is manual translation between two systems, and two that should not exist at all. If you build one SUSAN agent, you have automated the historical accident. Congratulations. The mess now has an API. The better move is to break the role into task verticals. What kind of work is this? Is it translation? Is it reconciliation? Is it formatting? Is it research? Is it exception handling? Is it judgment? Is it approval? Is it communication? Is it data movement? Is it reporting? Then you look across the company. And this is where it gets interesting. You often find that seven different people in seven different roles are all doing some version of the same small task. They are taking output from one system, cleaning it in Excel, changing column names, fixing dates, adding a comment and putting it into another system, or they are reformatting a weekly report, or importing a file, or checking three sources because no single system is trusted. In other words, humans are middleware for machines. They are not making a strategic decision. They are passing a bucket of data from one digital well to another digital well, and occasionally wiping mud off the side. That is not an insult to the employee. It is an indictment of the system, and it is exactly the kind of work agents can absorb quickly when the implementation is done correctly. Not one agent for one person, one agent that does 20% of the work for seven people, because that 20 is the same pattern hiding inside seven different job descriptions That is where the ROI starts to make sense The public literature points in the same direction even if it uses less colorful language Process mining reads system event logs from tools like ERP CRM ticketing and workflow platforms to reconstruct how business processes actually run. Task mining watches desktop level activity to see how individuals move through applications. Process intelligence tries to connect the two, including handoffs, exceptions, and work that accumulates in the gaps between systems. That last phrase matters, the gaps between systems. That is where half the company lives. The ERP says the invoice was approved. The CRM says the customer is active. The spreadsheet says the invoice is wrong. The email says legal has a concern. The shared drive has the newest version, unless Michelle renamed it Final Final Two. The dashboard says green because nobody had the courage to create yellow. This is not edge case work. This is work. And this is why generic AI rollouts disappoint. A chatbot sitting on top of a messy operating model does not fix the operating model. It just gives the mess a nicer interface. Here is my read. The companies that win with AI agents will not be the ones that ask, how many people can we replace? They will be the ones that ask, what work exists? Why does it exist? Where does it repeat? And which parts of it can be handled as a shared capability across the firm. That is a very different management conversation. It means the unit of analysis is not the employee, it is the task. More precisely, it is the task pattern, translation tasks, reconciliation tasks, reporting tasks, data movement tasks, intake tasks, triage tasks, first draft tasks, follow-up tasks, exception routing tasks. Once you see the company that way, the agent strategy changes. You stop building agents around job titles and start building agents around repeatable work. You stop saying this agent replaces an analyst. You start saying this agent handles the first pass of vendor file normalization across finance, operations, and customer success. Less dramatic, much more useful. This also changes how you price and measure the work. The wrong metric is did we replace a headcount? That is a blunt instrument, And blunt instruments are popular because they fit onboard slides. The better metric is output. How many transactions did the agent handle? How many files did it normalize? How many exceptions did it route? How many reports did it draft? How many cases did it prepare for human review? What was the accuracy rate? What was the cycle time? How often did a person intervene? What did the person do with the time that came back? IBM's research on scaling a genic AI makes a similar point. Companies need KPIs that measure both workforce and agent contributions with human checkpoints for high complexity or high impact decisions That is not academic That is the operating model People and agents are not separate teams They are part of the same production system And once you accept that you stop thinking about AI as a tool rollout You start thinking about it as a migration That word matters. Migration. You do not rip the operating system out of a company on a Friday and install the AI version on Monday. That is not transformation. That is arson with a steering committee. You migrate piece by piece, find the overlapping task, build the shared agent, measure the output, keep the human in the loop where judgment matters. Watch the exception paths, improve the agent, then move to the next task pattern. Over time, the company changes shape. The org chart may look the same for a while. The work underneath it does not. Some roles get lighter. Some roles get broader. Some people move from producing the first draft to supervising the system that produces the first draft. Some teams stop being defined by the systems they manually connect and start being defined by the outcomes they own. This is why OpenAI and Anthropic are moving into the services layer. They know that sticky permanent adoption will not come from a better demo. It will come from understanding the customer's workflows deeply enough that the AI becomes part of how the company runs. Once an AI system understands your workflows, your exceptions, your reporting patterns, your approval paths, your internal language, and your messy little bridges between applications, it is no longer a vendor tool. It is infrastructure. And infrastructure is sticky. If I were sitting in your seat this quarter, I would do three things. First, stop asking your team for an AI use case list. Ask for a work pattern map. Show me the 10 repeated tasks that cut across departments. Show me the handoffs. Show me where humans are moving data from one machine to another. Second, run interviews before you run pilots. Not surveys. Interviews. Sit with the people doing the work and ask them to show you the ugly parts. The ugly parts are the opportunity. The clean process diagram is usually fan fiction. Third, measure agents by output, not replacement. Transactions handled, reports produced, exceptions routed, cycle time reduced, human review required, that is the scoreboard. The irony of this week's news is not that AI companies need consultants. The irony is that the consultants are showing us what AI actually requires. The future is not one bot per employee. The future is a new operating system for the business, assembled from the real work people already do, stripped of the historical accidents, and rebuilt around tasks that can finally move at machine speed. The model is only the engine. The workflow is the road. That is the YPO Tech Network AI Brief for Friday, May 8, 2026. I am Stephen Forte. If this was useful, send it to a fellow member. I will be back Monday with more. Until then, stay sharp.
