The Revenge of the Business Process

(it was a hard choice between this and "The Business Process Strikes Back")

Satya Nadella seems to be on a roll lately, at least when it comes to his notes on agentic AI adoption in the enterprise. I never assumed I would agree with the writings of the CEO of Microsoft, but here we are (just don’t go back in time and tell the 14-year-old-installing-linux-mandrake-me, okay? Things have changed - OKAY?).

The first post I’m referring to is his essay, "A frontier without an ecosystem is not stable." It currently has 66 million views on X, so odds are you've seen it. 

The line that stuck with me most was this one: "A company should be able to switch out a 'generalist' model without losing the 'company veteran' expertise built into their learning system. This is the key 'test' of your control and sovereignty in the era ahead." I actually pulled this exact quote out and used it in a recent webinar, the recording is available here: Agentic AI Needs Orchestration| Flowable Webinar.

If you didn't watch that webinar, here’s a few things I said:

• If your (human) company veteran would retire, after a long and happy career, what would you do? Well, you’d sit down with that person, and try to get as much knowledge out as possible. After all, that’s what makes your company’s business processes unique: the little quirks and years-of-experience built in, often organically.
• Running your business processes in the same place as your ‘rented intelligence’ is asking for a vendor lock-in and you’re basically renting back your own uniqueness. You pay for the tokens and for your own expertise.
• Historically, putting all your eggs in the same vendor-basket has never been a good idea. The old rule hasn’t changed.

Now last Monday, he made a new post, titled “The Reverse Information Paradox”. There’s a lot to like in that post, which strengthens the points above even more: 

“You essentially pay for intelligence twice, once with money, and again with something even more valuable: the proprietary knowledge you must reveal to make that intelligence useful. The better you want the model to perform, the more of that knowledge you have to feed it! [..] Every correction is distilled into institutional know-how. It's the kind of knowledge a competitor could never buy, and the kind that leaks almost imperceptibly: trace by trace, correction by correction, eval by eval.” 

If you ask me, this is not a new problem. We’ve institutionalized specific company knowledge and know-how into artifacts called business process / case diagrams for many years. These artifacts are understandable by people and AI agents alike and are documentation and runtime at once.  But it gets better: 

“Ensure the orchestration layer is decoupled from any single model. Ask yourself: If any one model you are using is taken away, do you still have the ability to operate and optimize for your evals using other models? Does your company “veteran” capability remain with you even if a given “generalist” model is taken away?” 

I’m very happy he wrote ‘orchestration layer’. I would have been happier if he said something like “business process automation solutions such as, oh I don’t know, Flowable”, but we can’t have it all at once.

You see, I'm feeling upbeat, because the market is discovering what we’ve been preaching for a while (and preaching without validation tends to become arrogance which leads to the dark side): to get the right context to an agent, the solution is the same as when we had the problem of bringing the context to humans: an orchestration layer that is deterministic and governed by design. But blindly applying the same principles would not be smart: humans are resourceful when the business process isn’t 100% perfect, they pick up the phone and call. They understand the legal realities and potential repercussions they operate in. AI Agents don’t, and by extension the context passing and business process need to be way more exact.

There’s another, slightly more implicit message in Nadella’s writing: if you want to own your own business process, you should be able to swap in a local model (or at least, for those things that need the privacy) and route to it seamlessly. In my – always humble – opinion, this is the future: using small, specifically-tailored LLM models for the specific task at hand, receiving the right amount of context (but not more) and adding only what it’s allowed to towards the orchestration layer.

The previous line was supposed to be the conclusion of this writing and ending with a whimsical thought. But just before I hit the post button, I spotted that Aaron Levie of Box posted a similar reasoning this morning, which I want to share here verbatim as it seems like we’re in agreement about many things: 

“Open weights rapidly absorbs frontier breakthroughs (and drives other breakthrough directions given the constraints), offering both lower cost intelligence and the ability to be post trained for specific workflows and domains.”

And, more importantly: “The Applied AI layer has a huge opportunity to combine frontier intelligence with open or cheap closed models to orchestrate workflows in any given domain. Due to evals, deep domain context, being trusted with enterprise data and workflows, this layer can maximize performance and cost combination.”

It’s always better if someone that isn’t selling a pure orchestration product is saying these things. You know, the old ‘asking a barber if you need a haircut’ adage. As a side-note: Aaron has many great posts, definitely go and check them out.

But back to the point: capturing company knowledge and getting the right data to the right people, systems (and now AI agents) in a secure and governed way is exactly what business process automation has been all about. Maybe ‘business process automation’ isn’t a fancy enough name in the AI age. But fancy or not, all those people who spent years doing the unfashionable work of mapping processes, cleaning data and getting governance right ... now find themselves holding precisely the thing everyone suddenly wants.

The lesson from Nadella, Levie, and frankly from the last two decades of enterprise software is the same: context and context orchestration are key. The models will keep getting better, cheaper, local and specific. What won't be interchangeabl, though, is the accumulated knowledge of how your business actually works, and whether it lives in an orchestration layer you control, or leaks away trace by trace somewhere else. That choice, more than anything else, will separate the companies that own their AI future from the ones renting it.