Back to resourcesAI

Who Will Really Own Your Processes?

Who Will Really Own Your Processes?
David Novák

Vendors are happy to deliver your AI transformation turnkey. It’s worth asking what you’ll be left with afterwards.

You notice it in the small things

You need to change something minor in invoice approval. Say, invoices under ten thousand no longer need to land on the CFO’s desk. You write a change request, the vendor estimates it at three person-days and slots it into the release plan, and if all goes well, the change ships in three months. Until then, approvals run the old way.

Or you want a report you know the system “can definitely do”. The dashboard offers twelve pre-built views and yours isn’t one of them. It ends the usual way: someone exports a CSV and stitches it together in Excel on Sunday evening.

Or you ask whether you could pull the system’s data in bulk, because you want to try running it through your own AI. You find out the API only returns the last ninety days and a full export is an “enterprise feature” you can negotiate with a sales rep.

Or the one person who knew how the system is actually configured leaves. From then on, only the vendor knows.

And on top of it all, the bill keeps growing year after year. More SaaS line items, pricier licences, and less and less ability to change anything on your own. In meetings, all of this goes under the heading of “IT problems” and gets escalated to the vendor.

Each of these moments looks like an operational detail worth no more than a sigh. Together, though, they form one and the same pattern. And it’s worth naming before turnkey AI transformation arrives at your company.

What’s going on in the background

All these small things share one thing: the process is formally yours, but in practice you don’t control it. On paper, everything checks out. You pay the licences, your people click around in the system, the auditor lists the process as yours. But formal ownership breaks down into four practical questions you can ask about any process that lives in someone else’s software today:

  • Access. Can you reach your asset directly, without an intermediary’s permission?

  • Change. Can you change it yourself, whenever you need to, at close to zero cost? Or do you write a change request and wait for the vendor’s release plan and price list?

  • Exit. What would it cost to leave? If the exit cost is prohibitive, you don’t own it. You’re renting, just with a worse notice period.

  • Knowledge. Who knows how the whole thing works? Your people, or the configuration of someone else’s system and the vendor’s consultant?

By these criteria, most companies I know already don’t own their processes today. It hasn’t hurt much so far, because everyone around them was in the same position. Where this mechanism leads when left to run for twenty years is shown by an entire industry.

The client you’re not allowed to call

Picture an insurance company manager. She opens the system and sees a client: his policies, his payments, his risks. The company has insured him for fifteen years, sends him annual letters and collects his premiums. And the manager knows she must not call him with an offer. Not because of the law. Because of the contract with the adviser who once brought the client in. The client is in her database, but he isn’t hers.

The insurance industry didn’t get here through one bad decision. In times of rapid growth, insurers couldn’t build their own sales networks fast enough, so they handed sales to external advisers and brokerage firms. A rational move, and again every following year: production grew, the numbers looked good. Nobody ever decided “we’re giving up our customers”. It just turned out, twenty years later, that the client relationship, and with it the trust and the negotiating position, belongs to distribution. The insurer was left as a product factory and a risk carrier. And it still hasn’t been properly untangled, because every attempt at a direct channel threatens the revenue that distribution brings in today. This isn’t one company’s failure; it’s a structural feature of the industry across Europe.

Notice that it broke down along exactly those four questions. Access to the client only with the intermediary’s permission. Changing the relationship only with his consent. Exit at the cost of your own revenue. And knowledge of the client, his trust, lives in the adviser’s head. The change requests and Sunday Excels from the opening are the same mechanism, just a few floors earlier. We have one advantage over the insurers: we’ve seen this film before.

Ownership isn’t yes or no

The insurer didn’t jump from “the client is ours” to “we’re not allowed to call the client” in one leap. It moved in small, individually rational steps. In the same way, process ownership isn’t a binary state. It’s a spectrum, and it’s worth thinking hard about which level you’re on and which direction you’re moving:

  • Level 0: the process lives in someone else’s application. You change it at the pace of the vendor’s release plan and buy AI for it as a credit-based add-on. This is where most companies I know stand today.

  • Level 1: you own the data. A continuous copy of everything that matters, outside third-party applications, in your own storage. Sounds trivial. After Slack (more on that in a moment), it no longer is.

  • Level 2: you own the view. Reporting and management decisions run on your data, not on whatever the vendor deems fit to show you in its dashboard.

  • Level 3: you own the logic. The rules, exceptions and decision steps of the process live in your layer. The third-party system degrades into a database with a price list, and a database with a price list can be put out to tender.

  • Level 4: you own the pace of change. You change the process when reality changes, not when a release ships. I call this level the adaptive layer.

This scale is why it makes no sense to brush off the ownership question with a shrug that “everything runs in the cloud anyway”. Between level 0 and level 4 there’s a chasm in what a company can do, and each level can be gained or lost separately.

Why this is urgent now

Between May and July 2026, all four big AI players announced essentially the same thing. OpenAI founded The Deployment Company, a joint venture with more than $4 billion from investors and founding partners including Bain, Capgemini and McKinsey (TechCrunch). A few days earlier, Anthropic announced an enterprise services firm with Blackstone, Hellman & Friedman and Goldman Sachs. At the end of June, Amazon added a billion-dollar initiative, and on 2 July Microsoft launched its Frontier Company: $2.5 billion and 6,000 engineers and salespeople it will send to work directly inside companies (GeekWire). All of them copied Palantir’s forward deployed engineers model (The New Stack). The Microsoft Frontier Company page says everything important out loud: “No pilots. Scale from day one.” Your company’s intelligence will “compound over time”. All on their stack. Translation: we’re not waiting for you to buy AI. We’re coming to build it for you.

Why such sudden eagerness to help? I read three incentives in it. First, the public internet has run out as training data for models, and the last large deposits of quality data sit behind corporate firewalls (Business Insider). Second, data in documents is no longer the most valuable thing. Models now learn from trajectories, records of how a person solves a task step by step in real software, and a whole industry has grown up around them (SemiAnalysis, Epoch AI). Your workflow, meaning where people click, how decisions are made, which exceptions get handled by hand, is exactly the training material for an agent that will one day take that work over. And unlike an employee, a trained agent gets sold to everyone. Third, analysts estimate hyperscalers will spend around $725 billion on AI infrastructure this year (Tom’s Hardware), and someone has to pay for it. Mark Zuckerberg put it bluntly: “If we end up misspending a couple of hundred billion dollars, I think that is going to be very unfortunate obviously. But what I’d say is I actually think the risk is higher on the other side.” (Windows Central) A company investing out of existential fear can’t afford a transformation that lowers your spending. It has to collect it. From you.

The proof is in the price list. Microsoft 365 Copilot costs $30 per user per month on top of the existing licence, M365 has meanwhile raised prices across plans, and since June Copilot Cowork runs on credits: you pay for the model, for loaded context, for every tool call (Windows Latest). Salesforce charges $125 per user per month for Agentforce, plus 10 cents per agent action (SaaStr). The more AI helps you, the more you pay. And the other jaw of the pincer: in May 2025, Salesforce changed the Slack API terms to prohibit bulk data export, long-term indexing and the use of data for AI (Computerworld). You may not pull years of your own company conversations and decisions out in bulk for your own AI. Data can’t leave; intelligence is bought inside, with credits.

On the spectrum from the previous section, the movement is clear: a turnkey transformation built on the vendor’s platform sets you in concrete at level 0. This time including the decision logic you dictate to the agents during rollout.

The same technology, in the opposite direction

Now for the important part. The same technology vendors will come to wire into your processes also, for the first time, makes the opposite move possible: taking your processes back, piece by piece. Business logic can be written and rewritten in days, by a person who understands the process, with a model doing the craftsmanship for them. Data from third-party applications can be continuously pulled into your own storage via API. That wasn’t true before: changing a process meant a project, a vendor and a roadmap.

I won’t claim AI will bring companies adaptability; I’ve seen enough pilots from the inside that delivered nothing, and my experience comes from a handful of companies, not a hundred. But for the first time, it can offer it. And in an environment where model capabilities change by the quarter and vendors rewrite their price lists on the fly, the ability to adapt quickly is roughly the only competitive advantage that doesn’t go out of date. Everything else your competitors can buy at the same price list as you, with the same embedded team.

In practice, this means inserting your own AI layer between yourself and the vendor and letting it climb the levels of that spectrum. Start where the value-to-risk ratio is best, which tends to be the management layer and reporting: data already flows there through exports no one will take away from you, and mistakes don’t hurt. That gets you levels 1 and 2. In parallel, pull data out of third-party applications while the APIs still let something through; Slack showed that windows close. And at whatever pace growing model capabilities allow, move business logic into your own layer, that is level 3, all the way to a pace of change you control yourself, level 4. Meanwhile the third-party system degrades into a database with a price list and gets pushed on price.

It isn’t an easy road. It will involve dead ends, discarded prototypes and lost illusions. But what can be bought comfortably can be bought just as comfortably by your competitors. A capability that grew slowly and painfully can’t be copied the next day.

For the record: part of how I make a living is helping companies build exactly these layers, so take my argument with that in mind. And I may well be wrong about the pace at which all of this plays out. But what I believe that insurance story gets right is the question every turnkey transformation offer deserves: how many of your processes do you really own today, and how many more will you own after it?

Sources