Renting capability is rational.
Rent the intelligence. Own the operation.
Nobody needs to build their own frontier model, their own cloud or their own army of AI engineers. Businesses are about to rent extraordinary amounts of intelligence: models, agents, specialist software, cloud runtimes and forward-deployed engineers. That is fine.
What changed
Agentic systems absorb more than data.
The danger is allowing the supplier to become the only place where the business knows how the work works.
That danger is new because the products are new. The agents now on offer are always on. They come with their own cloud computer, their own memory and connections to thousands of applications. The versions packaged for small businesses plug straight into the accounts, the shop and the payments system.
Systems like that accumulate more than records. They accumulate context, procedures, policy, memory, integrations and operating knowledge.
Take a 30-person company. It will quite reasonably choose packaged agents from the large platforms, or from the software it already uses, rather than build a stack of its own. That is sensible. But over two years those agents may quietly accumulate customer knowledge, sales procedures, pricing logic, operating exceptions, supplier context, workflow state and management routines.
The owner eventually discovers that how the company operates is encoded inside somebody else's product. That is a very different kind of lock-in from being unable to move your CRM.
Being unable to change vendors is the smaller danger.
The larger one is being unable to change your own operation without the vendor.
Switching AI provider will increasingly mean more than switching software. You may be moving accumulated context, operating procedures, memory, tool bindings, policy, evaluation behaviour and work history. The lock-in moves from the application into the operation.
The same risk, delivered by people
Consulting can create the same dependency.
Forward deployment, where a supplier's engineers work inside the customer to build the system, is becoming the standard way to implement AI. One model provider has just committed $100 million to training 10,000 engineers for exactly that work, and the largest consultancies are in its first cohorts.
Forward-deployed teams can deliver astonishingly quickly. If they leave behind something only they understand, speed has purchased dependency.
Gartner now predicts that by 2028, 70 per cent of enterprises will abandon agentic AI built by vendors' forward-deployed engineers, because the costs keep rising and they cannot evolve it on their own. The warning is not about agents. It is about a way of delivering them that leaves the customer unable to run what it paid for.
The fastest deployment is not the best deployment if you cannot change it without the people who built it.
Essay 001 criticised suppliers who hand over and exit before the capability has taken hold. A supplier you can never let go is the same failure from the other side.
What to keep
Define what the company must own.
Essay 013 asked where the next unit of capacity is best bought. If machine capacity can be bought almost on demand, the next question is what should never be outsourced.
Models can be rented. Compute can be rented. Agents can be rented. Expert implementation can be rented. Organisational capability should not be rented indefinitely.
- ModelsThe intelligence itself.
- Compute and runtimeWhere it runs.
- AgentsPackaged workers and their tools.
- Specialist helpThe people who build it with you.
The leaving line
- ContextCompany knowledge and its authoritative sources.
- MethodsWorkflows, skills and operating procedures.
- ControlDecision rights, policies and escalation rules.
- EvidenceEvaluations, traces, baselines and outcome measures.
- PortabilityDocumented interfaces, dependencies and replacement paths.
- CapabilityPeople inside who understand the system well enough to change it.
If everything above the line left tomorrow, everything below it should still stand.
Some of that list will look familiar. Essay 009 named the things that must resolve the same way whichever AI does the work, among them authority, business rules, authoritative state and evidence. That was a question of where they sit. This is a question of who holds them. An anchor inside a supplier's product is still an anchor. It is just not yours.
Essay 004 argued that the value lies in the operating system around the agent, not in the agent itself. That is exactly why it is the part to own. And essay 002 said the contest in enterprise AI is over who controls the agent estate. The platforms are now competing to be that layer, so leadership teams should decide who holds it consciously rather than discover it at renewal time.
Your AI provider can own the model. They should not own how your company works.
The test
Change what counts as a successful implementation.
Most implementations are signed off on two words: it works. That is not enough. The finish line is that we can operate it, understand it, change it and replace its components without losing what we have learned.
There is a simple way to check. If the supplier disappeared tomorrow, could you still run and evolve what was built? If the answer is no, the work is not finished, however well it performs today.
Where Praxis stands
That test applies to us first. We call ourselves an operating partner. A partner helps you run what you own. It does not keep it.
What we build for you is yours: the context, the workflows and procedures, the decision rights and policies, the evidence. We keep our general methods. Everything made for your company stays with your company.
So we hold our own work, ongoing engagements included, to the same question. If Praxis disappeared tomorrow, could you still run and evolve what we built? If the answer is no, we are not finished.
AI capability should compound inside the client, not inside the supplier relationship.
Use the best intelligence you can rent. But never rent the knowledge of how your own company works.
Begin
If your supplier left tomorrow, what could you still run?
One conversation, no pitch deck. Bring one system somebody else built for you, and we will work out what you own, what you rent and what would stop if they left.