A mid-sized consultancy signs an enterprise agreement with a model provider in March, for the engineering team’s coding assistant. In April, the design team buys a tool with image generation built in. In June, marketing adopts a writing platform at $30 a seat with an undisclosed model behind it. By August the firm pays that same model provider through three contracts, two of which it did not know it had.
Which of those three relationships is actually yours?
One. The other two belong to software vendors who bought model access wholesale and sold it back per seat. They negotiated the data terms, chose the model, hold the rate card, and can change all three next quarter without asking.
Every enterprise already understands this arrangement, because it runs the same one with people. You bring a contractor in through an agency for a six-week job, paying a premium to have someone on Monday instead of in eight weeks. When that person turns out to be central to the work, you move them onto your own paper: you now need the background check run to your standard, the confidentiality terms your clients signed off on, and a day rate you can defend to a partner.
Nobody looks back on the agency as a mistake. The mistake would be staffing your core delivery team through one for four years and calling it a hiring strategy.
Four Things You Give Up When You Rent
Buying model access through a software vendor is not one decision. It is four, and three of them are invisible at the point of purchase.
1. The Markup You Cannot See
Per-seat AI pricing works by hiding the meter. Two hundred seats at $30 a month is $72,000 a year, and nothing in that figure tells you whether the model spend behind your heaviest user was $4 or $60. The vendor knows. The cost basis is deliberately not itemized, so you cannot separate a well-priced tool from a badly-priced one, or an expensive workload from a cheap one inside your own company.
2. Terms You Did Not Negotiate
Dana in Legal can produce a signed agreement with the writing platform. What Dana cannot produce is the agreement between that platform and the model provider behind it, the one that actually governs where the prompt went, how long it was retained, and whether it trained anything. When a vendor swaps its underlying provider for margin reasons, your data terms change with it, and you find out in a changelog if you find out at all.
We have argued that where your files land is an architecture question rather than a policy one. Its harder neighbour is this: whose contract governs the model call?
3. Spend That Never Adds Up
The coding assistant, the design tool, the chat workspace, and the support summarizer may all sit on one provider, arriving as four invoices coded to four cost centers, aggregating into nothing.
A firm spending $400,000 a year with a single provider across five use cases should be one customer with volume, an account team, and a negotiated rate. Split across intermediaries, it is five small customers paying retail, and nobody in the building can even assemble the number that would start the conversation. The leverage exists; it sits in other companies’ contracts.
4. A Relationship You Cannot Take With You
The renter accrues nothing. Four years of heavy usage produces no rate, no account relationship, and no usage history to negotiate against; leave the tool and you leave the model access, the workflow, and the record of both.
This is lock-in that never announced itself. Ravi on the platform team did not choose a model provider. They chose a workspace, and the provider came attached, picked by someone optimising a margin they cannot see.
Rent to Learn, Own to Operate
Renting is not the error. Renting permanently, by default, for work that has become load-bearing is the error. The useful frame is a sequence rather than a choice.
1. Renting Is the Right Way to Meet a Model
A new model ships on a Tuesday and nobody yet knows whether it beats what you have at your work. Opening a provider account, approving a card, running a security review, and signing terms for a model you might use twice is a procurement tax levied before any evidence exists. Most pilots that die, die in that queue rather than on the merits.
The right first move is small: put the model in front of the three people whose judgement you trust, on real work, for a week. Some never graduate past that, which is a legitimate resting place. Renting forever is correct for a model you consume 40,000 tokens of a month, and for most open-weight models no vendor is waiting to sell you an agreement anyway.
2. Owning Is the Right Way to Run the Work
Once a model does work you would have to describe to a client or a regulator, the questions change. Where does the data go, does the prompt train anything, how long is it retained, what is the rate. Those have durable answers only when you are the provider’s customer.
Ownership also consolidates. Coding assistants, design tools, and the conversation workspace drawing on one account turns scattered spend into one negotiation, one usage history, and volume that counts. The threshold for moving is not a dollar figure; it is the moment someone senior would have to explain the arrangement out loud.
3. The Door Between Them Has to Swing Both Ways
Most organizations pick one posture and get stuck in it. A trial with no path to ownership means the evaluation ends and the real work restarts elsewhere. Ownership with no cheap way back means you stop evaluating, commit to one provider’s roadmap, and meet next year’s better model a year late.
What makes both modes safe is that moving between them costs a settings change rather than a migration. Rae’s team should be able to evaluate a new provider on rented access this month while three client engagements run on the firm’s own contract, in the same workspace and on the same ledger, then promote the winner without rebuilding anything.
Where ParleHub Fits
This is precisely the problem we created ParleHub to solve. We treat model access as a setting rather than as the product, so the sequence above is one you can run.
- Two ways in, and you can run both at once. Bring your own keys for Anthropic, OpenAI, Google Gemini, or Azure OpenAI, held in Azure Key Vault and never exposed. Or use models we host, charged by what you use, with no provider account at all. That set is curated rather than exhaustive, and it reaches past the big three to open-weight models most firms will never hold a contract with. That is the toe in the water: a model on real work this afternoon, no procurement cycle in front of it.
- When you get serious, the spend moves and the work does not. Connect your own key at the org level and model spend lands on your provider contract, at your rate, under the terms your legal team negotiated rather than the ones we did, which retires Terms You Did Not Negotiate. Projects, conversations, files, budgets, roles, and history stay put, because they live in ParleHub rather than in the provider. Changing provider is a settings change, and it runs in reverse just as cheaply next quarter when something new ships.
- Governance is identical whichever way the tokens are paid for. Project budgets, personal and organization-wide caps, the warning at 80%, the hard block on a send that would breach a budget, the token ledger, your own project code, and the CSV export behave the same on our models and on yours. One ledger breaks spend down by project, person, provider, and model, whoever’s meter it landed on, which is what The Markup You Cannot See and Spend That Never Adds Up come down to.
- The model is chosen per conversation, by a person. Nothing reroutes your traffic to a cheaper model on our behalf, because our margin does not depend on the answer.
Renting is the fastest way to learn whether a model is any good at your firm’s work, and forbidding it only moves the experiment onto personal accounts. The failure is letting the rental become permanent, until the model is embedded in client delivery and the only contract in the building belongs to a company you picked for its user interface.
Try every model you like. Own the one you come to rely on.
Want to evaluate a model this week and still own the relationship next quarter?
See how ParleHub handles model access and spend →