What to do when you have eight AI vendors

At some point somebody counts, and the number is eight. Or eleven. Two frontier providers, a cloud-routed path to one of the same models, a couple of open-weight deployments on rented GPUs, two coding assistants, and a research tool somebody expensed.

The instinct is to consolidate. Fewer vendors, fewer contracts, fewer things to manage. That instinct is usually wrong as a first move, and acting on it early tends to cost more than it saves.

Here is a sequence that works better.

First, accept that nobody made a bad decision

Every one of those relationships was created by someone solving a real problem. A team picked the model that worked for their workload. Another routed through a cloud provider because the enterprise agreement made it easier. Engineering bought an assistant because engineers asked for it.

This matters practically, not just diplomatically. If you approach sprawl as a governance failure, you will get defensiveness and you will lose access to the information you need. If you approach it as an information problem, which is what it actually is, people will help.

The failure was never the individual decisions. It was that nothing was adding them up.

Second, get the total before you touch anything

You cannot make good consolidation decisions without knowing what each vendor is actually used for, by whom, and at what effective rate.

Effective rate is the part people skip. List price per million tokens is close to meaningless as a comparison, because it ignores your input to output ratio, your cacheable prefix share, your retry rate, your batch eligibility and whatever discounts you already hold. Two providers with similar list prices can differ substantially on the work you actually do.

Getting the total also tends to produce the first real surprise, which is that the total is higher than the working figure. The gap is usually three things: model access billed through a cloud provider, GPU compute running self-hosted models, and AI features inside SaaS contracts you already had. None of those look like AI spend in any bill, so none of them were being counted.

Third, reclaim seats

This is out of order in the sense that it is not strategic. It is in order in the sense that it is the fastest money in the exercise and it requires no decision from anyone.

Per-seat AI tools get assigned generously because provisioning is easy, and reclaimed never because reclaiming is nobody's job. Pull the activity data and look for seats with no usage in the last 30 days. In most organizations that is a meaningful share.

Reclaim them, then set an automatic rule at 45 days so it does not accumulate again. Do this before renewal discussions, because it frequently changes what the right contract size is.

Fourth, consolidate routes before you consolidate products

Here is the finding that surprises people most often: the same model bought three ways, at three different effective rates.

Direct from the provider. Through a cloud marketplace, drawing down a commitment. Through a reseller. These appear in three unrelated bills, so almost nobody has compared them, and the spread is frequently material.

Consolidating onto the best available route changes nothing about your product. No team switches tools, no workload is migrated, no evaluation is re-run. It is a purchasing change, and it is usually the largest saving available that carries no risk at all.

Do this before you consider consolidating actual products.

Fifth, distinguish duplication from variety

Now you can look at products, and the distinction that matters is between duplication and deliberate multi-vendor use.

Two coding assistants used by two separate teams may be preference worth respecting. Two coding assistants where a third of engineering holds an active seat on both is duplication, and it is measurable.

Similarly, using different model providers for different workloads is often correct. Models genuinely differ on different tasks, and maintaining a second relationship strengthens your position at renewal and gives you somewhere to go when a provider has an outage or deprecates something you depend on.

The thing worth eliminating is two vendors doing the same job for the same people. That is a much narrower target than eight vendors down to three.

Sixth, sequence against the renewal calendar

Consolidation only realizes value at a renewal or a contract break. Attempting it mid-term generally means paying for both for a period, which erases the saving and creates a migration under time pressure.

Build the renewal calendar with notice periods. Then order the consolidation opportunities against it and take them one at a time. One migration at a time also means that when one goes badly, and one will, it is not happening simultaneously with another.

For each renewal, arrive with the usage history, the effective rate, the growth trajectory and the overlap analysis. That is the same information the vendor already has about you, and having it changes the conversation from their projection of your growth to a shared set of facts.

Seventh, stop the next one arriving unnoticed

None of the above prevents recurrence. In six months there will be a new tool, bought reasonably by someone solving a real problem, invisible until renewal.

The fix is detection rather than policy. A first charge from an unrecognized AI vendor in your billing data should raise a flag that reaches whoever owns vendor management. Not to block it, which would be counterproductive and would push purchases further underground, but so that it enters the picture at month one rather than month fourteen.

An approval gate makes people route around you. A detection signal does not.

The thing that does not change

Reducing vendor count is not the goal, and treating it as the goal produces bad decisions. A company running six AI vendors with complete visibility, verified effective rates, per-workload attribution and a renewal calendar is in far better shape than a company running two vendors blind.

Sprawl is uncomfortable because it feels out of control. But the discomfort is about the absence of a total, not about the number of relationships. Fix the total, and most of the pressure to consolidate turns out to have been pressure to understand.

Then consolidate the parts that genuinely duplicate, at the renewals where it pays, in whatever order the calendar allows.