One door to every AI model: a gateway for predictable AI costs

It’s the first week of the month, and the CFO is staring at the AI bill. It’s three times what it was last month. Who spent it? Did sales launch a big outreach campaign? Did the support bot get stuck in a loop? Did someone in analytics run reports through the most expensive model available? Nobody knows. The CRM has its own AI, the knowledge base has its own, the team chat has its own, and the invoices come from different vendors in different units.

This post explains how the AI model gateway in PRANUS works and why it keeps AI costs predictable. The short version: every service reaches AI models through a single door. That door decides which model answers and who pays for it, and it stops spending before it turns into a surprise.

Why AI bills became hard to predict

Over 2025 and 2026, most business software moved to usage-based AI pricing: per action, per conversation, per credit. Here’s what that looks like at some of the big vendors (published pricing as of November 2026):

Vendor	How AI is billed	Price
Salesforce Agentforce	credits per action, per conversation, or a per-user add-on	$0.10 per action; $2 per conversation (legacy customers); $125 per user per month
Intercom Fin	per resolved conversation	$0.99 per outcome
Zendesk	per automated resolution	$1.50–$2 beyond the included volume

Each of these models makes sense on its own. Together, they leave a business leader with one problem: you can’t forecast the bill. How many “actions” will you have next month? How many “conversations”? How many credits will that new agent burn through?

It’s telling that Salesforce changed its AI pricing model four times in two years: per conversation, then per action, then per user, then editions with credits bundled in. Analysts point to the same reason every time: customers want a spending ceiling. A Bain study of more than 30 software vendors found that about 65% have added a separate AI usage meter on top of seat pricing.

Our view is that predictability isn’t a feature of a price plan. It’s a property of how the system is built. If you can’t see who is spending what, no pricing plan will save you.

How the AI model gateway works

In PRANUS, no service talks to a model directly. The CRM, the knowledge base, team chat, the AI employees in AGENTUS, the personal assistant — everything goes through one gateway to every model.

Services ask for a job, not a model

A service never says “give me model X.” It says what it needs done: “complex work,” “quick short answer,” “speech transcription.” The gateway decides which model handles it, with which key, and on whose budget.

For the business, that means switching AI vendors doesn’t break anything. When a better or cheaper model comes out, it’s connected in one place and every service starts using it. No CRM rewrite, no retraining your team.

Routing rules at every level

You can set which model handles a job at any level:

a specific person;
a service;
a project;
a department;
the whole organization.

Rules are read from most specific to most general. For example, “complex work” might go to one model company-wide, while the legal team uses another that handles long documents more carefully. The first matching rule wins; the rest become fallbacks.

Fallback chains: one model fails, work keeps going

Model providers fail sometimes: overload, an exhausted quota, an outage. If the first model doesn’t respond, the request quietly moves to the next one in the chain. Your team doesn’t notice, and the log records which model stepped in.

Two more safeguards:

multiple keys per provider — if one key is rejected, the next key takes over immediately and the failed one is paused;
an empty answer counts as a failure — if a model returns nothing, the request moves on instead of sending a blank message to a customer.

The gateway also tracks provider health: what share of requests fail and how slow responses are. The model catalog is checked against each provider once a day. New models show up in the list, but a person decides when to put them to work. Nothing changes on its own.

Keys live in one place

API keys for AI models are stored only in the gateway, and only encrypted. No service, no agent and no employee’s browser ever sees them. If a key needs rotating, you rotate it in one place.

If your company already has its own contract with a model provider, you can bring your own key. Calls are then billed to you directly, while tracking and limits keep working as usual.

Predictable AI costs: tracking and caps

Now for the part the CFO cares about.

Every call is accounted for. The gateway calculates cost at the model’s actual prices and records it with context: which employee, which service, which job, which agent, which department. The “who spent it?” question from the opening becomes a single report.

Caps at any level. You can set a ceiling for a person, a department, a service, an agent or the whole organization, per day or per month. A new cap starts as a warning, so you can see where it would have kicked in. Once you’re confident the number is right, you switch it to a hard limit.

Every block names its cause. When a cap is hit, the employee doesn’t get a cryptic “error 429.” They get a plain message: whose cap was reached and how much has been used out of how much. For example: “Sales department monthly limit: 100 of 100 used. Contact your department head.” The manager knows exactly what’s going on.

Every agent has a budget. Each AI employee gets its own daily and monthly budget and a rule for what happens at the cap: stop, switch to a cheaper model, or hand off to a person. We cover how that’s set up in “AI agents without code.”

Your own model is just another provider

To the gateway, a local model running on your own server is a provider like any large cloud model. You can make it the default for certain jobs, for one department or for the whole organization, and keep cloud models as fallbacks. There’s also a mode where an organization runs on local models only.

That matters if you can’t send data to an outside cloud, or if you’d rather cap AI costs at the price of your own hardware. We break down when that makes sense, and what it costs, in “local AI models for business.”

Checklist: how predictable is your AI spend?

Answer seven questions about your company today:

You know how many tools in the company use AI and whose API keys they run on.
You can state last month’s AI spend by department.
You can name your most expensive employee or agent in AI terms.
Every AI agent has a spending cap.
If a model provider goes down for an hour, your bots keep answering.
API keys aren’t sitting in the settings of ten different people.
You could switch to a cheaper model in a day, not a quarter.

If you checked fewer than four, your AI spend is being run by chance. Not because you missed something, but because every tool brings its own AI and its own bill.

A step-by-step plan to get it under control

If you move to PRANUS, the gateway works from day one. Here’s the setup order we recommend:

Weeks one and two: watch only. All services run through the gateway, and limits are in warning mode. You get a real picture of spend by department and agent.
Set ceilings. Based on what you saw, give departments monthly limits and agents budgets, with 20–30% headroom over actual usage.
Tune routing rules. Point simple jobs at cheaper models. Save the expensive ones for complex work.
Check fallback chains. Make sure key jobs — customer replies, call transcription — have at least one backup model.
Turn warnings into hard limits. Once the numbers settle, the warnings become real ceilings.
Review the report monthly. Who spends the most, and is it paying off?

What this means for leadership

One door to every model changes how a company talks about AI. Instead of “nobody knows what this will cost,” it’s “here’s spend by department, here are the caps, here’s who hit a limit and why.” Instead of being locked into one vendor, you can switch models in one place. Instead of keys scattered across people, there’s one secure store.

PRANUS pricing is built on the same idea: you know your maximum up front. See the details on our pricing page. If you’d like to see what an AI spend report by department and agent looks like, book a demo and we’ll walk you through one.