Controlling AI Costs: Budgets, Wallets & Pricing

Most finance leaders can tell you exactly what a SaaS seat costs. Few can tell you what last month's AI usage cost — or why it doubled.

Why AI spend is unpredictable

AI costs behave differently from software subscriptions. There's no fixed seat price to budget against. Every message, every agent run, every tool call consumes tokens, and token consumption varies wildly by task, by model, and by how much context gets pulled in. A team experimenting with agents will naturally generate more variance than a team running the same five reports every week.

Add multiple AI providers into the mix — different vendors, different pricing tiers, different credit systems — and finance teams lose the thread entirely. Credit packs obscure the real unit cost. Markups get buried inside "premium" plan tiers. By the time an invoice arrives, nobody can reconstruct which team, which agent, or which model actually drove the spend.

The fix isn't to slow down AI adoption. It's to make the cost structure legible from the start: one currency, one visible price per token, and controls that stop spend before it becomes a surprise.

Transparent per-token pricing, one wallet

AgentWorks bills usage at cost plus a flat 10% — nothing more. There are no opaque credit packs to decode and no per-model markup that quietly charges more when an agent happens to use a pricier model. Every model draws from the same €-denominated wallet, so finance sees one number, not five vendor invoices with different units and different margins baked in.

This matters because it makes transparent token pricing the default rather than a negotiated add-on. You can see, per message and per agent run, exactly what something cost — not an estimate, not a rounded credit deduction, but the actual token spend plus the fixed markup. That level of visibility is what lets an ops or finance lead sign off on AI usage the same way they'd sign off on any other line item: with evidence, not trust.

Because every model on one platform settles against the same wallet, you're not reconciling separate bills from separate providers either. Whether an agent runs on GPT-5, Claude, Gemini, or Mistral, the cost lands in one place, in one currency, at one markup.

The AUTO router: paying only for the capability you need

A large share of unpredictable AI cost comes from using an expensive model for a task that didn't need it. A one-line classification, a short summary, a simple lookup — none of these need your most capable (and most expensive) model, but without a routing layer, they often end up there by default because that's what someone happened to configure.

AgentWorks' AUTO router sends each message to the cheapest model capable of handling it, evaluating across GPT-5, GPT-5 mini, Claude Opus, Sonnet and Haiku, Gemini Pro and Flash, and Mistral Large. Simple tasks route to lighter, cheaper models automatically; harder tasks still get the capability they require. You don't have to manually pick a model for every agent and hope you guessed right on the cost/quality tradeoff.

Where a specific agent genuinely needs a specific model — a legal-review agent that should always use your most rigorous model, for instance — you can pin that model per agent. The router is a sensible default, not a constraint. This matters most once you're running multi-agent pipelines, where a single workflow might touch a dozen models across a dozen steps; routing at that scale is where manual model selection stops being practical.

Budgets and hard caps that actually stop spend

Transparency tells you what happened. Budgets control what happens next. AgentWorks lets you set budgets at the org, team, or individual user level, with two enforcement modes: soft warnings that flag when spend is approaching a threshold, and hard limits that pause agents outright once a limit is hit.

That distinction matters operationally. A soft warning is right for a team that's exploring new use cases and needs visibility without friction. A hard limit is right for a cost center that must not exceed its allocation, full stop — no agent silently overspends into next month's budget because nobody was watching a dashboard. Optional auto top-ups give you a middle path: a limit that expands automatically only when you've decided in advance that it should.

On Team and Enterprise plans, per-team cost reporting turns this from a control into a management tool — you can see which teams are driving usage, whether that usage is trending toward or away from budget, and where to intervene before a limit is even reached. If you're evaluating the AI workforce platform for a multi-team rollout, this is usually the section that turns a "can we afford this" conversation into a "here's exactly how we'll control it" conversation.

Prompt caching: cutting cost on repeated context

A meaningful chunk of token spend in production agents isn't the user's question — it's the context around it. System prompts, knowledge base chunks, tool definitions: the same material gets sent again on every message, every run, every conversation turn.

Prompt caching addresses this directly. When the same context repeats, cached tokens are billed at a steep discount — up to roughly 80% cheaper on cache hits, depending on how much of the request is reused. For agents built on AI agent templates with a stable system prompt and a fixed knowledge base, this isn't a marginal saving; it's often the difference between a cost-effective agent and one that quietly burns budget on context it re-sends every time. It works the same way whether you're running a single agent from the chat workspace or a full pipeline of them.

How to forecast AI spend

Forecasting starts with the free plan's structure: €0 to start, with a one-time €5 credit to see real costs against real usage before committing to anything. From there, Pro at €39/month includes a €10/month balance, and Team at €49/seat/month includes the same €10/month balance per seat, with per-team reporting to track how that balance gets consumed. Enterprise pricing is custom, built around the budgets and reporting your organization already needs.

Any usage beyond the included monthly balance draws from the wallet at the same cost-plus-10% rate — no cliff, no re-negotiated tier, no surprise invoice. Because every agent can be priced fixed or per-run and everything settles against that one wallet, forecasting becomes a matter of multiplying expected usage by a known unit cost, not guessing at how five different vendor bills will interact.

Summary: AgentWorks controls AI cost with one transparent wallet billed at cost + 10%, a router that defaults every task to the cheapest capable model, budgets with hard limits and soft warnings at org/team/user level, and prompt caching that cuts repeated-context cost by up to ~80% — so spend is forecastable instead of discovered after the fact.

Frequently asked questions

Does AgentWorks mark up different AI models differently?

No. Every model — GPT-5, Claude, Gemini, Mistral, and their smaller variants — is billed at the same cost-plus-10% rate from the same wallet. There's no premium markup for using a more capable model beyond the actual token cost difference.

Can we stop an agent from overspending automatically?

Yes. Set a hard budget limit at the org, team, or user level, and agents pause automatically once that limit is reached. Soft limits are also available if you'd rather get a warning than an automatic stop.

How does the AUTO router decide which model to use?

It evaluates each incoming message against the models available on the platform — from lightweight options like GPT-5 mini, Claude Haiku, and Gemini Flash up to more capable models like GPT-5, Claude Opus, and Mistral Large — and routes to the cheapest one capable of handling that specific task. You can override this and pin a fixed model for any agent where consistency matters more than cost optimization.