Agentic AI Costs and Ownership:
8 Key Questions Answered for Rapidly Growing and Mid-Sized Businesses

How much does this actually cost, and what’s the ROI?

The question of cost and ROI comes first in almost every initial session regarding agentic AI. The appropriate approach is to provide a structure and an honest assessment, not a number. The cost has three parts:

• Analysis and design: working out the process and the decisions.
• Build and deployment: configuring, integrating, and testing the agent(s).
• Ongoing operation: infrastructure, monitoring, and periodic updates.

For a single, well-scoped first process, this is a project budget, not an enterprise program. ROI depends on the process. The returns come from shorter cycle times, fewer errors, lower cost per transaction, and capacity freed up without additional hiring.

The honest answer is that nobody can quote ROI responsibly until the process has been appropriately analyzed, because returns depend on how many transactions the agent(s) handle, at what level of autonomy, and where the time is consumed today. This is also why the first step is a focused assessment, not a platform purchase.

What is a token, and what does it cost in real dollars?

A token is a unit AI services use to process text. In English, one token represents approximately three-quarters of a word, although the exact number varies with the content and model. Providers typically charge for the tokens you send in and the tokens the model generates in response, often priced per million tokens.

Prices vary widely among commercial AI models, and they change often, so any figure can become outdated quickly. The more useful point is how the bill behaves. Individual AI requests may cost very little, but costs can accumulate as agents perform multiple operations across high transaction volumes. Your bill tracks how much work flows through the AI, not how much value that work creates.

An agent that reads every incoming order, checks it against policy, and drafts a response may use thousands of tokens per case. Multiply that by your daily volume and you have a cost line that rises every time your business does. When open-weight models are deployed on infrastructure you control, you avoid a third-party AI provider's per-token inference charges. Instead, operating costs are driven by computing infrastructure, maintenance, and the resources required to support the workload.

What does it actually cost to run an open-weight model ourselves?

There are three parts:

• Build: analyzing the process, configuring the agent, integrating it with your systems, and testing it on real cases.
• Infrastructure: the servers or private-cloud capacity the model runs on. For appropriately scoped business workloads, smaller open-weight models may operate efficiently on comparatively modest infrastructure.
• Upkeep: monitoring, periodic model updates, and adjustments as your rules change.

The build can represent a significant share of the initial investment, and much of it is process work you would need under any pricing model. Once the capability is running, the cost of each additional decision is primarily driven by computing resources rather than a third-party provider's per-token pricing. Depending on workload volume, infrastructure utilization, and operating requirements, this can provide a significant long-term cost advantage.

However, self-hosting is not automatically less expensive than using a metered AI service. The economics depend on the volume and consistency of demand, infrastructure costs, ongoing maintenance, and the resources required to operate the system. These costs should be evaluated for a specific business process and set of use cases rather than assumed in advance. With a well-utilized deployment, additional decisions may be processed without a third-party AI provider's per-token charges.

Are open-weight models really good enough? Aren’t they a step behind the big AI companies?

For open-ended tasks like research, complex analysis, or broad strategy, the leading frontier models may still have an advantage. However, the performance gap between frontier and open-weight models has narrowed significantly, and the difference varies by model and application.

For bounded operational work, such as classifying requests, checking information against policies, extracting data from documents, or drafting responses within established rules, appropriately selected open-weight models can provide more than sufficient capability for many business processes.

The key point is that you don’t need the most capable model, only one that’s good enough for the work. Whether a model is good enough isn’t a matter of opinion. You test it against real cases from your own operation, including the exceptions, before anything goes live.

How do we avoid AI vendor lock-in?

Separate the things you want to keep from the things you might want to change. The process design, decision rules, controls, test cases, and documentation should be yours, written in your own terms. The model and the platform should be replaceable.

In practice, that means favoring open-weight models you can run in your own environment, avoiding features that only exist in one vendor’s product, keeping your data in formats you control, and confirming in the contract that you own the configured solution. Then ask one test question of every option: if we needed to change vendors next year, what would we have to rebuild? The less the answer includes, the less locked in you are.

We don’t have AI specialists. Can a company our size actually do this?

Yes, but be realistic about which skills you need. Building and running a well-designed agentic process doesn’t require a research team. It does require three things:

• Process and decision analysis: knowing exactly how the work runs today and how you want it to run going forward.
• Technical setup: configuring and integrating the technology.
• Operating knowledge: someone who can monitor the agent and adjust it.

Most mid-sized companies have the business knowledge but not the time to analyze it, and their IT teams are already stretched. That’s why outside help is common for the first initiative, provided it’s set up so your team learns to run and extend the capability rather than depend on outsiders indefinitely. Training people to own it should be part of the scope from the start.

Will this replace our people?

For many small and mid-sized businesses, one of the most valuable opportunities presented by agentic AI is increasing operational capacity without increasing staffing at the same rate. AI agents can take on routine, repetitive judgments that consume skilled employees' time, allowing those employees to focus on resolving complex exceptions, managing customer relationships, and handling work that requires experience and judgment.

The impact on staffing will depend on the organization, the processes being transformed, and how leadership chooses to use the additional capacity. Some organizations may redeploy employees to higher-value activities, while others may reduce future hiring requirements or adjust staffing levels over time. How the initiative is communicated internally matters. Employees who understand how AI can improve their work are more likely to contribute the process knowledge essential for successful implementation.

How do we capture what our best people know before they retire or leave?

Most of that knowledge is judgment: how an experienced person prices an unusual order, decides which exceptions matter, or reads a customer. It rarely lives in any system. Capturing this institutional knowledge, which resides primarily and informally in the minds of your best subject matter experts, takes structured conversations, not a questionnaire.

The practical method is to walk through real cases with the expert and ask, at each step, what they looked at, what they decided, and why. Recording the exceptions and edge cases is where the value is. The result is a set of explicit decision rules and heuristics that a new hire can learn, an auditor can follow, and an agent can execute. Done well, it turns a retirement risk into a company asset.