What it costs to run an AI agent

Four parts to the bill

When a finance head asks what an agent will cost, the answer has four parts:

1. The platform: the plan that lets your team build and run apps. AksharAI plans are on the pricing page, priced by users and active apps.
2. Model usage: what the model provider charges for the calls your agents make. This depends on volume and on the model.
3. The build: the time to design, connect, test and launch the agent, by your team, ours, or both.
4. Review time: the hours your people spend approving and correcting the agent's work. It goes down over time, but it never goes to zero for work that matters.

Most surprises come from the second and fourth.

Estimating model usage from your own numbers

Model providers charge per token, separately for input (what you send) and output (what the model writes). A token is roughly three quarters of an English word; Indian scripts usually need more tokens per word.

To estimate, work out four numbers for one typical case:

• Input tokens per call: instructions, the document or message, and any retrieved passages.
• Output tokens per call: the extracted fields, the reply or the summary.
• Calls per case: many agents make several calls per case, for example read, check and draft.
• Cases per month: from your own records.

Then: monthly cost = cases × calls per case × (input tokens × input price + output tokens × output price).

A worked example with assumed numbers

A distributor in Rajkot wants an invoice matching agent. Assume, for illustration only:

• 1,000 invoices a month
• 3 calls per invoice (read, check, prepare the entry)
• 4,000 input tokens and 500 output tokens per call
• an input price of P and an output price of Q per million tokens, taken from your chosen provider's current price list

Input tokens a month: 1,000 × 3 × 4,000 = 12 million. Output tokens: 1,000 × 3 × 500 = 1.5 million. Monthly model cost = 12 × P + 1.5 × Q.

Put in your provider's current prices and you have a working estimate. Then add a margin, because real documents are longer than the sample, and some cases are retried.

Prices change often and differ widely between models. Check the provider's published price list on the day you estimate.

What moves the model bill

• Model choice. Larger models can cost many times more per token than smaller ones. Use the larger model only where the task needs it.
• How much context you send. Sending a whole forty page policy on every call is expensive. Retrieving the three relevant passages is cheaper and usually more accurate.
• Number of steps. Each extra check is another call. Some checks are better done in code, at no model cost.
• Caching. Some providers charge less for parts of the input that repeat across calls, such as long standing instructions.
• Retries and long tails. Unreadable scans and long email threads cost more than the average case.

The build

A focused agent on one process, with one or two system connections, is a modest project. Older systems without APIs, many document types or several channels make it larger. If our team does the work, it is priced at INR 199 or USD 2 an hour on 2good terms, with a 150 hour minimum, no advance and payment after you accept the work. The cost to build estimate gives a rough range from a few simple inputs.

Review time, the cost people forget

If an agent prepares 1,000 invoice entries a month and an accountant spends thirty seconds approving each routine one and five minutes on each exception, that is real time. Count it. It is still usually far less than doing the work by hand, and a sound business case includes it.

The hours saved calculator lets you put in your own task volume, time per task, the share an agent could take and the review time, to see what is realistically freed.

Other costs to plan for

• Hosting, if you run AksharAI in your own cloud or on premise, including servers for open models.
• Channel charges, such as fees for WhatsApp business messages, set by Meta and your provider.
• Speech services for voice agents.
• Upkeep: updating knowledge sources, rerunning tests when policies change, and adapting to new model versions.

Check the estimate against real numbers

Start with one process, measure real volumes and real review time for the first month, and compare with your estimate. Measuring whether an agent is worth keeping sets out a simple scorecard for that first period.