An AI agent is software that reads a task, decides which operations to run, calls your systems to run them, and checks its own result. It pays off in queues where many items arrive each week, mostly as documents, and where the rules bend at the edges. It fails when the engineering around the model is missing. The model is rarely the weak part.
Where agents earn their cost
Four traits mark a good queue.
- Volume. Hundreds or thousands of items a week, so a saved minute adds up.
- Documents. Each item arrives as a PDF, an email or a scanned form that a person reads before acting.
- Rules with exceptions. Most items follow the written rule. A minority need a judgement call.
- A number today. Someone can state how long the queue takes, what it costs and how often it errs.
Typical fits: invoice-to-purchase-order matching, loan document checks, support ticket triage, insurance claim intake, vendor onboarding.
| Queue | Volume | Document load | Exception rate | First-use fit |
|---|---|---|---|---|
| Invoice to purchase order match | High | High | Low to medium | Strong |
| Loan application document check | High | High | Medium | Strong, with a human lane |
| Support ticket triage | High | Low | Medium | Strong |
| Contract clause review | Low | High | High | Weak: needs a lawyer in the loop |
| Pricing exceptions | Low | Low | High | Weak: judgement is the work |
In each strong case the agent does the reading, matching and routing. A person keeps the decision. That split is also what a risk committee will approve. The NBFC loan origination case study shows the pattern: documents are read and cross-checked by software, and approval stays with credit staff.
Why enterprise agent projects stall
The causes are mostly engineering gaps, and they repeat. For a longer treatment, see why AI agent projects get cancelled.
- No test set. With no fixed list of cases that have known answers, each change to a prompt or model is a gamble.
- No spending limit. An agent that retries on failure can run up a model bill with nothing to stop it.
- No list of allowed actions. A security reviewer will reject an agent that can reach every internal API.
- No lane for people. Items that need judgement either get a confident wrong answer or sit in the queue.
- Integration left to the end. The agent logic takes a short time. Connecting to the ERP, the mail system and the document store takes the bulk of the project.
The five parts of a production agent
Whatever model or vendor you pick, the working design has the same five parts.
- Tool contracts. A written list of the operations the agent may call, with their parameters. Nothing outside the list runs. The list is the security model, and a reviewer can read it in an afternoon.
- Two lanes. The agent scores its own confidence on each item. High scores go straight through. Low scores go to a person with the context already gathered. The cut-off is a setting the business owns, and it is not buried in code. See human-in-the-loop design.
- An event log. For every run, store the input, the documents retrieved, each tool call and the output, so any decision can be replayed. See what to capture in an AI audit trail.
- An evaluation set that gates releases. Nothing is released until the full set has been scored, as a build waits for its tests to pass. See how to build an AI evaluation set.
- Budgets per run. A ceiling on model spend, on steps and on time. When a run hits the ceiling, it stops and hands the item to a person.
Build these five first. The model then becomes a part you can swap when a better or cheaper one appears, and the system you own stays the same.
Indian rules that shape the design
Agents that read customer records handle personal data, and India's data protection law may apply. The log in part 3 is the evidence you will need when anyone questions a decision. See the DPDP Act and AI systems for the duties in plain terms.
A tool contract in practice
Take invoice-to-purchase-order matching. The contract might list six operations, and no others.
| Operation | Reads or writes | Needs a person? |
|---|---|---|
| Read the invoice PDF and extract fields | Reads | No |
| Look up the purchase order and goods receipt | Reads | No |
| Compare quantity, rate and tax | Reads | No |
| Post a matched invoice to the payable queue | Writes | No, when confidence is high |
| Flag a mismatch with the fields that differ | Writes | Routes to a clerk |
| Release a payment | Writes | Always, and never by the agent |
A reviewer can approve or reject this page in minutes. The same page tells the build team what to test, and it tells the operations team what the agent can never do.
What to measure
Agree four numbers before the build starts.
- Straight-through rate: the share of items finished with no person involved.
- Accuracy against the fixed test set, recorded for each release.
- Handover quality: when the agent passes an item on, does the reviewer get the extracted fields, the source page and the reason, or only a pile of text?
- Cost per completed item, against the cost of the process it replaced.
Read straight-through rate only beside accuracy. An agent can push the first number up simply by guessing more boldly.
How to start
- Choose one queue that already has a backlog and an owner who wants it fixed.
- Record the baseline: items per week, minutes per item, error rate.
- Write the tool contract with five or six operations at most.
- Build the human lane and the log before you tune the model.
- Run the agent beside the existing process for a set period, then compare the numbers.
When the numbers move, you have evidence for the next budget request. If you are still deciding between building and buying, see build versus buy for enterprise AI.
Questions buyers ask
Where do AI agents produce returns?
In queues with steady weekly volume, documents as input and decisions that follow rules but have exceptions: invoice matching, loan checks, ticket triage, claims intake.
What should an agent never do alone?
Anything irreversible or high in value without a person's approval: releasing a payment, approving credit, changing a master record. Set those as human-lane actions in the tool contract.
Does the choice of model matter?
Less than the surrounding design. With tool contracts, an evaluation set and a log in place, you can test a new model against the same cases and switch if it scores better.
Where Brihat Infotech fits
Brihat Infotech builds agents with tool contracts, human lanes and event logs as part of its AI agents and workflow automation work. Start a conversation.