An enterprise AI agent is software that takes a task from a queue, reads documents, calls your systems, and either finishes the task or passes it to a person. It earns its place where volume is high, inputs are documents, and decisions follow rules with many exceptions. This guide sets out where to start, seven controls every deployment needs, and the rules that shape the design. Each section links to a longer article.
What counts as an agent here
A chatbot answers questions. An agent acts: it opens a record, extracts the figures from an invoice, checks them against an order, posts an entry or routes the case. Because it acts on your systems, an agent needs the controls of any system that can change records: permissions, a log, a way to stop it.
This is the hub for the subject. AI agents in the enterprise covers which use cases pay off; this guide covers the controls, the rules and the rollout around them, and links to one article for each.
Where to start: score the queue
Pick the first use case with a scoring table. Give each queue a score from 1 to 3 on each row.
| Row | Score 3 when |
|---|---|
| Volume | The queue gets many items every day |
| Inputs | Most inputs are documents or free text |
| Rules | A written policy exists, with known exceptions |
| Baseline | You can state today's time, cost or error rate |
| Reversibility | A wrong action can be corrected cheaply |
| Owner | One business leader owns the queue and wants it changed |
Choose the queue with the highest total, and drop any queue that scores 1 on baseline or owner. A result you cannot measure will not secure a second budget. Good first candidates include loan-file checks, invoice matching, claims intake and ticket triage. AI use cases in Indian manufacturing lists plant-side examples.
Seven controls every agent needs
1. An evaluation set
Collect real cases with verified answers before the system exists. They cost little to collect, and without them every later argument rests on impressions. See building an AI evaluation set.
2. A retrieval check
When an agent answers from your documents and gets it wrong, inspect retrieval first. Chunks split mid-sentence, tables flattened into text and an old policy outranking the current one are common causes. RAG beyond the demo covers the checks.
3. An action list
List the actions the agent may take, and refuse everything else. Read access to a record differs from permission to change it. Give each action its own permission and a limit, such as an amount ceiling.
4. A cost cap
Set a budget per task and per day. An agent in a loop can spend a month's allowance in an afternoon. A cap that pauses the agent and alerts a person turns a surprise bill into a ticket.
5. A human lane
Route uncertain cases to a person, and show the evidence rather than a verdict. Measure whether reviewers still disagree sometimes. Human-in-the-loop design gives the design and four measures.
6. An audit log
For each action, record the input, the data retrieved, the model and version, the rule applied, the action taken and the person who reviewed it. The test is whether you can reconstruct a decision after the model has been replaced. The contents of an AI audit log lists fields.
7. Rollback and a stop switch
Give a named person a way to pause the agent, and write down how to reverse its last actions. Test the stop switch before go-live, as you would a fire drill.
Why agent projects get cancelled
Projects are rarely cancelled because the model is weak. They stall for engineering reasons: no evaluation set to show progress, no cost cap, an unrestricted action list, no review lane for people, and integration scoped as a side task when it is the bulk of the effort. The article on AI agent projects that get cancelled walks through each.
Check your plan against the seven controls. Any control with no owner is a risk with no budget.
Which rules bind the design
Keep three levels apart.
- Voluntary frameworks. The NIST AI Risk Management Framework, released in January 2023 and intended for voluntary use, organises work into four functions: govern, map, measure and manage. Source: NIST AI Risk Management Framework. ISO/IEC 42001 is a management-system standard in the same space.
- Foreign law that reaches you if you sell abroad. The EU AI Act requires human oversight of high-risk systems that works in practice; human-in-the-loop design quotes Article 14.
- Indian law and sector rules. The Digital Personal Data Protection Rules and the Reserve Bank's outsourcing and payment-data rules fix where data can be held and who can inspect the vendor, so they shape the architecture before any model is picked. See the DPDP Act and AI systems, security and IP for offshore engineering teams and AI governance for regulated Indian companies. Public-sector buyers should also read government software procurement under GeM and GFR.
Build, buy or mix
Buy the generic tools, such as meeting notes and office assistants. Build where the agent works inside the process that sets you apart. A third route is often missed: license a platform and write only the thin layer that is yours. Ask whether the process is your edge or just a habit. Build versus buy for enterprise AI sets out the tests.
A worked example
A lender used agents for the file preparation of consumer loans. The agents read documents, checked them against outside sources, scored the file and routed it, and a named underwriter kept the approval. AI in loan origination explains the design, and the NBFC loan origination case study gives the figures.
What to ask a vendor
- Show the evaluation set and the scores for the last three releases.
- List every action the agent may take, with its limit.
- Where does each data item sit, and who can inspect the provider?
- What stops the agent when a cost cap is reached?
- Show a sample audit record for one decision.
- How do I pause the agent, and how do I reverse its last actions?
- After the project, who owns the test cases, the prompts and the code?
Specific answers, with documents behind them, separate a deployable system from a demonstration. Choosing a software development company in India adds general questions for any supplier.
How the rollout runs
Move through four stages, and do not skip one.
- Shadow. The agent runs on live cases and writes its answer to a log. People continue to do the work. You compare the two.
- Assist. The agent prepares the file and a person decides every case.
- Partial automation. Cases above the confidence threshold proceed with sampling by a reviewer. The rest go to the human lane.
- Review and widen. Reset thresholds from real outcomes, then add the next queue.
The second stage is where most of the benefit appears, and it needs no change of authority.
Starting plan
- Choose one queue that your operations team wants changed, using the scoring table.
- Write today's figure for time, cost or error rate.
- Assemble the evaluation set from live cases.
- Ship an agent with an action list, a cost cap, a human lane and a log.
- Run it beside the existing process, not in place of it.
- Compare results every fortnight and move the threshold from the data.
- Move volume to the agent only after the comparison holds for a full cycle.
Questions buyers ask
Does an agent need a large language model?
Not always. Many queue tasks need a simple rule engine or a document-reading model with no text generation. Pick the least capable tool that meets the evaluation set.
Where should the data sit?
Where your regulator and your contracts allow, which is a design decision before model choice. Payment data has its own storage rule, covered in the security and IP article.
How do we know the agent is getting better or worse?
Run the evaluation set before each release and after any model change, and track reviewer overrides in production.
Who is accountable when the agent makes a mistake?
A named person in the business who owns the queue. The audit log lets them find the cause.
Where Brihat Infotech fits
Brihat Infotech builds AI agents and automated workflows with policy limits, exception lanes, an action audit trail, a stop switch and release gating on evaluation, and custom AI systems with evaluation harnesses. It uses ISO/IEC 42001, the NIST framework and EU AI Act tiers as design references, not as audited conformance.