Services for businesses in Lucknow, India and beyond
AI agents for one repeated task, with a written stop when the case is unclear.
An AI agent is software that can take a series of steps toward a task: read a request, look something up, draft a reply, or update a record. It is not a chatbot with the keys to the business. The useful version has a narrow job, a list of systems it may touch, and a person who still owns anything that cannot be undone.
Sunbeam designs and builds AI agents from our studio in Lucknow for teams in India and for clients elsewhere, including the Gulf, the UK and the US. A first agent covers one workflow. Widening it is a later decision, made after you have seen how the first one behaves on real examples.
The business case
What better looks like.
- A task that no longer waits on someone remembering the next step
- A review queue for actions that should not run unattended
- A log of what the agent did, escalated, or refused
What it is
One job, with a stop when the job is unclear
The agent receives a trigger: a message, a form, a new file, or a status change. It gathers the context it is allowed to see, proposes or performs the next step, and writes down what it did. If the case is outside the examples you agreed, it stops and hands the work to a person. That refusal is part of the product, not a failure of the model.
AI automation is the wider practice of putting a model inside a workflow. An agent is the case where the software chooses among several steps. A single classification, a reminder, or a field copy is usually a rule, and we will say so when a rule is the better tool.
A written boundary
The proposal states what the agent may read, what it may change, and which actions stay with a person.
Tools, not a blank prompt
It calls the systems you name: a mailbox, a CRM, a document set. It does not invent access it was not given.
A log
You can see what ran, what was escalated, and what failed, without asking an engineer to reconstruct it.
An evaluation set
We test it on examples from your work before it meets a customer or a wider staff group.
Who it is for
Teams with a repeated task and real examples
This fits a front desk, an operations queue, a support inbox, or a sales follow-up that already has a pattern. You can point at last month's cases. Someone on your side can say whether an answer was acceptable. It fits poorly when the task changes every time, when a wrong action would move money or a legal position without review, or when nobody will look at the first weeks of output.
An AI receptionist is one shape of this work: it can take a message, answer from a short approved list, and book or record a request. Changing a booking, taking a payment, or giving advice that your profession regulates stays with a person unless you have explicitly accepted that case.
Triage
Sort incoming work into the queue that already exists, and leave the unusual items for a person.
Drafts
Prepare a reply or a summary a staff member edits before it is sent.
Follow-up
Notice a step that has gone quiet and prepare the next nudge, with a record of what was sent.
A single specialist agent
One role, not a general assistant that is expected to run the company.
How we work
Define the task before choosing the model
We start with the workflow as it happens, the systems involved, and a set of real examples including the awkward ones. We write the actions that may run unattended and the actions that must stop. Only then do we choose a model and the tools it may call. A pilot uses your examples and a review queue. A wider release waits until the failure cases are understood.
What is the job, in one sentence?
If it takes a paragraph of exceptions, it is more than one agent, or it is not ready.
Which examples count as done?
We agree a set and use it as the test. A rehearsed demo is not that set.
What must a person still do?
Payments, deletions, regulated advice, and anything you cannot reverse without cost stay on that list until you move them.
Who watches the first weeks?
A named person reviews escalations while the context is still fresh.
Typical stack
A workflow, a model call, and a record of both
The usual shape is a small service that receives the trigger, calls only the tools in the proposal, and stores the transcript and the outcome. Next.js fits the review screen your team uses. Go fits the worker that runs the steps and retries a failed tool call without hiding it. Postgres holds the log. The model is a dependency you can change. It is not where your records live.
Review
Next.js
The queue where a person sees what the agent did, edits a draft, or takes the case back.
Steps
Go
Workers that call the allowed tools, enforce the stop conditions, and record failures.
Record
Postgres
The cases, the actions, and the escalations, so the system can be audited later.
Model
A hosted model
Chosen for the task and the data you are willing to send. Swapping it later should not require a rewrite of the workflow.
Timeline ranges
A narrow agent is weeks. A wider one waits for evidence.
These are planning ranges, not a promise. The short end assumes you can provide examples, a person to review them, and access to the one or two systems involved. Missing access, an undefined task, or a demand that the agent act without review moves the work into the longer range, or into a recommendation not to build it yet.
1 to 2 weeks
Task and boundary
The job, the examples, the systems, and a written list of actions that stay with a person.
3 to 6 weeks
One agent on real cases
A review queue, a log, and an evaluation against the examples you supplied.
After the pilot
Widen or stop
A second workflow, fewer review steps, or a decision that a rule would have been enough.
Process
How an agent goes into use
You can stop after the written boundary or after the pilot. Both are useful. Unattended actions are a separate decision.
01
Watch the task
We use your recent cases, not a hypothetical customer.
02
Write the boundary
What it may read, what it may change, and when it must stop.
03
Run the evaluation set
The agreed examples, including the ones that should be refused.
04
Open a review queue
A named person sees the output before it affects a customer, unless you have accepted a narrower exception.
05
Read the failures
Escalations and bad drafts tell you whether the boundary or the source material is wrong.
06
Hand over the log
Where it runs, who can turn it off, and how a later engineer reads what it did.
Who this is for
For teams ready to improve an important part of the business.
- Front desks and inboxes with a repeated pattern
- Operations teams that can supply real cases
- Sponsors who will name a person to review the first weeks
Frequently asked questions
Useful questions before you begin.
Is an AI agent the same as automation?
Automation includes rules that do not need a model. An agent chooses among steps. If the task is a single predictable update, we will recommend the rule.
Can the agent send messages or change records on its own?
Only for actions listed in the proposal. Payments, deletions, and regulated advice stay with a person unless you explicitly accept that case.
Do you train a private model?
A standard agent calls a hosted model and your own tools. Training a private model is a different project and is rarely the right first step.
Where do you deliver this?
The studio is in Lucknow. The same team works with clients across India and in other countries.
Related capabilities
One connected engineering team.
Automation and AI
Business automation, AI automation, AI agents and receptionist pilots for one workflow. Sunbeam is based in Lucknow and works across India and worldwide.
Knowledge systems
Knowledge systems and document Q&A for teams that need answers cited from their own files. RAG pilots and production setups from Sunbeam in Lucknow.
Starter pilots
Small fixed-scope pilots: booking, intake, lead follow-up, dashboards, QR ordering, document Q&A, an AI receptionist, or one AI agent. No prices listed.
Considering whether a custom approach is justified? Read our build-versus-buy decision guides, explore industry workflows, or review how ownership and handover work.
Start with the business problem, not the technology.
Tell us what is getting in the way. We will help identify a useful next step.