Who this is for
Businesses that have already automated the fixed-rule work and are left with tasks a person handles by checking two or three systems and using judgement: qualifying an enquiry against eligibility rules, chasing a document, reconciling an order against a payment, preparing a quote from a specification. This page sits under marketing and business automation services, and it is the most advanced form of the work described on AI automation services.
If you have not yet automated the deterministic part of the process, an agent is the wrong purchase. Agents are more expensive to build, harder to test and slower to trust than a rules-based workflow, and they should only take over the part that genuinely needs judgement. Start with workflow automation and add an agent to the step that rules cannot express.
Agent, chatbot or workflow
| Workflow | Chatbot | Agent | |
|---|---|---|---|
| Decides its own steps | No, the path is fixed | Only within a conversation | Yes, within the tools it is given |
| Typical job | Form to CRM to WhatsApp | Answering visitors | Qualify, look up, prepare, act |
| Testing effort | Low | Medium | High, and never finished |
| Failure mode | Stops and alerts | Says something wrong | Does something wrong, repeatedly, quickly |
| Right choice when | The rules are known | People need answers | The judgement is real and the actions are reversible |
What I build agents for
- Enquiry qualification. The agent reads a message in Nepali or English, checks the CRM for an existing record, checks a sheet or system for eligibility, drafts the reply, proposes a next action and books a slot if the person asks for one.
- Document chasing. For processes that need paperwork, the agent tracks what is outstanding per case, sends the right reminder, reads what comes back, checks it is the right document and updates the record.
- Order and payment reconciliation. Comparing what the store says, what the bank statement says and what the sheet says, flagging the differences a person should look at rather than fixing them silently.
- Research and preparation. Assembling a briefing before a sales call from the CRM history, the website, past emails and public sources, so the person walks in prepared.
- Internal operations. Assigning work by workload and skill, escalating what has stalled and preparing the weekly review, which is the sort of thing built into an internal tool or dashboard.
What an agent is never allowed to do
- Move money. No payments, refunds, transfers or credit notes. It can prepare one for a person to approve.
- Send to a customer unreviewed, until the measured error rate on a real test set justifies it, and never for complaints, medical, legal or financial content.
- Delete anything. Records are archived or flagged, never removed.
- Act outside its tool list. An agent can only do what the workflow explicitly exposes to it. There is no general access to your systems, no shell, no arbitrary browsing.
- Run without a budget and a stop. Every agent has a step limit, a token limit, a daily spend ceiling and a pause switch anyone can use.
How a build runs
- Write the job description. Literally: what a new employee would be told to do, what they may decide alone, and what they must ask about. If that cannot be written in one page, the task is not ready for an agent.
- List the tools and the limits. Every read and every write the agent may perform, with the boundary on each: which records, which fields, which value ranges.
- Build the test set. Fifty to a hundred real past cases with the correct outcome, including the awkward ones. The agent is scored against these before anyone sees it, and rescored after every change.
- Shadow mode for two to four weeks. The agent proposes; a person still does the work; we compare every decision. Nothing goes live until the comparison is boring.
- Limited live running on the lowest-risk slice of cases, with every action logged and reversible, and an alert when the agent takes an unusual path.
- Widen slowly, guided by measured accuracy, not by enthusiasm. Reviews are quarterly, because model updates change behaviour and the test set has to be re-run.
What you receive
- The agent running in your own automation account, on your own model API keys.
- The written job description, tool list and limits, in language your team can audit.
- The scored test set and the accuracy record, so changes are measured rather than assumed.
- A full decision log: what the agent saw, what it chose, what it did, and how to reverse it.
- Spend ceilings, alerts and a pause switch, plus documentation and a handover session.
Pricing pointer
Agent setup is quoted on request and sits at the upper end of automation builds because testing, shadow running and guardrails are most of the effort, not the model call. Maintenance and quarterly accuracy reviews are quoted on request, and model usage is billed to your own account. If what you actually need is a team assistant that answers from your documents rather than one that takes actions, that is cheaper and lives on custom GPT and AI assistant development.
Next step
Bring one task, the systems it touches and twenty real past cases. I will tell you within four business hours whether it needs an agent, a plain workflow or neither, and what the guardrails would have to be. I turn this work down more often than I take it, which is the point.
