Who this is for
Businesses with a repeated task that needs judgement but not much of it: sorting 200 enquiries a week by service and urgency, turning supplier invoices into sheet rows, answering the same 30 customer questions, writing first drafts of proposals from a call transcript. If a competent new hire could do the task after a one-page instruction, a model can usually do it with the same instruction and a person checking the output. If the task needs relationships, negotiation or accountability, it stays with people.
This page is the overview. Specific builds have their own pages: AI chatbot development for website and WhatsApp bots, AI agents for business for multi-step agents, custom GPT and AI assistant development for team assistants on your documents, and AI consulting and tool setup for choosing and rolling out tools. Ranking in AI answers is a different job and lives under AI search optimization.
What I automate with AI
| Task | What the model does | What stays with a person |
|---|---|---|
| Enquiry triage | Reads the form, email or WhatsApp message; tags service, urgency, language and location; routes to the right person; drafts a reply | Sends the reply, or edits it first |
| Data extraction | Pulls names, amounts, dates and line items from PDFs, invoices, CVs and screenshots into a sheet or CRM | Approves rows flagged as low confidence |
| Summaries | Turns call recordings, meeting notes and long email threads into a summary with actions | Confirms actions and owners |
| Drafting | Writes first drafts of proposals, quotes, job descriptions, captions and ad variants from a template and the record | Edits and sends |
| Classification and QA | Checks content against a brand or compliance checklist; scores leads; flags reviews that need a response | Handles the flagged items |
| Customer answers | Answers questions from your own documents on the website or WhatsApp, and hands off when unsure | Takes the hand-off |
What an AI automation engagement includes
- Task selection: a short review of your repeated work to pick the one or two tasks where AI pays back fastest.
- Prompt and instruction design written from your real examples, with a test set of past cases to measure accuracy before launch.
- Workflow build in n8n, Make or Apps Script connecting the model to your forms, inbox, CRM, sheets and messaging.
- Guardrails: confidence thresholds, an approval step, a “do not answer” list, and logging of every input and output.
- Model and provider choice based on cost, language quality in Nepali and English, and where the data may be stored.
- Cost estimate per month from expected volume, so there are no surprise API bills.
- Usage policy for the team: what may be sent to the model, what may not, and who reviews.
- Documentation, training and an optional monthly review of accuracy and cost.
How I build an AI workflow
- Collect 50 to 100 real examples of the task with the correct outcome. Without them there is nothing to test against, and I will not ship a workflow that has only been tried on made-up data.
- Write the instruction the way you would brief a new employee, then test it on the examples and record the error rate.
- Decide the threshold. Above a confidence level the workflow acts; below it, a person decides. Early on the threshold is high and almost everything is reviewed.
- Build the workflow around the model call: trigger, data gathering, model step, checks, action, log, alert.
- Run in shadow mode for one to two weeks: the model produces output, a person still does the task, and we compare.
- Go live with review and lower the review rate only as the measured accuracy justifies it.
What to decide before automating with AI
- Data. Which customer data may leave your systems and reach a model provider, and whether that is acceptable for your industry, especially healthcare and finance.
- Accountability. Who is responsible when the model is wrong. The workflow must make that person’s review easy, not hide the output.
- Language. Nepali quality varies by model; I test with your actual Nepali and romanized Nepali messages before choosing.
- Cost ceiling. A monthly API budget and an alert when it is approached.
- What not to automate. Complaints, medical advice, legal commitments and pricing exceptions are routed to people by rule, not by model judgement.
What breaks, and how I prevent it
- Confident wrong answers: the “do not answer” list, retrieval from your own documents only, and a hand-off when the question is outside them.
- Drift after a model update: the test set is re-run on every provider change and the workflow is pinned to a model version.
- Runaway cost from loops or long inputs: token limits, retries capped, daily spend alerts.
- Prompt injection from customer messages: instructions and data are separated, and the model cannot trigger actions the workflow does not expose.
- Silent failures: every run logged, every error alerted.
Proof
I use AI steps in my own reporting, lead triage and content workflows, and I have built AI-assisted operations and task tools for an agency in Kathmandu and for a Nepali financial-services company’s tool rollout. Screenshots and a walkthrough are available on a call.
Pricing pointer
Setup is quoted on request; a single AI workflow with one model step and an approval queue sits in the middle of the range because of the testing involved, and an agent that reads several systems and takes actions sits at the top. Maintenance and accuracy reviews are quoted on request, and model API usage is billed to your own account. For the general non-AI version of these builds, see workflow automation; for sheet-heavy work, document and data automation.
Next step
Bring one task and twenty real examples of it to a call and I will tell you whether a model can do it, at what accuracy, and what it would cost per month. The rest of the automation services are on the automation hub.