Question 1

What does AI Sewa actually build?

Accepted Answer

Data pipelines and search over your company's information, models set up for your business (hosted or private), connectors into your existing software, and agents that carry out multi-step work in those systems.

Question 2

What is an AI agent?

Accepted Answer

A program built around a language model that can use tools, such as searching your documents, reading a CRM record or creating a ticket, to complete a task. It can only use the tools and data it has been given.

Question 3

What is an agentic team?

Accepted Answer

Several agents, each with one narrow job, passing work between them. One reads the incoming request, another looks up the customer, another drafts the reply. Narrow agents are easier to test and to trust.

Question 4

Will it work with our existing software?

Accepted Answer

Almost certainly. We connect through APIs where they exist and through the database or file exports where they don't.

Question 5

Do we have to replace anything?

Accepted Answer

Usually not. The point is to make the systems you have do more, not to migrate you to new ones.

Question 6

Can it run privately?

Accepted Answer

Yes. Models can run in your own cloud account, on dedicated servers or on-premise. It costs more than calling a public API, so we'll tell you whether your data actually needs it.

Question 7

Can you train a model on our company information?

Accepted Answer

Yes, though most of the time you don't need to. Retrieval over your documents handles most question-answering. Fine-tuning is worth it when you need a model to write, classify or decide the way your team does.

Question 8

Does our data go to public AI services?

Accepted Answer

Only if the architecture you choose uses one, and you'll know before we build it. Private deployments keep sensitive data inside infrastructure you control.

Question 9

Can agents act on their own?

Accepted Answer

Within the permissions you set. We usually start with a person approving every action, then remove approvals for the low-risk ones once the agent has a track record.

Question 10

How do we start?

Accepted Answer

A conversation about one process that costs your team time. We map it with you, say honestly whether AI is the right fix, and if it is, propose a small first build.

Question 11

Will our data be used to train someone else's model?

Accepted Answer

Not by default on the major enterprise APIs. OpenAI and Anthropic both state that business API data isn't used for training unless you opt in, though providers may keep it for a limited period for abuse monitoring. The questions that usually matter more are where the data is processed, how long it's kept and which sub-processors see it. Where those answers aren't acceptable, a self-hosted model is the fix.

Question 12

RAG or fine-tuning: which do we need?

Accepted Answer

Start with retrieval (RAG) in most cases: it answers from your current documents, cites its sources and updates the moment a document changes. Fine-tuning helps when you need a fixed output format, a house style or heavy domain vocabulary, and it costs more to train and keep current. We decide with an evaluation set, not by preference.

Question 13

How do you stop the model making things up?

Accepted Answer

We can't promise it never will, so we design around it. Answers are grounded in retrieved documents and cite them, the system is allowed to say it doesn't know, and we measure accuracy against a test set of real past questions before launch and after every change. Anything consequential goes to a person for approval.

Question 14

Why do so many AI pilots never reach production?

Accepted Answer

The common reasons are unready data, demos that never touched real systems, unclear success measures and costs nobody modelled. Gartner, for example, predicts that 60% of AI projects without AI-ready data will be abandoned through 2026. We start with the data and the process, agree the measure up front, and build the first version against your real systems.

Question 15

What is prompt injection and how do you defend against it?

Accepted Answer

It's text, often hidden in an email, web page or document, that tries to make the model ignore its instructions. OWASP ranks it the top risk for LLM applications. Defences are layered: agents get the narrowest permissions possible, untrusted content is kept apart from instructions, tool calls are validated, and risky actions wait for a person.

Question 16

What is the Model Context Protocol (MCP)?

Accepted Answer

An open standard for connecting AI models to tools and data sources, now maintained under the Linux Foundation's Agentic AI Foundation. We use it where it fits, and treat every MCP server like any other third-party dependency: reviewed, permissioned and logged.

Question 17

Does the EU AI Act apply to us if we're outside the EU?

Accepted Answer

It can, if your AI system is used in the EU or its output affects people there. As of October 2026: transparency duties (such as telling people they're talking to an AI) apply now; obligations for high-risk uses listed in Annex III, such as credit scoring and hiring, are scheduled for 2 December 2027. Dates have moved before, so check the current text and take legal advice for your case.

Question 18

Who owns the work at the end of an engagement?

Accepted Answer

You do: code, prompts, evaluation sets, pipelines, configuration and any fine-tuned weights, in your own accounts. We document it and plan the handover so your team can run and change it without us.

Question 19

What happens after go-live?

Accepted Answer

Someone has to watch it. We set up monitoring for accuracy, cost and failures, re-run the evaluation set whenever a model, prompt or data source changes, and review edge cases with your team. That can be us on a support arrangement or your own team after handover.

Question 20

What drives the running cost?

Accepted Answer

With a hosted model you pay per token, so cost follows usage, prompt length and model size. With a self-hosted model you pay for GPUs whether or not they're busy, which tends to make sense only at steady, high volume or when data rules require it. We estimate both from your real volumes before you commit.

Tell us what you're working on

You don't need an AI strategy before talking to us.

Questions people ask first