AI Integration & Automation

Practical AI, built into how you already work.

Practical AI sits in software you already run

AI integration, here, means putting a narrow capability into an application your staff already open: draft a reply on an enquiry, extract fields from a supplier PDF, answer a policy question from files you actually keep, classify a ticket so a person sees the right queue. It does not mean a new chat site with a new password that nobody opens twice. If the useful button is not on the screen where the work already happens, the project is not finished.

LavisTech is a software shop in India. We add these features to line-of-business systems — often ASP.NET — and we refuse work that is only a prompt pack with no home. The model is a service you call. Auth, roles, logging, and a person who still owns the output are the product. For the wider picture of what pays off, start with AI for business. This page is how we run the build when you already have an application, or when you are about to admit you do not.

Person in the loop is not a slogan

A person in the loop means a named role must see the draft, the extract, or the suggested next step before anything consequential is sent, posted, or paid. Consequential means a customer sees it, money moves, stock moves, or a record that an auditor might ask about is written. Until those cases are boringly correct, the system creates drafts. It does not “just send it”.

That is slower to demo than autonomy. It is why the system is still on in month four. We implement the click in the same permission model the screen already uses. A warehouse login does not approve a credit. A model call does not bypass that. If a vendor wants the Tally password so the agent can “sort accounts”, that is not a design. That is a junior hire with your books and no supervision.

Read-only first is the default. A summary on a case file, extracted lines on an invoice form, a cited answer under a question box. Writes that create a draft come second. Writes that complete the job without a click come last, if they come at all, and only for a path you can describe without waving.

RAG is not a chatbot, and a chatbot is not an agent

People use these words as synonyms. They are not. A chatbot answers and stops. Retrieval-augmented generation — RAG — is how you stop the model inventing your policy: search your current files, hand the passages over, answer from those, show the source. An agent is a loop with tools: look something up, decide, write back. The comparison that matters is in AI agents versus chatbots. Buy the verb in your sentence, not the word on the slide.

“Tell me the notice period in the current handbook” is RAG plus a chatbot skin. “Draft a reply to this delivery complaint from the order record” is generation on supplied facts, still with a send click. “Post the invoice and email the supplier if it matches the PO” is agent-shaped, and we will still start with extract-and-draft. If the steps never vary and the input is already a form, you may not need a model. You need a workflow. That distinction is the whole of AI versus automation.

When a workflow is enough

If you can draw the path without a box that says “interpret this email”, do not pay for interpretation. Leave request under the balance, notify the manager, escalate after two days — that is automation. Invoice arrives as a structured XML from a supplier you already know — that may be a map, not a model. We would rather you keep the money and the complexity. Models earn their keep on messy input: scans, free text, the PDF with a stamp over the total.

A good test: after you write the happy path, write the three exceptions that already happen every week. If those exceptions are still rules — manager away, amount over a limit, attachment missing — you are in workflow. If the exception is “we cannot tell what the customer is asking until a person reads it”, then extraction or classification in front of the same workflow is the AI slice. Do not let a vendor roll those into one “platform” so you cannot see which part you bought.

What we will not build for its own sake

We will not put a chat widget on a brochure site and call it transformation. We will not fine-tune a model on every email since 2016 so it can “know the business”; that is how you cite the wrong year. We will not connect a shared drive with no owner and no permissions and hope retrieval respects the MD’s salary file. We will not start an agent that emails customers in week one. We will not replace Tally because a demo wrote a haiku about GST.

If the application cannot accept a button — no API, no way to add a screen, a desktop package you cannot query — you are in legacy and cloud modernisation first. That is not a sales upsell. It is the same sequence we use when adding AI to an application you already run. Paying integration rates to wrap a system you cannot reach is how pilots die in week two, when the notebook demo meets production logins.

A first project that can finish

Pick one task. High volume helps. A wrong answer that is cheap to catch helps more. Invoice fields into a draft bill, first-line replies to “where is my order”, staff questions against a curated handbook, routing of inbound mail into two or three queues. Write the definition of done before anyone chooses a model. “Staff like the chat” is not a definition. “Hours in that mailbox fell, and the exception rate is visible” is.

We ask four things in the first meeting. Where do the facts come from on Tuesday if the file changed on Monday? What must never happen without a person? Which number will we refuse to move after thirty days? Who on your side owns the exceptions? Vague answers mean you are buying a workshop. We will sell a short paid discovery. We will not sell a six-month platform that discovers the same things in week two.

Bring one ugly document and one real question. The scan, the stamp, the second page that is only a stamp. Demos use clean PDFs. Your mailbox does not. That file is the project more than the model card.

What the weeks look like

Scope	Typical build	Shape
Single narrow task	3 to 6 weeks	Extract or draft into one system you already run
Assistant over curated files	6 to 10 weeks	RAG, citations, an owner of the index
One tool, still read-only or draft-write	Adds weeks after the assistant is trusted	Lookup an order, fill a form, stop
Multi-step agent with writes	3 to 6 months	Permissions, caps, logs, reversibility — not a smarter model

Two costs sit outside the table. Running cost: usage is charged per call. A drafting feature on tens of documents a day is usually small. An unbounded loop the whole company can start is visible. Evaluation: deciding whether the output is right often takes as long as the feature. Budget a person to sample. Unowned evaluation is how you argue forever about “quality” with no number.

We do not invent a rupee total for “AI for SMEs”. The model bill is rarely the surprise. The surprise is dirty files, missing APIs, and a process three people perform three ways. Those belong in discovery, said plainly, not in a surprise change request in week five.

Grounding, permissions and Monday

Grounded means the model was handed this invoice, this policy, these three notes, and told to use only that. It does not mean “we connected the drive”. A dump since 2016 is how you retrieve the superseded rate card. Curation is the work. If the handbook changes on Monday, the old file must leave the index on Monday. Someone has that job after we leave. If you cannot name them, the citations will rot and staff will stop asking.

Permissions are the same as the screens. If a clerk cannot open a folder, retrieval must not return it. We implement tools as ordinary methods with the authorisation you already have. The model can only call what that user could click. Conservative, and the reason the first wrong suggestion is an embarrassment rather than a posted bill.

Tally, Excel, WhatsApp — only where they are true

Many of the firms we see still post in Tally and chase exceptions on WhatsApp. Excel is the real stock list more often than anyone writes down. AI does not replace those on day one. An extract can draft a bill your accounts package already understands. A classifier can turn a forwarded WhatsApp export into a ticket — if you actually export, and if someone owns the queue. A model cannot reconcile a book it cannot see. If Tally is closed, locked, or only available as a printout, say that. Access is the first sprint, not the demo.

GST lines are a common failure. Treatment is a rule, not a vibe. We validate extracts against a schema and leave doubtful tax fields for a person. We will not let a model invent a HSN. That is not caution for its own sake. It is how you stay out of a notice you will not enjoy.

What failure looks like, so you can avoid buying it

Scope was a category: “use AI in customer service”. Nothing integrated: a clever answer someone still copies. No measure: the debate never ends. Nobody owned exceptions: the feature was switched off after the first odd draft. Data worse than admitted: week two, every time. We would rather lose a pitch than start that project with your name on it.

The path that usually works is dull. Weeks one to eight: one task, read-only or draft, real users, a log. Then one read tool if the log says people always go and look something up. Then one draft write. Autonomy is a later purchase you may never need. If a proposal cannot describe those first eight weeks without write-access, they have not priced permissions. They have priced a word.

How you know it worked

Pick the number before the build. Hours keying invoices. Time to first reply on a routine ticket. Times someone asks the same person for the same policy. After a month, it moved or it did not. We will look at the log with you. We will not accept “the board liked the demo” as a pass.

If you are still deciding whether any of this is your problem, read the business guide linked above, then the pages on RAG and on agents versus chatbots. If you already know the task and the screen it should live on, send that sentence and one ugly file through contact. We will tell you whether it is a three-to-six-week slice, a RAG assistant, a workflow with no model, or modernisation first. That answer, in writing, is the start. A platform tour is not.