Name the step somebody does by hand today, or there is nothing here worth buying.
This page is about one narrow thing. A person in your office repeats a step every day that a machine could do. Opening the vendor bill and typing six numbers into Tally. Reading an enquiry and deciding which salesman gets it. Pulling a delivery address out of a WhatsApp message.
That step is the unit of work. Not a strategy, not a platform. If you cannot point at the person and the step, nothing we set up will be measurable afterwards, and you will not know whether it worked.
AI automation means dropping a model into a process that already runs, so it can do one step that needed a person because the input was messy. The bill arrives in a different layout every month. The enquiry is a paragraph of free text. A rule cannot read either, and a model can, roughly and quickly.
Where it earns its place.
The kind of step this takes away
Each of these is a job somebody in an Indian office does now, on a screen, between other work.
- Reading a vendor invoice or a delivery challan and pulling out the fields
- Sorting incoming email or enquiries by what they are actually about
- Turning a long WhatsApp thread into an order with an address
- Drafting the reply a person edits and sends
Where a plain rule is the better answer
This is the part vendors skip. A model is slower than a rule and costs money on every call. It also gives a slightly different answer on Tuesday. Where a rule fits, use the rule.
- When payment is received, send the receipt: a rule, not a model
- When stock falls below the reorder level, raise a request: a rule
- When the input is free text, a photo or a scan that keeps changing shape: a model
- When the decision needs judgement you can describe but not write as conditions: a model
Most of what we are asked to automate turns out to be four rules and one model. That proportion is worth knowing before anybody signs anything.
The queue in the middle
Nothing goes straight out at the start. The model produces a draft and a person approves it, which is both the safety net and the way you find out how good it actually is.
- Every result lands in a queue with the original beside it
- Corrections are counted, and the rate either falls or it does not
- Confident cases release automatically once the numbers earn it
How we set it up
- Pick the step
- Collect real samples
- Prototype
- Approval queue
- Measure
- Widen
- Choosing one step, with the person who does it in the room
- Running it against your own documents, including the ugly ones
- Counting minutes and corrections before and after
- Wiring the result into the system that already holds the record
How it goes wrong
Worth reading before the demo, because the demo always works.
- A vendor changes their invoice layout and accuracy falls quietly
- The model is confidently wrong and nobody is checking any more
- The running cost per document was never compared with the minutes saved
- The step was automated for a process that should have been removed
The fix for most of these is the same: keep counting corrections after go-live, not only during the pilot.
When this is the right choice.
- A step happening many times a day on messy input: documents, email, free text, photos
- A queue of work that piles up and is always a day behind
- Input that arrives in shapes you cannot control, from vendors or customers
- A person whose job has a boring half and a skilled half, where only the boring half goes
When it is not.
- A step that follows a clear rule, where the rule is cheaper to run, easier to test and does not drift
- A process nobody has written down, because automating it means settling arguments no one has had
- Work done a few times a month, where you spend more setting it up than it will ever return
- Work where a mistake reaches a customer or the ledger before anybody has seen it
Questions we are asked about it.
How do we know if a step is worth automating?
Count how many times a day it happens and how long it takes. Then look at the input. If it arrives in a predictable shape, a rule will do it for less. If it is a scan, a paragraph or a photo, this is where a model earns its keep.
Will it make mistakes?
Yes, which is why the first version puts everything in front of a person for approval. What matters is the rate, whether it is falling, and whether a mistake is caught before it reaches a customer or an account. We show you the corrections, not a demo.
Does our data go to OpenAI or another company?
That depends on which model we use and it is your decision, taken before anything is built. A hosted model is cheaper and easier, and a model on your own server keeps the documents in. We put both options in writing.
What does it cost to keep running?
There is a per-document charge from whichever model is used, on top of hosting. The honest test is whether that charge is smaller than the minutes it removes, and we work that out during the pilot rather than after.
Can it work with Tally and the software we already run?
Often, and it turns on whether your software will let anything in. Tally needs a connector rather than a direct call. Where a system is shut tight, the result gets imported instead.
Is this the same as an AI agent?
No. Here the path is fixed and the model does one step inside it, which makes it predictable and easy to test. An agent chooses its own steps and acts on your systems, which is more useful in a few cases and riskier in most.
Not sure AI automation is the right choice?
Tell us what the software has to do and who opens it. If something else fits better, we will say so, and say why.
- No obligation
- We reply the same working day
- Your details stay private