Cases outside the agreed conditions
An automation works only under the conditions it was built for: which cases, arriving in what shape, from which systems. We name those conditions with you, and every case outside them goes to a person.
AI automation services in India
We measure your manual process first, then pilot on a real sample of your cases to find the share that still needs a person. Anything the system isn't sure about goes to a person, at a threshold you set.
What stays with a person
An automation works only under the conditions it was built for: which cases, arriving in what shape, from which systems. We name those conditions with you, and every case outside them goes to a person.
Everything we build returns a confidence signal for each item: how sure the system is. Items below a cut-off, the threshold, go to a person, and you set the threshold, not us.
Some rules settle this before any model exists. The Reserve Bank of India (RBI) lets a bank use AI in video-based customer identification (V-CIP), yet only specially trained bank officials may operate it.
The number you don't have yet
Nothing in your current process shows that share, because no system has tried your cases yet.
ISO/IEC 22989:2022, the international standard for AI concepts and terms, defines automation in clause 3.1.7 as working without human intervention "under specified conditions". You can read its definitions on ISO's Online Browsing Platform. So an automation percentage means little unless it says which cases it covers.
Your systems' record of each step is called an event log. The Process Mining Manifesto, from the IEEE Task Force on Process Mining, warns that a log holds only a sample of behavior. It adds that in lower-quality logs people can go around the system, and that outliers may be cleaned away as noise. Our conclusion, not the Manifesto's: a log also misses cases that haven't arrived yet, so the exceptions in your past records are the least to expect, not the real number.
Wrong answers are tested by our security team under generative AI security testing, part of AI security. Wiring the hand-off to a person into your ERP or CRM is on AI automation and integration, and the dependencies a hosted model adds to software you already run are on AI integration. Or see all AI development.
How we work
Before any case skips a person, the pilot shows you how many still need one.
We agree with you which cases the system should handle.
Volume per week, time per item and error rate, starting from your own records.
We run the system on real cases and measure two things: accuracy, and the share that still needs a person.
At first the threshold sits high, so almost everything routes to a person.
You decide when it comes down. Where the work has real consequences, a person approves each item until the share they approve is high enough to relax that check, as AI agent development explains.
The evidence behind the method
Quotes are each publisher's own. None of it describes SecWiz's work.
The Process Mining Manifesto asks analysts to work under an "open world assumption": "the fact that something did not happen does not mean that it cannot happen".
The BPI Challenge 2019 page from the International Conference on Process Mining (ICPM) describes a real purchase-to-pay event log: 1,595,923 events across 42 activities, performed by 627 users, of whom 607 are human and 20 are batch users. Sometimes no user was recorded. One unblocking step can be done "by a user, or by a batch process at regular intervals". The data holds roughly four types of flow for line items, and the page warns the complexity "goes further". The case ID combines the purchase document and the purchase item.
Google's Rules of Machine Learning (last updated 25 August 2025) open with "Rule #1: Don’t be afraid to launch a product without machine learning." Rule #2: design and implement metrics first, tracking as much as you can in your current system. Rule #3: a simple heuristic gets a product out the door, a complex one is unmaintainable, and machine learning comes once you have data.
NPCI, set up by the RBI and the Indian Banks’ Association to run retail payments and settlement in India, publishes a UPI Deemed Approved percentage: the share of transactions where credit confirmation from the beneficiary bank is not received online. NPCI adds that if the beneficiary account is not credited online, the beneficiary bank processes the transaction manually under RBI guidelines. For August 2026, NPCI's Top 50 Member Performance shows 0.00% to 0.76% across the top fifty beneficiary banks.
Each rule binds only the entity it names.
Digital Payments – E-mandate Framework, 2026 (RBI/DPSS/2026-27/396, April 21, 2026), for payment system providers and participants processing recurring transactions, domestic or cross-border, on cards, PPI or UPI. It repeals the earlier e-mandate circulars.
Commercial Banks – Know Your Customer Directions, 2025 (RBI/DOR/2025-26/169, November 28, 2025, updated December 29, 2025), for commercial banks. Paragraph 27 applies only where a bank opts to undertake V-CIP:
Paragraph 23, second proviso, applies to every commercial bank, but only for the e-KYC exceptions it describes: where e-KYC fails for a benefit or subsidy claimant under section 7 of the Aadhaar Act, 2016, through injury, illness or old age, an official carries out due diligence using offline verification or another officially valid document. The case joins the concurrent audit (paragraphs 12 and 13) and a centralised exception database.
RBI/2019-20/67, "Harmonisation of Turn Around Time (TAT) and customer compensation for failed transactions using authorised Payment Systems" (September 20, 2019, in effect October 15, 2019), for banks and other operators and system participants.
Not on their own. The Process Mining Manifesto, by the IEEE Task Force on Process Mining, says event logs contain only sample behavior and "should not be assumed to be complete". A record of past cases shows what did happen. It cannot show a kind of case that has not arrived yet, or a step people took around the system. Your records are still where the manual measurements start: volume, time per item and mistakes. The share that stays with a person comes from piloting on a real sample, as the steps above describe.
They route to a person. Which cases those are depends on a threshold you set. Keeping a person approving items one at a time, and deciding when that can stop, is covered on AI agent development.
Not by Google's account. Its Rules of Machine Learning tell readers "Don’t be afraid to use human editing either." Google advises launching without machine learning until there is data, and names human editing as acceptable. That is Google's advice about building with machine learning in general. It says nothing about which steps in your own process need a model.
Then the share is harder to measure, and the Process Mining Manifesto explains why. It lists incomplete event data among the hurdles, including events that do not point to the case they belong to and events with no time attached. At the lowest of its event log levels, recorded events "may not correspond to reality", and the Manifesto says logs "for which events are recorded by hand typically have such characteristics". A process with no record of its cases cannot have its share measured from history, so the numbers have to come from measuring the manual work and a pilot.
Yes, by our security team rather than the people who built the system. That work is described on generative AI security testing.
This page is about how many of your cases a system can close on its own. How much authority a system is given is covered on AI agent development.
Let's talk
Tell us what the process is and how it runs today. Project work is delivered remotely from India during business hours. We reply within one working day.