03 Delivery
How we build one
Five steps, and the first is allowed to end the project by concluding that rules would do the job better.
Step 1: Check the field is not already there
If the routing input exists as a field somewhere, rules are cheaper, faster and need no review lane.
Week 1Step 2: Pilot the classification
100 to 300 real items with the correct routing recorded by your team, and accuracy reported per category rather than as one average.
Weeks 1 to 3Step 3: Set the lane boundaries
Per category, from the cost of a misroute. A cheap-error category gets a wide straight-through band; an expensive one may have none.
Week 4Step 4: Build the exception desk and fallback
Ownership, a service level, escalation, and a rule that routes everything to people if the provider is unavailable.
Weeks 4 to 8Step 5: Close the correction loop
Every reassignment recorded and fed back into the evaluation set, so accuracy stays true and problem categories surface.
Ongoing
- Every routing decision records what the model was shown, what it concluded, its confidence, the lane it took and who acted, which is what makes a disputed route answerable later.
- Where the workflow writes into the books, the proviso to Rule 3(1) of the Companies (Accounts) Rules, 2014, administered by the Ministry of Corporate Affairs, has required a non-disableable audit trail since 1 April 2023.
- Where items carry personal data, the Digital Personal Data Protection Rules, 2025, notified on 13 November 2025 by the Ministry of Electronics and Information Technology, apply, with the main obligations commencing on 13 May 2027.
04 Honesty
The line we hold on decisions
A model can prepare a decision. It should not make one that somebody has to answer for. Three cases where the human stays in the seat rather than in the loop.
- An approval. Extracting the value, checking it against policy and routing to the right approver is appropriate. Approving it is not, because an accountable person has to own that outcome.
- A rejection that affects somebody. A declined claim, a rejected application or a refused refund reaches a person before it reaches the customer. The model may recommend and must not conclude.
- An escalation nobody can appeal. Any lane with no human touch anywhere needs a sampling review, because a category the model quietly gets wrong at scale is worse than one it gets wrong loudly.
Where this goes wrong in practice
A straight-through lane set too wide during a quiet month, then left in place when the document mix changed. It is why the lane mix and the misroute rate are on the dashboard from day one.
05 Reference
Terms used on this page
- Confidence-based routing
- Sending an item down one of three lanes according to model confidence: automatic, assisted with a suggestion attached, or fully human.
- Exception desk
- The staffed queue receiving items the model could not route confidently, with ownership, a service level and escalation.
- Straight-through rate
- Written STP rate: the share of items routed with no human involvement.
- Misroute rate
- The share of items that reached the wrong owner and had to be reassigned, measured after go-live rather than assumed from the pilot.
- Correction loop
- Feeding every reassignment back into the evaluation set, so accuracy stays measured and problem categories become visible.
- Anchoring
- The tendency to accept a confident pre-filled suggestion. The reason the fully human lane deliberately presents nothing.
- Data Protection Board of India
- The adjudicating body under the Digital Personal Data Protection Act, 2023. The Schedule to that Act sets fixed rupee ceilings rather than a share of turnover: up to ₹250 crore for failing to take reasonable security safeguards against a personal data breach, up to ₹200 crore for failing to notify the Board and the people affected, and up to ₹50 crore for a breach of any other provision. That is why consent, retention and access control are build decisions here rather than paperwork.
06 Questions
AI workflow automation in Banaras: FAQs
By the number of queues, the daily volume in each, and how many lanes the routing needs, none of which changes with the city. One queue with classification, extraction and confidence routing is the starting unit for a team in Banaras. Queues added after the first cost less each. Every quote states IncorpX professional charges; model usage is billed by the provider at actuals, and any government fees are billed separately at actuals.
5 to 9 weeks including the pilot. The pilot measures classification accuracy on your own items and produces the number that decides whether this is worth building at all.
No. Pilot sampling, lane design, integration, reviewer training and monitoring all run over calls and screen shares from our base in Chennai, Tamil Nadu.
When the routing rule depends on something nobody has typed yet. If the value or category already exists as a field, rules are cheaper and more reliable. If it is inside a PDF or a paragraph of free text, the item has to be read first.
Three lanes instead of two: high confidence straight through, medium confidence to a person with the model's suggestion pre-filled, and low confidence to a person with nothing pre-filled. The middle lane usually carries the most volume and most of the saving.
Because a confident but wrong pre-fill gets accepted more often than it would have been chosen. For categories where an error is expensive we remove the anchor rather than trusting people to resist it.
No, and that is the point of the lanes. A model at 88% accuracy with a good confidence signal produces a process that is right almost always, because what it is unsure about never routes itself.
Everything routes to the human lane. That fallback is built before the happy path, because a queue that halts on a third-party outage is a problem you introduced yourself.
It can prepare one. Extracting the value, checking it against policy and routing to the right approver is appropriate; deciding is not, because an accountable person has to own that outcome.
Every routing decision records what the model saw, concluded, how confident it was, the lane and who acted. Where the workflow writes into the books the Rule 3(1) audit trail obligation applies, in force since 1 April 2023. Where the items carry personal data, the Digital Personal Data Protection Rules, 2025 apply, with the main obligations commencing on 13 May 2027.
Next step
Name the queue somebody reads all day just to sort it.
We will pilot the classification on your own items and come back with the accuracy per category, the share that could route straight through, and a fixed quote.
Read by Dhanush Prabha, our CTO, not a form queue. Reply usually within one working day, in your time zone.
Nebin Binoy 
