Operations Triage With AI — What It Routes, What Your Team Still Decides
By Akshit Agrawal, Co-founder, Bitvyas · Published 10 November 2026 · Facts as at 10 November 2026
At Bitvyas, an AI operations-triage workflow reads every incoming request, tags it, sends it to the right owner, answers the routine ones from your own documents and warns before a deadline is missed. Your team keeps priority calls between departments and anything unusual. It is process 6 of the eight in our hub post.
In many companies, requests still arrive by email, WhatsApp, a form and a word at someone's desk, and the person who asks most often sets the priority. This post maps the process step by step: where AI takes over, and what stays with a person. It is one of the eight processes in our hub post on what to automate with AI first. In that scoring, operations triage scores 6 of 8. The volume is high and the rules are clear, but data scores low until phone and in-person requests are also logged.
What are the key facts for automating operations triage?
Three points to settle before you start. All are as of 10 November 2026.
| Topic | Fact | Source |
|---|---|---|
| Time lost to inboxes | McKinsey Global Institute's 2012 report estimated that the average "interaction worker" (managers and professionals) spends 28% of the workweek managing email. It is a 2012 estimate; measure your own team's time before you rely on it. | McKinsey Global Institute, The social economy, July 2012 |
| Project risk | Gartner predicted in June 2025 that over 40% of agentic AI projects will be cancelled by the end of 2027, citing escalating costs, unclear business value or inadequate risk controls. A narrow first workflow with a clear owner is the answer to all three. | Gartner press release, 25 June 2025 |
| India personal data | Requests often contain names, phone numbers and account details. Under the DPDP Rules 2025 (published 13 November 2025), most data-fiduciary duties commence on 13 May 2027. We checked this in a secondary summary; read the Gazette notification before you plan around it. | Secondary summary of G.S.R. 843(E) |
What does request handling look like when it is done by hand?
Someone reads the shared inbox between other jobs. A WhatsApp message asks for a gate pass. A form asks for a vendor-payment status. A manager mentions a broken printer in the corridor. Each request lives in a different place, and nobody has one view of what is open.
Three patterns repeat:
- Priority follows volume of noise. The person who chases hardest gets served first, not the request that matters most.
- Routing depends on memory. One colleague knows that invoice queries go to finance. When they are away, the request sits.
- Deadlines surface late. A request that has a two-day service window is noticed on day three.
This is a sorting and remembering problem, which software does well.
What can AI do on its own?
The repeatable middle: reading, tagging, routing, answering, reminding and recording.
- Collect everything in one queue. Email, forms, WhatsApp Business messages and chat land in one list, whatever time they arrive.
- Read and tag. For each request it identifies the type (finance query, facilities, IT, HR, vendor), the department and any deadline or reference number mentioned.
- Route to the owner. Your routing table decides who gets it. If the owner is on leave, it goes to the named backup. If nothing in the table fits, it goes to a person, not to a guess.
- Answer the routine ones. Where the question is covered by your own documents (leave policy, expense limits, how to raise a purchase order, office hours), it replies with the answer and the source document. If the document does not cover it, it routes the request instead.
- Ask for what is missing. "Which floor?" or "Please attach the invoice" goes back automatically, so the owner receives a complete request.
- Track the clock. Each request type has a target time. The workflow warns the owner before it is breached and tells the manager when it is.
- Spot duplicates and patterns. Five requests about the same fault are merged into one. A rising number of the same query is reported.
- Close the loop. It tells the requester when the work is done and records what happened.
A sample routing table
Illustrative only. Every business writes its own.
| Request type | Routed to | Target response | AI answers directly? |
|---|---|---|---|
| Expense or reimbursement query | Finance operations | 1 working day | Yes, from the policy |
| Vendor payment status | Accounts payable | 1 working day | Yes, from the ledger status |
| Office or facilities fault | Facilities lead | Same day | No, routes with details |
| Access or equipment request | IT | 1 working day | Partly, form and checklist |
| Leave and policy question | HR | 2 working days | Yes, from the policy |
| Complaint, legal wording or an upset tone | Named manager | Same day | Never |
What does your team still own?
Anything that needs judgement or authority.
- Priority calls between teams. When finance and operations both need the same person this week, a manager decides.
- Anything unusual. A request that matches no rule goes to a person, with the thread attached.
- Exceptions and approvals. Spending, access to sensitive systems and policy exceptions.
- The routing table itself. The rules are yours. AI applies them but should not invent them.
- People matters. Complaints, grievances and sensitive requests are never auto-answered.
What should you check before it starts answering for you?
Where the answers come from, and what personal data passes through. Set both up before the first reply goes out.
Source documents. Give the workflow only approved, current documents. An old policy PDF will produce an old answer. Assign an owner for each document and review a weekly sample of replies.
Personal data. A request that mentions an employee's health, salary or a complaint should be routed to a named person and not summarised into a wider channel. In India, the DPDP Rules phase in over 2026 and 2027 (see the table above). In the UK and EU, the usual data protection duties already apply. Keep access to the queue limited to the people who need it, and keep a record of who saw what.
Where it runs. Ask where the data is processed and stored, and whether it is used to train anyone's models. Get the answer in writing.
How do you measure whether it is working?
Take a baseline before you start, from a two-week log:
- Median time from request to first owner action.
- Share of requests that reached the right owner first time.
- Share of requests closed inside their target time.
- Requests answered without human effort, sampled and checked for correctness.
- Hours the team spends sorting and forwarding, before and after.
We do not publish client results without permission, so we do not quote a percentage improvement here. The baseline is how you see your own.
Where should you start?
Start with one inbox and three to five request types. Write the routing table first. If your managers cannot agree who owns a request type, fix that before automating anything. Log phone and in-person requests too, or the data will always be incomplete. Keep a person reviewing a weekly sample of routed and answered requests, and add WhatsApp and more request types once the team trusts the first queue.
If operations is not your most pressing process, the eight-process hub helps you choose. A related inbound process is covered in sales follow-up with AI, and requests from clients in client document collection with AI. Costs and payback are in what AI automation costs. Our operations use case shows how Bitvyas sets this up with the inbox, forms and tools you already use.
Sources
- McKinsey Global Institute, The social economy: Unlocking value and productivity through social technologies, July 2012 — https://www.mckinsey.com/industries/technology-media-and-telecommunications/our-insights/the-social-economy
- Gartner, Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027, 25 June 2025 — https://www.gartner.com/en/newsroom/press-releases/2025-06-25-gartner-predicts-over-40-percent-of-agentic-ai-projects-will-be-canceled-by-end-of-2027
- DPDP Rules, 2025, G.S.R. 843(E), 13 November 2025 (commencement dates read in a secondary summary; check the Gazette) — https://dpdp.myndsolution.com/wiki/guides/key-dates-and-commencement/index.md
Related
- What to automate with AI first: 8 business processes mapped
- Sales follow-up with AI
- Client document collection with AI
- What AI automation costs, and how to work out payback
- Operations use case
Figures as of 10 November 2026. General information, not legal advice.
See it working
Keep reading
Still doing this by hand?
Bitvyas builds AI that takes this kind of work off your team. It reads the documents, does the matching and drafts the next step, inside the tools you already use. A person still approves anything that matters.
Tell us what your team repeats