How to Choose an AI Automation Company in India: A Scorecard You Can Check
By Akshit Agrawal, Co-founder, Bitvyas · Published 13 November 2026 · Facts as at 13 November 2026
To choose an AI automation company in India, score each shortlisted vendor on six things you can check in one meeting: a named process, work inside your existing tools, a human approval step, where your data is processed, a measured baseline, and written exit terms. Disclosure: Bitvyas is an AI automation company, so we are one of the vendors you may be comparing.
The market is crowded, and the labels are not reliable. Gartner's June 2025 analysis, as reported by SD Times, estimated that of thousands of agentic AI vendors only about 130 were real; it calls the rest "agent washing", meaning existing chatbots, RPA or assistants renamed as agents. So this post does not ask you to trust a brand. It gives you questions with answers you can verify. It sits in our "buying and running AI" group, next to what AI automation costs and the hub on what to automate with AI first.
What are the key facts to know before you shortlist?
All as of 13 November 2026. We read the regulatory points in secondary summaries; check the primary text before you rely on them.
| Topic | Fact | Source |
|---|---|---|
| Vendor claims | Gartner estimated in June 2025 that only about 130 of thousands of agentic AI vendors are real; the rest are "agent washing". | SD Times report of Gartner, 30 June 2025 |
| Project risk | Gartner predicted that over 40% of agentic AI projects will be cancelled by the end of 2027, citing rising costs, unclear business value and weak risk controls. | SD Times report of Gartner, 30 June 2025 |
| Incident reporting | CERT-In's directions of 28 April 2022 require covered entities to report cyber incidents within 6 hours of noticing them and to keep system logs for 180 days within Indian jurisdiction. | Mondaq summary of the directions |
| India personal data | The DPDP Rules 2025 were notified on 13 November 2025. The security-safeguards and notice rules (Rules 3 and 6, among others) start on 13 May 2027. | SCC Online summary of the Rules |
Why does "who is real" matter more than "who is biggest"?
Because the failure modes Gartner lists, cost, unclear value and weak controls, are all things you can test in a conversation before you sign. A vendor that cannot show a working version of your process, or cannot say what a person approves, is selling a label.
What are the six criteria?
Each criterion has a question, what a good answer sounds like, and what to ask to see.
1. Can they name your process, step by step?
A good vendor describes your process in the order your team does it, and says which steps AI takes over and which stay with a person. Ask them to write it down in one page before any contract. If the answer is "we will find opportunities once we start", the scope is unclear. Our own format is the one used in every post on this blog: a business process, where AI takes it over, what stays with a person.
2. Does it work inside the tools you already run?
Ask which of your systems it connects to, and how: your accounting package, email, CRM, WhatsApp Business, document store. Ask to see one posting or one message arriving in your tool, not in a separate dashboard. A tool your team has to log into separately becomes one more inbox to forget.
3. Where does a person approve?
For anything that moves money, files a return, signs or sends to a client, there should be an approval step with a named role. Ask what the workflow does when it is unsure: a good answer is "it stops and asks", not "it makes its best guess". Ask whether every action is logged with who approved it.
4. Where is your data processed, and who can see it?
Ask in writing: where data is processed and stored, whether it is used to train anyone's models, who at the vendor can access it, and how long it is kept. Ask how they would handle a security incident, given the CERT-In 6-hour reporting window for covered entities. If you hold personal data, ask how they will meet the DPDP safeguards that start on 13 May 2027. You are looking for specific answers, not a reassurance page.
5. Do they measure a baseline before they start?
A vendor who is serious will log your current numbers for a couple of weeks first: hours spent, turnaround time, error rate. Without that, "it saved time" cannot be checked, and payback cannot be worked out. Be wary of a percentage improvement quoted before anyone has seen your process. Ask for the method, not the headline.
6. What happens if you leave?
Ask who owns the workflow definitions, prompts and rules, whether you can export your data and logs, what the notice period is, and what you pay to run it after the build. The ongoing run cost, review time and upkeep are part of the real price; our cost and payback post breaks them out.
How do you turn the six criteria into a score?
Give each vendor 0, 1 or 2 per criterion: 0 for no answer or a vague one, 1 for a written answer you could not yet verify, 2 for a written answer you saw working or could check. Weight the criteria by what your business can least afford to get wrong. The weights below are an example, not a standard.
| Criterion | Example weight | 0 | 1 | 2 |
|---|---|---|---|---|
| Named process | 3 | "We'll discover it" | Generic outline | One-page map of your steps |
| Your tools | 2 | Separate portal only | Connector promised | Seen working in your tool |
| Human approval | 3 | Not described | Described | Shown, with logging |
| Data location | 3 | Marketing page | Written answer | Written, specific, checked |
| Baseline | 2 | None | Offered | Method and sample shown |
| Exit terms | 2 | Not discussed | Stated | In the contract draft |
Multiply, add up, and compare. A vendor that scores 0 on human approval or data location should not go further, whatever the total.
What should you ask for in the first meeting?
- A one-page map of one of your processes.
- A live or recorded example in a tool like yours.
- The data answers, in writing.
- A named person who will review the workflow weekly on your side and theirs.
- A pilot limited to one process and a fixed period, with success defined before it starts.
What stays with you, however good the vendor is?
You choose the process, you own the rules, and a named person on your team approves what matters. A vendor can build and run the workflow, but it cannot take responsibility for your decisions. Gartner's cancellation forecast is about projects with unclear value and weak controls; the two fixes are a narrow first process and an owner. If you are unsure where to start, the hub post helps you pick. For a concrete example of a mapped process, see operations triage with AI or sales follow-up with AI.
If you would like Bitvyas to score itself against these six, get in touch and we will answer in writing. We would rather you compared us with others. See also the operations and reconciliation use cases.
Sources
- SD Times, Gartner: More than 40% of agentic AI projects will be canceled in the next few years, 30 June 2025 (reporting Gartner's 25 June 2025 release; Gartner's own page could not be fetched) — https://sdtimes.com/agentic-ai/gartner-more-than-40-of-agentic-ai-projects-will-be-canceled-in-next-few-years/
- Mondaq, 6 hours to report cyber incidents to CERT-In and more compliances (summary of CERT-In directions of 28 April 2022; read the directions on cert-in.org.in) — https://www.mondaq.com/india/it-and-internet/1270362/6-hours-to-report-cyber-incidents-to-cert-in-and-more-compliances
- SCC Online, MeitY notified Digital Personal Data Protection Rules 2025, 14 November 2025 (summary; check the Gazette) — https://www.scconline.com/blog/post/2025/11/14/meity-notified-digital-personal-data-protection-rules-2025/
Related
- What to automate with AI first: 8 business processes mapped
- What AI automation costs, and how to work out payback
- Operations triage with AI
- Sales follow-up with AI
- Operations use case
Figures as of 13 November 2026. Bitvyas is a vendor of AI automation. 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