Most writing about AI in business software falls into one of two camps: it will run your company, or it is a gimmick. Neither matches what we see building an ERP that small and mid-sized Indian businesses use every day. AI is genuinely useful for a specific kind of work, and genuinely risky for another kind, and the line between them is easier to draw than the hype suggests.
Here is where AI earns its place inside an operations ERP today, where it doesn't yet, and what to ask any vendor (including us) before you trust it.
A simple test: is the AI reading, or deciding?
Almost every useful AI feature we have built does the same thing: it turns unstructured text into structured data, and then a person or a fixed rule decides what happens next. An email becomes a lead with a company name, contact and enquiry type. A 60-page tender PDF becomes a checklist of eligibility clauses. A vendor invoice becomes a draft bill.
The features we would not build, or have deliberately pulled back from, are the ones where the model's output is the decision itself, especially when the decision moves money, creates a legal record or is hard to undo.
| Task | What AI does well | Who signs off |
|---|---|---|
| Inbound email | Extracts fields, classifies the enquiry | Review queue for low-confidence mail |
| Tender documents | Pulls out eligibility clauses, flags risks | Bid team marks each clause, a person files |
| Vendor invoices | Reads the PDF into bill fields | Bill lands as a draft for approval |
| Letters and proposals | Drafts body text | Always a draft; your team sends it |
| GST filing | Nothing | Fixed computation, your CA, a person marks filed |
Where AI earns its place
1. Triaging the shared inbox
For many Indian SMBs the real CRM is a shared inbox: enquiries, RFQs, support requests, vendor quotes and payment advice all arrive at the same address. Sorting it is repetitive work that suits a language model.
In DexVee, inbound email is read by an AI parser that pulls out the company, contact details, GSTIN and the kind of enquiry (wireless, networking, cabling, support and so on), and rates its own confidence as high, medium or low. What matters more is what happens at the edges. An email with low confidence, or one where the system can't identify a company or a contact email, is routed to a draft review queue rather than creating a lead for "Unknown". Mail from a known supplier quoting to you is kept out of the sales pipeline even when it looks lead-shaped.
The same thinking applies to helpdesk tickets. Each new ticket is classified into one of seven buckets: genuine support, a sales enquiry, an RFQ, a renewal, a product question, an upgrade or a continuing deal. The result is shown as a suggestion on the ticket ("This appears to be a sales opportunity: convert to deal?") with a one-line reason. The classifier is told to default to support when unsure, and classification runs in the background, so a slow or failed AI call never blocks a ticket from being created. A person clicks convert, or doesn't. See how this fits into helpdesk and support.
2. Keeping the CRM clean at the door
Most CRM data-quality problems start at intake: a customer replies from a new address without the deal number, and someone creates a second lead. DexVee's email matcher handles exactly that case. When an email carries no deal or ticket number and doesn't belong to a known thread, the model is shown the email and a short list of that customer's recent and open deals and tickets, and asked whether it continues one of them or is a new matter.
Three design choices make this safe enough to automate. The model can only pick from the candidate list it is given, and is told to treat the email text as untrusted data and ignore any instructions inside it. A match needs a minimum confidence before it is filed. And every decision is stored so staff can confirm or correct it; those corrections are fed back to the model as ground truth for later emails from the same sender.
It is unglamorous, and it pays for itself quietly: fewer duplicate leads and a cleaner history on each account.
3. Reading tender documents
Government and PSU tenders are long, repetitive and full of clauses that can disqualify you: turnover thresholds, OEM authorisations, certifications, past-experience requirements. Too often they are found two days before the deadline.
In DexVee, AI extracts every eligibility requirement from the tender text into a checklist, with a category (financial, technical, experience, compliance and so on), a mandatory flag based on the tender's own wording, and the source clause reference. Each requirement starts as pending until someone marks it eligible, not eligible or waived, and the tender's overall eligibility rolls up from those human answers.
A separate AI analysis scores the tender on value, timeline and category fit, lists risks with mitigations and produces a pass, fail or warn compliance view. It also estimates a win probability, and this is where restraint matters. On the daily Bid Manager board that estimate is folded in at half weight to just one of six dimensions, win likelihood, which carries 15% of the total score. The code comment puts the reason plainly: the AI figure is an estimate from the tender text, whereas competitor records are fact. Filing itself stays with your team on the portal, where a digital signature is needed. More on the workflow in tenders and bid management.
4. Turning paperwork into records
Vendor invoices, customer purchase orders and bank statements arrive as PDFs in a dozen layouts. Reading them into the right fields is a strong use of AI, provided the output is treated as a draft.
DexVee reads a vendor invoice PDF into bill fields (vendor, invoice number, dates, place of supply, tax registration, line items) and creates the bill as a draft, which a person approves before anything is paid. Bank statement PDFs are parsed into transaction rows; the matching of those rows to invoices and bills is then done by deterministic rules and shown as suggestions in a review queue. As the code puts it, no category is better than a wrong one.
5. Drafting, never sending
Drafting is the most visible AI feature and the easiest to overrate. DexVee can draft the body of a formal letter on your letterhead from a short prompt, using the company and bank details on file, and is told not to invent bank details it hasn't been given. It can suggest terms-and-conditions clauses that look missing or contradictory for a document, as suggestions only. It can summarise a lead's history before a call.
When an inbound enquiry is complete enough, DexVee can create a proposal automatically, but it is always saved as a draft for someone to review and adjust. The mail-ingest pipeline does not send outbound email at all. Nothing an AI drafts reaches a customer without a person deciding it should.
Where AI doesn't earn its place yet
Autonomous financial decisions
Approving a bill, allocating a receipt across invoices, deciding which vendor to pay first: these decisions are hard to reverse and depend on context a model doesn't see, such as a dispute that never reached email. A model that is right most of the time is still wrong sometimes, and in accounts payable "sometimes" means money sent to the wrong place. We keep these decisions with people and with rules you can read.
Statutory filing
GST returns need to be exact, not plausible. DexVee builds monthly GST filing packs by fixed computation from your invoices and raises an in-app alert; you download the pack and send it to your CA yourself. After filing, a person clicks "Mark filed" so the next pack skips those invoices. There is no AI step in that chain, and we think that is correct today.
Judgements a rubric already captures
Sometimes the honest answer is that you don't need a model at all. DexVee once scored sales visit notes with an AI prompt. That scoring now runs on deterministic signal detection that mirrors the original rubric, which is cheaper, faster and gives the same answer every time. AI is kept for the part that genuinely needs synthesis: the weekly narrative that summarises the trends for managers.
Questions to ask any ERP vendor about its AI
These questions separate useful AI from demo AI, whichever vendor you are evaluating:
- What happens when the model is unsure? Look for confidence levels and a review queue, not a silent best guess.
- What does it create on its own? Drafts and suggestions are fine. Paid bills, sent emails and filed returns are not.
- Can I see why? A one-line reason on a suggestion is the difference between trust and guesswork.
- What happens when the AI service is down? Core work such as creating a ticket should never wait on it.
- Is customer email treated as untrusted input? Anything that reads external text should not follow instructions hidden in it.
- Can my team correct it, and does the correction stick?
The practical takeaway
The best AI in an operations ERP today is the kind your team barely notices: the inbox is sorted, the tender checklist is filled in, the vendor invoice waits as a draft, and nothing important happens without someone signing off. If you'd like to see how this works on your own email, tenders or invoices, talk to us.