Unilever Cut Order Processing to 90 Seconds With AI Agents — Here's the Same Pattern for Smaller Founders

Unilever just showed that the slowest, most boring part of running a business, typing someone else's order into your system, can be cut from 15-20 minutes down to roughly 90 seconds. The company didn't do this with more staff or a bigger ERP system. It did it with AI agents that read the order the way a person would, then enter it the way a person would, just thousands of times faster and without getting tired on order number 400 of the day.

Unilever, working with Microsoft, built a system called NOVA, a network of five AI agents running on Microsoft Foundry. NOVA's job is to handle the messiest part of Unilever's order-to-cash process: orders that don't arrive in a clean, structured format. Instead they come in as emails, PDF attachments, Excel spreadsheets, and even photographs of handwritten paper order forms sent in by retailers and distributors.

Across 11 pilot markets, NOVA took the first stage of order processing, reading the order and entering it correctly, from 15-20 minutes of manual work down to about 90 seconds. Microsoft reported that 95% of these orders moved through completely touchless, with a human stepping in only when something genuinely needed judgment. Overall, the process ran roughly eight times faster than before. Francis Castro, Unilever's head of technology for customer operations, was cited in connection with these pilot results.

It's worth understanding the scale of the problem Unilever was solving, because it's the same problem most smaller businesses have, just at a different size. Unilever processes roughly 20 million orders a year. Of those, around 2 million still arrived through unstructured channels, an email with an attached price list, a scanned paper form, a spreadsheet with a distributor's own formatting. Every one of those needed a human to open it, read it, figure out which SKUs and quantities were meant, cross-check prices, and key it into the system by hand.

This is not a Unilever-specific problem. Any Indian business that takes orders through WhatsApp photos of handwritten chits, email attachments from a distributor, or an Excel sheet forwarded by a sales rep is sitting on exactly this same bottleneck, just at a smaller scale. A kirana distributor, a garments wholesaler, a B2B spare-parts supplier, an FMCG stockist, all of them lose hours a day to the same manual re-typing work Unilever was trying to eliminate.

NOVA's five agents each do one part of the job, which is why it works reliably instead of guessing. One agent extracts the raw information from whatever format the order arrives in, text from an email, text pulled off a scanned image, data from a spreadsheet. A second agent validates that information against the product catalogue and price list, catching mismatched SKU names or unrealistic quantities. A third checks it against customer master data, confirming this is a real, known account with the right delivery terms. A fourth flags anything that doesn't fit cleanly. A fifth enters the validated order into the system of record.

The reason only 5% of orders needed a human is that the system is deliberately built to escalate uncertainty rather than guess through it. If a handwritten quantity is unreadable, or a product code doesn't match anything in the catalogue, the agent doesn't force an entry, it routes that one order to a person. That's the part most businesses get wrong when they try automation on their own: they either automate everything and let bad data slip through, or they're too nervous to automate anything and keep doing it all by hand.

The real win isn't speed for its own sake. It's freeing people from the data-entry part of the job so they only spend their time on the one order in twenty that genuinely needs a human decision.

The pattern is scale-independent, but Indian data shows most smaller businesses haven't actually built it into their operations yet. A CII-Indian School of Business study surveying 2,919 MSMEs found that 65.8% of them already use some AI tool, but only 21% use ERP or any customised business software to run their operations. That gap is exactly where the Unilever pattern becomes relevant: businesses are experimenting with AI chatbots or writing tools, but not yet connecting AI to the actual order, inventory and accounting workflow.

Akshay Garkel, a partner at Grant Thornton Bharat, pointed to the same opportunity, noting that AI can automate customer servicing, order processing, inventory management and accounting for smaller businesses, letting them handle higher volumes without a proportional increase in headcount. That's precisely the mechanism Unilever demonstrated at a much larger scale.

The data on this gap is quite stark. A 2026 Dun & Bradstreet India study of 210 MSME leaders found 76% were using AI in some form, but only 5% had made it central to how they actually operate. Separately, Adobe's 2026 India findings, reported on October 16, 2026, found only 6.7% of Indian enterprises had scaled AI across multiple workflows, while 59% remained stuck somewhere between a pilot project and full deployment. Interestingly, 77% of companies said they'd increased AI investment year over year, yet only 10% were comprehensively measuring the return on that investment.

The reasons cited for this stall are consistent across the research: affordability, fragmented data sitting in different spreadsheets and systems, limited in-house technical skills, and genuine uncertainty about whether the investment will pay off. These are solvable problems, not fundamental barriers, which is also why industry players have started addressing them directly. On October 28, 2026, Google Cloud held an event in India bringing together more than 600 founders, promoters and technology executives, showcasing no-code AI agents built specifically to help SMBs move from experimenting with AI toward actually running their operations on it.

The agent pattern translates directly. An extraction step reads whatever comes in, a photographed order, an emailed PDF, a forwarded spreadsheet, and pulls out the product names, quantities and customer details. A validation step checks those against your actual price list and stock levels, whether that lives in Tally, Zoho, a Google Sheet or a custom dashboard. A third step checks the customer against your existing records so repeat buyers get their usual terms automatically. Anything that doesn't match cleanly, an unclear handwritten number, a product name that isn't quite right, gets flagged for a quick human look instead of being guessed at. Everything else goes straight into your order system.

The founder doesn't need Unilever's budget or Microsoft Foundry to get a version of this working. What they need is the same three things Unilever's NOVA relies on: a way to read unstructured input reliably, a set of business rules to validate against, and a clear line for when a human needs to step in. Built properly, this can take a 5-10 minute manual order entry down to well under a minute, with the owner or a staff member only looking at the exceptions instead of every single order.

This is exactly what ODIV's ai-workflow-automation service is built for. We sit down with you, map how orders (or claims, or onboarding forms, or whatever repetitive paperwork is eating your team's day) actually arrive today, across email, WhatsApp, PDF, spreadsheet, whatever the mess looks like, and build an AI agent workflow that reads it, validates it against your real data, and only brings a human in for the genuine exceptions. If your order intake runs through WhatsApp, we can also set that channel up properly with ODIV Engage so incoming orders land in one organised place instead of a scattered phone chat.

Our engineers build this using modern AI coding environments like Lovable and Claude Code, combined with conventional software engineering to make sure the result is secure, properly connected to your existing systems, and maintainable after launch, not a fragile demo. That combination is what lets us deliver this kind of custom automation in a fraction of the time and cost of a traditional hand-coded development project. If a 15-20 minute manual task in your business is quietly costing you hours every day, start a chat with us on WhatsApp and we'll walk you through what your version of this would look like.