AI-Based Store Audits: How They Work, What They Can and Cannot Check, and What They Cost in India (2026)

What an AI store audit is, and what it replaces

A conventional store audit is a person with a checklist. An area manager visits, ticks items, takes a few photos and writes a report the store hears about days later. In an Indian brand network one area manager typically covers eight to fifteen outlets (a working range, not a statistic), so any store is properly audited perhaps once a month and self-audited on trust in between. The result is familiar: a visit report as a WhatsApp album, a checklist ticked at 21:40 for a store that closed at 21:00, a festive window four days late in a third of the network.

An AI-based audit changes two things. The evidence is captured by the store itself, every day, as live photos or recordings tied to a time and a place, so coverage goes from a sample to the whole network. And a model reads that evidence against the standard before any human does, so the area manager's time goes to the twelve stores that failed rather than the two hundred that were fine. The AI does not decide the standard; the brand does. It does not fix anything; a named person does. Three separate technologies sit under the phrase, and an RFP that conflates them fails.

The three kinds of AI store audit

Each kind reads a different input against a different standard, for a different buyer. A CPG brand auditing its shelf in a general trade store and a fashion brand auditing its own window are not buying the same thing. Shelf image recognition is the most mature kind, and the right answer when the question is "how many of my facings are on that shelf in Nagpur". Photo verification against a docket is what most own-store chains mean by AI audits; BorentisOps calls its model Drishti, trains it on the brand's own dockets, and is the wrong tool for a general trade shelf. Conversation AI audits what the other two cannot see: whether the advisor pitched the launch, handled the objection and asked for the booking. Borentis does it in Hindi, English and Hinglish; the wider vendor set is at /learn/in-store-conversation-intelligence-vendors-map/.

What AI can and cannot check: a capability map

The honest answer to "can the AI check this" depends on the evidence, not the model. The table sorts common audit items by the weakest evidence that reliably answers them, where reliably means a verdict the store manager accepts without arguing. If the evidence needed is video and you only have photos, mark the item unverified rather than let a vendor score it. In the human rows the photo is a record and a person is the check.

Evidence rules that make an AI audit hold up

An AI verdict is only as good as the photo it read, and a photo is only as good as the guarantee that it was taken here, now, by this person. Disputes between a store and head office are about the evidence, not the model. Five rules settle most of them; a tool that cannot meet all five is a WhatsApp group with a nicer interface. The field tricks they defeat are at /learn/how-to-stop-backdated-store-photos-and-fake-checklists/. Recording a conversation adds a consent rule: a DPDP notice at the entrance and the counter, recording limited to the sales floor, retention stated, as at /learn/dpdp-consent-notice-retail-store-template/.

• Live capture only. The photo or recording is taken inside the app at the moment of the check; the gallery is disabled. Last week's window sent as today's is the commonest fake, and the cure is to make it impossible.
• Time stamp from the server, not the phone. The phone clock can be changed; the server time is recorded against the store's hours, so a checklist completed at 21:40 for a store that closed at 21:00 is flagged.
• GPS lock to the store. Capture is allowed inside a geofence around the registered location; outside it, the item is marked off-site. Basement showrooms and malls need a fallback such as the store's Wi-Fi or a QR code on the wall.
• Anti-backdating. A check belongs to the slot it was due in. A morning item captured at 15:00 is late, not complete, and the record keeps both times. Nothing is edited after the fact; a correction is a new entry.
• Audit trail on the verdict. Every AI pass or fail keeps the photo, the docket version, the model's confidence and any human override with a name, so a disputed window is settled with both pictures open.

The workflow: capture to score to ticket to verified closure

The audit is a loop, not a report. The eight steps are the same for a shelf photo, a window photo or a recorded conversation; only the model in step three differs.

1. Standard published. The planogram, this fortnight's docket, the SOP checklist or the sales playbook, with a version and an effective date.
2. Capture at the store, on schedule. The store manager, or the field rep for general trade, captures the required photos or conversations inside the app at the slot they are due, with live capture, server time and GPS lock; the app works offline and syncs later.
3. First-pass scoring by the model. Each photo is compared with the standard and each conversation with the playbook: a pass or fail per item with the reason, a confidence, and a store score from the brand's own weights, with fatal items that cap the score.
4. Human review of exceptions. The area manager or the VM team sees only the fails and the low-confidence items, photo and docket side by side, and confirms or overrides with a logged name.
5. Ticket to a named owner, with an SLA by type: price boards within the hour, a dusty display car the same day, a signage letter within the week.
6. Escalation on SLA. A ticket past its SLA moves up one level automatically, photo attached. Nothing is sent to customers at any step; this is an internal loop.
7. Verified closure. The ticket closes only with a new live photo scored against the same standard. "Done hai, sir" on a call does not close a ticket; a passing photo does.
8. Score to the scorecard and the digest. Audit score, time to fix and open tickets roll into the store scorecard beside sales and into an end-of-day digest, as in /learn/store-operations-kpis-for-brand-outlets/.

Before and after, in working ranges

The table gives the ranges a photo-verification programme should move for a brand running its own stores. They are working ranges from field programmes the author has seen and from pilot design assumptions, not a survey; each row states its basis. The after column assumes the evidence rules above are enforced; without them coverage reaches 100 percent on paper and nothing changes in the store.

What it costs, and how to work out your own number

The calculator below works the cost side for an own-store chain: what the manual audit programme costs in people's time against what an AI first-pass programme costs in software, with the monthly difference and the payback for your store count. It does not claim the sales uplift from better execution, which is real but harder to attribute; treat any saving it shows as the floor of the case.

The inputs. Stores is the number of outlets in the programme. Visits per month is how many audits each store receives today, typically one full and one partial. Minutes per manual audit is on-site time plus write-up per visit; 90 to 180 minutes is a working range. Loaded cost per hour is the area manager's cost to company plus travel and allowances, divided by working hours; a working range in Indian brand networks is Rs 400 to Rs 1,200 per hour. AI verification cost per store per month is the software price; vendors that publish prices quote per store or per location, and a working range for own-store photo verification with tickets and dashboards is Rs 800 to Rs 3,000 per store per month, with enterprise tiers above that when conversation capture is added. Ask for the year two price in writing.

Two costs sit outside the calculator: devices, if the network issues phones rather than using the store manager's Android, and the two to four weeks in which the store learns that the photo must be live and the ticket closes only with another photo. Both are where programmes fail.

A 30/60/90 day pilot plan: shadow, assist, auto-triage

Run the pilot in ten to twenty stores across two area managers, including your best store, your worst store and a basement store where GPS is weak. Each phase has an exit test. If the sales conversation is in the pilot, run it in the same stores from day 31 with the consent notice up from day one, and read the two scores together: a store at 92 on the window and 58 on the pitch has a coaching problem, not a facility problem. The selling-side method is at /learn/how-to-run-a-retail-sales-execution-audit-in-30-days/.

1. Days 1 to 30, shadow. Stores capture their opening routine and window photos daily; area managers visit as usual without seeing the AI verdicts. At month end, compare the model's verdicts with the area manager's findings. Exit test: agreement on nine photo-checkable items in ten, every disagreement reviewed with both pictures open. Most disagreements here are the photo, not the model, so fix the photography guide first.
2. Days 31 to 60, assist. Area managers see the verdicts and the exceptions list before each visit and go where it points. Tickets are raised from confirmed fails and closed only with a passing photo. Exit test: median time to fix under three days, disputes under one in twenty, every fatal item at 100 percent capture. Read the first scorecard with the store managers, not to them.
3. Days 61 to 90, auto-triage. High-confidence passes are accepted without a human; high-confidence fails go straight to a ticket; only low-confidence verdicts reach the area manager, whose visit becomes a coaching visit. Exit test: area manager time on audits down by a third, coverage above 90 percent of store-days, one launch or festive docket verified across the pilot stores within 72 hours of go-live.

Vendor landscape by use case

The table sorts vendors by the kind of audit their public pages described on 30 September 2026; it states claims, not test results. Borentis publishes this page and makes BorentisOps; its row follows the same rule. If your execution is on shelves in other people's stores, shortlist from the first five rows and ask about SKU coverage for your packs. If you run your own stores, shortlist from the rest and run the shadow month with your own dockets. If you sell through an advisor and want the pitch audited as well as the window, BorentisOps at /solutions/products/borentisops/ is the tool here that does both; the nearest platforms are compared at /learn/borentisops-vs-wooqer-vs-pazo-vs-amply/, and the 72-item checklist the AI would score is at /learn/store-operations-audit/.