Stage 1 of 4
Bring the records together
Pull the relevant numbers from the systems your team already uses, including browser-only back offices.
Read
- Site A
- 30 units available
- Site B
- stock constrained
- Sales ledger
- RM41/day vs RM309/day
AI agents that run your operation overnight.
They read every site, cost the next move, and hold each decision for your approval.
21:14
Operobot
Today
Nightly read finished. Two sites reconciled, one product line flagged:
- Site A: 30 units available
- Site B: stock constrained
- Sales ledger: RM41/day vs RM309/day
Held for you
Awaiting approvalRM351/day of RM481 estimated21:14
Why an agent, and not a report
01Every night, decisions go unmade because no one had time.
Stock, purchasing, capacity, exceptions. The numbers are already there, spread across the business. Nobody gets to them in time, and the report only tells you what happened after the month is over.
Works inside the systems you run
The decision loop
02One decision loop, from issue to verified outcome.
The evidence, the options, the approval, the action, and what happened next. All in one place.
Sample agents
03One agent for every decision your team makes repeatedly.
Showing 7 of 30
Ran in pilot Stock
P-1Transfer before you buy
Move stock you already own before raising an order.
- Signal
- Checks stock and sell-through for the same line across every site, nightly.
- Decision
- Moves existing stock to where it sells before any new purchase, as a small first move.
- Action
- Writes the transfer, notifies both sites, suppresses the duplicate order.
- Measure
- Confirms arrival and checks sell-through again at both sites.
From the record
30 units moved between two sites before any purchase; RM114/day recoverable, estimated.
Seven agents from the dated multi-site pilot.
Start with this agentRan in pilot Stock
P-2Clear a stalled line
Stop restocking what has stopped selling, and free the cash.
- Signal
- A line that has stopped selling at one site while still holding stock and display space.
- Decision
- Stop restocking it and clear it from display; redeploy the space.
- Action
- Clearance task to the site; the reorder is suppressed.
- Measure
- Cash freed, and whether the line moves once cleared.
From the record
RM2,981 of stock on a stalled line identified for clearance in the pilot; cash freed, estimated.
Seven agents from the dated multi-site pilot.
Start with this agentRan in pilot Purchasing
P-3Purchase guard
Hold spend that duplicates stock already owned elsewhere.
- Signal
- Reads every purchase request against stock on hand at every other site.
- Decision
- Holds spend that duplicates stock on hand; prices the cheaper route instead.
- Action
- Routes the viable option, flags the rest for approval.
- Measure
- Closes when the purchase is avoided or made.
From the record
A RM960 purchase rejected because the stock was already owned at another site.
Seven agents from the dated multi-site pilot.
Start with this agentRan in pilot Purchasing
P-4Reorder against cover and lead time
Raise the order when nothing else covers it, sized to the cover left.
- Signal
- Days of cover on each line against the supplier's lead time.
- Decision
- Raises the order only when no transfer covers it, sized to case packs and cover.
- Action
- Drafts the purchase order for approval; notifies the receiving site.
- Measure
- Stock arrives before cover runs out; cover checked again on arrival.
From the record
48 units, about two cartons, raised with 7 days of cover left; RM237/day recoverable, estimated.
Seven agents from the dated multi-site pilot.
Start with this agentRan in pilot Capacity
P-5Asset tuning in capacity-capped batches
Decide which assets to adjust tonight, and how far, within the crew's capacity.
- Signal
- Each asset's output against its own normal range, ranked by the money at stake.
- Decision
- Which to adjust tonight and by how much; your written rules decide the direction and the size.
- Action
- A batch of tuning tasks, capped to what the crew can do, each removable before approval.
- Measure
- Output against the band on the following nights.
From the record
One asset adjusted slightly in the pilot; RM130/day recoverable, estimated.
Seven agents from the dated multi-site pilot.
Start with this agentRan in pilot Yield
P-6Yield opportunity ranking
Rank every asset by what it earns against the typical asset at its site.
- Signal
- What each asset earns against the typical asset at its own site, for the period.
- Decision
- Diagnose each shortfall: retune it, swap the product, or inspect the asset.
- Action
- The ranked list with a diagnosis per line to the site lead.
- Measure
- Take against the median after the change.
From the record
Ran nightly in the pilot; every recommendation on this page was checked against this ranking before it reached a person.
Seven agents from the dated multi-site pilot.
Start with this agentRan in pilot Exceptions
P-7Exception watch
Escalate only what passes your own rules, so alerts stay worth reading.
- Signal
- Compares every cost, order, and service line against its own three-week normal.
- Decision
- Escalates only what passes your rules; the rest is logged, not sent.
- Action
- Sends the linked record to whoever owns it, the same evening.
- Measure
- Closes, reverses, or keeps the exception open.
From the record
Tuned to fire once or twice a day; an alarm that cries wolf gets ignored.
Seven agents from the dated multi-site pilot.
Start with this agentHigh-SKU retail Demand
R-1Forecast agent
Predict what each product sells at each store, every day, including the long tail.
- Signal
- Till-level sales per product per store, with promotions, weather, holidays, and price changes.
- Decision
- The forecast each store orders against, with slow sellers handled by their own statistics.
- Action
- Publishes the baseline forecast to ordering; flags the lines where it and the planner disagree.
- Measure
- Forecast error per line against last year's memory.
Published figure, not ours
20 to 50% fewer forecast errors; 30 to 50% less fresh waste (McKinsey, RELEX).
One agent per use case in a published retail implementation study.
Start with this agentHigh-SKU retail Stock
R-2Replenishment agent
Turn the forecast into orders that respect case packs, shelf space, and lead times.
- Signal
- Forecast against stock records, supplier lead times, order cutoffs, case packs, and shelf capacity.
- Decision
- How much each store orders of each product, and the safety stock per line.
- Action
- Drafts the orders for approval; a store override is logged rather than lost.
- Measure
- Spoilage and availability per line.
Published figure, not ours
10 to 40% less spoilage; 20 to 30% less safety stock (RELEX, McKinsey).
One agent per use case in a published retail implementation study.
Start with this agentHigh-SKU retail Pricing
R-3Pricing and markdown agent
Price the long tail; mark down perishables while they can still sell. People keep the known-value items.
- Signal
- Demand response to price, competitor prices, days to expiry, and shrink logs.
- Decision
- Prices on the long tail and markdowns by days to expiry; humans govern the items shoppers remember.
- Action
- Price changes to the shelf labels, held for approval.
- Measure
- Margin and fresh waste on the line.
Published figure, not ours
+2 to 5 points of margin; 33 to 47% less fresh waste (Competera, Wasteless). Person-level pricing excluded.
One agent per use case in a published retail implementation study.
Start with this agentHigh-SKU retail Availability
R-4Shelf monitoring agent
Compare what is on the shelf with what should be, and raise the fixes in order.
- Signal
- Shelf photos from robots, fixed cameras, or staff phones, against the planogram.
- Decision
- Which empty spots, misplaced products, and wrong price tags to fix first.
- Action
- Tasks to store staff, with the photo.
- Measure
- On-shelf availability and price-tag accuracy.
Published figure, not ours
2 to 4 points of on-shelf availability (vendor-claimed); price-tag verification is the consistent quick win.
One agent per use case in a published retail implementation study.
Start with this agentHigh-SKU retail Checkout
R-5Checkout assist agent
Catch missed scans and recognise produce at self-checkout, and prompt the shopper first.
- Signal
- Camera flags over self-checkout: missed scans, barcode switching, produce on the scale.
- Decision
- Prompt the shopper to self-correct, or call a member of staff.
- Action
- The prompt, then the intervention.
- Measure
- Error rate and interventions per hundred transactions.
Published figure, not ours
More than 75% of self-checkout errors self-corrected (Kroger).
One agent per use case in a published retail implementation study.
Start with this agentHigh-SKU retail Loss
R-6Loss prevention agent
Watch existing cameras for theft behaviour and send only the alerts worth acting on.
- Signal
- Existing camera feeds: concealment gestures, shelf sweeps, skipped scans for accomplices.
- Decision
- Which alerts reach the floor now and which are logged.
- Action
- Alert to staff with the clip attached.
- Measure
- Shrink on the covered categories.
Published figure, not ours
About 60% shrink reduction on covered categories (vendor-reported, directional).
One agent per use case in a published retail implementation study.
Start with this agentHigh-SKU retail Availability
R-7Stock-out and phantom-inventory agent
Spot empty shelves, and correct the record that says they are full.
- Signal
- Camera gap alerts, and lines with stock on record but no sales for longer than their normal.
- Decision
- Restock it, or count it and correct the record so reordering triggers again.
- Action
- Restock or count task to the site; the record is corrected once confirmed.
- Measure
- Availability and sales of the line afterwards.
Published figure, not ours
Out-of-stocks cost about 4% of grocery sales (industry rule of thumb); fixing phantom inventory is where camera ROI compounds.
One agent per use case in a published retail implementation study.
Start with this agentHigh-SKU retail Offers
R-8Personalisation agent
Predict what each household buys, target the offer, and pick the substitution.
- Signal
- Loyalty and transaction data, and out-of-stocks at pick time.
- Decision
- Which offer to each household, and which substitution for an item that is out.
- Action
- Offers to the channel; substitutions at pick.
- Measure
- Conversion and basket size.
Published figure, not ours
+16 to 27% conversion or basket lift; about 95% shopper approval of AI-picked substitutions (Instacart).
One agent per use case in a published retail implementation study.
Start with this agentHigh-SKU retail Customers
R-9Customer contact agent
Handle order status, refunds, and refill calls end to end, and route the rest.
- Signal
- Inbound contacts: order status, delivery problems, refunds, store questions, refill lines.
- Decision
- Resolve it, or route it to a person.
- Action
- Handles the contact, verifies the caller against the system of record.
- Measure
- Contacts handled without staff, and time to first response.
Published figure, not ours
Plan on 40 to 55% of contacts handled without staff; vendors advertise 70 to 80% (Zendesk measured median 41%).
One agent per use case in a published retail implementation study.
Start with this agentHigh-SKU retail Logistics
R-10Logistics agent
Route the trucks, predict real lead times, and cut the safety stock that stale averages force.
- Signal
- Delivery routes, shipment tracking, and actual supplier lead times against the fixed averages.
- Decision
- Routing, predicted arrival, and the lead time each line should plan against.
- Action
- The route plan; updated lead times into replenishment.
- Measure
- Transport cost and lead-time accuracy.
Published figure, not ours
5 to 10% fuel and 10 to 30% transport cost from route optimisation; about 15% logistics cost (McKinsey).
One agent per use case in a published retail implementation study.
Start with this agentHigh-SKU retail People
R-11Scheduling agent
Forecast traffic in 15-minute steps and build schedules that respect the rules and the people.
- Signal
- Store traffic in 15 to 30 minute increments, by role, against labour law and preferences.
- Decision
- Staffing by role and shift; which store tasks to push and verify.
- Action
- Builds the schedule; pushes and checks store tasks.
- Measure
- Labour cost and wait times.
Published figure, not ours
About 15% average labour-cost savings; one grocer cut wait times 30% (industry benchmark).
One agent per use case in a published retail implementation study.
Start with this agentHigh-SKU retail Pharmacy
R-12Pharmacy agent
Pre-fill prior authorisations, process routine refills, route exceptions. A pharmacist decides.
- Signal
- Rejected claims that need prior authorisation; refill requests on the phone line.
- Decision
- Pre-fill the paperwork; process the routine refill; route everything else to a person.
- Action
- Submission, refill, or routing. The final prescription decision is always a person's.
- Measure
- Turnaround and abandoned calls.
Published figure, not ours
83% reduction in prior-authorisation turnaround (CoverMyMeds); AI never makes the final prescription decision.
One agent per use case in a published retail implementation study.
Start with this agentHigh-SKU retail Finance
R-13Accounts payable agent
Match supplier invoices to orders and receipts line by line, post the clean ones, route the rest.
- Signal
- Supplier invoices against purchase orders and receiving records, including direct-store-delivery invoices.
- Decision
- Post untouched, or route the exception to the right person.
- Action
- Three-way match and posting; deductions and chargebacks recovered.
- Measure
- Cost and days per invoice; exception rate.
Published figure, not ours
Best-in-class US$2.78 and 3.1 days per invoice against US$10.89 and 10.9 days average.
One agent per use case in a published retail implementation study.
Start with this agentManufacturing Production
M-1Production set-point agent
Recommend set-points and rebuild the schedule against real constraints, in minutes.
- Signal
- Hundreds of interdependent settings against output, yield, and energy; changeovers, tooling, and manpower.
- Decision
- Set-point recommendations, and the sequence that clears the week's constraints.
- Action
- Recommended set-points and a rebuilt schedule to the planner, held for approval.
- Measure
- Output, yield, and OEE against the line's own normal.
Published figure, not ours
10 to 20% production output improvement described as a well-supported target (Deloitte); +15 to 25% OEE (vendor claims).
One agent per department in a published study of large Malaysian manufacturers.
Start with this agentManufacturing Quality
M-2Quality inspection agent
Inspect every unit on every shift, isolate what fails, and draft the CAPA.
- Signal
- Camera classification of every unit at line speed; process drift before it produces scrap.
- Decision
- Pass, hold, or isolate the lot; which drift to act on first.
- Action
- Isolates non-conforming product; drafts the deviation and CAPA for a person to sign.
- Measure
- Defect escapes and scrap.
Published figure, not ours
Defect escapes −94%, defect rate −87% in named vendor cases; human inspection 70 to 80% against AI vision 95 to 99% (vendor claim).
One agent per department in a published study of large Malaysian manufacturers.
Start with this agentManufacturing Maintenance
M-3Maintenance agent
Repair in the next planned window instead of on the night it breaks.
- Signal
- Vibration, temperature, current, and control-system data against each asset's own normal.
- Decision
- Repair in the next planned window, or run to the next check; which spares to hold.
- Action
- Work order and spares check, held for approval.
- Measure
- Unplanned downtime and maintenance cost on the asset.
Published figure, not ours
30 to 50% less downtime, 10 to 40% lower maintenance cost (McKinsey); PETRONAS reported RM73.1 million at 14× return in year one.
One agent per department in a published study of large Malaysian manufacturers.
Start with this agentManufacturing Supply chain
M-4Supply planning agent
Project the shortfall, draft the replenishment, check the lead time, route it for a yes.
- Signal
- Demand by SKU, customer, and channel; supplier distress, port congestion, and tariff change.
- Decision
- Dynamic safety stock per line, and what to replenish when a shortfall is projected.
- Action
- Drafts the replenishment, checks supplier lead time, routes it for approval.
- Measure
- Forecast error, inventory, and lost sales from stock-outs.
Published figure, not ours
Forecast error −30 to 50%; inventory −20 to 30%; lost sales from stock-outs down by up to 65% (McKinsey).
One agent per department in a published study of large Malaysian manufacturers.
Start with this agentManufacturing Procurement
M-5Procurement agent
Classify the spend, benchmark the price, find the mismatch, and put buyers back to negotiating.
- Signal
- Spend across the ERP, contract terms and expiry dates, prices against commodity indices.
- Decision
- Should-cost, which suppliers to query, which invoice, order, or contract mismatch to chase.
- Action
- Drafts the RFQ and quote comparison; flags duplicate suppliers and payments.
- Measure
- Savings against benchmark; variance recovered.
Published figure, not ours
15 to 22% more savings opportunities identified than rules-based classification (vendor claim); procurement controls 50 to 70% of spend with 1 to 2% of headcount.
One agent per department in a published study of large Malaysian manufacturers.
Start with this agentManufacturing People
M-6Workforce agent
Screen at volume, forecast absenteeism, and answer SOP questions in the worker's own language.
- Signal
- Applicant volume for operator roles, shift absenteeism and attrition signals, permit and levy dates.
- Decision
- Who to shortlist and when to interview; which shifts are at risk next week.
- Action
- Screening and scheduling; multilingual SOP and training answers on the floor.
- Measure
- Time to hire and vacancy days on bottleneck lines.
Published figure, not ours
Time-to-hire −47% (SAP-cited, vendor claim); treat anything above 50% sceptically.
One agent per department in a published study of large Malaysian manufacturers.
Start with this agentManufacturing Finance
M-7Accounts agent
Capture and three-way match invoices, validate e-Invoicing, catch duplicates and mis-postings.
- Signal
- Supplier invoices against orders and receipts; MyInvois e-Invoicing fields; payment history.
- Decision
- Post, hold, or route; which anomalies are duplicate payments or mis-postings.
- Action
- Matching and posting; supplier statement reconciliation; variance narratives for the close.
- Measure
- Cost and days per invoice; late-payment penalties.
Published figure, not ours
US$15.97 to US$2.36 per invoice, an 85% reduction (Ardent Partners 2025); 17.4 to 3.1 days per invoice.
One agent per department in a published study of large Malaysian manufacturers.
Start with this agentManufacturing Sales
M-8Quotation and pricing agent
Draft compliant quotes from spec sheets and past bids, and flag the deals that erode margin.
- Signal
- Spec sheets, past bids, and margin by customer, SKU, and region.
- Decision
- The quote, and whether a deal breaches the margin guardrail.
- Action
- Drafts the quote and technical response for a sales engineer to send.
- Measure
- Quotation turnaround, win rate, and gross margin.
Published figure, not ours
Recovering 0.5 to 1.0 point of gross margin on RM500 million of revenue is RM2.5 to 5 million (study estimate).
One agent per department in a published study of large Malaysian manufacturers.
Start with this agentManufacturing Orders
M-9Order management agent
Enter customer POs from PDFs and email, answer status questions, route complaints to CAPA.
- Signal
- Customer purchase orders arriving as PDFs and emails; status enquiries; complaints.
- Decision
- Enter and acknowledge, answer the ETA, or route the complaint to quality.
- Action
- Order entry into the ERP; acknowledgements and ETAs; complaint classification.
- Measure
- Re-keying errors and response time.
Published figure, not ours
+14% issues resolved per hour, +34% for novice workers (Brynjolfsson, Li and Raymond, QJE, peer-reviewed).
One agent per department in a published study of large Malaysian manufacturers.
Start with this agentManufacturing Compliance
M-10Compliance and safety agent
Watch existing CCTV for unsafe acts, assemble the audit evidence, draft the non-conformances.
- Signal
- CCTV for PPE, restricted zones, and vehicle proximity; obligations across ISO, IATF, DOSH, and customs.
- Decision
- Which safety alerts to escalate now; which audit gaps to close first.
- Action
- Alert to the floor; evidence packs and draft non-conformances; ESG data collection.
- Measure
- Incidents and audit findings.
Published figure, not ours
−62% safety-vest incidents (NSG) and −86% vehicle incidents (Piston), vendor-named customers. Any CCTV use needs a PDPA impact assessment.
One agent per department in a published study of large Malaysian manufacturers.
Start with this agentAfter approval
04One approval. The work gets done everywhere it needs to.
This is where a chat assistant stops. Opero carries the approved decision into the systems and teams that need to act, then checks that it actually happened.
- 21:14:01Approval recordedApproved in Telegram and logged against the decision.
- 21:14:02Systems checkedOpero checks for conflicts or duplicate work before making any change.
- 21:14:04Task createdThe approved action is entered into the operating system and given a reference.
- 21:14:05Team notifiedThe people responsible on the floor and in the warehouse are told what changed.
- 21:14:06Records updatedThe decision is reflected in the relevant ledger and dated to when it was raised.
- +24hOutcome checkedOpero goes back to the source records to confirm the stock arrived and measure what changed.
Is this your operation
05Built for operations where the same decisions come up again and again.
Opero works best when a few things are already true.
Requirements
- More than one site or system
- Two outlets, several branches, or simply a back office and a ledger that no one has time to read together.
- Decisions that repeat
- Stock, purchasing, capacity, exceptions. The same judgment calls, week after week.
- Someone who can make the call
- One accountable person who can approve the decision in a channel they already use.
- No API required
- Browser-based systems, spreadsheets, CSV exports. If your team already works from it, Opero can usually work with it too.
Start read-only. Prove the first decision.
Return date agreed upfront.
Meet your operator
Choose the first part of the operation to work on and agree what a useful result looks like.
Map the operation
We learn the records, rules, exceptions, and who has authority to approve what.
Observe
Opero reads, but changes nothing. You can compare its recommendations with the decisions your team actually made.
First decision
One real recommendation, with the evidence and expected cost, held for approval.
Act and check
Once approved, Opero carries it through and comes back to measure what happened.
Questions operators ask first
06Clear boundaries before anything goes live.
Opero starts by watching. It only gets permission to do more once it has shown that it can make the right call consistently.
Observe
Reads the operation. Sends nothing. Changes nothing.
Recommend
Shows what it found, the options, and what each one is likely to cost.
Approve
Anything consequential waits for a person to say yes.
Limited execution
Opero can only make changes in the systems and categories you have explicitly enabled.
Earned autonomy
Routine decisions can become more automatic over time, if you choose.
Some decisions can stay approval-only forever. In our pilot, automatic approval is still switched off.
Questions
If there is something we have not covered, email hello@opero.ai. For deployment, access, and data controls, see the security page.
Does Opero ever act without approval?
Not on arrival. A fresh install sends nothing, writes nothing, and auto-approves nothing. Routine categories can earn auto-approval after at least 95% human agreement across at least 20 decisions on at least 8 nights, and the consequential ones can require your approval forever. In our pilot, auto-approval is still switched off.
What if our systems do not have an API?
No. It is the normal case. Usually a login is enough: Opero drives the back office the way a manager does, through a real browser. Google Sheets and CSV exports count as records too.
What data leaves our premises?
Each customer has a separate installation. Settings, connected-system logins, and operating data live with you and are not shared between customers. Opero reads only the systems you connect, and customer data is not used to train a model shared with other customers. The security page states only the controls the current product supports.
What happens when Opero gets it wrong?
Nothing, until someone says yes. Every recommendation points to the original records and can never alter a number; where a figure is an estimate, it is labelled as one. Your disagreement is logged against the decision, and a category only earns more room by being right, decision after decision.
How do we stop it?
A kill switch pauses the system from the dashboard, and changes to your systems can be switched off per site at any time without removing read access. Every recommendation, approval, change, and reversal sits in a log that can only be added to, with who did it and through which channel.
How is it priced?
Per site. The scope, the boundary, and the price are agreed with you before anything is connected.
Who builds this
07Built by an operator and a researcher.
One of us spent years running the kind of operation Opero reads. The other builds the systems that read it.
Chief executive
Runs the kind of operation Opero was built for.
Built and operates a multi-site claw machine arcade business, where stock, pricing, purchasing, and capacity decisions were previously made by hand every night.
Credentials
- Built an eight-figure business within its first year
- Operates one of the largest businesses in its category
- Still runs the operation day to day
Chief technology officer
Builds the decision engine behind Opero.
Works in machine learning and computational neuroscience, with published research and top results in large international AI competitions.
Credentials
- Ranked 1st among 15,000+ participants in an RSNA Kaggle AI competition
- 8 first-author, peer-reviewed papers in AI and computational neuroscience
- Silver medalist, International Chemistry Olympiad
- AI and computational modelling lead, Stanford iGEM team
Start with a real operating week
Send us one week where stock, purchasing, capacity, or another operating decision went wrong. We will work through the records with your operator and come back with one costed decision, the evidence behind it, and what Opero would have done. No integration required.