Blog
The Autowhat AI blog.
What we think about running operations with AI: transport, plants, cold chain, finance, field teams and public systems.
- Write the Tolerance Down Before You Inspect AnythingQuality disputes are usually about an unwritten tolerance, not a measurement. Write a plus-or-minus per dimension on the spec, fix where and how it is measured, and treat labels as exact-match. AutoQC checks length and width to ± 1 cm, height at the barcode to ± 0.5 cm and the MRP label exactly, from four locked-order photos.
- Why We Spend Weeks on Site Before Writing CodeSite operations software fails in hand-offs nobody wrote down: late guests, vendor-run kitchens and vehicle pools, night rosters, VIP surges. That is why our developers spend 1.5 to 2 months on site shadowing real shifts before building Sampark, then deliver in waves starting with the visitor spine.
- Why the Same Three Brokers Win Every LoadWhen rates are gathered by phone, the brokers called first win most, so they get called first again. Concentration is fine if evidence supports it. Inviting every relevant vendor per load and recording bids, wins, win rate and drop-offs turns habit into a choice the desk can defend.
- Why Sales and Logistics Argue About OTIFSales and logistics argue about OTIF because they count different units, start the clock at different points and treat damage differently. The fix is not a better spreadsheet but one written definition computed in one place, written on-time rules, and an owner and reason for every late trip.
- Why Your Production Plan Needs Frozen DaysA plan that changes every hour looks responsive on screen but causes chaos on a bakery floor, where flour, packaging, people and dispatch are already committed. We argue for frozen days near term, firming days after them and a fluid plan beyond, with every break of the freeze made deliberately and recorded.
- Why Every Grain Grading Number Needs a RangeA grading result without a range looks precise but hides how sure the method is. GrainGrade, in development, is designed to show every number with its range: three independent count estimates, flagged if they disagree, with counting error measured at 3.5% against a 3% target. Ranges make borderline lots visible before they become disputes.
- When Every Steel Shop Hits Its Number, and Margin SlipsIntegrated steel plants measure each shop well, but margin leaks between shops: a converter hits tap-to-tap while the caster breaks sequence. Steel Twin traces every heat from charge to coil on heat_id and ranks holds, constraints and energy variances across shops by cost in one action queue the shift can act on.
- WhatsApp Is Already Your Transport Operating ChannelDrivers, branches and customers already run transport on WhatsApp, and new apps often go unused. We think the fix is not a new channel but records behind the chat: PODs matched to LRs, status answered from GPS, approvals confirmed with a PIN, while the books stay in the office.
- What We Check Before an AI-Made Post Goes LiveGenerating a post takes seconds; approving it well still needs a person. Check facts a model cannot verify first: prices, offers, claims, stock and dates. Then check the image and read the caption once for voice. Fix repeated tone issues in brand settings and skip cosmetic debates.
- What an AI Assistant Should Not Do in ROC FilingAn assistant can classify a client’s request, pre-fill known facts, interview for gaps and validate DIN, PAN, CIN and dates. It should not file on MCA, judge legal correctness, guess missing values, override failed checks or close a ticket without an SRN. The time saved should go into professional review.
- What AI Should Never Do in a Loan FileAn AI agent in a loan file should draft, not decide. It should never sanction a loan, state a figure it cannot cite, hide a policy deviation, pull data beyond consent, or vary its method between files. Inside that boundary it can assemble the file, run consistent analysis and draft a cited memo for a credit officer.
- What AI Must Never Change in an Audit TemplateAI can draft an audit checklist from a sentence, but it can also invent citations, soften critical items and turn evidence rules into suggestions. We argue the model should never touch clause references, critical items or evidence minimums, which is how AutoAudit’s AI studio is governed.
- What a Photo of Wheat Cannot Tell YouA phone photo can count wheat grains and spot broken, shrivelled, damaged, immature and insect-bored grains and foreign matter. It cannot see moisture, protein, gluten, aroma or hidden insects. GrainGrade, in development, takes moisture from a meter and is designed to state these limits on every report.
- What a Supply Chain Digital Twin Is Actually ForA supply chain digital twin is useful when it changes a decision next week: where to pre-position stock before a monsoon week, which RDCs to serve if a plant stops, whom to allocate when a carrier fails placement. It needs observed transit, dwell and batch-level stock, which is why it should follow the control tower.
- What a Sales Voice Agent Should Never SayThe guardrails matter more than the features. A sales voice agent should never invent a price, ask the budget twice, claim to be human, keep calling someone who said stop or handle a complaint alone. AI Calling holds each of these lines and scores compliance on every recorded call so you can check, not trust.
- A Voice Agent Is Only as Good as Your Price ListVoice quality is the easy part of an AI calling agent. Retailers keep ordering only if the price, scheme and minimum are right. AI Calling answers only from your price list and knowledge base, sends unknowns to a callback and scores price accuracy on every call, so treat the price list as master data with a review step.
- Why Ultra-cold Alerts Need a Ladder, Not an SMSAt −80 °C, an alert that is sent but not acknowledged is the real risk. An escalation ladder treats each alert as a task with a clock and keeps climbing roles until someone acts. It should trigger on drift and dry-ice exhaustion risk, not only on excursions, and record who did what.
- Try the Furnace Recipe Change on the Twin FirstMost furnace improvement ideas are never tried because every trial risks a heat. Furnace Twin, a physics model calibrated on every heat, lets teams compare reductant, basicity, temperature and feed changes side by side, then send only the best scenario to a two-heat trial. Trials become tests of a prediction, not gambles.
- The True Cost of an Unbilled LRAn unbilled LR is not just a paperwork delay. Diesel, tolls and advances are already paid, the customer's credit clock has not started, and evidence for detention and short-delivery fades. We would make the POD reach the biller first, then track billing lag and unbilled value every week.
- Track the Trip, Not the Truck: A Shipper’s ViewVehicle GPS was built for fleet owners, but shippers mostly hire trucks they do not own. We argue the unit of tracking should be the trip: a device assigned at the gate, a live record of one consignment, and a device collected at arrival, with fixed GPS and SIM feeds on the same trip.
- Test Collections Changes Against a Control GroupBefore-and-after charts mislead in collections because salary dates, portfolio mix, team changes and regression to the mean move results too. Test each change on one bucket against a random control group, decide the metric and threshold first, and compare promises kept, complaints and breaches as well as cure rate.
- Students Should See What an AI Tutor RemembersTutor memory lets an AI schedule revision on each student’s real gaps, but hidden memory makes students afraid to be wrong. We argue memory must be visible, editable and erasable by the student, with no ID numbers or OTPs in prompts, which is how Teach-Back handles it.
- Start Warehouse Software With the Step That HurtsSwitching on every warehouse step at once makes problems impossible to untangle. Start with the step that costs the most each cycle, usually the purchase list PDF or bag allocation, run a full cycle, then add modules. Warehouse OS writes every step to one record, so the box on the van still traces back to the purchase list line.
- Stale Plant Data Should Look StaleA plant screen that shows old data in the same way as live data invites wrong decisions. Steel Twin joins seven systems that update on different clocks, keeps each event’s source timestamp and quality flag so stale data is never hidden, and replays every shift with what the controller could actually see.
- Your Site Staff Will Not Install Another HR AppHR software for site teams usually stalls at the download: sign-ups, passwords, storage and English-only screens. We argue adoption is the first feature, so the employee side should live in the chat staff already use, in their language, while the office keeps a proper portal.
- Before You Buy New Sensors, Use the Ones You Already HaveMonitoring projects often start by buying new loggers, but most cold rooms and freezers already have probes on a controller readable over BACnet/IP or Modbus TCP. We would connect those first, add escalation and an audit trail, check calibration, and buy hardware only for the gaps real data shows.
- Tracking Only Some Managers Skews the PictureWhen portfolio manager data is collected by hand, coverage is set by analyst hours, not by the research question. A fixed list can hide flows to smaller managers and distort rankings. We argue coverage should be a deliberate choice made after seeing the whole universe, and stated in every output.
- Why Recon Software Should Be Read-only FirstReconciliation software must see everything but should change nothing by default. Read-only access limits the damage of a bad rule, keeps the check independent and eases security approval. Run in parallel before switching, and open any write path one purpose at a time, under maker-checker and one audit log.
- Your Recon Rules Should Not Live in One HeadReconciliation pain is often blamed on spreadsheets, but the real risk is matching rules that live in one person’s head: unreviewable, drifting and gone when they leave. Write rules down, type them, version them and put them through maker-checker, and turn recurring breaks into suggested rules.
- Five Questions Before Accepting a −80 °C ShipmentAccepting an ultra-cold shipment is a decision, not a formality. Before signing, review both logger traces, seal and dwell at every handoff, dispatch sign-offs, transit alerts and who closed them, and how fast the product reaches the freezer. If any answer is no, quarantine and raise an NCR.
- Quality Control Should Block the Truck, Not File a ReportMost quality checks end in a report that is read after the goods have shipped. We argue a check is real only if it sits in the path of the goods, like a kitchen pass. AutoQC gives every piece one state; packing, labelling and loading all check it, and only “QC passed” lets a piece onto the truck.
- The Product Master Matters More Than the 3D TwinA 3D warehouse twin is only as accurate as the product master behind it. In Warehouse OS every purchase list line is matched to SKU, packing, weight and zone before it counts, and mismatches are flagged, not guessed. Fix one row per SKU and packing, checked weights and real zones first; then the twin earns its place.
- Your Product Catalogue Is the Content BriefFor a product brand, the catalogue is the content brief. Listings already hold the product, photos, price and selling points; a person only needs to choose the product, the goal and the channel. Catalogue-led posts make a steady base, with story and festival posts on top, and good listings make better posts.
- Plant Logistics Needs a Control Tower, Not a ReportPlant logistics teams rarely lack reports; they lack an early signal of what needs a decision and who owns it. A control tower ranks exceptions on one batch-keyed record, attaches context such as RDC stock cover, and depends on honest inputs: staged carrier files, several ping sources and written KPI rules.
- Stop Retyping the Visit Request: One Email, Every DeskAt most industrial sites one visit email is retyped by rooms, travel, transport, safety and the gate, and each copy can be wrong. We think the fix is to read the email once, have a person check the extraction, and let each desk approve its part of one shared request.
- One Bad Row Should Not Stop the Freight PlanFreight plans run late not because of spreadsheets but because one bad row stops the whole model. FreightPlan validates eight CSV datasets row by row, rejects bad rows with the reason and fix, loads the rest, shows the effect of an upload before committing it and always returns a plan with the exact shortfall.
- Why the Nearest Auditor Is Rarely the Right OneAssigning field audits by distance and phone calls ignores capacity, domain expertise, rating and SLA history, and leaves no record of why someone was chosen. We argue for hard filters first, then a scored decision with visible weights, which is how AutoAudit routes every requirement.
- In Market Data, a Blank Is Not a ZeroA blank cell in a research dataset can mean a real zero, an entity that filed nothing, or a collection that failed. Treating all three as zero quietly distorts analysis. We set out rules for any research dataset: explicit “no data” status, early format conversion, expected-versus-collected counts and lineage.
- Start KYC Automation With One Journey, One MetricKYC automation stalls when it tries to fix every product at once. Pick one journey with real volume, a written policy and a clear pain, agree one metric, and go live within weeks with read-only defaults. The second journey reuses the channel, document reading, controls and core connection.
- Why a KYC Agent Should Start Read-OnlyAn AI agent in KYC should start read-only: reading documents, matching details and asking for what is missing carry little risk, while writes to the core change records. Widen write access one API at a time, keep maker-checker below threshold, a kill switch per agent and one audit log.
- Stop Judging B2B Outreach by Open RatesSales reviews often judge outreach on opens and read receipts because they are easy to see. We argue channels should be compared on replies, meetings and wins, which needs every send, read and reply joined to the lead and its stage in one record.
- The Hidden Cost of Calling Round for Truck RatesCalling round for truck rates looks thorough but produces quotes that are not comparable, favours whoever answers first and leaves no record. Sending one complete load message to every vendor with a closing time, keeping every bid in one place and checking awards against what the lane cost fixes that.
- A Health Dashboard Must Show How Old Its Numbers AreMany public-health dashboards show old and fresh numbers in the same font, so stale ICU or surveillance figures look reliable. We argue every number should carry its source and timestamp and missing reports should show as gaps, because honest screens prevent bad decisions and keep trust.
- A Health Alert Without an Owner Is Just a Red CellPublic-health command centres are good at turning things red, less good at deciding who acts. We argue every exception needs evidence, a suggested action, one named owner and a clock, delivered to district officers on their phones with approvals logged, so shared visibility becomes action.
- Hardship Calls Need a Person, Not a ScriptAutomated reminders and voice agents suit borrowers who forgot or need a date. Hardship, disputes and restructuring need a trained person with judgement and authority. Automation’s job there is to spot the signals early and hand over cleanly, with the conversation summary and account history, so agents have time for the calls that matter.
- Give Teachers Time Back Before Asking for DataSchool monitoring depends on data from teachers who already spend much of the day on paperwork. We argue a rollout should give time back first, with voice attendance, automatic filing and no pilot penalties, because data produced while doing the job is more accurate.
- Furnace Answers Should Cite Their SourcesA fluent assistant that gives an unsourced answer about a smelting furnace is a risk. Furnace Twin answers plain-language questions from plant records, the calibrated physics model and a literature library, cites each source and uses a local language model. Good citations need good records, so shift handovers are captured by voice.
- Forest Fires Should Not Be Answered in Arrival OrderFire-season desks usually answer alerts in the order the inbox, chat group or phone delivers them. We argue the queue should be ranked on fire weather, wind, fuel, people and wildlife in the path, and incident age, with the reasons visible so officers can trust and challenge the rank.
- After the Flood: Make the Review a ReplayAfter a city floods, the review usually runs on memory and registers, so it becomes an argument. A minute-by-minute replay of rain, channel levels, pumps, complaints and crews on one timeline shows where the warning window went, protects field staff, and ends in one agreed change before the next storm.
- Why a Fleet Video Wall Should Show Exceptions, Not MapsA live map of every truck looks like control but rarely tells anyone what to do. We think a fleet video wall should lead with exceptions: late trips, e-way bills expiring, detention risk, pending PODs, each with a threshold and an owner, mirrored as WhatsApp alerts. The map belongs to one side as a detail view.
- What Counts as Firm Demand Is a Policy ChoiceEvery production plan starts from demand, but most plants never write down which orders count as firm. That unwritten rule causes planning arguments, wrong buys and meaningless plan-vs-actual gaps. We argue for an agreed rule by stage and channel, held as a setting, with exceptions made visibly.
- Fire Season Reviews Should Run on Evidence, Not MemoryMany forest divisions rebuild the fire season review from chat groups, emails, registers and memory. We argue the review should run on evidence captured during the season: timed incidents, patrol tracks and photos, a running season report and replays of the worst fires.
- Every Freight Sourcing Rule Has a PriceCarrier caps, minimum commitments and depot habits are all reasonable, and all cost money. FreightPlan writes sourcing rules down as constraints, separates physical limits from policy choices and prices a carrier-share policy live, so logistics and finance can decide whether each rule is worth what it costs.
- Why Every Claim Payout Needs Two PeopleAutomate claims up to the payout, then require two people. The payout is the one step that cannot easily be undone, so it needs the strongest control. Rules flag leakage on arrival; a maker and checker see the documents, policy checks, triage reason and flags, and every decision lands in one audit log.
- Stock Corrections Are Not a Distributor’s FailureStock in a distributor system always drifts from the shelf. If brands treat every correction as a failure, distributors stop reporting and stock data gets worse. We argue for welcoming corrections with reasons, reviewing them for patterns by SKU and distributor, and using audits to check the trail.
- Why the Pending Reason Matters More Than DispatchSales teams track dispatches, but the orders that did not go, and why, say more about a distribution network. Pending can mean a stock-out, a route gap, a credit hold or a wrong order, each with a different owner. We argue for making the reason as easy to send as a dispatch, then reading reasons weekly.
- Your Customer Should Not Have to Call for an ETAStatus calls pass through sales, dispatch and the driver, and the answer gets worse at each step. We argue customers should get a no-login tracking link with an honest ETA based on actual speed, while the desk keeps the exceptions and acts on them first.
- Credit Deviations Belong in the Draft, Not in CommitteePolicy deviations are normal; surprises in committee are not. When deviations are found in the meeting, files wait and members stop trusting memos. Flag every deviation in the draft with the policy reference, the cited figure and the analyst’s case, in a fixed section, so the committee decides rather than investigates.
- The Company Master Is a CS Practice’s Best AssetMost fields on an ROC e-Form are facts a practice already knows about the client. Re-typing them for every filing invites avoidable rejections. One checked master record per company, updated when each ticket closes, lets forms start pre-filled, catches errors at entry and keeps history in one place.
- The College Helpdesk Problem Is the BrochureColleges blame the phone for admissions-week chaos, but repeated questions usually mean the brochure is buried, split, stale, ambiguous or missing a line. We argue for treating every student question as data about your documents and running a weekly loop: group repeats, fix the document, retire old versions.
- Why a College Chatbot Should Say “I Don’t Know”A confident wrong answer about a fee or date does more harm than no answer. We argue a college chatbot should answer only from the college’s own documents, cite the page, say “I don’t know” when nothing matches and hand the question to the office with the full chat. Measure honesty, not just coverage.
- Clean the Lead List Before You Write One EmailB2B teams polish cold copy and paste in whatever list they have. We argue the list matters more: bad addresses, duplicates and messy phone numbers cause silent failures and can harm your inbox. Merge, dedupe, verify and normalise first, then personalise and send.
- Claims Stall on the Second Document RequestMuch claims delay comes from the document loop, not hard decisions: generic lists, documents checked days later, then a second request. Ask for what each claim needs, collect it on WhatsApp, read it on arrival and say what is missing at once. Adjusters then spend their time on judgement, and status calls fall.
- Children’s Data Comes Before the DashboardSchool monitoring builds a state-wide record of children, including faces and dropout risk. We argue safeguards must be architecture, not promises: data in India, consent, on-site face matching, logged access, tokenised names, government ownership and human sign-off on decisions about a child.
- Attendance Proof Protects Honest Staff TooVerified attendance is often sold as a way to catch absentees. We argue the honest employee gains most: a timed check-in with selfie and geofence, recorded late reasons and task proof settle payslip disputes fairly, as long as collection is proportionate and explained plainly.
- Why an AI Tutor Should Never Do Mental MathsLanguage models explain well but are unreliable at arithmetic, and in commerce and finance the numbers are the subject. We argue an AI tutor should take every figure from a visible calculator or rate table, which is how Teach-Back works, while never doing graded work for the student.
- An AI Warning Is Not an Alarm, and Shouldn't Pretend to BeAI that predicts which freezer will breach next is useful, but it is an estimate, not a fact. We keep alarms on configured limits, escalating until acknowledged, and let AI warnings sit beside them to move work into daylight and suggest causes. AI should never suppress or delay an alarm.
- AI Agents at the Plant Gate: What They Should DoAI agents suit the plant gate because it is full of repetitive checks: ANPR gate events, safety checklists, photo proof and driver questions. They should handle volume and consistency while people keep judgement calls such as releasing a vehicle, attributing delays and resolving conflicting carrier data.
- What We Won't Let an AI Agent Do in a Transport OfficeThe useful question about an AI agent in a transport office is what it may never do alone. Ours never invents a number, never decides access, never moves money without a confirmed reply and a PIN above a threshold, never starts in Auto and never works without a log. That is what lets owners trust it.
- A Silent Pump Is Not a Running PumpA pump or rain gauge that stops reporting is often drawn as running or dry, which hides the biggest risk in a city control room. Show “no telemetry” as its own state, set a freshness rule, give every silent sensor an owner and review blind spots after each storm.