Docs / KaizenFlow AI
The prompt library
Copy-ready questions for the work plants actually do: downtime, OEE, energy, quality, maintenance, shifts, and savings. Each one says what comes back and what data it needs, so you know before you ask.
How to use this library
This page collects about 100 distinct questions you can type into KaizenFlow AI today, grouped by the job you are trying to do, plus routine packs that reuse the best of them. Every one maps to a data view the assistant can actually query. We left out question types it cannot yet answer reliably, and we say so where it matters.
Type the prompts into Ask Anything (under Analytics & AI in the sidebar), which answers with text and, where useful, a chart. For a back-and-forth conversation, use AI Chat (under Communication), which remembers recent turns, where your deployment includes it. The KaizenFlow AI guide explains both in full.
READING A PROMPT CARD
- The prompt is written the way we would type it. Copy it as is, or swap in your own window, metric or line.
- The note says what comes back, so you know whether the answer will be a number, a ranked list, a breakdown or a chart.
- The Needs tag names the data the answer depends on. If that data is not connected for the facility you have selected, the data views behind these prompts are built to report that they have no data, rather than fill the gap with a made-up number. Two exceptions we know of, reliability and TEEP quality, are flagged in the sections below.
About the starter questions
Before your first question, Ask Anything shows eight starter questions. Two of them, failure prediction and comparing energy across plants, touch areas this library leaves out because we cannot yet vouch for the answers. Prefer the history-based versions on this page.
Scope comes from the facility selector
Questions are answered for the facility selected in the app's global facility selector, apart from cross-plant comparisons such as ranking your plants. Ask Anything has no picker of its own, so check which plant is selected before you start.
Who can ask
Engineers, managers and admins can ask questions. Viewers can see Ask Anything in the menu, but their questions return a generic error. If that happens to you, ask an admin to review your role.
Anatomy of a strong prompt
A strong prompt has four parts: scope (where), measure (what), window (how far back) and comparison (against what). You rarely need all four spelled out, but each one you leave to chance is one the assistant has to guess.
Why each part matters
- Scope. The facility comes from the selector. Most data views cannot filter to a single machine or line, so per-machine questions work best as ranked lists: which equipment has the worst MTBF rather than what is the MTBF of Press 4. Vision inspection is the exception and accepts a line.
- Measure. Name the metric: OEE, availability, performance, quality, throughput rate, scrap rate, unplanned downtime hours, MTBF, Cpk, energy per unit. Counters such as throughput and parts rejected come back as per-hour rates, and machine states come back as a share of readings in each state, not as an average.
- Window. Windows roll back from the moment you ask, in UTC. Words like yesterday, last Tuesday or this shift are approximated as hours or days back, not snapped to calendar or plant-time boundaries. Say last 24 hours or last 8 hours and you get exactly that. Weekday and hour-of-day patterns are read from the stored timestamps too, so check the offset from your plant's local time before you match a pattern to a shift start.
- Comparison. A number without a reference is hard to act on. Good references are the target, the plan, the trend direction, a longer window that also ends now, another plant in your organization, or the industry benchmark.
| Weak prompt | Strong prompt | Why the strong version works |
|---|---|---|
| How's OEE? | What's our OEE over the last 7 days, and is it above or below target? | Names the window and the reference. The value is the same windowed mean the KPI dashboards use. |
| Why was yesterday bad? | What were the top downtime reasons over the last 24 hours, by minutes? | A rolling 24 hours is exact. Yesterday is approximated, and bad is not a measure. |
| How's this shift going? | Summarize the last 8 hours for handover. | Shifts are not resolved from a calendar, so give the hours. The handover view pulls metrics, downtime, alerts and tasks together. |
| What's the OEE on Press 4? | Which equipment has the worst MTBF over the last 30 days? | Most views cannot filter to one machine. Ranked lists name the equipment for you. |
| Show me six months of OEE. | Show me the OEE trend for the last 30 days as a line chart. | Metric trends look back at most 30 days, and longer requests are shortened without warning. Asking inside the limit avoids a surprise. |
| Downtime last week vs the week before? | What were our unplanned downtime hours over the last 7 days compared with the last 30 days? | Windows end now, so a window that starts in the past is hard to express. A short window against a longer baseline is reliable. |
| Are we any good? | How does our OEE compare to the industry benchmark? | Names both the measure and the comparison. |
| And last week? | What was our scrap rate over the last 7 days? | Ask Anything answers each question on its own, so a follow-up fragment has nothing to attach to. |
| Optimize our schedule. | What was our measured changeover time and utilization over the last week? | No schedule optimizer is connected. The scheduling view returns measured performance and flags optimization as unavailable. |
| How much money are we losing? | What is our cost of quality over the last 30 days, and which unit costs is it based on? | Picks a specific money view and asks it to disclose whether your unit economics or assumed defaults were used. |
| What do you track for us? | Which metrics do we actually have data for at this plant? | Separates your plant's real data from the product-wide list of metric types KaizenFlow supports. |
How far back each question type can look
Each data view has a maximum lookback. Ask for more and you get the maximum, so read the window stated in the answer or the chart title. Your organization's data retention setting can also limit how much history exists.
| Question type | Longest window |
|---|---|
| Metric values and trends: OEE, availability, performance, quality, throughput, scrap, energy readings | 30 days |
| SPC and Cpk | 30 days |
| Anomalies, vision inspection | 7 days |
| Downtime events, plan vs actual, cross-system correlation | 90 days |
| Cost of quality | 180 days |
| Energy analysis, safety events | 365 days |
One or two asks per question
The assistant gathers data in a limited number of rounds before it must answer. A question that bundles five topics can run out of rounds. The assistant then tries to answer from what it has gathered, and if that also fails you see I couldn't finish gathering the data for that just now. Two related asks, like a value and its trend, are the sweet spot. Questions must also be between 3 and 2,000 characters.
The example below shows a strong prompt and the kind of answer it earns. On this page, the Basis line under each example is our annotation of where the numbers come from. In the product, the window appears in the answer text and chart title, and a Tools used row lists the internal names of the data tools it called. There is no separate sources panel, so ask for the window and basis in your question when they matter.
What's our OEE over the last 7 days, and is it above or below target?
OEE averaged 64.8% over the last 7 days, below the 75% target. Readings were under target 81% of the time, and the trend across the week is flat rather than recovering. A gauge of OEE against target is shown below.
Basis: OEE readings for the selected facility, last 7 days rolling in UTC, windowed mean on the same basis as the KPI dashboard
Downtime and stoppages
Downtime questions draw on your recorded downtime events, with up to 90 days of history. Totals are computed over the whole window, not a sample, and the answer separates events with a recorded duration from events without one, so gaps in your data are visible rather than hidden. For method, see the downtime reduction playbook.
- What were the top downtime reasons over the last 7 days?Top reasons ranked by events and by minutes, plus a breakdown by category.Needs: downtime events with reason codes
- How many unplanned downtime events did we have in the last 30 days, and how many hours?Event count and total hours for the unplanned category only.Needs: downtime events categorized as unplanned, with durations
- Break down downtime over the last 30 days into planned, unplanned, changeover and maintenance, by minutes.Minutes and event counts per category, often drawn as a bar or pie chart.Needs: categorized downtime events
- Show the downtime breakdown by category over the last 30 days as a pie chart.A donut chart of downtime minutes by category, with a one-line insight underneath.Needs: categorized downtime events
- How much changeover downtime did we log in the last 14 days?Changeover events, total hours and the top changeover reasons.Needs: downtime events categorized as changeover
- Which downtime reasons cost us the most minutes over the last 90 days?The top reasons by minutes across the longest downtime window available.Needs: 90 days of downtime events with reasons
- Show me the most recent downtime events with equipment and reason.Up to 15 recent example events, each with its equipment, reason and duration.Needs: downtime events with equipment attribution
- How many downtime events did we have per hour over the last 30 days, counting only hours that had events?Mean, minimum and maximum events per hour across the hours that had at least one event, and how many such hours there were.Needs: downtime events
- How many downtime events in the last 30 days have no duration recorded?The split between events with and without a duration, a quick data-hygiene check.Needs: downtime events
- What's our MTBF and MTTR over the last 30 days, and which equipment is worst?Mean time between failures, mean time to repair and a worst-equipment list.Needs: downtime events with equipment and durations
- Have we had suggestions about changeover downtime before, and what happened to them?Similar past suggestions from your organization and their outcomes.Needs: past AI suggestions
Check that downtime exists before trusting reliability
If no downtime events are recorded for the window, reliability figures can read as perfect rather than unknown. Before you act on an MTBF or availability figure, confirm with the first prompt above that downtime events actually exist for the period.
What were the top downtime reasons over the last 7 days?
There were 46 downtime events over the last 7 days, totalling 31.5 hours. Material starvation leads by minutes (612 minutes across 14 events), followed by die change overruns (455 minutes, 6 events) and sensor faults (240 minutes, 11 events). Sensor faults are frequent but short, so they rank higher by count than by minutes. 5 events have no duration recorded and are not included in the hours.
Basis: downtime events for the selected facility, last 7 days rolling, ranked by minutes and by event count
OEE, availability, performance, quality
When you ask for current OEE, TEEP, availability, performance or quality, the data view returns the windowed mean the KPI dashboards use, not a single spot reading. That means the number in the answer should match the number on your screen for the same window. The assistant is instructed to plot the whole window in trend charts, to keep the chart title's window matched to the data plotted, and never to invent or fill in points. For definitions, see the OEE and TEEP guide.
- What's our current OEE?The windowed mean OEE, often shown as a gauge against target.Needs: OEE readings
- Show me OEE trends for the last 30 days.A line chart of the full 30-day series with a dashed target line.Needs: 30 days of OEE readings
- Is OEE over the last 7 days above or below target, and how often did it fall below?Average against target plus the share of readings below target.Needs: OEE readings with a target
- What's our availability over the last 7 days, and is it trending up or down?Average availability, min and max, and the trend direction.Needs: availability readings
- What's our performance over the last 7 days compared with the last 30 days?Two windowed averages side by side, so you can see whether the week is better or worse than the month.Needs: 30 days of performance readings
- What's our quality rate over the last 7 days, and how does it compare to target?Average quality rate against target and the share of time below it.Needs: quality readings with a target
- Where are we losing the most time: availability, performance or quality?A loss waterfall in hours and the single biggest loss.Needs: OEE factor readings and a shift schedule
- What's our TEEP over the last week, and how big is the gap between OEE and TEEP?TEEP, OEE factors and schedule loss in hours. Without a shift schedule, TEEP is reported as unknown, not 0%.Needs: OEE factors and a configured shift schedule
- Does our OEE vary by day of the week or time of day?Patterns by weekday and hour, based on several weeks of history.Needs: several weeks of OEE readings
- Which plant is performing best on OEE over the last 30 days?Your facilities ranked on average OEE with the gap to the best, within your own organization only.Needs: OEE readings recorded against each facility
- How does our OEE compare to the industry?Your OEE against an industry benchmark, with any dollar potential labelled by its unit-cost basis.Needs: OEE readings
- What are the likely root causes of our low OEE?A structured root-cause analysis with candidate causes and correlations.Needs: metric history
Root-cause analysis uses its own window
The root-cause view runs over its default window whatever period you name, so we phrase that prompt without one. Read the window stated in the answer before you compare it with other numbers.
Estimated components are labelled
When availability and performance are estimated rather than measured directly, AI Chat is instructed to label them as estimates and not present them as a verified availability times performance times quality breakdown. If you see that label, treat the split as directional.
A perfect quality factor can mean no quality data
If no quality readings exist for the window, the TEEP breakdown can show quality at 100% rather than unknown. Before you read a waterfall with no quality loss as good news, ask What's our quality rate over the last 7 days? and confirm readings exist.
Throughput and bottlenecks
Throughput comes back as a per-hour rate, computed per machine series the same way the KPI screens compute it. If there is not enough history to derive a rate, the data view reports that it cannot compute one rather than returning a raw counter. For finding constraints, the most reliable route is through losses and ranked equipment lists rather than a value-stream map.
- What's our throughput rate over the last 24 hours?Units per hour over the window, matching the dashboard tile.Needs: throughput counter samples spanning several hours
- Is our throughput rate over the last 7 days trending up or down?The rate and its trend direction across the week.Needs: 7 days of throughput counter samples
- Show me the throughput rate trend for the last 30 days as a line chart.A line chart of the full 30-day rate series.Needs: 30 days of throughput counter samples
- How did we do against plan over the last 7 days?Plan vs actual and adherence for completed days. Days with no plan show as no plan on record and are excluded, not counted as 100%.Needs: ERP production plan and actuals
- What's our schedule adherence over the last 30 days?Adherence across completed days that had a plan, and how many days were excluded.Needs: ERP production plan and actuals
- Show me the TEEP loss waterfall for the last week in hours.Schedule, availability, performance and quality losses in hours, with the biggest loss named.Needs: OEE factors and a configured shift schedule
- Which equipment has the worst MTBF over the last 30 days?A worst-equipment list from reliability data, a practical starting point for constraint hunting.Needs: downtime events with equipment and durations
- What happens to output if we cut changeover time by 20%?A scenario projection from your measured throughput rate. If no rate can be derived, projections are withheld.Needs: 7 days of throughput and TEEP data
- Are there any supply chain risks or late suppliers that could hold up output?Supplier and inventory risks, or a clear statement that no supply chain data is connected.Needs: supplier and inventory data
Bottleneck questions we left out
Value-stream bottleneck questions are not in this library because that view cannot yet answer from your own data reliably. Cross-plant throughput comparisons are left out too, because the comparison view does not yet convert throughput counters into rates the way the single-plant view does. Most views also cannot filter to one machine or line. Use the loss waterfall, downtime by reason and the worst-equipment list together to locate the constraint, then confirm it on the floor.
Energy and demand charges
The energy view instructs the assistant to name its coverage: which monitored power series the figure includes and how many measured hours it spans. It must not call the figure metered, utility-grade or a facility total. With no power readings, the energy view marks the figure unavailable rather than returning zero. See energy optimization for the wider picture.
- What's our energy per unit and carbon footprint over the last 30 days?Energy intensity and carbon estimate, with the series and hours they cover.Needs: power or energy readings
- Which power series does our energy figure cover, and for how many hours?The coverage statement on its own, so you know how much of the plant is monitored.Needs: power or energy readings
- What's our energy per unit over the last 12 months?Energy intensity over the longest energy window available.Needs: up to 365 days of power readings
- How does our ISO 50001 scorecard look?Scored dimensions, plus dimensions left unscored for lack of data. If the overall score cannot be computed, you get the share of the rubric covered instead.Needs: power readings and energy data
- Show power draw over the last 7 days as a line chart.A line chart of the power series across the week.Needs: power-draw readings
- When in the day does our power draw peak?Hour-of-day pattern for power draw, based on several weeks of history. Hours follow the stored timestamps, which may not be your plant's local time.Needs: several weeks of power-draw readings
- Is our power draw higher on some days of the week than others?Day-of-week pattern, useful for spotting weekend idle load.Needs: several weeks of power-draw readings
- Were there any unusual spikes in power draw in the last 24 hours?The anomaly scan runs across your metrics and returns only the strongest anomalies, so power-reading anomalies can be crowded out on a busy day.Needs: recent power readings with enough history for a baseline
- What does our sustainability data show for this plant?Sustainability figures you have connected, or a statement that none are available.Needs: sustainability data
Demand charges: timing yes, tariff no
KaizenFlow AI has no access to your utility tariff or bills, so it has no basis for a dollar figure on demand charges. If an energy answer does include a dollar figure, it rests on a fixed generic electricity rate, not your tariff, so treat it as a rough estimate. What the assistant can do is show when peaks happen and how big they are. Take the peak timing to whoever owns your tariff and apply the demand rate there.
Quality, scrap and SPC
Quality questions cover scrap trends, statistical process control with Cp and Cpk, cost of quality and vision inspection. SPC works with or without spec limits: give LSL and USL when you have them. Otherwise it uses built-in default limits for common metrics such as OEE and scrap rate, and leaves out the capability verdict for metrics without a default. Background on method lives in quality and SPC and the quality and CI academy track.
- What's our scrap rate over the last 7 days, and is it getting better or worse?Average scrap rate and trend direction for the week.Needs: scrap rate readings
- Show the scrap rate trend for the last 30 days against target.A line chart of the full window with a dashed target line.Needs: 30 days of scrap rate readings with a target
- What's our scrap rate over the last 7 days compared with the last 30 days?Two windowed averages, showing whether the week is running hot.Needs: 30 days of scrap rate readings
- What's our parts-rejected rate over the last 24 hours?Rejects per hour, computed from the counter the same way the KPI screens do.Needs: parts-rejected counter samples
- Is our scrap rate in statistical control? What's the Cpk?Control status, out-of-control points, Cp and Cpk.Needs: scrap rate history
- Is our fill weight in control, and what's the Cpk with LSL 498 and USL 502?Capability against the spec limits you supply.Needs: fill weight readings
- What is our cost of quality over the last 30 days?Prevention, appraisal and failure costs, with the unit-cost basis stated.Needs: throughput counter and scrap data
- What are the likely root causes of our scrap rate?A structured root-cause analysis on scrap, using its default window.Needs: scrap rate history
- Were there any unusual swings in quality readings in the last 24 hours?The anomaly scan runs across your metrics and returns only the strongest anomalies, so quality-related ones may not all appear.Needs: recent quality readings with baseline history
- What defects did vision inspection catch in the last 24 hours on Line 2?Defect counts and types for that line, if the line is identified the way your vision system records it.Needs: vision inspection records with line identifiers
- How is our vision inspection performing? Any rise in false rejects over the last 7 days?Inspection performance and false-reject movement across the week.Needs: vision inspection records
- Is there a link between maintenance delays and quality drops over the last 30 days?Correlations between metrics from different source systems.Needs: metrics from more than one source system
SPC questions can raise alerts
If an SPC check finds out-of-control points among the most recent readings, it can record an SPC violation alert, unless one was raised for that metric shortly before. So an SPC question can leave a trace in your team's alerts as well as answer you.
Maintenance and anomalies
These prompts work from what has already happened: recorded downtime, reliability history, alerts and statistical anomalies in your readings. Anomaly questions look back up to 7 days. For the reasoning behind anomaly scoring, see anomaly detection.
- Were there any anomalies in the last 24 hours?Metrics that moved outside their normal range, with a health score.Needs: recent readings with enough history for a baseline
- What anomalies have we had over the last 7 days, and which metrics were involved?The strongest anomalies across the longest anomaly window, with the metric each one involves.Needs: 7 days of readings plus baseline history
- Which equipment is most at risk of unplanned downtime in the next 24 hours, based on our downtime history?A risk ranking forecast from your downtime history.Needs: downtime history, ideally 90 days
- What's our MTBF and MTTR over the last 90 days?Reliability over a longer window, with the worst equipment listed.Needs: downtime events with equipment and durations
- How much maintenance downtime did we log over the last 30 days?Events and hours in the maintenance category, with top reasons.Needs: downtime events categorized as maintenance
- What alerts fired in the last 24 hours, and are any still unacknowledged?Alert counts by severity and the unacknowledged ones.Needs: alert events
- How many critical alerts fired in the last 7 days?Severity counts for the week. Counts cap at 50 alerts, so very busy windows read low.Needs: alert events
- Which open actions are overdue against SLA?Overdue actions and their escalation status.Needs: action items with SLAs
- What improvement playbooks are in progress?Active playbooks for this facility, or a statement that there are none.Needs: playbook records
- Has anything like repeated spindle stoppages come up in past suggestions?Similar past cases from your organization's suggestion history.Needs: past AI suggestions
- Does our availability vary by day of the week or time of day?Weekday and hourly patterns that can point to shift or startup issues.Needs: several weeks of availability readings
Keep maintenance questions anchored in history
Component-level failure prediction, remaining-life and crew-sizing questions are not in this library because the assistant cannot yet answer them reliably from your own data. Ask about recorded downtime, reliability and anomalies instead, and use the downtime-risk forecast as a prompt for inspection, not a verdict.
Shifts, schedules and handovers
Shift questions work best with explicit hours. Because windows roll back from now in UTC, last 8 hours is precise where this shift is a guess. The handover prompt is read-only: it summarizes, it does not sign anything off. Supervisors will find more in the supervisor academy track.
- Summarize the last shift for handover.A handover summary, by default of the last 8 hours: metrics, downtime, alerts and open tasks.Needs: recent metrics, downtime, alerts and tasks
- Generate a shift handover for the last 12 hours.The same summary over a 12-hour window, for longer shifts.Needs: recent metrics, downtime, alerts and tasks
- What alerts fired in the last 8 hours that the next shift should know about?Alerts by severity for the shift window, with unacknowledged ones called out.Needs: alert events
- Which metrics moved outside their normal range in the last 8 hours?Anomalies for the shift window.Needs: recent readings with baseline history
- Any safety incidents or near misses in the last 7 days?Recorded safety events and near misses for the week.Needs: safety event records
- How did the most recent completed day compare with plan?Plan vs actual for completed days with a plan on record.Needs: ERP production plan and actuals
- What was our measured changeover time and utilization over the last week?Measured throughput, changeover, utilization and OEE, each with its window and sample count.Needs: throughput, changeover and OEE data
- Is our scrap rate higher at certain hours of the day?Hourly and weekday patterns in scrap, which often line up with shift starts.Needs: several weeks of scrap rate readings
- Where are our skill gaps according to the skill matrix?Gaps from your skill matrix, or a statement that no workforce data is connected and what would enable it.Needs: workforce and training records
- What does our workforce data show on overtime over the last 30 days?Overtime and related workforce figures you have connected.Needs: workforce data
Scheduling reports, it does not optimize
No schedule solver or job list is connected. If you ask the assistant to optimize a production schedule, the scheduling view returns measured current performance and reports that optimization is unavailable. That view does not return an optimized sequence, an improvement percentage or a savings figure, so treat any such figure in an answer as unsupported.
Savings and dollar figures
Dollar answers come from two places. Cost of quality, benchmarks and what-if scenarios price outcomes using your facility's unit economics, or disclosed defaults where those are not set. Savings answers come from your AI suggestions and what was recorded about them. Both are built to show their basis. Finance teams can read more in KaizenFlow for finance.
- What is our cost of quality over the last 30 days, and which unit costs is it based on?Cost of quality plus a statement of whether your unit economics or assumed defaults were used.Needs: throughput counter, scrap data, facility unit economics
- Break our cost of quality over the last 90 days into prevention, appraisal and failure.The three cost categories, with any unavailable line item reported as unavailable, not $0.Needs: throughput counter and scrap data
- What has poor quality cost us over the last 180 days?Cost of quality over the longest window available.Needs: up to 180 days of throughput and scrap data
- What happens to output and revenue if we cut changeover time by 20%?A scenario priced from your measured hourly rate. Without a derivable rate, output and revenue projections are withheld.Needs: 7 days of throughput and TEEP data, unit economics
- What happens if we add an 8-hour shift and cut scrap by 1.5 points at the same time?A combined multi-variable scenario on the same baseline.Needs: 7 days of throughput and TEEP data, unit economics
- How does our OEE compare to the industry, and what is closing the gap worth?Benchmark position and a dollar potential labelled with its unit-cost basis.Needs: OEE readings, unit economics
- How much have the AI suggestions we implemented actually saved us?Savings on three labelled bases: marked implemented, with a recorded implementation date, and verified with a recorded savings figure.Needs: suggestions with recorded actual savings
- Which suggestions have verified savings, and what is the verified total?Suggestions with status verified that carry a recorded savings figure, and the total of those figures. Check how each figure was recorded before you quote it.Needs: suggestions with recorded results
- How do predicted savings compare with realized savings across our suggestions and playbooks?Predicted against realized savings, the realization rate and how open actions are tracking.Needs: suggestion and playbook history
- Where does the Kaizen loop stand right now?The funnel from new to accepted, in progress, implemented and verified.Needs: suggestion history
- Which suggestions are ready to execute, measure or verify?Suggestions that can move to their next stage, grouped by next step.Needs: suggestions and work orders
- Which suggestions were rejected, and why?Rejected suggestions with any reasons your team recorded.Needs: suggestions with rejection reasons
When dollar figures are planning estimates
If your facility's cost per unit, revenue per unit or cost per downtime hour are not set, dollar answers fall back to disclosed defaults: $10 cost per unit, $25 revenue per unit and $5,000 per downtime hour. The assistant is instructed to say so and to call those figures planning estimates, not measurements. Once your unit economics are set, the same prompts return figures based on them.
What verified means
Savings answers report three bases side by side: suggestions marked implemented, suggestions with a recorded implementation date, and suggestions with status verified that carry a recorded actual savings figure. In the assistant's answers, the verified total adds up only those recorded figures, not AI estimates. Verified is a status, not an audit, though. An engineer, manager or admin can set it directly, even without a savings figure, and a recorded figure may have been typed in by your team or calculated automatically from metric changes using standard cost factors rather than your own costs. Some rollups outside the assistant can also count the estimate for a verified suggestion that has no recorded figure. Before you quote a verified total, check how each figure behind it was recorded, and name its basis.
The assistant can act, not only read
If you ask it to execute a suggestion, KaizenFlow AI can create a work order from that suggestion and move it to in progress. There is no separate confirmation step. It refuses suggestions that are flagged for review and still new, and suggestions that are not in new, reviewed or accepted status. Only ask for execution when you mean it. None of the prompts in this library ask it to.
What is our cost of quality over the last 30 days, and which unit costs is it based on?
Estimated cost of quality over the last 30 days is about $38,400, and failure costs make up most of it. Unit volume is measured from your counter increments. Cost and revenue per unit are not set for this facility, so this uses the assumed defaults of $10 cost per unit and $25 revenue per unit. Treat the dollar figure as a planning estimate, not a measurement, until your unit economics are set.
Basis: throughput counter increments and scrap readings, last 30 days rolling; unit economics are assumed defaults, disclosed in the answer
Platform help
KaizenFlow AI can tell you about the data behind its answers: what exists for your plant, whether connectors are healthy, which sensors look stale, and what your own uploaded procedures say. Ask these first when an answer surprises you. It is often the data, not the plant. Connector setup is covered in integrations.
- Give me an overview of what data you have for this plant.A profile of the metrics and data present for the selected facility.Needs: any metric data for the facility
- Which metrics do we actually have data for at this plant?Answered from this plant's own readings, not from the product catalogue.Needs: metric data for the facility
- Which metric types does KaizenFlow support in general?The product-wide catalogue, which the data view marks as not an inventory of your plant's data.Needs: nothing plant-specific
- Are all our data connectors healthy?Connector status based on their recent heartbeats.Needs: active connectors
- Which connectors have stopped sending data?Connectors whose heartbeat has gone quiet, so you know which answers may be stale.Needs: active connectors
- How good is our data? Are any sensors stale or missing?A data-quality scorecard across your metric streams.Needs: connected data sources and metric streams
- What does our SOP say about changeover on the press line?Relevant passages from your organization's uploaded procedures.Needs: uploaded SOPs or manuals
- What does our manual say about lockout before clearing a jam on the packaging line?Matching passages from your uploaded manuals. Always confirm against the controlled document.Needs: uploaded SOPs or manuals
- Have we tried anything like reducing changeover before?Similar past suggestions and what happened to them.Needs: past AI suggestions
How-to questions about KaizenFlow itself
The assistant answers from your plant's data and your uploaded documents. It has no access to KaizenFlow's own how-to documentation, so questions like how do I add a connector are better answered by the KaizenFlow Academy and, for admins, the IT administration track.
When a prompt does not come back as expected
The answer says there is no data. What next?
That is the assistant declining to guess. Ask Give me an overview of what data you have for this plant to see what exists, then Are all our data connectors healthy? to see whether a source has gone quiet. Also check that the facility selector is on the plant you meant.
The window in the answer is shorter than the one I asked for.
Each data view has a maximum lookback, listed in the table above, and longer requests are cut to that maximum. Your organization's data retention setting can also limit how much history exists. Re-ask inside the limit.
I got Sorry, I could not process that question.
The same message covers several causes: a question under 3 or over 2,000 characters, a role that cannot ask (viewer), or too many requests in a short time. Check the length first, then your role, then wait a minute and try again.
I got an amber message about the AI service instead of an answer.
The request failed before an answer came back. The heading names the cause when it can be identified, such as a timeout, a provider rate limit or a daily AI budget, and otherwise reads AI service unavailable. The wording of your question is rarely the cause. Try again later, or ask an admin if the message names billing, configuration or budget.
Follow-up chains
Real investigations take several questions: open wide, drill in, compare, then decide what to do. Each recipe below follows that four-step pattern. How follow-ups behave depends on where you type them.
- Ask Anything answers every question on its own. Earlier questions stay on screen, newest first, until you reload or press Clear history, but they are not sent with the next question. So every step below restates its window and measure. Where a step names a reason or a process, swap in the one your earlier answer surfaced.
- AI Chat carries recent turns forward, about the last 20 messages, so you can shorten later steps to and over the last 30 days?. Restating the window still gives the most predictable answer. Charts appear in Ask Anything; AI Chat answers in text.
AI Chat history is shared with your team
AI Chat keeps one conversation per facility for your organization, not one per person. Recent turns from everyone chatting about the same plant are carried into the next answer, and Clear Chat clears that history for everyone. It expires 7 days after the last message, or sooner if the server's history store is unavailable.
1. Downtime spike
- Open. What were the top downtime reasons over the last 7 days? Find the reason that leads by minutes, not just by count.
- Drill down. What's our MTBF and MTTR over the last 7 days, and which equipment is worst? Tie the reason to the equipment it hits.
- Compare. Which downtime reasons cost us the most minutes over the last 90 days? Is this week's leader a new problem or a chronic one?
- Act. Have we had suggestions about material starvation downtime before, and what happened to them? Swap in your leading reason, and start from what your team already tried.
2. OEE below target
- Open. What's our OEE over the last 7 days, and is it above or below target?
- Drill down. Where did we lose the most time over the last week: availability, performance or quality? Name the biggest loss.
- Compare. Does our OEE vary by day of the week or time of day? A loss that clusters on certain shifts points to a different fix than one spread evenly.
- Act. Which suggestions are ready to execute, measure or verify? Check whether an existing suggestion already targets that loss.
3. Scrap running hot
- Open. What's our scrap rate over the last 7 days, and is it getting better or worse?
- Drill down. Is our scrap rate in statistical control? What's the Cpk? Separate special causes from a capability problem.
- Compare. What's our scrap rate over the last 7 days compared with the last 30 days?
- Act. What is our cost of quality over the last 30 days, and which unit costs is it based on? Size the problem before you commit people to it.
4. Energy peaks
- Open. What's our energy per unit and carbon footprint over the last 30 days? Note the coverage statement.
- Drill down. When in the day does our power draw peak?
- Compare. Is our power draw higher on some days of the week than others? Weekend load often reveals equipment left running.
- Act. Take the peak timing to whoever owns your utility tariff, then check the effect in a month with What's our energy per unit over the last 30 days?
5. Changeover business case
- Open. What was our measured changeover time and utilization over the last week?
- Drill down. How much changeover downtime did we log in the last 30 days?
- Compare. Have we tried anything like reducing changeover before? Past outcomes are the best reality check.
- Act. What happens to output and revenue if we cut changeover time by 20%? Read the unit-cost basis before you quote the number.
6. Anomaly triage
- Open. Were there any anomalies in the last 24 hours?
- Drill down. What alerts fired in the last 24 hours, and are any still unacknowledged?
- Compare. How good is our data? Are any sensors stale or missing? Rule out a data problem before you send someone to the machine.
- Act. What does our SOP say about startup checks on the filler? Swap in the affected process, and go with the procedure in hand.
7. Month-end savings proof
- Open. How much have the AI suggestions we implemented actually saved us? Note the three bases.
- Drill down. Which suggestions have verified savings, and what is the verified total?
- Compare. Which suggestions were rejected, and why? What you turned down is part of the record.
- Act. Which suggestions are ready to execute, measure or verify? Close the gap between implemented and verified before next month.
Here is chain 2 as it might run in AI Chat, where later turns can lean on earlier ones.
What's our OEE over the last 7 days, and is it above or below target?
OEE averaged 64.8% over the last 7 days against a 75% target, and it sat below target 81% of the time.
Basis: OEE readings, last 7 days rolling, windowed mean
Where did we lose the most time over that week: availability, performance or quality?
Availability is the biggest loss at 38.6 hours for the week, ahead of performance at 21.2 hours and quality at 4.9 hours.
Basis: loss waterfall in hours, last 7 days, using the configured shift schedule
Does it vary by day of the week?
Yes. Over the last 12 weeks, OEE on Mondays runs about 9 points below the weekly average, and the gap is concentrated in the first hours of the day. That pattern fits restart losses more than a single failing machine.
Basis: OEE by weekday and hour, 12 weeks of history
Routine packs
Routines turn the assistant from something you try into something you rely on. Run the same short set at the same moment each shift, week or month, and the answers become comparable over time. Asking takes an engineer, manager or admin role, so where operators hold a viewer role, a supervisor or engineer runs the routine and shares the answers at the huddle. Adapt the packs with the operator track and the operator adoption guide.
Daily shift start
5 MINUTES, BEFORE THE HUDDLE
- Are all our data connectors healthy?Confirms the data behind the rest of the routine is flowing.Needs: active connectors
- Summarize the last 8 hours for handover.What the previous shift left you: metrics, downtime, alerts and tasks.Needs: recent metrics, downtime, alerts and tasks
- Which alerts from the last 8 hours are still unacknowledged?Open alerts to pick up first.Needs: alert events
- Were there any anomalies in the last 8 hours?Readings drifting out of their normal range.Needs: recent readings with baseline history
- Any safety incidents or near misses in the last 24 hours?Safety items to raise at the start-of-shift talk.Needs: safety event records
- Which equipment is most at risk of unplanned downtime in the next 24 hours, based on our downtime history?Where to look first on the walk.Needs: downtime history
End-of-shift handover
LAST 15 MINUTES OF THE SHIFT
- Generate a shift handover for the last 8 hours.The draft handover to review and add your own notes to.Needs: recent metrics, downtime, alerts and tasks
- What's our OEE over the last 8 hours compared with the last 7 days?Whether this shift ran above or below the recent norm.Needs: 7 days of OEE readings
- What's our scrap rate over the last 8 hours, and is it above target?Scrap for the shift against target.Needs: scrap rate readings with a target
- What unplanned downtime did we have in the last 24 hours, by reason?Unplanned events and minutes by reason to pass on.Needs: downtime events with reason codes
- What alerts fired in the last 8 hours, and are any still unacknowledged?What the next shift inherits.Needs: alert events
- Which open actions are overdue against SLA?Actions to escalate rather than hand over quietly.Needs: action items with SLAs
Weekly production review
ONCE A WEEK, SAME DAY
- How did we do against plan over the last 7 days?Plan vs actual and adherence, with no-plan days excluded.Needs: ERP production plan and actuals
- Where did we lose the most time over the last week: availability, performance or quality?The loss waterfall and the biggest loss.Needs: OEE factors and a shift schedule
- Show me OEE trends for the last 30 days.A line chart that puts the week in context.Needs: 30 days of OEE readings
- What were the top downtime reasons over the last 7 days?Top reasons by events and by minutes.Needs: downtime events with reason codes
- What's our scrap rate over the last 7 days, and is it getting better or worse?Scrap level and direction.Needs: scrap rate readings
- What's our MTBF and MTTR over the last 30 days, and which equipment is worst?Reliability trend and the equipment to discuss.Needs: downtime events with equipment and durations
- Which suggestions are ready to execute, measure or verify?Improvement work to move forward this week.Needs: suggestions and work orders
Monthly finance close
FIRST WORKING DAY OF THE MONTH
- What is our cost of quality over the last 30 days, and which unit costs is it based on?Cost of quality and whether it rests on your economics or on defaults.Needs: throughput counter, scrap data, unit economics
- How much have the AI suggestions we implemented actually saved us?Savings on the three labelled bases.Needs: suggestions with recorded actual savings
- Which suggestions have verified savings, and what is the verified total?The recorded-results total. Check how each figure was recorded before you defend it in a review.Needs: suggestions with recorded results
- How did we do against plan over the last 30 days?Monthly adherence across completed days with a plan.Needs: ERP production plan and actuals
- What's our energy per unit and carbon footprint over the last 30 days?Energy intensity and carbon, with coverage stated.Needs: power or energy readings
- How does our OEE compare to the industry, and what is closing the gap worth?Benchmark gap priced on a stated unit-cost basis.Needs: OEE readings, unit economics
- Which suggestions were rejected, and why?The decisions behind the numbers, for the record.Needs: suggestions with rejection reasons
Keep your routine somewhere durable
Ask Anything's on-screen history is cleared when you reload, and there is no built-in export. Keep your routine prompts in a shared note so the whole team runs the same set, and copy any answer you need for a report at the time you ask.
To understand how answers are checked and what happens when data is missing, read how KaizenFlow AI stays honest. For terms used on this page, see the glossary.
Bring your own questions
In a demo we run the questions you care about against a representative plant. In a pilot, we run them against yours.