KaizenFlow AI user guide

THE PROBLEM

A modern plant produces more data than anyone has time to read: machine states, part counters, downtime reasons, scrap, energy, plans, work orders. The answer to "why did we miss plan this week" is usually spread across five screens and two systems, and someone has to stitch it together by hand.

Plants are data-rich and answer-poor. KaizenFlow AI does the stitching. You ask in plain English. It queries your plant's own data in KaizenFlow, does the arithmetic over the window each data tool covers, and tells you what it found, including when it found nothing.

THE SURFACES

KaizenFlow AI is not one feature. It is a set of surfaces that read the same data under the same rules.

THE RULE UNDER ALL OF IT

Proof, not promises. The assistant is instructed to take every figure from its data tools and never make data up. The tools are built to say "no data" instead of returning a comfortable default, and to label dollar figures that rest on assumed costs as planning estimates. Where that is not yet fully true, this guide says so in place, not in fine print. The full account of the checks lives in How KaizenFlow AI earns trust.

Ask Anything and AI Chat draw on the same 44 data tools. Here is what they cover. Where the data a topic needs is missing, the tool says so.

Two to start with. The prompt library has many more, by role.

• Give me an overview of what data you have for this plantThe fastest way to learn what the AI can answer for youNeeds: any data
• Show me OEE trends for the last 30 daysUsually a line chart over the full window, with a dashed target line where a target is setNeeds: OEE readings

Three things that shape every answer

• Time is rolling, not calendar. Windows count back from now in UTC. The AI is not given today's date, so "yesterday" or "last Tuesday" becomes an approximate number of hours back, not a calendar day in your time zone. Say "the last 24 hours" or "the last 7 days" and the window is unambiguous.
• Machine-level questions are limited. Most tools report at facility level and cannot filter to one machine or line. Equipment appears in ranked lists and recent downtime examples. Only vision inspection filters by line.
• Each tool has a maximum lookback. A longer request is shortened without warning, so check the window the answer states. The limits are in Data it reads.

Every Ask Anything answer has up to three parts:

1. The written answer, formatted with headings, bullets and tables where useful.
2. Charts, when one helps: line, area, bar, scatter, pie (drawn as a donut), gauge or table. A gauge shows the value against its target and turns green when the target is met. Line, area, bar and scatter charts draw a target, when there is one, as a dashed line. Any chart can carry a one-line insight underneath. A chart with no data points is not drawn, though an empty table can still show its title.
3. Tools used: a row of chips naming the data tools that ran, for example query_metrics or get_downtime_events.

Reading the basis

There is no separate "sources" box. You read the basis of an answer from three places: the Tools used chips tell you what data was queried, the window is stated in the answer's text or chart title (the model is instructed to match chart titles to the plotted data), and the scope is the facility selected when you asked. If an answer does not state its window, ask again and name one. In the examples below, the Basis line is our annotation assembled from those three places; the product does not print it as one line.

When data is missing or a figure cannot be backed

Each Ask tool is tested in our build against an empty plant: it must return nothing it cannot back, or say plainly that the data is absent. Three tools are still on a known-exceptions list (the predictive-maintenance, value-stream and crew-sizing tools named above), and the build fails if any other tool returns figures for an empty plant without declaring that the data is absent. The test checks that the absence is declared, not that every figure beside it is neutral, which is how the two soft spots below still pass. You will see answers like these:

• "No data found for this metric and period." or "No downtime events found."
• TEEP and utilization reported as unknown, not 0%, when no shift schedule is configured.
• Projections withheld, and cost line items reported as unavailable rather than $0, when they cannot be computed.
• Dollar figures marked as planning estimates when your facility's unit economics are not set, using $10 cost per unit, $25 revenue per unit and $5,000 per downtime hour.

You

What were the top downtime reasons over the last 7 days?

KaizenFlow AI

Over the last 7 days the plant logged 46 downtime events totalling 38.2 hours. 41 events have a recorded duration; the other 5 are counted but not included in the hours.

By time lost, the top three reasons are material shortage (12 events, 11.4 h), tool change (9 events, 7.9 h) and hydraulic fault (3 events, 6.1 h). Unplanned stops are 58% of events but 71% of minutes, so they are fewer and longer than planned ones.

Chart: bar, "Downtime hours by reason, last 7 days".

Basis: Tools used get_downtime_events. Window: rolling 7 days back from now, UTC. Scope: selected facility, all equipment. Method: totals over every event in the window.

You

What's our TEEP this week, and where is the biggest loss?

KaizenFlow AI

TEEP and utilization are unknown for this facility, not 0%. No shift schedule is configured, so scheduled time cannot be separated from unscheduled time.

What can be measured over the last 7 days: OEE is 63.8% and quality is 94.3%, taken from the scrap rate. This facility does not stream availability and performance separately, so they are an estimated split of what remains (about 82.3% each), not measurements. What is solid: availability and performance together cost more than quality does.

The loss waterfall in hours also needs a shift schedule. To get TEEP and the full waterfall, configure this facility's shift schedule.

Basis: Tools used get_teep_breakdown. Window: rolling 7 days. Scope: selected facility. Missing input named: shift schedule. Availability and performance flagged as estimated.

Follow-up questions

AI Chat keeps the recent conversation for each facility, so follow-ups such as "that one" carry context. Ask Anything answers every question on its own.

You

Which equipment had the worst reliability over the last 30 days?

KaizenFlow AI

Plant-wide there were 22 unplanned failures: MTBF 30.3 hours, MTTR 38 minutes. The machine with the most failures was CNC-07: 9 unplanned failures, MTBF 79.1 hours, MTTR 52 minutes.

Basis: Tools used get_reliability_metrics. Window: rolling 30 days. Scope: selected facility. Method: unplanned stops count as failures; equipment ranked by failure count.

You

What have the stops on that one been about?

KaizenFlow AI

Taking "that one" as CNC-07. Its 9 unplanned stops over the last 30 days were 5 tool breakages, 3 low-coolant stops and 1 spindle alarm. This comes from the per-machine breakdown in the reliability data, which covers unplanned stops only; planned stops such as changeovers on CNC-07 are not separated out, because downtime reports cannot be filtered to one machine.

Basis: Tools used get_reliability_metrics. Window: rolling 30 days. Scope: selected facility, CNC-07's top reasons by count, unplanned stops only.

You

Have we tried anything on tool breakage before?

KaizenFlow AI

Two earlier suggestions match. Changing inserts on a tool-life count instead of on failure is marked implemented. A switch to a premium insert grade was rejected, and the recorded reason is cost.

Basis: Tools used find_similar_past_cases, get_suggestion_history. Scope: your company's past suggestions.

In Ask Anything, the second question above would arrive with no context. Restate it in full instead: "What were CNC-07's top unplanned stop reasons over the last 30 days?"

When there is no answer

In Ask Anything, an amber box names why the AI service could not answer, such as "AI service timed out" or "AI provider rate limit". A red "Sorry, I could not process that question." means the request itself failed, most often because it was under 3 or over 2,000 characters, came from a Viewer role, or hit the request limit. The message does not say which, so check those three in that order. "Could not reach the AI service" means your connection dropped. None of these is ever replaced by a made-up answer.

KaizenFlow has nine AI specialists. They work in AI Suggestions, the background analysis that proposes improvements, not in Ask Anything or AI Chat. Each brings the methods of its discipline and a narrower set of tools.

How a run is routed

1. A run starts. Every 4 hours when there is new data, when the anomaly scan asks for a focused look, or on demand.
2. Triage picks specialists from your data, not from words. A specialist joins when the facility has the metrics its lane depends on. Capacity always joins, as do categories where your team has accepted valuable suggestions before.
3. Your settings apply last. Specialists can be switched off per company in Settings under AI Engine, so your company may run fewer than nine. Capacity cannot be switched off.
4. Specialists work in parallel. Each gathers data with its own tools and drafts findings.
5. One synthesis merges the drafts. Cards matched as the same opportunity collapse to the strongest one, so that saving is counted once. The match works on category and wording, so two cards that describe one fix differently can both survive; check look-alike cards before adding up their savings. A quality check can send the run back for up to two more passes.
6. Cards pass the write-side checks. By default, every card is validated and scored before it is saved, and a card that falls short is not saved. Where a card shows its working, figures that contradict their own arithmetic are corrected, and some provably impossible statements are removed.

Why an ensemble here, and one assistant in Ask

Finding improvements is a survey problem. A downtime engineer and an energy engineer see different money in the same plant. A focused lens with its own methods avoids generic advice, and the merge step stops two lenses claiming the same saving.

Answering a question is a direct problem: you want one answer quickly, from whatever data it touches. So Ask and Chat use one assistant with every tool, not a relay between specialists.

Access follows your KaizenFlow role, not your job title: Engineer, Manager and Admin accounts can ask; Viewer accounts cannot. Each prompt below maps to a question the data tools can answer.

Operator

Less time reconstructing the last shift. Asking needs the Engineer role or above, so where operators hold Viewer accounts, a supervisor can run these at handover and share the answer. Training: operator track.

• Summarize the last shift for handoverMetrics, downtime, alerts and tasks from the last 8 hoursNeeds: recent metrics, downtime, alerts
• What alerts fired in the last 24 hours, and are any still unacknowledged?Alerts by severity and type, with how many are unacknowledgedNeeds: alert events
• Any safety incidents or near misses in the last 7 days?Recorded incidents and near missesNeeds: safety event records

Supervisor

The downtime picture and overnight exceptions in one question each, before the start-of-shift meeting. Training: supervisor track.

• What were the top downtime reasons over the last 7 days, by events and by minutes?Top reasons, categories and recent examples with equipmentNeeds: downtime events with reason codes
• Were there any anomalies in the last 24 hours?Readings that broke from their recent baselineNeeds: recent metric history
• Which open actions are overdue against SLA?Escalation status for open action itemsNeeds: action items with SLAs

Quality and continuous improvement

Ask whether a process is in control, what poor quality costs, and what the team already tried. Training: quality and CI track.

• Is our scrap rate in statistical control over the last 30 days? What is the Cpk?Control status and Cp/Cpk; out-of-control points can raise an SPC alertNeeds: scrap rate history
• What is our cost of quality over the last 30 days, and which figures use assumed unit costs?Prevention, appraisal and failure costs, assumptions labelledNeeds: throughput counter, scrap, unit costs
• Which suggestions were rejected, and why?Rejected suggestions, with a reason where one was recordedNeeds: suggestion history

Engineer

Rank where reliability work pays back, and find what the plant's own documents say. Training: engineer track.

• What is our MTBF and MTTR over the last 30 days, and which equipment is worst?Reliability metrics and a worst-equipment listNeeds: downtime with equipment and durations
• Does our OEE vary by day of the week or time of day?Recurring patterns across several weeksNeeds: several weeks of OEE history
• What does our SOP say about changeover on the press line?Relevant passages from manuals your team uploadedNeeds: uploaded SOPs and manuals

Plant manager

The loss structure, plan adherence and a quick sizing of an idea before committing people to it. Training: plant manager track.

• Where are we losing the most time: availability, performance or quality?Loss waterfall in hours and the biggest lossNeeds: metrics and a shift schedule
• How did we do against plan over the last 7 days?Plan vs actual; days with no plan on record are excludedNeeds: ERP plan and actuals
• What if we cut changeover time by 20%?Projection from the measured throughput rate, or withheldNeeds: 7 days of throughput, unit economics

Executive and finance

Separate claimed savings from verified savings, and compare plants on the same definitions. Training: executive track; see also KaizenFlow for finance.

• How much have the AI suggestions we implemented actually saved us?Marked, dated and measured counts reported separatelyNeeds: suggestions with recorded results
• Which plant is performing best on OEE over the last 30 days?Comparison across your own facilitiesNeeds: OEE for more than one facility
• How does our OEE compare to the industry?Benchmark position; assumed economics labelledNeeds: facility OEE

IT administrator

A health check on the data feeding the AI; bad answers usually start with a stale connector. Training: IT administration track.

• Are all our data connectors healthy?Connector status and heartbeatsNeeds: active connectors
• How good is our data? Are any sensors stale or missing?Data-quality scorecardNeeds: connected metric streams
• Give me an overview of what data you have for this plantWhat this facility holds, not the product catalogueNeeds: any facility data

Run the checks in order; most problems are role, window or missing data.

"Sorry, I could not process that question. Please try again."

1. Check your role. This is what a Viewer sees on every question.
2. Check the length. Questions must be 3 to 2,000 characters. Pasted logs often exceed this.
3. Check the pace. A burst of requests from you or your organization can hit the rate limit.

Fix: get the Engineer role or above, shorten the question, or retry after a minute. If the message instead reads Could not reach the AI service. Please check your connection and try again., the request never arrived: check your network and reload.

"Still loading your facility, please try again in a moment."

1. Wait for the app to load. Ask uses the facility selected across the app and has no picker of its own.

Fix: reload, confirm the facility selector shows your plant, and ask again.

An amber box instead of an answer

1. Read the heading. It names the cause: AI billing limit reached, AI provider rate limit, AI configuration problem, Daily AI budget reached, AI service timed out, or AI service unavailable.
2. Match the owner. Billing, configuration and budget need an administrator. Rate limits and timeouts usually clear on their own.

Fix: retry shortly for rate limits and timeouts. A daily budget resets at the next UTC day; its limit is set in server configuration, not in Settings. The page never substitutes demo answers.

"I couldn't finish gathering the data for that just now."

1. Count the topics. A question spanning downtime, quality, energy and plan needs many data rounds. The AI gets a fixed number of rounds, and a data step that runs past its time limit ends in this message.

Fix: split it into one question per topic, each with a named window.

The spinner runs for a long time

1. Know what is happening. Analyzing your question with AI tools... stays until the whole answer is ready; nothing is streamed.
2. Check the scope. Broad questions trigger several tool rounds, each with its own time limit.

Fix: narrow the question. If it still hangs, reload and ask a smaller version.

"No data found for this metric and period." or a missing figure

1. Ask what data exists. Give me an overview of what data you have for this plant.
2. Check connectors. Are all our data connectors healthy? A stale connector explains most gaps.
3. Check the window against retention. Metric readings are kept only for your organization's retention period, and some questions cap shorter.

Fix: connect or repair the source (see integrations). By design, the data tools are tested against an empty account and must return nothing they cannot back, or say the data is absent, so missing energy data reads unavailable, not 0. The known exceptions are listed under Best practices and Other known limits.

You

What's our energy per unit and carbon footprint this month?

KaizenFlow AI

That figure is unavailable. No power or energy readings are connected for this facility, so I can't compute energy per unit or carbon. Connecting a power series through OPC-UA or MQTT would enable it.

Basis: energy analysis returned no data for this facility; no figure reported

The number does not match my dashboard or report

1. Compare windows. Ask counts rolling hours back from now in UTC, not calendar days in plant time.
2. Check for a cap. A 6-month OEE request is answered over 30 days without a warning. Read the window stated in the answer.
3. Check the facility. In a multi-plant organization, confirm with your admin which facility each connector's data is assigned to.
4. Check the unit. Throughput and part counts are per-hour rates, not raw counter totals.

Fix: restate the question with a window that matches the report.

TEEP or utilization shows UNKNOWN

1. Check the shift schedule. Without one, scheduled time cannot be computed, so the AI reports UNKNOWN rather than 0%.

Fix: configure the shift schedule. The OEE and TEEP guide explains why TEEP needs it.

A dollar figure looks too round

1. Look for the planning-estimate note. Without your unit economics, answers use $10 cost per unit, $25 revenue per unit and $5,000 per downtime hour.
2. Check for a rate. With too little throughput history, projections are withheld rather than guessed.

Fix: set the facility's cost per unit, revenue per unit and cost per downtime hour.

My follow-up ignored the previous answer

1. Check the page. Ask answers every question independently.

Fix: restate the metric, window and scope, or use AI Chat for multi-turn drill-downs.

Other known limits

• With zero downtime records, reliability reports 100% availability. Read it as "no downtime data".
• Alert totals stop at 50 in a busy window. Ask for a shorter window.
• Root cause analysis ignores a requested window today.
• A chart with no data points is not drawn, and AI Chat draws no charts.
• Your interface language is not passed to the AI and its instructions are in English, so expect answers in English.