The dashboard nobody has opened since the launch demo, and the one screen a shift cannot run without

Signs the screen on the wall is not a command centre

Each of these starts as a good dashboard. What goes wrong is not the chart, it is that nothing on the screen tells anybody what to do next.

1. 01Nobody has opened it since the launch demo

   The link still works and the data still refreshes. Look at the access logs and the last real visit was the week it went live, which was the week everyone was told to look at it.

2. 02It shows yesterday to people deciding today

   The load lands at six in the morning and the decisions that matter get taken at eleven. A number that is correct and late is worth the same as a number that is wrong.

3. 03Forty tiles where four would do

   Every request from every stakeholder got its own chart, because adding one was easier than saying no. The tiles that matter are now somewhere in the middle of the third row.

4. 04The floor has muted the alerting

   An alert that fires on an ordinary Tuesday teaches people to ignore it. By the time a real breach fires, ignoring it is the habit rather than the exception.

5. 05The real command centre is a WhatsApp group

   A supervisor who knows everything, a phone and a group chat. It works, it is faster than the dashboard, and it walks out of the building when that person takes leave.

6. 06Nothing on the screen has a name against it

   Every tile reports how things are going. None of them says who will do something about it, by when, or what happens if they do not.

If three or more of these are true, adding another chart makes it worse. What is missing is a decision surface, not another view of the same numbers.

What the floor says, and the tile it turns into

Nobody asks for a command centre. They describe a shift that went wrong and a phone call that should have come two hours sooner.

Left is how an operations lead describes the week. Right is what has to exist on the screen for that week to go differently. Illustrative phrasings used to explain the method, not a client's incident log.

Five surfaces, one discipline

Different floors and different breaches, held to the same rule: a tile earns its place by changing what somebody does in the next hour.

1. Shift floorThe screen a supervisor runs the day from01

   Open exceptions, who holds each one, how long it has been open and what is about to breach. Totals live somewhere else. This screen is about the things that are not going to plan while there is still time to change them.

2. Wall displayLegible from across the room, to somebody walking past02

   A different design problem from a desk screen. Large type, few states, no hover, and nothing that needs a click before it means anything. It should be boring on a good day, which is the only way it is unmistakable on a bad one.

3. Exception deskA queue you claim from, not a list you scroll03

   Work arrives as items with an owner and a clock. Claiming one takes a single action and takes it off everybody else's screen, so two people stop working the same problem without knowing it.

4. Decision surfaceThe action taken where the signal appeared04

   Reassign, hold, release, override, call. The control sits beside the number that caused it. Every system you have to open in a second tab is a place where the response stops and the exception ages.

5. HandoverThe state one shift hands to the next05

   What is open, what is claimed, what breached and what was decided about it, carried across the change of shift by the screen instead of by a conversation that nobody wrote down.

What comes off the screen

The hard part of this work is subtraction, and it is the part that gets skipped, because adding a chart is easier to invoice for than taking one away.

1. 01Watch a shift before drawing anything

   Time on the floor, standing behind the person who will use the screen. What they open, what they ignore, and what they pick up the phone about instead.

2. 02Write the decision, not the metric

   Each candidate tile gets one sentence: who looks at this, what they do when it moves, and by when. A tile that cannot finish that sentence does not go on the screen.

3. 03Set the threshold with the people who breach it

   The line that turns a reading into an exception is an operational decision, not a statistical one. It gets agreed with the supervisor and written down where they can find it and change it.

4. 04Give every exception one owner and a clock

   Not a team and not a distribution list. A person, or a named role on the roster. The clock starts when the exception appears, not when somebody happens to notice it.

5. 05Cut everything that survived on politeness

   The tiles that exist because a stakeholder asked come off, move to the reporting surface where they belong, and the person who asked is told where to find them.

6. 06Run it for a fortnight and count what nobody touched

   Any tile nobody acted on in two weeks is either wrong or unnecessary. We fix it or we take it off.

Where this ends and the neighbouring work begins

Three things get confused with each other on most first calls, so it is worth being plain about which one you are actually buying.

1. Not this pageAntvia is the product you report fromBI

   A lakehouse and the BI layer over it: dashboards, exploration and the pack that goes to the board. If what you need is a reporting layer over governed data, that is the product, and this page is not it.

   See Antvia ↗
2. Not this pageData agents watch a metric and explain the movementAgents

   An agent notices that something moved, works out what is driving it and writes the explanation. That is a producer of signals. A command centre is where a signal becomes somebody doing something about it.

   Read about data agents ↗
3. This pageThe human surface the shift acts onFloor

   One screen, exceptions rather than totals, an owner and a clock against each one, and the action available in the same place. Built for the person on shift, not for the person writing the monthly.

4. TogetherYou usually need more than one, in a particular orderOrder

   The signal often comes from an agent, the history usually lives in the reporting layer, and the acting happens here. Week one says which of the three you need first, and it is not always this one.

From breach to record

The path one exception takes, from the moment a threshold is crossed to the moment somebody can prove what was done about it.

1. SignalSomething crosses a line

   A threshold agreed in advance, on data fresh enough that the decision it serves is still open. Not a daily average, and not a total.

2. ClaimOne person takes it

   The exception lands in a queue with an owner and a clock. Claiming it is one action, and it disappears from everybody else's screen the moment it is claimed.

3. ActThe response happens here

   The response is available on the screen the exception appeared on, so nobody has to log into a second system to answer a question the first one raised.

4. EscalateSilence is an event

   If the clock runs out with nothing done, it moves to a named person on the roster. Nothing is allowed to expire quietly and nothing is closed by being scrolled past.

5. RecordWhat was done, and by whom

   Every claim, action and escalation is written down as it happens. That record is what makes the handover short and the argument about last Tuesday shorter.

What happens overnight, when nobody is watching

Most floors are not staffed round the clock, and most alerting behaves as though they are. Three levels, agreed in daylight, so the night is not one long judgement call.

1. Level 1Hold it for the morning

   Recorded, queued and waiting on the first screen the early shift opens. Most things belong here, and putting them here honestly is what makes the other two levels credible.

2. Level 2Wake a named person

   One name on a rota, not a group address that everybody assumes somebody else is reading. Agreed with the people who will be woken, before the first night it happens.

3. Level 3Stop, automatically

   The small set of cases where continuing is worse than halting. The system takes the safe action itself and tells somebody it did, rather than waiting to be told.

Which breach sits at which level is your operation's decision, not ours. We write it down during the audit week rather than discovering it after the first bad night.

The wall, and the ten minutes at shift change

Two parts of this work get designed last and decide whether the rest of it holds.

What the first six weeks look like

Three phases, always, and the interesting part is what has to be true by week three.

1. Week 1The audit, fixed fee

   Two calls, and a shift followed end to end: what the floor looks at, what it ignores, and where the decisions actually get made. You get a one-pager, yours whether or not we go further.

2. Week 2The first tiles are chosen

   Not forty. The handful of exceptions that change what somebody does in the next hour, each with an owner, a threshold and an escalation written against it before anything is built.

3. Week 3The floor can open it

   Software a real supervisor can run a real shift from, with your own data in it. This is the week we find out which of our tiles was wrong, which is precisely why it is not week eight.

4. Weeks 4 to 6It becomes the screen they work from

   The actions move onto the surface, the escalation path goes live, and the handover starts running off the screen instead of off a conversation. Our median time to a first production-grade artefact is six weeks.

5. Week 7 onwardsOperate

   Quarterly reviews and on-call governance. Thresholds drift as the operation changes, and a threshold nobody revisits is how a floor ends up muting its alerting all over again.

Where week three falls in the six

The point of week three is that a busy person disagrees with us while disagreeing is still cheap.

Six weeks is our median time to a first production-grade artefact, measured across all of our engagements. Week three matters more than usual here, because a command centre is only really tested by somebody who has no time to spare for it.

We have not published a named command centre case yet, so these are firm-wide delivery figures and we will say on the call which of them came from work like yours.

The ones that come up on almost every first call.

• When the spreadsheet is the real system of record, integrate itMigrating the finance model into the warehouse usually stalls. Landing the file as a governed source is faster, with a clear line for when it stops being enough.
• Two dashboards, one metric, two different answersTwo teams, two revenue figures, one warehouse. The tool is almost never at fault, and fixing it is less a modelling problem than an argument nobody wants to have.
• Attribution disagreements are an argument about the model, not the dataMarketing and finance rarely disagree about the events. They disagree about the credit rule, and a rule is something somebody chooses and owns.