Your energy dashboard has not saved you anything

A facilities director showed us a dashboard last year with eleven months of fifteen-minute electricity data across nine buildings. Beautiful charts. Trend lines, heat maps, a league table. We asked one question: what did you change because of it?

Nothing. Not one setpoint, not one schedule, not one procurement decision. Eleven months of faithfully recorded data and zero recovered cost.

This is the normal case, not the embarrassing exception. The monitoring industry sold a generation of buildings on visibility, and visibility was delivered. What was never delivered was the thing visibility was supposed to enable.

The frustrating part is that the technology works when it is used. The US Department of Energy's Smart Energy Analytics Campaign tracked energy management and information systems across 96 commercial organisations, nearly 6,000 buildings and 518 million square feet of gross floor area, and reported median energy savings of 4 per cent for energy information systems and 9 per cent for fault detection and diagnostics, at a one- to two-year simple payback. Those are medians, which means roughly half the participants did worse — and every participant had signed up to a campaign whose whole point was using the software. The buildings that never joined anything are the ones with the dead dashboard.

How a monitoring deployment stalls, in sequence

Nobody decides to waste the investment. The stall is procedural, and it happens in a recognisable order. If you have lived through it, you will recognise the months.

Month one: the installation succeeds, and that is the problem

Meters are commissioned, points are mapped, the portal goes live, and a training session is run for six people. Everyone involved has now discharged their responsibility and the project closes green. Note what has been proven: that data arrives. Nothing has been proven about whether the data changes anything, because no test for that was written into the project. This is where the accountability gap opens — the deployment had a sponsor, a budget line and a completion date, and the use of it has none of the three.

Month three: the first real finding appears, and lands nowhere

Someone notices that a building draws two-thirds of its weekday peak at three in the morning. This is a genuine finding. Now watch it travel.

The energy manager spotted it, but does not control the plant. The facilities manager controls the plant, but is measured on comfort complaints and uptime, not on kilowatt-hours. The controls contractor could change the schedule, but is on a reactive call-out arrangement and will raise a variation. The finance business partner could fund the variation, but wants a savings estimate, and the energy manager cannot produce one with an error bar on it. The building's occupier is on a service charge that passes the energy cost through, so nobody in the building has a personal reason to care.

Every one of those people is behaving rationally. The finding dies of organisational friction rather than disagreement: nobody disputes it, it simply has no owner, and an observation without an owner is a rumour.

Month six: the logins stop

A dashboard that has never produced an outcome is a dashboard nobody opens. The alert emails go to a distribution list and become a filter rule. Where the system alerts on thresholds — and most do — the noise makes this inevitable: an alert that fires on a hot Tuesday because the chillers worked harder on a hot Tuesday teaches the recipient to ignore alerts. Within a quarter, the people who trust the system are the people who installed it.

Month twelve: the renewal

The subscription comes up. Nobody can point at a recovered cost, but nobody wants to be the person who switched off the monitoring, so it renews and the cycle repeats. This is the quiet failure mode of the whole category: not cancellation, but indefinite renewal of something that has never been asked to justify itself.

Four different jobs, sold as one product

Part of the confusion is vocabulary. "Energy analytics" describes four distinct jobs with different inputs, outputs and actors. A platform that does the first well is often bought in the belief that it does all four.

The four are sequential, and each is worth little without the next. Monitoring without analytics produces charts nobody can interpret. Analytics without diagnostics produces a ranked list of buildings with no instruction attached. Diagnostics without verification produces work orders whose value is never established, which is exactly how a maintenance budget gets cut in the next round. Most deployments buy job one, receive job one, and are disappointed that jobs two through four did not arrive. That is not a vendor conspiracy; it is a category that never separated its terms.

Why the chart does not become a decision

Underneath the sequence, three failures recur, in roughly this order of frequency.

1. The chart shows consumption, and consumption is not actionable

A building used 412 MWh last month. Is that good? Nobody on the screen can say. Consumption only becomes actionable when it is compared against something that accounts for why it moved — weather, occupancy, production, operating hours. Without that, every month looks like weather noise, because mostly it is.

This is not a novel observation. The EPA's benchmarking programme makes the same point in its design: the ENERGY STAR score exists precisely because raw consumption is not comparable, and the score is "normalized for weather and operating characteristics" before any building is ranked against another. The useful question is never "how much did we use", it is "how much more did we use than the conditions justify". That requires a model. Most dashboards do not have one, and those that do rarely say how well it fits.

2. Nobody is accountable for a number that has no error bar

Ask an operations manager to defend a savings claim to a CFO and watch what happens. If the number comes from a dashboard that will not say how confident it is, what it excluded, or what it did when a meter dropped out for six hours, the manager will not stake their credibility on it. So the chart stays a chart.

A number an operator will act on is a number they can survive being wrong about — which means it arrives with its own provenance. That is our whole position on measurement integrity: a reading is measured or refused, and a gap is shown as a gap with a reason rather than quietly totalled as a zero.

3. The finding has no owner and no clock

"Chiller 2 is running overnight" is an observation. It becomes value when it is assigned to a person, with a date, and a measured before-and-after. Dashboards are excellent at the observation and have nothing to say about the other three.

What an actionable finding looks like

An alert says a threshold was crossed. A finding says what is wrong, where, what it costs, who can fix it, and how you will know afterwards that it was fixed. The difference is not tone — it is the number of fields.

Five worked examples, all found in buildings that already had monitoring. In every case the fix belongs to the customer or their contractor: we measure the deviation, name the plant, and then measure what the intervention was worth. We do not write setpoints and we do not touch the control system.

Out-of-hours base load

What you see: the overnight and weekend floor of the load profile sits at a large fraction of the occupied-hours peak, week after week, with no weather dependence.

Why it is actionable: it needs nothing but a timestamp and a kilowatt-hour reading, and unlike most findings the counterfactual is obvious — an empty building has a defensible target draw, and everything above it is a list of things left running.

What turns it into an instruction: decompose the overnight floor by whatever sub-metering exists, and where there is none, attribute it using the step changes at known switching times. The output is not "your base load is high"; it is "between 22:00 and 05:00 this building draws X kW that nothing is asking for, and a step at 22:30 accounts for roughly half of it."

Simultaneous heating and cooling

What you see: heating energy and cooling energy rising together in the same hours, typically in shoulder seasons, typically in a reheat system.

Why it happens: a supply air temperature set low enough to satisfy the worst zone, with every other zone reheating back up. Control sequences to prevent exactly this are standardised — ASHRAE's Guideline 36 exists so that engineers can adopt sequences already proven to perform rather than write their own. Many buildings predate it, or had it specified and then overridden during a comfort complaint some winters ago.

What turns it into an instruction: quantifying the overlap in energy terms per hour, then per year, so the reheat penalty can be set against the cost of a controls engagement. Without that arithmetic it is a chart; with it, a business case.

Schedule drift

What you see: the daily start time creeping earlier, or a zone that no longer stops. Typically the residue of a one-off override — a weekend event, a commissioning test, a contractor who put something in hand and never put it back.

Why it is missed: it is invisible in monthly totals and in a year-on-year comparison, because it drifts slowly and the weather moves more than it does. It shows up only when you overlay the load profile on the declared occupancy schedule — which requires knowing that schedule, a piece of information that in most estates lives in a person rather than a system.

Economiser failure

What you see: mechanical cooling running during mild outdoor conditions when free cooling should be available, or the mirror image, an economiser stuck open and dragging in hot humid air. In tropical and subtropical climates the second case is the expensive one.

Why it is a good example: a seized damper actuator raises no alarm, because no alarm exists for it. The only evidence is the relationship between outdoor conditions and cooling energy — which is to say the fault is detectable only analytically, from data the building has been logging all along.

A stuck valve showing as flat consumption

What you see: a load that should modulate and does not. Consumption that is too stable — the same figure hour after hour, across conditions that ought to move it.

Why almost nothing catches it: every threshold alert in the building looks for a value that is too high, and a plausible constant sits inside every threshold, forever. The monitoring system is working perfectly and recording a fault it has no concept of. Treat a suspiciously invariant series as a data-quality refusal rather than a reading — see why null is not zero.

The economics: the capital was spent years ago

Here is the part that gets missed. The buildings with the dead dashboards are usually the best candidates, not the worst. They already have the expensive part: instrumentation, a historian, points that have been logging for years. What they lack is the cheap part — the analysis layer that turns those points into a ranked list of things worth doing.

The payback arithmetic follows from that. Metering, cabling, gateways and commissioning are capital, procured on a capital cycle, depreciated over years. Analysis is an operating cost applied to an asset that already exists, so its payback is not a function of hardware lead times. Berkeley Lab's meta-analysis across three decades and 1,500 North American buildings found a median simple payback for commissioning existing buildings of 1.7 years, with a 25th–75th percentile range of 0.8 to 3.5 years, and median primary energy savings ranging from 5 per cent under utility programmes to 9 per cent for monitoring-based commissioning and 14 per cent for projects outside utility programmes. Commissioning is not continuous analysis, but the shape is the same: most of the value is in finding what the building is already doing wrong, not in buying it anything new.

We routinely find that a portfolio's existing meters, with no new hardware at all, support:

• a weather-normalised baseline per building, good enough to certify against IPMVP Option C;
• an occupied-hours split that shows what the building costs when nobody is in it;
• plant-level efficiency trending where a chiller has both a power meter and a flow meter;
• a defensible Scope 2 figure, because the underlying kilowatt-hours are already half-hourly.

None of that needs a procurement cycle. It needs someone to read the points that already exist and refuse to guess at the ones that do not. That is the entire premise of our work on estates that already have monitoring.

The test for a monitoring investment

One question, and it is not about data quality or dashboard design:

In the last quarter, how much recurring cost was removed because of something this system showed us, and can we show the before and after?

If the answer is zero, adding more sensors will not change it. The constraint is downstream of measurement. It is the step where a measured deviation becomes a named action with a verified result — and that step is a product, not a chart.

An audit you can run this week

None of this requires a vendor, a trial or a budget. Five days, one person, using what you already own.

1. Inventory what actually logs, not what is supposed to. Not the BMS points list — the points that produced data in the last 30 days, at what interval, with what gaps. A point that stopped reporting in March and was never noticed is your first finding, and it is free.
2. Pull one clean year from the building with the best history. Fit a baseline against weather and operating hours, and check how well it fits before anything else. A model that cannot explain the past will not adjudicate the future.
3. Plot the overnight floor for every site, side by side. One chart, no normalisation needed. Rank the estate by out-of-hours draw as a proportion of occupied peak; the top three are your candidate list.
4. Attribute the floor at the worst site. Sub-meters if they exist, switching-time step changes if they do not. Get to named plant, not to a building.
5. Write one finding down properly and assign it. A person, a date, an expected effect, and the measurement you will make afterwards to settle whether it happened. Then make that measurement. One completed loop tells you more than another year of dashboards.

If step five stalls — and it stalls in most estates — you have learned the useful thing: the constraint is not instrumentation. No sensor purchase fixes an unassigned finding.

What to ask a monitoring vendor before you renew

Renewal is the only moment of leverage most buyers get. Six questions, and the quality of the answers matters more than the answers.

• "Show me a finding this system produced in our estate, and what happened to it." Not a feature; an outcome, with a date and a name attached. If there is none in twelve months, that is the answer to every other question.
• "What does the platform do when a meter stops reporting?" The correct answer is that it says so, and refuses to total the period as though the missing intervals were zero.
• "What is my coverage percentage for last month, per site?" If the system cannot produce this instantly, it does not know what its own numbers were computed from, and neither do you.
• "Can I export the underlying intervals and reconcile them against my utility bills?" If the only way to check a figure is to ask the vendor, the figure is testimony rather than evidence.
• "Against what is my consumption normalised, and how good is the fit?" A platform comparing this month to last month is comparing two different weather events.
• "When we make a change, how does this system establish what it was worth?" Listen for a baseline method and a treatment of uncertainty. Listen for whether it will decline to claim a saving it cannot substantiate. A system that always produces a number is not more capable; it is less honest.

The last one separates the categories. Savings are, as the IPMVP puts it, "the absence of consumption or demand" — they cannot be measured directly, only inferred by comparing before and after with adjustments for changed conditions. A product that hands you a savings figure without saying what it adjusted for has made a choice on your behalf and not told you.

The wider picture, briefly

This is not only a building-level problem. The IEA's Energy Efficiency 2025 reports global efficiency progress improving to 1.8 per cent in 2025, up from 1 per cent in 2024, while noting that recent average progress "remains well below the global doubling target agreed at COP28". National efficiency is the sum of a very large number of buildings in which somebody either did or did not act on something they could already see.

The data is already there. That is genuinely the good news — it means the payback period on the analysis is measured in weeks, because the capital was spent years ago. What has to change is not the instrumentation. It is that a finding acquires an owner, a date, and a number afterwards that somebody is willing to defend.