Trust sets the pace for AI agents in energy efficiency

Energy efficiency depends on how fast a team can find, quantify and act on a loss. AI agents shorten that path, and trust decides whether the result gets used.

Key Highlights

  • Industrial steam systems offer significant efficiency potential, but manual data gathering and analysis can keep worthwhile investigations from getting started.
  • Analytics tools and agents can help engineers answer questions ranging from simple averages to comparisons and root-cause investigations.
  • Trustworthy results should be deterministic, transparent, repeatable and saved for colleagues to review, with a person confirming findings before action.
  • Plants can measure program reach through time to finding, cases worked and reuse of analyses across assets, alongside energy and cost savings.
  • Consistent asset structures and reusable calculations can help teams turn individual findings into ongoing, site-wide energy optimization.

Energy costs and sustainability reporting pressure are pushing processing plants to chase every percent of efficiency. The U.S. Department of Energy notes that system improvements can often reduce the energy costs of a typical industrial steam system by 10% to 15%. And with the current artificial intelligence (AI) push, plants are also collecting more process data than ever. The natural assumption is that this is now an AI problem, that enough data and a capable enough model will surface the inefficiencies on their own, and that the results can simply be trusted.

Behind that assumption sits concrete work, from pulling the relevant data together to building a key performance indicator (KPI) that reflects actual consumption and comparing periods against each other. Each step is straightforward on its own. Together they take hours, and those hours compete with everything else on an engineer’s day. For a long time, that meant analyses which would have found real savings simply never got started.

Energy efficiency is not only about how much energy a process consumes, but also about the speed at which a team can see it, quantify it and act on it. An answer nobody can check saves no time.

The effort per case decides how many get solved

Energy inefficiency has many possible root causes, and they sit in different parts of the plant. In 2006, steam accounted for about 40% of the energy used in U.S. industrial process applications, and the chemicals industry used about 47% of its onsite energy, excluding feedstocks, to generate steam, according to the U.S. Department of Energy. A distillation column can run at a higher reflux ratio than the separation actually requires, meeting specification while consuming more reboiler steam than necessary. The overall heat declines as fouling builds up. A mass balance across the steam network can drift, pointing to leaks, measurement errors or abnormal consumption. The same unit producing the same product can show different energy per ton across shifts or batches. Startups and shutdowns can take longer than necessary, with the duration translating directly into energy consumed.

While the cases differ, the approach does not. Engineers identify the relevant tags and time periods, search through years of data for values that exceed a threshold, recurring patterns or correlations between parameters. They compare periods against each other, build the calculations behind a KPI or a transfer coefficient, monitor the result going forward and share the findings across teams.

Each of these is a separate analysis with its own assets, tags and logic. What decides how many of them actually get carried out is the effort each one takes. Setting up an analysis can take weeks when data connections must be made, a project has to be set up or a data scientist has to be involved. Even with the data available, building it by hand in a spreadsheet takes hours. Under those conditions only a few cases get selected. Where the engineer responsible for the unit can do it directly, both small and large cases become worth investigating. Software with fast search, visualization and correlation analysis shortens each of these steps, and it puts the analysis in the hands of the people who know the process best.

What can an analytics agent answer?

The natural question at this point is whether AI can simply take over the work directly, without an engineer defining each step. Analytics tools already shorten each of these steps. The 2026 MIT Sloan article “Agentic AI, explained” cites associate professor John Horton, who describes AI agents as going beyond generative AI by acting and making decisions the way a human might. An analytics agent is an AI agent that answers questions about process data by running defined analytics tools, rather than reading raw historian data itself. It carries out the same sequence just described, from identifying the relevant data to monitoring the result, and runs it behind a plain language question. The engineer confirms the finding and decides on the measure. That shortens the path from question to finding further.

The questions an analytics agent can answer form a ladder, and each rung has its own value.

  • Single value: What was the average steam flow to a reactor last week? Each answer saves the time that would otherwise go into pulling and averaging a trend by hand.
  • Combined conditions: How much reboiler steam was used while a column ran above target reflux? These are often questions that previously never got asked.
  • Comparison: How does energy consumption per ton differ between shifts on the same line? The result is a quantified gap, where before there was only a general impression of better or worse.
  • Root cause: Why does a column suddenly need a higher reflux ratio to hit the same specification, and what has shifted recently? Beyond the loss avoided, the reasoning behind the finding gets captured along the way, rather than staying with whoever happened to investigate it.

Each rung is useful on its own, and the next one builds on it. Findings from any of them can be turned into monitors that alert the team when the same thing happens again. The same progression points toward energy optimization, as long as the answer at each stage can be relied on.

What makes an analytics agent’s answer trustworthy?

Trust is one of the limits of AI in industrial analysis. A general AI assistant can be pointed at historian data, which raises the question of why an analytics layer is needed in between. The reason sits in the output. An answer that exists only in a chat window is hard to validate, share or build on, and trust in AI for complex analysis stays limited when results cannot be verified, repeated or audited.

An analytics agent shortens the path from a question to a finding, but that time is only saved if the finding can be relied on without repeating the analysis by hand. Checking an answer with no visible steps means redoing the work.

For an analytics agent, this sets out a few requirements:

  • Is the analysis deterministic? The same question on the same data should give the same result tomorrow, and the analysis should be explainable, not a black box.
  • Can each step be seen? Tools, assumptions and analyses are transparent, and specific parts are editable without rerunning the prompt.
  • Is the analysis carried out by defined tools for time series data, following the same steps an engineer would? The model then works with their results. Because these tools do the analytical work first, the model receives condensed results instead of large volumes of time series data, which also reduces token consumption.
  • Can a colleague reopen the result? Findings, monitors and calculations are saved, so the same result can be reopened the next day by the same engineer or a colleague.
  • Does a person confirm the finding? Before any action is taken, a human confirms it.

These questions matter most for small errors. A time zone offset, a wrong unit or a mis-mapped tag is found within minutes when the question is simple. Inside an open-ended root cause answer without visible steps, the same error stays hidden and shapes the conclusion. Trust builds with each answer that holds up, and one confident answer that turns out wrong can undo it.

For energy, this has a practical side. Efficiency measures get implemented when the result behind them is trusted. Plants that follow ISO 50001 have to show energy performance improvement against an energy baseline, using energy performance indicators, and a number that goes into an energy audit or a report to a regulator needs a derivation that can be traced.

What this looks like in practice

Analytics tools for process data have already brought analyses that needed weeks of setup down to hours, for example in optimizing the operation of heat exchangers or in balancing and monitoring steam networks. Early use of analytics agents indicates that the time can drop further. A multi-step aggregation analysis took 15 to 20 minutes with an agent, against 1.5 to 2 hours in a spreadsheet, and the engineer used the result to adjust the process. A report to a regulator was prepared by asking simple questions, and the engineer checked the numbers before they were used.

Measuring value beyond kWh saved

Efficiency programs report kWh saved and cost saved. Three further measures show how much of the potential is used:

  • Time to finding: the time from a first suspicion to a confirmed finding.
  • Cases worked: the number of cases a team works through.
  • Reuse across assets: how far an analysis can be scaled, from one trend to another, from one asset to the next and across a site.

The first two depend on who can run an analysis. When the process experts who know a unit can search the data and ask their questions themselves, more cases get looked at, and each case is a chance to find an inefficiency.

Reuse across assets has two conditions. The analysis has to give the same result for the same input, and its steps have to be transparent, so that engineers know what is being applied and can check what comes out. A deterministic, explainable analysis meets that. The assets also need a consistent structure, so that the same calculation finds the right data on each of them. An analysis built once for one heat exchanger can then be applied to the other heat exchangers on the site. The engineer still confirms each finding and decides on the measure.

The direction is continuous energy optimization, with more cases looked at and results that hold up. Energy efficiency then becomes a routine of the plant, with saved analyses that keep running and carry over to the next asset.

Plants collect more process data than ever. How much of it leads to an efficiency gain depends on how many people in the plant can work with it, and how far their findings can be trusted and scaled.

References

Stackpole, Beth. “Agentic AI, explained.” MIT Sloan, Ideas Made to Matter, February 18, 2026. John Horton, associate professor at MIT Sloan, is quoted. 

U.S. Department of Energy, Advanced Manufacturing Office. “Improving Steam System Performance: A Sourcebook for Industry,” Second Edition. DOE/GO-102012-3423, revised October 2012. 

U.S. Department of Energy, Industrial Technologies Program, Best Practices: Steam. “Save Energy Now in Your Steam Systems.” DOE/GO-102006-2275, January 2006.

About the Author

Charlotte Fischer

Charlotte Fischer

Industry expert at TrendMiner

Charlotte Fischer is an industry expert at TrendMiner and a biochemical engineer by background, with more than eight years of industry experience. TrendMiner provides industrial analytics for process data, including a deterministic AI agent.

Sign up for our eNewsletters
Get the latest news and updates