Executive Summary 

Artificial intelligence is moving from pilot projects to core infrastructure across maintenance, repair, and operations (MRO) materials management. Unlike retail or finished-goods supply chains, MRO inventory is defined by extreme part variety, low and intermittent usage, long and unpredictable lead times, and a direct link between stocking decisions and asset uptime or safety. These characteristics have historically made MRO one of the hardest domains to plan and one of the last to receive serious investment in advanced planning technology. That is changing. 

Over the next one to two years, organizations should expect AI to take hold first in the areas where it augments experienced planners and buyers: cleaning and classifying item master data, improving forecasts for both planned and unplanned demand, recommending criticality-based stocking policies, flagging cycle-count exceptions, and drafting routine purchase requisitions. Over a five-year horizon, these capabilities are likely to become networked and increasingly autonomous - self-classifying item masters, demand signals drawn directly from equipment condition data, inventory optimized across entire enterprises rather than single sites, largely self-verifying storerooms, and agentic systems that execute low-risk purchases without a human initiating each transaction. 

None of this displaces the MRO materials professional. The parts that fail catastrophically, the emergency purchase during an outage, the judgment call about whether a duplicate part record is truly a duplicate will remain, and continue to remain, decisions that require accountable human review. This paper works through five core MRO process areas: master data management, demand management, inventory management, warehouse operations, and purchasing and procurement, before closing with a view of why keeping a qualified human in the loop is not a temporary limitation of the technology, but a permanent design requirement for asset-intensive industries. 

Why MRO Is a Different Problem 

Most commercial AI supply-chain narratives are written for retail and consumer goods, where demand is high-volume, patterns repeat weekly or seasonally, and a stockout costs a lost sale. MRO materials management operates under nearly opposite conditions. A typical industrial site can carry tens or hundreds of thousands of unique stock-keeping units (SKUs), a large share of which move only a handful of times a year, or not at all until the day a critical asset fails. Demand is driven by two very different forces: planned maintenance schedules that are knowable well in advance, and unplanned breakdowns that are, by definition, not. Getting the stocking decision wrong on the wrong part is not merely a service-level statistic; it can mean an idle production line, a grounded aircraft, or a safety incident. 

This is precisely the environment in which modern AI techniques - probabilistic forecasting for intermittent demand; natural-language processing for messy catalog text; computer vision for physical verification; and large language models for drafting and summarization - have matured enough to be useful, while still requiring careful governance. The sections below take each core MRO process in turn. 

1. Master Data Management 

Item master data is the foundation everything else in this paper depends on: a forecast, a stocking policy, or a purchase order is only as good as the part record behind it. MRO catalogs accumulate decades of inconsistent naming conventions; duplicate entries created because a part could not be found under its existing record; and free-text descriptions that vary by the person who typed them. Cleaning this data manually has traditionally required large one-time consulting engagements that begin decaying the moment they finish. 

Near-Term (12-24 Months) 

AI-assisted tools can already parse free-text descriptions, normalize abbreviations and units of measure, and surface likely duplicate records for review - turning a multi-year manual cleanup into a continuous, largely automated review queue. Early adopters report meaningful reductions in duplicate inventory and faster resolution of “we already own this part” questions, though these results come from vendor case studies and should be treated as directional rather than independently benchmarked. 

5-Year Horizon 

Within five years, expect classification and enrichment to shift from a periodic cleanup exercise to a real-time function: new parts are classified and checked for duplicates automatically at the point of creation, using extracted specifications pulled from OEM manuals and catalogs rather than a human retyping them. Standards bodies are also actively working on AI-assisted mapping between classification systems, which should reduce the translation burden when data moves between ERP, EAM, and supplier catalogs. 

Where the human stays in the loop 

Automated matching will always produce false positives - two parts that look identical in text but differ in tolerance, coating, or certification that matters for a safety-critical application. A subject-matter expert needs to approve merges and classifications for critical spares before they become the system of record, not after. 

2. Demand Management: Planned and Unplanned Demand 

MRO demand planning must answer two different questions with two different methods. Planned demand (parts consumed during scheduled overhauls, preventive maintenance, and known project work) is forecastable from maintenance plans and historical job usage. Unplanned demand (parts consumed after a breakdown) is driven by failure rates that are inherently uncertain and, for many spares, occur so infrequently that traditional moving-average or seasonal forecasting methods do not apply at all. Industry estimates commonly suggest that a large share of MRO parts - frequently cited figures run as high as 70–80% - exhibit this kind of intermittent, “lumpy” demand pattern, which is why specialized statistical methods exist specifically for this problem. 

Near-Term (12-24 Months) 

Machine-learning forecasting is being layered on top of these intermittent-demand statistical methods, and predictive-maintenance signals (vibration, temperature, oil analysis, and other condition data) are starting to feed directly into parts-demand forecasts rather than living in a separate maintenance system. Organizations piloting AI-driven condition monitoring have reported meaningful reductions in unplanned downtime and maintenance cost in published studies, which should translate into fewer true emergency parts orders over time. 

5-Year Horizon 

Within five years, the boundary between “predictive maintenance” and “demand planning” should largely dissolve. A sensor detecting early-stage bearing wear should be able to trigger a parts reservation and a work order in the same motion, effectively converting a category of what used to be unplanned demand into planned demand with a shorter lead time. Forecasts should also become more transparent about their own uncertainty, giving planners a probability range for a critical spare rather than a single number. 

Where the human stays in the loop 

A forecasting model can tell a planner what usually happens, but it cannot know about the plant shutdown scheduled for next quarter; the engineering change that just made a part obsolete; or the one-time project that will spike usage. Planners need to remain the final check on the forecast for critical, long-lead, and expensive items, and organizations should resist the temptation to fully automate replenishment for parts where a stockout has safety or major financial consequences. 

3. Inventory Management 

Inventory decisions in MRO are a balancing act between two costs that are hard to compare directly: the carrying cost of a part that may sit on a shelf for years, and the cost of not having a critical part when an asset goes down. Criticality, not just historical usage, must drive the stocking policy which is why blanket inventory-reduction initiatives so often backfire in MRO environments. Read that again! 

Near-Term (12-24 Months) 

AI-assisted tools can score parts by criticality using a combination of asset importance, lead time, and failure consequence, and recommend differentiated safety-stock levels rather than a single service-level target applied to everything. The same techniques are being used to flag likely obsolete stock (parts tied to decommissioned equipment or superseded designs) for review before they are written off or, conversely, mistakenly discarded. 

5-Year Horizon 

Over five years, expect inventory optimization to move from a single-site exercise to a networked one: multiple plants or business units sharing visibility into each other’s stock of the same critical part, with AI recommending when to transfer rather than buy, and simulating “what if this site’s stock were pooled with that one” scenarios continuously rather than in an annual study. Vendor-reported results in this space (15–30%-range inventory reductions alongside maintained or improved service levels) are promising but should be treated as case-study evidence, not a guarantee, until validated against your own network. 

Where the human stays in the loop 

Criticality classification is a judgment call, especially at the margins. A part that looks like a low-usage, low-cost commodity item can be catastrophically important if it protects a single point of failure. Materials management and maintenance/reliability engineering need to jointly own and periodically re-validate criticality rankings. This should never be delegated entirely to a model trained on historical usage alone. 

4. Warehouse Operations 

MRO storerooms tend to be smaller and more specialized than distribution-centre warehouses, but they still carry a heavy manual burden: cycle counting across a huge and low-velocity SKU base, verifying that the part put away is actually the part that was ordered, and picking accurately from crowded shelving that mixes fast-moving consumables with rarely touched critical spares. This is the area where MRO-specific proof points are currently thinnest. Most available evidence comes from adjacent domains such as production-line logistics and infrastructure inspection rather than dedicated spare-parts storerooms, so near-term expectations here should be treated as a projection based on adjacent-industry results rather than an established MRO track record. 

Near-Term (12-24 Months) 

Computer vision is being used for automated cycle counting and put-away verification in shared industrial environments, and autonomous mobile robots are already handling parts-to-line and inbound-to-storage movement inside manufacturing plants. Voice- and AR-assisted picking, and remote-expert video assistance for technicians, are established in adjacent maintenance contexts and are a natural extension into storeroom picking and kitting for job packages. 

5-Year Horizon 

Within five years, larger MRO storerooms should look more like lights-on, continuously-verified operations: vision systems maintaining a near-real-time count rather than a periodic cycle count; robotics handling routine put-away and staging; and kitting for planned maintenance jobs assembled with minimal manual list-checking. Smaller or highly specialized storerooms (a single critical-spares cage at a remote site, for example) are less likely to justify this level of automation and will likely remain manually run. 

Where the human stays in the loop 

Automated counting and verification systems still need a person to resolve exceptions - a part with damaged packaging, an incorrect label, or a receiving discrepancy against a purchase order. Quality holds and receiving discrepancies on safety-critical parts should always route to a person, not be auto-resolved. 

5. Purchasing and Procurement 

MRO procurement carries its own peculiar difficulty: much of the spend is reactive, low-dollar, and high-transaction-volume (the classic “tail spend” problem), while a small number of transactions — an obsolete part needed urgently, a single-source OEM negotiation, an emergency air-freight buy during an outage — carry outsized cost and risk. Generative and agentic AI are being marketed aggressively into this space, and 2026 analyst commentary treats agentic procurement as one of the defining themes of the year, alongside explicit warnings about the governance maturity required to deploy it safely. 

Near-Term (12-24 Months) 

AI copilots are being used to draft purchase requisitions from maintenance work orders, suggest suppliers for hard-to-find or obsolete parts, summarize supplier contracts, and classify spend for tail-spend consolidation. In practice this is largely a drafting and recommendation function today: a human buyer still reviews and releases the transaction. 

5-Year Horizon 

Within five years, expect agentic systems to autonomously execute low-risk, rules-based purchases — routine replenishment below a defined dollar threshold, for instance — end to end, while sourcing strategy, supplier relationship management, and anything involving a critical or sole-source part remain human-led. Analysts are already flagging the risk of “agent sprawl,” where multiple uncoordinated purchasing agents operate with unclear ownership and accountability; organizations that get ahead of this with clear governance will be positioned to capture the benefit without the downside. 

Where the human stays in the loop 

Every agentic purchasing action needs a named, accountable human owner and a defined spending-authority boundary above which the agent must stop and ask. This is not optional risk-aversion — it is the same internal-control discipline that already governs human purchasing authority, extended to a new kind of buyer. 

6. Why the Human Stays in the Loop 

It is tempting to read the sections above as a countdown to automation. That is the wrong frame. In every process area, the pattern is the same: AI compresses the time spent on data assembly, pattern recognition, and routine execution, and returns that time to people for the judgment calls that actually require it — is this really a duplicate part, is this forecast trustworthy for a critical spare, is this the right moment to authorize an emergency buy. Analysts covering enterprise AI adoption in 2026 have begun explicitly warning against what one report calls the “human-in-the-loop fallacy”: the assumption that minimizing human review is itself a sign of AI maturity. The more accurate framing is that oversight needs to scale intelligently with risk, not disappear as automation increases. 

Three reasons keep the human central to MRO materials management specifically, more so than in lower-stakes commercial applications: 

• Accountability. A missed critical spare or a wrong duplicate merge is not a statistic — it can be an asset outage or a safety incident. Someone with the authority and expertise to be accountable for that outcome needs to have reviewed the decision, not just the system that made it. 

• Exception density. MRO is disproportionately made up of exceptions — the part that doesn’t fit the pattern, the supplier that only makes sense for this one contract, the criticality ranking that needs updating because of a plant change. Models trained on historical data are, by construction, weaker exactly where MRO judgment matters most. 

• Trust and adoption. Materials professionals will not use — and should not blindly trust — a system whose recommendations they cannot inspect or override. The organizations getting real value from AI in this space are the ones treating it as a tool that makes a skilled planner or buyer faster and more consistent, not one that replaces their judgment. 

Practically, this means building explicit autonomy tiers into every AI deployment rather than treating “human in the loop” as a single on/off switch: some decisions the system can simply act on and log; some it should recommend and a person approves; and some — anything touching a safety-critical part, a sole-source supplier, or a spend threshold above a defined level — should always require a person to initiate or explicitly sign off, not just be able to override after the fact. Defining those tiers, and revisiting them as the technology matures, is itself a core materials management responsibility over the next five years. 

Summary: A Five-Year Roadmap 

The table below condenses the near-term and five-year outlook for each process area, alongside where human ownership stays central. 

Conclusion 

MRO materials management has spent decades as an underserved corner of supply chain technology, largely because its problems — extreme variety, intermittent demand, and safety-critical consequences — did not fit the tools built for high-volume retail supply chains. The current wave of AI, particularly probabilistic forecasting, natural-language processing for messy catalog data, and increasingly capable computer vision and agentic systems, is the first generation of technology genuinely suited to this problem. The organizations that benefit most will not be the ones that automate the fastest, but the ones that are most deliberate about where automation earns trust and where a qualified person needs to stay firmly in the loop.