Where AI Fits in Supply Chain (And Where It Doesn't): The Middle Steps Framework
Most conversations about AI in supply chain are happening at the wrong altitude. Either it is going to automate everything and half the profession is redundant, or it is overhyped and the fundamentals have not changed. Neither framing is useful, and neither is how the best supply chain organisations are actually thinking about it.
The more productive question is a narrower one: where does AI earn its keep in supply chain decisions, and where does deploying it create more noise than signal?
The answer, when you map AI capability against how supply chain decisions actually get made, is surprisingly precise. AI does the middle steps brilliantly. It struggles badly at the beginning and the end. Understand that boundary clearly, and you can deploy AI in a way that genuinely sharpens outcomes. Ignore it, and you will spend a lot of money on tools that produce confident-sounding answers to the wrong questions.
At Trace, we are not approaching AI cautiously. We are pointing it directly at the work our clients need done, because the leverage is real and the opportunity to compress timelines and sharpen analysis is significant. But we are doing it with a clear view of where it earns its keep and where experienced practitioners are irreplaceable. That view shapes everything in this article.
The Ten-Step Arc of a Supply Chain Decision
Before arguing about where AI fits, it helps to be explicit about the arc itself. Whether you are a COO at a major retailer, a procurement director in the public sector, or a supply chain lead at an FMCG manufacturer, the sequence is recognisable.
Step 1: Define the real problem. Not the presenting symptom. The actual problem. Inventory write-offs are a symptom. The problem might be forecast accuracy, supplier lead time variability, or a category management gap. Getting this right requires judgment, curiosity, and the ability to hold a room of people with competing interests and surface the truth.
Step 2: Gather stakeholder requirements and constraints. What does the business actually need from the outcome? What are the non-negotiables: budget, timeframe, risk appetite, industrial relations constraints, board sensitivities? This requires relationship capital, political awareness, and the ability to read what is not being said in a meeting.
Step 3: Set the decision criteria. How will we know a good answer from a bad one? What are we optimising for, and what are we willing to trade off? This is a human conversation because it is fundamentally about values and priorities, not data.
Step 4: Collect, cleanse, and structure data. Pull together the relevant data sets, identify the gaps, normalise formats, and build a clean analytical foundation.
Step 5: Analyse, identify patterns, and surface anomalies. Run the numbers. Identify where performance is strong and where it has degraded. Find the correlations that are not obvious to the human eye buried in millions of rows.
Step 6: Model scenarios and test sensitivities. What happens to the answer if demand spikes 20%? If the primary supplier exits? If lead times blow out by six weeks? Build the range.
Step 7: Generate options and draft recommendations. Take the analysis, synthesise it into credible options, and build the structured case for each.
Step 8: Make the decision. Own it. With a name on it.
Step 9: Implement. Change management, sequencing, stakeholder communication, supplier conversations, workforce transitions. The messy, relationship-intensive work of making something actually happen.
Step 10: Review, learn, and course-correct. Hold the outcome accountable against the intent. Adjust. Feed the learning back into the next cycle.
Why AI Owns Steps Four to Seven
Steps four through seven are where AI's core capabilities, namely pattern recognition, speed, scale, and tirelessness, are genuinely transformative. The clearest way to demonstrate this is to look at the specific problems it solves in each step.
Step 4: Data Collection and Cleansing
In supply chain, data fragmentation is one of the most stubborn operational problems. A typical mid-size Australian retailer or distributor will have purchasing data in one ERP, warehouse performance data in a separate WMS, supplier lead time history scattered across buyer inboxes and spreadsheets, and customer demand signals split between a POS system and an e-commerce platform that do not talk to each other. Pulling all of that together, normalising formats, resolving duplicate supplier codes, and building a single analytical foundation used to consume weeks of analyst time and still produced inconsistent results.
AI-assisted integration tools can ingest from disparate source systems, apply transformation logic, flag data quality issues for human review, and produce a clean, structured foundation faster and more reliably than any manual process. For a procurement programme covering several hundred suppliers and a few thousand SKUs, this is not a marginal efficiency gain. It is the difference between being able to run the analysis at all and being unable to.
The practical implication for clients is straightforward: what previously sat inside a twelve-week diagnostic can now be delivered at equivalent depth in six, with the remaining time invested in implementation, capability transfer, and making sure the change actually sticks.
The important caveat is that AI cannot tell you whether the data reflects reality. If the ERP has been coded inconsistently by different buyers over three years, or if stock-on-hand figures are unreliable because cycle counting has lapsed, the AI will work confidently with bad data. The practitioner who set the decision criteria at step three needs to have flagged those data risks before the analytical work begins. That is a human responsibility.
Step 5: Pattern Recognition and Anomaly Detection
This is where AI genuinely sees things that humans miss, and the supply chain applications are specific and significant.
Demand forecasting is the clearest example. A demand planning AI running across three years of weekly sales data across 40,000 SKUs will surface seasonal patterns, inter-SKU cannibalisation effects, and slow-moving inventory clusters that a human analyst would take months to identify, and could easily get wrong given the combinatorial complexity. More importantly, it will do this at SKU level, not just at category level, which is where the actual inventory holding and service decisions are made. Australian FMCG and retail businesses that have moved to AI-assisted demand planning consistently report meaningful improvements in forecast accuracy, with corresponding reductions in both stockouts and excess inventory. A five-percentage-point improvement in forecast accuracy across a large SKU base carries substantial working capital implications.
Supplier performance monitoring is a second area where AI pattern recognition produces results that humans operating manually cannot replicate at scale. A procurement team managing 200 suppliers across a complex categories portfolio faces a genuine bandwidth problem when it comes to continuously tracking on-time-in-full performance, price variance against contracted rates, lead time drift, and quality reject rates across all 200 vendors. The result is that supplier performance problems are typically identified late, after they have already caused operational disruption.
AI tools that continuously monitor supplier performance data can flag drift in delivery reliability weeks before it becomes a stockout, identify systematic overbilling against contracted rates, and surface which suppliers are trending toward non-compliance before the relationship deteriorates. For organisations managing significant supplier networks in retail, FMCG, or construction and infrastructure, this is a category shift in procurement intelligence. It gives a procurement team the visibility that previously required a much larger headcount to maintain, and it points their attention at the exceptions that actually matter.
Inventory anomaly detection addresses one of the most persistent and expensive problems in supply chain operations. Slow-moving and obsolete inventory (SLOB) typically accumulates gradually and is often not identified until a financial year-end stock count forces a write-down. AI tools running continuously across inventory data can identify SKUs where coverage is building relative to demand trends, flag items where sales velocity has dropped below the replenishment trigger, and surface potential obsolescence risks months earlier than a manual review cycle would. For an organisation carrying tens or hundreds of millions in inventory, the working capital benefit of earlier identification and intervention is direct and measurable.
Step 6: Scenario Modelling
Scenario modelling has historically been one of the most resource-intensive steps in any supply chain programme, constrained as much by computational time as by analytical skill. A distribution network design exercise involves evaluating combinations of facility locations, transport mode splits, inventory positioning strategies, and service level commitments across a large and variable demand base. Building a model capable of evaluating even a few dozen scenarios meaningfully used to require weeks of specialist work, which meant the sensitivity analysis was inevitably limited and the confidence intervals on the recommended design were wider than they should have been.
AI and advanced optimisation engines have changed this. A network design exercise that previously generated and evaluated 30 to 40 scenarios can now evaluate thousands, testing the optimal design against a range of demand futures, fuel cost trajectories, and labour market assumptions in a fraction of the elapsed time. The practical impact is not just faster analysis. It is better decisions, because the recommended design has been stress-tested against a far wider range of plausible futures. We can now model distribution scenarios across an entire network in hours or days, not weeks. That compression goes directly into more time on implementation and execution, which is where client value is actually realised.
The same principle applies to procurement scenario modelling. When evaluating a major category sourcing strategy, the question is never just who is the cheapest supplier today. It is what the total cost of ownership looks like at different volume commitments, across different contract lengths, under different risk scenarios including supplier insolvency, logistics disruption, and regulatory change. AI-assisted scenario tools can model these trade-offs at a granularity and speed that changes what is analytically feasible within a normal project timeframe.
S&OP and integrated business planning is a third domain where AI delivers tangible value. Many Australian organisations run S&OP processes that are more administrative ritual than genuine decision-making forum. The bottleneck is usually that building the demand, supply, and financial reconciliation for the monthly cycle consumes so much analyst time that little is left for the actual discussion. AI-assisted planning tools that automate the reconciliation work and surface key exceptions for human discussion are shifting the S&OP meeting from data presentation to genuine scenario analysis, which is what it was always supposed to be.
Step 7: Options Synthesis and Recommendation Drafting
Given a well-structured analytical output from steps four through six, AI can generate a coherent options paper, surface the key trade-offs between alternatives, and draft the narrative structure of a recommendation. This is not the same as making the recommendation. It is compressing the distance between raw analysis and a structured decision document from days to hours, freeing up senior practitioners to focus on sharpening the argument, testing the logic, and pressure-testing the assumptions rather than building the document architecture from scratch. That is where consultant time should go, and AI is creating the space to put it there.
Why Humans Must Own Steps One to Three
The failure mode for most AI deployments in supply chain is not that the tools are bad. It is that organisations skip the front end, or assume AI can handle it.
Step 1: Problem Definition Is Irreducibly Human
AI can help you analyse the problem you hand it. It cannot tell you whether you have handed it the right problem. An AI tool asked to optimise inventory will optimise inventory. Whether inventory optimisation is actually the binding constraint on your supply chain performance, or whether the real issue is forecast methodology, procurement lead times, or an incentive structure that rewards the wrong behaviour, is a diagnostic question that requires human judgment, experience, and the willingness to challenge the brief.
What this looks like in practice: a large Australian retailer presents with persistently high inventory write-offs. The tempting move is to deploy an inventory optimisation tool. The actual problem, on investigation, is that the buying team is making range decisions based on historical sales data that systematically undercounts online channel demand. The inventory optimisation tool would have worked diligently on a problem that was not the constraint. The write-offs would have continued.
Or consider a distribution network where transport costs are running above benchmark. The presenting problem looks like a carrier contract issue. The actual problem is a warehouse slotting configuration that produces excessive pick path distances, driving up labour time and creating a ripple effect on dock scheduling that causes late departures and missed delivery windows that the carrier then charges for. AI applied to the transport contract would not have found that.
The most expensive supply chain mistakes do not happen in the analysis. They happen when a team spends six weeks solving the wrong problem with impeccable rigour.
Step 2: Stakeholder Requirements Involve Politics, Trust, and Relationship Capital
Understanding what the CFO is genuinely concerned about, as opposed to what she said in the steering committee, requires a human in the room who can read the dynamic, ask the awkward question, and build enough trust to get the real answer. AI cannot attend a discovery workshop and sense when the head of operations is sandbagging the data, or when the procurement director's stated preference for a single-source strategy is actually being driven by a supplier relationship that the organisation has not surfaced in the brief.
This matters because the requirements that shape a good solution are often held by people with strong incentives not to share them fully. The warehouse manager who does not want the slotting review because he knows it will expose years of ad hoc decisions. The buyer who does not want the category analysis because it will surface the off-contract spend she has been approving. A skilled practitioner can work around this. An AI tool cannot.
Reading the room in a stakeholder workshop, knowing which recommendation a leadership team will actually execute versus the one they will nod at and ignore, building trust with a procurement director who has been burned by consultants before: these are not soft skills. They are the hardest skills in consulting, and they are now the most valuable.
Step 3: Decision Criteria Reflect Organisational Values
When two options trade off cost against service level, or short-term cash against long-term supplier relationships, the weighting applied to those trade-offs is a leadership decision. It reflects what the organisation actually cares about, and that cannot be delegated to an algorithm.
A concrete example: an AI-assisted network design might identify a configuration that minimises total landed cost by consolidating two distribution centres into one. The model is correct. But the CFO's real constraint is that the lease on one of those DCs cannot be exited for four years without a material break cost, and the board has already approved a capital plan that assumes it remains operational. The decision criterion that makes that option impractical was never in the model because it was never surfaced in the problem framing. That is a human failure at step three, and no amount of analytical quality at steps four through seven can compensate for it.
Why Humans Must Own Steps Eight to Ten
The back end of the arc is where AI's limitations become most acute, and where the consequences of misunderstanding those limitations are most serious.
Step 8: Decisions Require Ownership
AI can produce a recommendation. It cannot be held accountable for it. In supply chain, where a wrong call on a major procurement contract or a network footprint decision can have multi-year financial consequences, accountability is not optional. Someone has to put their name on the outcome, defend it to the board, and own the consequences. That person needs to understand why they made the call, not just what the model suggested.
There is a growing pattern in Australian organisations, visible particularly in large procurement and logistics transformations, where AI-generated recommendations are being treated as decisions rather than inputs. The risk is not only governance failure, though it is that too. It is that the organisation loses the ability to learn from its own choices, because no one has exercised the judgment muscle that makes future decisions better. When a machine handles the data extraction, the scenario modelling, and the first draft of the analysis, what is left is the hard stuff: the decisions that actually define outcomes. That judgment cannot be outsourced.
Step 9: Implementation Is a Human Endeavour
A major supplier consolidation programme involves more than selecting the winning suppliers on an analytical scorecard. It requires managing the exit of incumbent suppliers who may hold knowledge, tooling, or secondary capacity the organisation has not fully accounted for. It requires negotiating transition plans that protect continuity while moving volume. It requires managing internal stakeholders who have existing relationships with the exiting suppliers and may resist the change. It requires sequencing the transition to avoid service disruption during peak trading periods. Each of those activities requires a skilled practitioner who can read the room, adapt the approach, and navigate resistance without escalating conflict.
The same applies to warehouse and distribution network changes. Standing up a new DC configuration, transitioning 3PL providers, or implementing a new WMS requires change management at an operational level: training workforces, managing the parallel run period, handling the inevitable exceptions that the model did not anticipate, and maintaining service levels throughout. AI can support this work by tracking milestones and flagging dependencies. It cannot lead the work itself.
Where supply chain consulting used to front-load value in the diagnostic and back off at implementation, the real value is increasingly in the back half: the capability transfer, the change leadership, and making sure the outcome actually sticks. AI accelerates the front and creates the space to invest more in the back. That is a better service for clients, and it is how the best engagements are being structured now.
Step 10: Learning Requires Human Reflection
Supply chain organisations that improve over time are those where senior leaders genuinely interrogate what happened, why, and what they would do differently. An AI system will faithfully track performance against KPIs. It will not tell you that the forecast error is being driven by a cultural resistance to sharing commercial intelligence between the sales and supply chain teams, and that the fix is a different conversation, not a better algorithm.
Post-implementation reviews are consistently the most neglected step in the cycle. Organisations that build compounding capability over time are those that invest in genuine reflection and feed those lessons back into the next cycle of problem definition. That is steps one through three again, and it is entirely human.
The "Less Is More" Principle in AI Deployment
Alongside the middle-steps framework sits a second challenge that does not get enough airtime: the proliferation problem.
Organisations are now deploying AI across supply chain functions at pace: demand planning tools, procurement intelligence platforms, inventory optimisation engines, logistics visibility platforms, workforce scheduling systems, and generative AI for analysis and reporting. The result in many cases is not a smarter supply chain. It is a fragmented one.
Each tool has its own data model, its own interface, and its own logic for generating recommendations. The result is a proliferation of signals that frequently conflict. The demand planning system says increase safety stock. The inventory optimisation tool says reduce it. The procurement platform flags a supplier risk that has not been integrated into either model. The people in the middle of all this are not more empowered. They are more confused.
The organisations getting the most from AI in supply chain are not the ones with the most tools. They are the ones with the fewest.
A single, well-integrated demand planning AI that is properly trained on clean data, well understood by the team, and connected to the downstream planning process will generate more value than five AI tools operating in silos. The discipline is in choosing where to concentrate AI investment and having the organisational will to resist the pressure to adopt every new capability that crosses the desk.
Three questions worth asking before adding an AI tool to the stack:
Does it connect to the existing data foundation? Standalone tools that operate on their own data extract are almost always inferior to tools integrated with the core ERP or planning environment. A procurement intelligence platform that cannot read from the same supplier master as the ERP will generate insights that are immediately challenged on data integrity grounds, which is how those tools end up going unused.
Does it cover a step where AI genuinely earns its keep? Using an AI tool to support problem definition or decision-making is a misapplication. Using it to compress data analysis and scenario modelling is the right deployment. The question is not "how do we use AI in procurement?" It is "which steps in our procurement decision process are currently the bottleneck, and is AI the right tool to address them?"
Can the team actually explain what it does? If the practitioners using the tool cannot articulate the logic, even at a high level, the organisation has substituted one form of uncertainty (not knowing the answer) for a more dangerous one (not knowing why the tool produced that answer). When an inventory optimisation recommendation drives a buying decision on a high-value SKU, the buyer needs to be able to explain to the category manager why the system generated that output, and to override it when the underlying assumptions do not hold.
What This Means for Supply Chain Leaders
The middle-steps framework has practical implications for how supply chain and procurement leaders should approach AI investment.
Invest in human capability at the front and back end. As AI handles more of the analytical middle, the value of genuinely senior judgment at steps one to three and eight to ten increases. AI will increasingly commoditise the outputs that consulting firms have charged a premium for: the benchmarking decks, the first-pass analysis, the data crunching. The organisations and advisers that remain valuable are those who combine sharp analytical tools with the judgment, domain expertise, and implementation capability that no algorithm replicates. The Trace Consultants team is structured on exactly this basis.
Define AI use cases by step, not by function. Rather than asking "how do we use AI in procurement?" ask "where in the procurement decision arc are we currently most constrained?" In most organisations the bottleneck is not the analysis. It is the problem definition at the front, or the decision and implementation accountability at the back. More AI in the middle will not fix either of those.
Build an integrated data foundation first. The organisations that get the most from AI-assisted supply chain analysis have consistently done the hard work of building clean, connected data. AI applied to poor data produces confident wrong answers faster. The foundational work of data governance and integration is unglamorous, but it is the prerequisite for everything that follows.
Set a tool ceiling and hold to it. Decide how many AI tools will operate across the supply chain function and resist pressure to exceed it. The ceiling forces prioritisation and integration, which are the two disciplines that separate supply chain teams that benefit from AI from those overwhelmed by it.
How Trace Consultants Can Help
Trace Consultants works with Australian organisations to design and implement supply chain and procurement improvements that deliver measurable outcomes, not just analytical outputs. We use AI where it earns its keep, and we deploy experienced practitioners at the front and back end where judgment, relationships, and accountability matter most. Shorter timelines on the analytical work means more time invested where client value is actually created: in the decisions, the implementation, and the capability transfer.
Problem definition and diagnostic work. Before any analytical work begins, Trace practitioners invest in rigorous problem definition: challenging the brief, aligning stakeholders, and setting decision criteria that reflect what the organisation is actually trying to achieve. This is where most supply chain transformations either earn their investment or waste it. See our planning and operations and strategy and network design capabilities.
AI-assisted analysis at scale. For the analytical middle steps, including data consolidation, pattern recognition, scenario modelling, and options synthesis, Trace uses AI tooling to compress timelines and raise quality. Our technology capability supports clients in selecting, integrating, and extracting value from AI tools in supply chain and procurement environments.
Procurement intelligence and sourcing. Our procurement team uses AI-assisted spend analysis, supplier benchmarking, and market intelligence to compress the analytical work in major sourcing programmes, then brings experienced practitioners to the decision and negotiation table where AI stops being useful.
Implementation and change management. The back end of the arc, covering decision, implementation, and review, is where Trace's project and change management capability comes in. We stay through implementation, because that is where supply chain value is either realised or lost.
Explore our supply chain and procurement services →Speak to an expert at Trace →
Where to Begin
If your organisation is actively evaluating AI investment in supply chain, three starting points are worth considering.
First, map your current decision processes against the ten-step arc above and identify where the actual bottlenecks sit. In most organisations the constraint is not the analysis. It is the problem definition at the front, or the decision and implementation accountability at the back. More AI in the middle will not fix either of those.
Second, audit the AI tools you already have before adding more. Are they integrated? Can the practitioners using them explain the logic? Are they generating consistent signals or conflicting ones? Consolidation frequently delivers more value than additional investment.
Third, identify the one or two places in your supply chain where better, faster analytical capability in steps four through seven would most change the quality of decisions. Then invest there, with an integration-first approach and a clear accountability structure wrapped around the output.
The Bottom Line
AI is not going to run your supply chain. It is going to do some of the most time-consuming and error-prone analytical work in your supply chain faster, more consistently, and at greater scale than any team of people can. That is genuinely valuable, and organisations that have not harnessed it yet are leaving real efficiency on the table.
The organisations winning with AI in supply chain are not the most automated. They are the ones that have been precise about where to point it: at the analytical middle, with experienced practitioners anchoring both ends. The supply chain director who hands a leadership team a scenario model they can act on this week rather than next quarter. The procurement team that can see the category risks hiding in their spend data that nobody had time to find manually. The workforce planner working from a demand forecast they actually trust.
The middle steps belong to the machine. The first three and the last three belong to people. Get that boundary right and AI becomes one of the most powerful tools in the supply chain arsenal. Get it wrong and it is an expensive source of confident-sounding noise.
