AI in Inventory Management: Deciding Where the Human Goes

The $1.73 trillion nobody has fixed

IHL Group put the global cost of inventory distortion — out-of-stocks and overstocks together — at $1.73 trillion a year in its 2025 research. In the same year, retailers implemented roughly $172 billion in improvements, which brought distortion down to about 6.5% of global retail sales from a higher share in previous years.

Both numbers matter, and they say different things. The second says the software works: a hundred and seventy-two billion dollars of effort moved the needle in a measurable direction. The first says it did not move it very far. After a decade of forecasting platforms, demand sensing, RFID and machine learning, the problem is still measured in trillions.

That gap is worth sitting with before buying anything. If better algorithms alone were going to solve inventory, the curve would look different by now. Something else is binding.

$1.73T

Annual global cost of out-of-stocks and overstocks

IHL Group, 2025

10%

Supply chain leaders who would trust AI to decide alone

RELEX, State of Supply Chain 2026

Firms Gartner expects to allow unsupervised planning decisions by 2030

Gartner, September 2026

The automation debate is already over

RELEX surveyed 514 supply chain leaders across retail, manufacturing and distribution in January 2026. The findings are unusually clear for survey data. Confidence in AI is rising — 67% said theirs had grown over the previous year. Adoption is real: 47% already apply AI to inventory and supply optimisation, which is not a pilot-stage number. And yet only 10% would trust AI to make supply chain decisions on its own, while about 54% explicitly prefer a model where AI recommends and a person makes the final call.

Gartner, surveying 243 senior leaders at companies above $500 million in revenue in late 2025, found 83% had already spent at least $3 million on supply chain planning automation. Its September 2026 prediction: by 2030, only 5% of organisations using planning automation will let technology make even 10% of planning decisions without human involvement.

Read those together. The money is committed. The confidence is climbing. The autonomy is not arriving. Buse Aras, a director analyst in Gartner's supply chain practice, put the reason plainly: "investment alone does not create AI readiness."

This is not caution for its own sake. It is a reasonable read of where the technology is strong and where it is brittle. Which means the design question is no longer whether a human stays involved. It is where — and a human placed in the wrong spot is worse than no human at all, because it produces the appearance of oversight without the substance.

Where the model genuinely wins, and where it does not

Most of the disappointment in this category comes from applying the model to decisions it was never going to be good at, then concluding that AI does not work for inventory.

The model is better

High-volume, history-rich, repetitive

→ Tens of thousands of SKU-location-week combinations at once, where a planner realistically covers a few hundred with any attention
→ Weak signals a person would not weight correctly: weather, local events, cannibalisation between similar SKUs, day-of-week effects per store
→ Recalculating safety stock and reorder points continuously instead of once a quarter, which is how most of them are actually set
→ Consistency. It applies the same logic at 2am on a Sunday as it does on Monday morning, and it does not get tired at SKU 900

The model is worse

Sparse, novel, asymmetric

→ Anything with no history — new products, new stores, new channels. The cold-start problem is real and no amount of model sophistication removes it
→ One-off events it has never seen: a competitor closing nearby, a regulatory change, a port disruption, a viral moment
→ Decisions where being wrong is asymmetric — an over-committed seasonal buy costs far more than an under-committed one, and the model optimises symmetric error by default
→ Knowing that the on-hand number it is optimising against is wrong. It has no way to tell

Notice that the right-hand column is not a list of things AI will fix in two years. Cold start is structural. Asymmetric cost is a business judgement about risk appetite. Data quality is an operational problem in warehouses and stores, not a modelling problem. These are the decisions that need a person, and they will keep needing one.

The decision ladder

The practical move is to stop treating "inventory decisions" as one category. They are at least four, and each warrants a different split between machine and person.

Tier	Decision	Who decides	What the person does	Cadence
1	Daily replenishment on stable, high-velocity SKUs	Machine	Reviews exceptions only — never the full list	Daily exception queue
2	Safety stock levels and reorder points	Machine, inside guardrails	Sets and approves the guardrail bands, not each value	Weekly review
3	Promotion forecasts and new-product introductions	Model proposes, person decides	Accepts or overrides, with the reason recorded	Per event
4	Seasonal buys, network moves, supplier commitments	Person decides, model informs	Owns the call and the consequence	Per decision

Gartner's own framing lands in the same place: routine decisions such as replenishment and order prioritisation have the most room for automation, while decisions about facility placement and inventory levels carry heavier consequences and will keep requiring people.

Two failure patterns follow from ignoring the ladder. Automate everything, and tier four eventually produces a buy that nobody sanity-checked. Route everything to a planner, and tier one buries them in a queue they cannot read — which brings us to the part most implementations get wrong.

Why the human in the loop usually fails

"Human in the loop" has become a reassurance rather than a design. In practice it fails in two opposite directions, and both are common enough that you should assume one of them is happening unless you are measuring.

Failure one: rubber-stamping

Give a planner four thousand recommendations a day and an Approve All button, and they will use it. This is not negligence; it is arithmetic. Nobody evaluates four thousand decisions on their merits. The oversight becomes nominal — a signature on a page that was never read — and the organisation believes it has a safeguard that does not exist.

The tell is an override rate near zero. Teams usually read that as trust. More often it is capacity.

Failure two: override addiction

The opposite problem, and the more expensive one. A planner who does not trust the model overrides constantly, often on instinct, often on the SKUs they personally know. Forecast accuracy drifts below what the model would have delivered untouched, and because the overrides feel like diligence, nobody questions them.

The tell here is a high override rate paired with outcomes no better than the unoverridden baseline — which you only know if you kept the baseline.

Measure this from week one

Two numbers tell you whether your oversight is real: what share of recommendations get overridden, and whether the overridden decisions actually beat what the model proposed. Log the model's recommendation even when it is overridden, so the counterfactual exists. Most teams that start measuring find the answer uncomfortable in one direction or the other — which is the point. If you are not tracking both, you do not have human oversight. You have a user interface.

The problem underneath: your on-hand count is wrong

In one of the more useful studies in retail operations, Nicole DeHoratius and Ananth Raman examined nearly 370,000 inventory records across 37 stores of a single retailer. They found 65% of them inaccurate — the number in the system did not match what was on the shelf. The work was published in Management Science in 2008.

That study is old, and the obvious objection is fair: RFID, cycle counting, better point-of-sale integration and warehouse automation have all improved matters since. What has not happened is anyone credibly declaring the problem solved. Inventory record inaccuracy remains a live operational issue in most multi-site retailers and distributors.

For an AI system this is the quiet killer. A replenishment engine optimising against an on-hand figure that is wrong does not hesitate or flag uncertainty. It produces a confident, precise, wrong order — and it produces it faster and at greater scale than the process it replaced. Automation does not fix bad data. It industrialises it.

The same study pointed at the fix, and it is not algorithmic. Inventory auditing practices reduced inaccuracy; store environment complexity and distribution structure increased it. Those are operations and process levers, which means the work sits with the people running the stores and the warehouses, not with the data science team.

Before any model goes near your inventory

Establish three baselines and write them down. Inventory record accuracy — audit a representative sample and get a real percentage, not an assumption. Lead time variance by supplier — the average is almost useless; the spread is what safety stock exists to absorb. Your actual stockout and overstock cost in currency, not in a dashboard percentage. Without these three you cannot prove the system helped, and you will not be able to tell a model problem from a data problem when something goes wrong.

Designing the oversight layer

If the human is staying — and every piece of evidence says they are — then the oversight layer is a real engineering component with requirements, not a checkbox at the end of the project. Six things make the difference between oversight that works and oversight that is theatre.

Exceptions, not approval queues

A person should see roughly 2% of decisions, surfaced because they are unusual — outside the forecast's confidence band, unusual order size, a supplier with recent delivery problems — not because the queue reached them. The reviewable volume is the design constraint.

Confidence travels with the number

A forecast without an uncertainty range invites false precision. Low-confidence predictions should route up the ladder automatically. A model that says "420 units, wide range, little comparable history" is more useful than one that says 437.

Overrides are training data

Every override captured with a structured reason code becomes two things at once: an audit trail for whoever asks later, and a signal about where the model is systematically wrong. Free-text reasons are better than nothing but far harder to learn from.

Guardrails are hard limits

Maximum order quantity, maximum days of supply, spend ceiling per category, no single order above a threshold without approval. The model cannot exceed these however confident it is. This is what makes tier-two automation safe enough to leave running.

Accuracy by segment, not in aggregate

A single headline accuracy figure hides the segments where the model is failing badly, and those are usually the expensive ones. Track error by category, by velocity band, by store cluster, over time — and alert on drift rather than reviewing it quarterly.

A documented fallback

If the system is unavailable for three days, what happens to ordering? Teams that cannot answer this have built a dependency, not a capability. The fallback does not need to be elegant. It needs to exist and to have been tested.

A ninety-day rollout that proves something

The most common sequencing error is going live across the catalogue and discovering six months later that nobody can say whether it helped, because no one recorded what "before" looked like.

Baseline everything

Current forecast accuracy by category. Inventory record accuracy from a real audit sample. Lead time variance by supplier. Stockout and overstock cost in currency. This is unglamorous and it is the only thing that lets you prove value later or diagnose a failure.

Days 1–15

One category, shadow mode

The model runs and recommends nothing. You compare its forecast against what actually happened, and against what the planner decided. Pick a category with decent history and real volume — not the easiest one, and not the hardest.

Days 16–40

Tier one live, exceptions routed

Automate daily replenishment on that category with guardrails in place and exception routing switched on. Measure override rate and override outcomes from the first week. Expect to tune the exception thresholds two or three times before the volume is reviewable.

Days 41–65

Tier two, then decide on evidence

Bring safety stock and reorder points under automation inside bands the planners set. Then compare against the day-one baseline and decide expansion on what the numbers show — including the possibility that one category is not ready and the next one is.

Days 66–90

What good looks like at day ninety

You can state your forecast accuracy before and after, by segment. You know your override rate and whether overrides are beating the model. Planners are reviewing a volume of exceptions they can actually read. Guardrails have been tested by at least one recommendation hitting a limit. And you know which tier of decision you are not going to automate, and why. That last one is a finding, not a failure.

Frequently asked questions

Will AI replace demand planners?

The evidence points the other way. RELEX found only 10% of supply chain leaders would trust AI to decide alone, and 54% prefer AI recommending with a human deciding. Gartner expects just 5% of firms with planning automation to allow unsupervised decisions by 2030. What changes is the job: less time producing forecasts, more time setting guardrails, handling the decisions the model cannot make, and judging when to override. That shift is real and worth planning for — it is a different skill set from spreadsheet-based planning.

How much forecast accuracy improvement should we expect?

Be sceptical of any fixed number, including the large ones in vendor material. The honest answer is that it depends almost entirely on your starting point. A team forecasting in spreadsheets on a quarterly cadence has a great deal of headroom. A team already running a tuned statistical model with good data has much less, and the gain may come from coverage and cadence rather than per-SKU accuracy. Baseline your current accuracy by category before anyone quotes you an improvement, so you can hold the number to account afterwards.

Do we need clean data before we start?

You need to know how clean it is, which is different. The DeHoratius and Raman study found 65% of inventory records inaccurate at one retailer, and record inaccuracy remains common. You do not have to fix it all first, but you do have to measure it, because an optimiser working from a wrong on-hand figure produces confident wrong orders at scale. If record accuracy is poor in a category, that category is a data and operations project before it is an AI project.

How do we stop planners from rubber-stamping recommendations?

Control the volume they see and measure what they do. If a planner is shown everything, approval becomes a formality — nobody evaluates thousands of decisions on their merits. Route exceptions only, target a volume that can genuinely be read, and track override rate alongside override outcomes. An override rate near zero usually signals capacity, not confidence. Logging the model's recommendation even when it is overridden gives you the comparison that settles the question.

Our catalogue is small. Is AI worth it?

Often not, at least not first. The model's advantage is scale and consistency across thousands of SKU-location combinations. With a few hundred SKUs and experienced planners, better data discipline, tighter supplier lead times and properly recalculated safety stock frequently deliver more than a forecasting model would, at a fraction of the cost and risk. The honest test is whether your planners are at capacity. If they are comfortably covering the catalogue with attention, the constraint is somewhere else.

What does a project like this take to build?

For a single category through shadow mode to tier-one automation, ninety days with a small team is realistic when the data is accessible. The variable that moves the timeline most is not modelling — it is data access and record accuracy, which is where most schedules slip. Expect the oversight layer itself, exception routing, guardrails, override logging and accuracy monitoring, to be a meaningful share of the engineering effort rather than an afterthought. It is the part that determines whether the system is still running in year two.

Thinking About AI for Inventory or Planning?

We build AI systems that make operational decisions with oversight that actually works — exception routing, guardrails, override logging and accuracy monitoring, not a dashboard. Book a free 30-minute call with our engineering team. NDA first, no obligation.

Book Free Strategy Call →Our AI Services

#InventoryManagement#SupplyChainAI#DemandForecasting#HumanInTheLoop#AIStrategy#MLOps#RetailTech#Vikgol

Vikgol Engineering Team

AI Engineering & Delivery · Vikgol

The Vikgol engineering team has shipped 90+ AI, web and cloud projects for startups and enterprises across the US, UK, UAE and India. We are an OpenAI Select Partner. We build systems that make decisions in production, which means we spend as much time on evaluation, guardrails and observability as on the model itself.