GenAI in the Trough of Disillusionment: Why Vendors Oversell
Generative AI in supply chain landed in the trough of disillusionment because the technology arrived faster than the organisations buying it could get ready for it. Gartner put GenAI for procurement in the trough in July 2025, and its September 2026 spending forecast still had GenAI firmly in the trough across industries. By August 2026, 55% of chief supply chain officers said they were unclear on the ROI of their AI investments while 67% of supply chain digital spend was going to AI. Spending up, certainty down. Part of that gap is vendors overselling. Most of it is buyers skipping the unglamorous groundwork the technology needs to pay back.
What does the trough of disillusionment actually mean for a supply chain buyer?
It means the technology stopped being a story and started being a line item that someone has to defend. Gartner’s curve is a map of expectations, not capability, so nothing about GenAI got worse between the peak and the trough. What changed is that the first budget cycle closed and finance asked what came back.
The headline numbers are grim. MIT’s Project NANDA reported that about 95% of enterprise GenAI pilots produced no measurable P&L impact (52 organisations interviewed, 153 senior leaders surveyed, over 300 public deployments reviewed, research run January to June 2025). S&P Global Market Intelligence’s 2025 survey, as reported by CIO Dive, found the share of firms scrapping most of their AI initiatives jumped to 42% from 17% the year before, with the average organisation killing 46% of proofs of concept before production.
We hold MIT’s research in high regard, and we still would not lean on the 95%. The sample is small for a claim that sweeping, and measurable P&L impact is a hard bar for any programme to clear within months of a pilot. Read the method before you repeat the number.
Our own experience points the same way. Two agentic AI programmes we ran at large multinationals delivered on their business cases. The journey was not easy, and it made it obvious why so many pilots fail. It is rarely the models, the infrastructure or the harnesses that let a programme down. It is poor data quality, an AI and technology architecture nobody designed, organisational inertia, leadership that wants results before the foundations are in, and users who quietly worry the system is there to take their job.
Where our programmes worked, they sat on top of years of effort that had already dealt with most of those problems. The result was thousands of hours of manual work removed, and that freed time went into opportunities that had been left on the table because they were too difficult or looked too small to be worth a senior person’s afternoon. When an analysis costs a few dollars and a few hours instead of a week, a business can afford to chase every crumb.
Those conditions are not nice to have. Organisations that invest in them will get what GenAI promised. Organisations that skip them will keep feeding the failure statistics.
Why did generative AI hit the trough in supply chain faster than elsewhere?
Because supply chain work is mostly deterministic, and GenAI is not. A marketing team can absorb a plausible-sounding paragraph that turns out to be wrong. A replenishment system cannot absorb a plausible-sounding quantity. The tolerance for confident error in our function is close to zero once money moves, so the use cases where GenAI shines get pushed to the edges of the process, and the edges are where the small budgets live.
Three structural things make it worse here than in most functions.
Our data is fragmented across ERP, WMS, TMS, the 3PL’s system and a spreadsheet, and the version everyone trusts is usually the spreadsheet. Our processes contain undocumented human judgment, the “Bob knows what to do” step we wrote about in the agentic AI reality check. And our decisions are already governed by planning systems with their own optimisers, so a language model arrives as a fifth opinion in a room that could not agree with four.
How do vendors oversell generative AI? The seven mechanics
None of these will surprise anyone who has sat through a GenAI pitch.
| # | Mechanic | How it shows up in the pitch | What it costs you | The question that kills it |
|---|---|---|---|---|
| 1 | Demo data laundering | A flawless walkthrough on a curated dataset with clean item masters and one system of record | Months of unbudgeted data remediation discovered after signature | “Run this on an extract of our data, that we pick, this week” |
| 2 | Capability transfer | Frontier-model performance implied for a product that is a thin wrapper plus prompts | You pay platform pricing for orchestration you could have built | “Which model, which version, and what exactly do you add on top of it?” |
| 3 | Use-case bundling | Twelve use cases in the deck, two of them real, priced as one platform | Shelfware; the ten you never deploy are still in the subscription | “Price the two we will deploy in year one, alone” |
| 4 | Pilot-scope shrinkage | The pilot lands on the cleanest category, the friendliest region, the best-documented process | A success metric that will not reproduce anywhere else in the business | “Include our worst-documented category in the pilot scope” |
| 5 | Benefit double-counting | Savings attributed to GenAI that a process change or a data cleanup delivered | An ROI case finance will not sign off twice | “What is the control group, and who owns the benefit if we do the cleanup alone?” |
| 6 | The productivity proxy | Hours saved presented as money saved, with no headcount or capacity decision behind it | A cost saving that never appears in any P&L line | “Which cost line goes down, or which work do the freed hours go to?” |
| 7 | Roadmap collateral | Today’s gap covered by a Q3 release that will close it | You bought a roadmap and got a beta | “Put the dated capability in the contract with a credit attached” |
Most of these are not lying. That matters for how you respond. A vendor with a working demo on curated data genuinely believes it will work on yours, because the failure is in your data and neither of you priced it.
Two of the seven deserve a more honest reading than a vendor-bashing post would give them.
Mechanic 5 cuts both ways. The data cleanup a consultant scopes as pre-work for a GenAI programme is often a cleanup that would never have happened otherwise. And once a VP gets a poor answer from the agent they asked for, the people feeding it suddenly care about data quality in a way no governance memo ever achieved. “Hallucination” stops being a convenient excuse for bad master data. So yes, challenge the double-count in the business case, but do not pretend the cleanup would have happened on its own.
Mechanic 6 is real as a pitch tactic and misleading as a verdict. We have not yet seen an organisation restructure with AI as the genuine root cause, whatever the press releases say, and no department we know of has handed a whole function to an agent. So hours saved mostly do not become headcount saved. They turn into other things that matter: better engagement scores, people moving to employers that are better equipped with AI, and teams doing work that was previously thought impossible. That is value. It just is not a cost line, and a business case that claims it is will fall apart in the second budget cycle.
What is generative AI actually good for in supply chain right now?
Language work, retrieval and drafting. Not decisions. The honest 2026 tier list, based on what we see reaching production rather than what reaches a demo stage:
| Tier | Use case | Why it works or doesn’t | Buy or build |
|---|---|---|---|
| Works today | Contract and tariff-clause retrieval, supplier document Q&A | Bounded corpus, human reads the answer, errors are visible | Buy, but only if the corpus is yours and stays yours |
| Works today | Drafting supplier comms, RFx documents, exception write-ups | Human edits before it leaves the building | Mostly build on a general assistant |
| Works today | Summarising and triaging exception queues | Prioritisation is a judgment task with cheap errors | Buy inside a planning suite you already own |
| Conditional | Master data cleansing and classification suggestions | Real gains, but needs a review workflow and a rollback | Buy with a human-in-loop gate, never straight-through |
| Conditional | Natural-language query over planning data | Depends entirely on the semantic layer beneath it | Do not buy the chat layer before the data layer |
| Not yet | Autonomous replenishment or PO changes from an LLM | Confident error with working capital attached | See agentic AI for the bounded-autonomy pattern |
| Not yet | Demand forecasting by LLM | The maths that works is ML, not language models. See what the accuracy claims leave out | No |
Notice the shape of the “works today” rows. They all have a human reading the output before anything commits, and they all replace typing rather than deciding. That is the current frontier, and it is a genuinely useful frontier if you stop pricing it like a transformation programme.
How should you buy generative AI during the trough?
Start by asking whether you should be buying at all. Our bias is toward building AI capability inside the organisation rather than paying a consultant to implement an AI SaaS product. A lot of what is sold as an AI platform is a general model, some prompts and an “AI” badge that mostly exists to lift the licence price. The capability you build yourself stays when the contract ends.
When you do buy, buy hard, and now. A trough is the best negotiating position a supply chain buyer has had on AI software since the peak, because the seller’s pipeline is thinner and their board still wants logos. Three things we would put in writing.
Pay for outcomes on a named process, not seats on a platform. If the vendor will not price against a metric that already exists in your reporting, they do not believe the metric will move. Second, get the data remediation scoped and priced by them, before signature, on your extract. The cost does not disappear because it was omitted from the proposal. Third, put the roadmap items in the contract with dates and service credits, because roadmap collateral is the mechanic with the least downside for the seller and the most for you.
One more thing that costs nothing. Run the pilot on a process you would be willing to keep running manually, so walking away is a real option. Pilots on processes you have already dismantled are not pilots.
What should you measure before you fund the second pilot?
The numbers finance will ask for anyway, captured before you start rather than reconstructed afterwards.
| Measure | Baseline captured when | Why it settles the argument |
|---|---|---|
| Cycle time on the specific task | Before pilot, 4 weeks of it | Stops “it feels faster” from becoming the business case |
| Error or rework rate on the task output | Before pilot | GenAI trades speed for a new error profile; you need the old one |
| Where the benefit lands, named and owned | Before pilot | Either a cost line that goes down, or the specific work the freed hours move to. Kills mechanic 6 without pretending hours are worthless |
| Human review time per output | During pilot | The hidden cost that turns a 90% automation into a 40% one |
| Adoption and engagement of the team using it | Before and during pilot | Tells you whether the people the tool depends on are using it, or routing around it |
The review-time row is the one teams skip. An output that needs checking is not the same as an output you can trust, and the checking time is real work that lands on someone’s week. It is entirely possible for a “fully automated” summary flow to consume more analyst hours than the manual version it replaced, and nobody notices unless someone counts the reading.
What the trough is actually telling you
Every supply chain technology programme we have seen succeed started with someone admitting what the current process costs and what the organisation was missing to fix it. The trough of disillusionment is the industry doing that out loud. The technology was ready sooner than the data, the architecture and the people around it. That part is fixable, and it is the buyer’s to fix.