Right now, the fastest way for a logistics company to get an AI agent running is to hire someone to build it. That's produced a wave of consultancies and forward-deployed-engineer shops standing up bespoke agents one engagement at a time: real, often actually useful software. It's also structurally different from an AI agent orchestration platform a company can keep extending after the invoice is paid, and the difference matters more than the sales pitch usually lets on.
Homodeus is a clean example of the pattern, not because it's unusual but because it's explicit about it. Founded in 2024 by João Panizzutti and marketed as an "Operational AI Lab for the Enterprise," it sells project-priced engagements: an AI ROI audit around $12,000-18,000, a discovery sprint around $35,000, a "production pod" at $90,000-150,000, managed AI ops afterward at $11,000-25,000 a month. It builds custom agents across roughly 19 industries, with a client base weighted toward freight and logistics in the São Paulo market. By its own description, each engagement produces "agents, memory, and an ontology per area, creating a digital twin of how a company decides," with human review on sensitive output like bid drafts.
Worth being precise here. Homodeus isn't claiming to skip governance, and it isn't a vaporware landing page. There's a named founder, a live product description, and a specific delivery model. What isn't independently verifiable is the performance data it publishes: 94% faster quotes, a 35% cost reduction from its control-tower work, 50% fewer stockouts, an 18% fuel saving, a claimed 95% project success rate. Every one of those numbers traces back only to Homodeus's own site. No named customer, no third-party case study, no funding or Crunchbase record turned up anywhere in independent research. That doesn't make the work fake so much as it makes the discipline clear: treat vendor-published outcomes, Homodeus's or anyone else's, as claims rather than verified results before betting a workflow on them.
The real question isn't the engagement. It's what's left afterward.
A digital twin built for one production pod describes, by construction, how a company's systems connected at the moment that engagement happened. Six months later an ERP module gets upgraded, a second facility goes live on a different WMS, a carrier changes its API, and the twin is stale until somebody re-engages the people who built it. Gartner's framing of the last two decades of supply chain software makes the same point at an industry level: business intelligence dashboards in the 2000s, control towers in the 2010s, command centers and digital twins in the 2020s, and only now, orchestration platforms that prescribe and perform a decision rather than modeling it once and drifting out of date. A digital twin is real progress over a dashboard, but it stops short of being infrastructure the company owns and keeps extending on its own.
IDC's research points at the same gap from a different angle: by 2029, an estimated 45% of Global 2000 companies will run agentic-AI-driven orchestration across their supply chain ecosystems, and IDC's stated reason is that ERP-anchored, Tier-1-only visibility leaves companies blind to risk across their extended supplier networks. A bespoke build scoped to one production pod carries that same blind spot by design. It models what the engagement covered and nothing past that boundary.
Legacy sprawl doesn't hold still for a one-off model
This would matter less if logistics systems were static. They're not. The TMS market alone is projected to grow from $18.5 billion in 2025 to $37 billion by 2030, and WMS from $4.57 billion to over $10 billion in the same window, meaning the systems underneath any AI engagement are themselves being replaced and re-integrated on an ongoing basis. Seventy percent of Fortune 500 companies still run mainframe systems for critical infrastructure, with an estimated 220 billion lines of COBOL still executing today. None of this is for lack of investment. 55% of supply chain leaders are increasing tech spend this year, with 19% committing more than $10 million, and after a multi-year pullback, supply chain tech VC funding is climbing back toward its old highs. Gartner expects spend on supply chain software with agentic AI features to grow from under $2 billion in 2025 to $53 billion by 2030. All of that money is going into systems that a knowledge graph modeled once, at the start of an engagement, has no way to keep tracking.
When durability doesn't happen, even at scale
It isn't only small consultancy engagements that skip this. Amazon announced its Blue Jay multi-arm warehouse robot with real fanfare in October 2025, backed by a $200 billion AI infrastructure budget, and quietly discontinued it within months. Even at Amazon's scale, an AI initiative that isn't built as durable, compounding infrastructure gets shelved rather than iterated on. DPD's customer service chatbot made headlines for the opposite failure mode in January 2024, when it swore at a customer and wrote a poem calling its own employer "the worst delivery firm in the world". A production AI agent shipped without the governance to catch that before it reached a customer, and the fix was the only one available for ungoverned software: turning it off.
Compare that to McKinsey's account of a last-mile delivery operator running more than 10,000 vehicles: a $2 million investment in virtual dispatcher agents produced $30-35 million in savings. The difference between that outcome and Amazon's or DPD's isn't a smarter model. It's whether the thing was built to be validated, governed, and kept current, or shipped once and left to drift.
What actually compounds
This is the same distinction we've built Hintas around, and it shapes our own delivery model too: even our enterprise engagements are built to leave the customer holding a knowledge graph they own and can keep extending, not a project deliverable that ages out the day the contract ends. We extract that graph from a company's own sources of truth (OpenAPI specs, SOPs, runbooks, database schemas), the same way we've described agent memory as a first-class primitive rather than something bolted on after the fact. We validate it against staging on an ongoing basis, not just at go-live. And we expose it to any MCP client through two tools instead of one per endpoint, so the graph keeps working as new systems get added instead of being scoped to whatever the original engagement covered. Human-in-the-loop approval gates the actions that matter, whether the agent came out of a six-week engagement or a platform built to last years.
Vertical AI wins because it starts from a company's actual constraints instead of generic capability, but only if the model of those constraints outlives the person who built it. That's the thread running through this whole series: visibility platforms that stop at the alert, voice agents that stop at the phone call, and bespoke builds that stop at the invoice. The gap in all three cases is the same one. Something has to hold the company's operational knowledge somewhere durable enough that the next agent, whoever builds it, doesn't start from zero.
If you're interested in early access, reach out at hintas.com.