September 2026 · 5 min read
Where AI ROI actually shows up
unrecorded work
Key Definitions
Invisible work Processes that exist in reality but never on a cost line — no owner, no SLA, no system record. Missed carrier calls, unlogged phone updates and quarterly-if-at-all price monitoring are typical forms.
Baseline-first valuation Estimating the cost of doing a process fully manually — as the baseline — before arguing about ROI. For work with a zero baseline, any value an agent creates is pure increment.
The measurable AI ROI of 2026 sits in work that was never systematized. Evans Transportation used agents to catch 100K+ inbound carrier calls it used to miss; one C.H. Robinson agent pulled 318K tracking updates out of phone calls in a month; Duvo delivered EUR 2.8M+ in annualized savings for Rohlik in three months, from processes that had never been systematically run. Not a coincidence — the same structure repeated three times: AI’s most valuable work sits exactly where ROI ledgers cannot see.
Three cases, one structure
Evans Transportation (US 3PL): before AI call agents, roughly half of inbound carrier calls were missed. Agents have now answered 100K+ calls — identifying carriers by MC number, confirming safety and setup status, screening spam and bad actors, running an initial rate qualification. Manual order entry dropped from 100–120 orders per person per day to one or two touches (Inbound Logistics, Aug 2026, self-reported by Evans).
C.H. Robinson (3PL, company-reported Oct 20 2025): CTO Mike Neill says one AI agent captured 318,000 freight tracking updates from a single type of phone call in September — data previously invisible to its systems — feeding another agent that updates the platform and the predictive ETAs. Its Always-On Logistics Planner workforce spans 30+ connected agents.
Duvo × Rohlik (retail, Anthropic case study): Rohlik previously monitored thousands of commodity-linked SKUs quarterly at best, often not at all. Duvo’s agents monitor continuously, build negotiation cases and initiate supplier outreach: EUR 1.45M in annualized savings in the first week across 120+ SKUs and 15+ suppliers; EUR 2.8M+ in three months; 40%+ of team capacity freed on average.
The common thread: every return came from work that was never done, never doable, never recorded — not from speeding up work people already did. Missed calls had no cost line; phone-based tracking updates had no database field; quarterly price checks had no SLA.
Our judgment: get the baseline right, or the ROI debate is fake
Standard ROI compares before vs after. Invisible work has no before: never recorded, never assigned, never costed. No baseline in the ledger, so the return reads as zero. Hence 2026’s most common contradiction: operations says “we now do what was impossible before,” while finance says “ROI cannot be calculated.”
Our judgment: fix the baseline before arguing about ROI. For every candidate process ask three questions: ① does it have an owner on the org chart? ② does it have a cost line or SLA? ③ has any system ever recorded it? Three no’s equal invisible work equal a zero baseline. For zero-baseline work, any value an agent creates is pure increment — the ROI denominator is the cost of doing it fully manually, not 0, and not “it had no cost because nobody did it.”
This also explains why pilots fail to scale: pilot ledgers use “speed up an existing process” as the denominator, while scale value hides in invisible work. The more faithfully you screen pilots with the old ledger, the more you miss the scenarios worth scaling.
Action list: repair the ROI ledger with baselines
① Inventory invisible work
Ask every process owner to list flows that exist but have no cost line — missed calls, unrecorded updates, quarterly-if-at-all monitoring, verbal handoffs.
② Value baseline-first
How many person-hours, how much latency, what opportunity cost to do it fully manually? Baseline first, returns second.
③ Put the baseline into the ROI math before deciding
The denominator is the cost of doing it fully manually, not 0; zero-baseline work is booked as pure increment.
④ Report incremental value over agent cost
Most invisible work pays out as new value, not headcount savings — the wrong metric understates the whole program.
For measurement: in call handling, track answer rate × first-resolution rate × conversion or loss prevention; in data capture, new records ingested × downstream usage; in monitoring, the decision-latency delta from quarterly to daily. None of these exist in the old ledger — they are where AI ROI actually lands in 2026.
OOMeta AI
OOMeta’s own operation is evidence for this framework: one human plus many agent units, creating value mostly in the region where no person had capacity — signal monitoring, content production, audit loops. We use the same baseline-first frame when evaluating agent scenarios for clients: find work with no cost line, no owner and no SLA first; argue about returns after.
Schedule a DiagnosticSources: Inbound Logistics Aug 2026 Evans case (via Marketscale; Evans self-reported) https://www.marketscale.com/industries/transportation/evans-transportation-says-ai-agents-answered-100000-inbound-carrier-calls-in-months · C.H. Robinson press release (Oct 20 2025, vendor-reported) http://chrobinson.com/en-us/about-us/newsroom/press-releases/2025/ch-robinson-agentic-supply-chain-advance-2025 · Anthropic case study, Duvo (vendor case) https://claude.com/customers/duvo · Vooma × Evans case study https://www.vooma.com/case-studies/evans
FAQ
Why does a standard ROI ledger systematically understate AI returns?+
Standard ROI compares before vs after. But missed calls, unrecorded tracking updates and quarterly price checks never had a cost line, so there is no before in the ledger — the return reads as zero.
What are the Evans Transportation numbers?+
The US 3PL’s AI answered 100K+ inbound carrier calls in months; roughly half used to be missed. Manual order entry dropped from 100-120 orders per person per day to one or two touches (Inbound Logistics Aug 2026, self-reported).
What are the C.H. Robinson numbers?+
Company-reported (Oct 20 2025): one AI agent captured 318,000 freight tracking updates from a single phone-call type in September, data previously invisible to its systems. Its digital workforce spans 30+ connected agents.
What does the Duvo/Rohlik case show?+
Anthropic’s case study: Duvo’s agents delivered EUR 2.8M+ in annualized savings at Rohlik Group in three months, across processes that had never been systematically run; the first week alone closed EUR 1.45M from price monitoring.
How should an enterprise apply this framework?+
Two steps: inventory invisible work — processes with no cost line, no owner and no SLA — then value each candidate baseline-first, putting the cost of doing it fully manually into the ROI math.
Related Articles
Tripadvisor’s 2.5-month voice agent: design first
Tripadvisor: 8 people, 2.5 months to a live voice agent (vendor case). Sentiment 90% vs 71% human. Our take: agentic CX learns from every call.
Sysco put $500M of AI savings in its guidance
Sysco (Sept 9): $500M in AI-driven savings by FY2029, $100M in FY2027 guidance, EPS targets raised. AI savings count only when they enter financial guidance.
Citi’s Arc: the largest measured agent deployment
Citi’s Arc runs agents as a central OS: 180k staff, 40k devs on Devin, 100k+ agentic hours weekly, legacy migration from 12 months to 4 weeks.
1,100-person firm, 50+ agents: ROI is a discipline
ABC Legal’s 50+ agents cut costs ~50% (vendor-reported). Transferable: recommend, review, labeled data, eval gate, automate; per-agent efficiency accounting.