Agentic AI ROI by Industry: What Mid-Sized Businesses Can Realistically Expect

What ROI can you realistically expect from AI?

Across the market, IDC’s 2024 study (sponsored by Microsoft) found an average return of about $3.70 for every $1 invested in generative AI, with top performers nearer $10.30, and value realised within roughly 13 months (IDC, 2024). That is the optimistic, vendor-sponsored headline. The honest counterweight: independent research finds strong returns are still the exception, not the rule.

Put the promise and the reality side by side and the picture is clear. McKinsey’s 2025 State of AI found that only about 39% of organisations attribute any EBIT impact to AI so far, and most of those put it at under 5% of EBIT, with roughly 6% counting as high performers (McKinsey, 2025). Deloitte’s 2024 State of Generative AI found that while about 74% say their most advanced initiative is meeting or exceeding ROI expectations, only about 15% report significant, measurable ROI (Deloitte, 2024). Both things are true at once: the ceiling is high, and most firms have not reached it yet.

Why ROI varies so much by sector

The sector matters less than the shape of the job. Returns cluster wherever the work is high-volume and rule-based. A repetitive, document-driven task pays back faster than a bespoke, judgement-heavy one, whatever industry it sits in.

This is why IDC (2024) ranks financial services first on gen-AI returns: underwriting, claims and onboarding are structured, repetitive and expensive, so a small percentage improvement is worth a lot. The same logic explains why legal document review and clinical documentation show up strongly, and why sectors dominated by one-off, relationship-heavy work show thinner, slower returns. When you read the table below, look less at the industry label and more at whether the winning use case is a job your own team repeats often enough to matter.

Agentic AI ROI by industry

Here is the sector-by-sector picture, with a sourced result and a realistic time-to-value for each. Every figure is attributed to a named source, and where a source is a vendor-sponsored survey we say so. Figures are indicative: they come from different studies, geographies and years, so use them to set expectations, not to forecast your own numbers.

A few of these deserve a note. In financial services, the widely-repeated "20x speed, 80% cost" claims for claims processing trace only to vendor marketing, so we have used named cases instead: Hiscox’s underwriting result and Allianz’s reported cut in claim processing time (Hiscox, 2024; Allianz, 2025). The 77% of financial-services firms said to report positive returns within a year comes from a Google Cloud survey, so read it as a vendor-sponsored figure (Google Cloud / National Research Group, 2024).

Healthcare is the clearest example of why the counterweight matters. The Permanente Medical Group’s roughly 15,700 saved documentation hours is a large, independently-reported result (Kaiser Permanente / NEJM Catalyst, 2024 to 2025), but an independent multi-site study found the per-clinician saving was far smaller, about 16 minutes per 8-hour shift (STAT, 2026). Both are real. The lesson is that aggregate hours saved across thousands of users can be large even when the per-person minute-level saving is modest.

In legal and professional services, the Bird & Bird case is genuine but comes from the vendor whose tool was used, so treat "79 to 3,600 documents an hour" as a vendor-reported client result (Luminance / Global Legal Post, 2018). The scale of investment is easier to stand up: Deloitte has committed around $3bn, KPMG around $2bn and EY around $1.4bn to AI (Bloomberg Tax, 2024 to 2025). In manufacturing and logistics, we have deliberately dropped the often-quoted "General Motors saves $20m a year" line, which we could not trace to any GM source, in favour of Deloitte’s and McKinsey’s sourced ranges, plus BCG’s finding that supply-chain AI can add roughly 2 to 4 percentage points of EBITDA (BCG, 2024). In HR, per-hire pound-savings figures almost always come from vendor ROI calculators, so we have kept only the attributed survey number.

How to read AI ROI numbers without getting burned

Most of the damage in AI business cases comes from taking a vendor’s best-case number as a planning assumption. A simple filter: if a figure has no named source, or the only source is the company selling the tool, treat it as a ceiling to aim at, not a baseline to budget on.

• Separate independent research from vendor surveys. McKinsey, Deloitte and government studies are measuring the whole market. A vendor’s case study is measuring its best customer. Both are useful, but only one tells you the odds.
• Watch the currency and the year. The much-quoted "$3.50 per $1" is IDC’s older figure; the current one is $3.70 (2024). A "£3.20" healthcare return is actually US$3.20. Small errors like these are how a stat drifts away from its source.
• Prefer aggregate, real-world deployments. "Saved 15,700 hours across 7,260 physicians" is more trustworthy than "saves 3 hours a day", because it is measured at scale rather than demoed once.
• Assume the average firm, not the top performer. The gap between the $3.70 average and the $10.30 top-performer figure is the groundwork. If you have not done it, plan for the average.

What actually determines your ROI

The sector benchmarks set the ceiling. Whether you get near it is decided by the same unglamorous factors every time: a job worth automating, clean-enough data, a fixed process, and people who actually adopt the tool. The technology is rarely the deciding variable.

This is the single most consistent finding in the research, and it is why we wrote a whole guide on why most AI projects fail. The short version: roughly 70% of what makes AI succeed is people and process, not the algorithm (BCG, 2024). The firms that hit the top-performer returns pick one contained, high-value job, map how it really works, fix the broken parts before automating, measure a baseline, and scale only what works. The firms that see nothing usually bought a tool before they had defined the job. If you want the full playbook, start with that guide, then come back to the sector table to size the opportunity.

How AI colleagues change the ROI math

If most of the return depends on adoption and on knowledge accumulating rather than leaking away, the shape of the tool matters. That is the idea behind Frntir’s AI Synths: a named AI colleague with persistent memory that owns a recurring job, rather than another system your team has to remember to open.

A Synth works inside the tools your team already uses, operates with bounded, explainable autonomy under a human boss, and keeps a full audit trail. That targets the two things the numbers above keep coming back to. It lowers the adoption barrier, because it lives where people already work rather than being a new destination, and it compounds, because the context it builds up (past decisions, what worked, who to ask) stays with the business instead of walking out with a spreadsheet. It does not remove the need to pick a good job and fix the process first, because nothing does. But on the jobs where AI already pays back, a colleague that owns the whole loop tends to close more of the gap between the average return and the top-performer return.