How the Roiva AI in Practice Index works

How the AI in Practice Index counts AI in use, its figures, governance, Responsible AI and people, how it names leaders, and how to correct it.

The Roiva AI in Practice Index scores AI that is running, measured and accountable, as a company's public evidence shows it:

  • running: how many distinct uses of AI the company describes, and how many are live;
  • measured: whether figures for its AI are on the public record;
  • accountable: whether it shows how it governs AI, the Responsible AI practices it has in place, and practices for the people AI affects: its staff (training, redeployment, a say in how AI is rolled out) and its customers (bias testing, human review or appeal of AI decisions).

This page explains what we count, how we weigh it, who is named, and how to correct us. One method covers every edition of the index.

What it measures

The index is built from public evidence: what a company has said about its AI, about how it governs AI and about AI's effect on its people, in SEC filings, earnings calls, press releases, investor presentations, its own website, vendor case studies and trade press.

Evidence a company sends us privately never counts toward the index, in any edition: not toward a score, a rank, a leader decision or a profile. Our evidence guidelines explain what happens to it. Being a Roiva customer, partner or prospect never changes a score either; see our independence policy.

It doesn't measure how well that AI works, what it earns or saves, or how advanced a company is. A company that discloses little may be doing more than it says. And every figure in the index is a claim as published, by the company or, for some figures, by its vendor or the press: none is audited or compared with a control group.

What we count

We look at each use of AI a company describes, such as claims triage, underwriting, a coding assistant or customer service. General statements about investing in AI, and AI that a company sells to its customers as a product, don't count as uses. Each use is assigned a standard use-case category, the same set of categories for every company.

How specific a use must be. A use of AI counts when its source says both that AI is used and what it does (a vendor case study or the press counts under the source rules on this page): they name a tool, model, agent or AI product, or the specific work the AI does, such as screening submissions, answering after-hours calls, flagging suspicious transactions or writing code. A company rebuilding a named business process around AI, with a stated change to that process, also counts: for example, "AI-native software delivery lifecycle transformation: 20–30% faster cycle times." AI named only for a business function ("AI in underwriting"), a goal ("to increase efficiency"), a program, team, training effort or investment counts as a general statement about AI, which earns no points.

The four editions published on November 16, 2026 (Mid-Size Insurers 2026, US Regional Banks 2026, US Freight & Logistics 2026 and US Industrial Machinery 2026) were researched and checked under an earlier reading of this test, which also counted AI named for a function together with a goal. Their scores and leaders stand as published. Every later edition applies this test from its first research pass.

We also look at how the company governs AI: board oversight, the people and policies responsible for it, the Responsible AI practices it says are in place, and what it does for the staff and customers AI affects.

A score out of 100 is built from six parts:

Part What we count Full marks at Weight
Breadth distinct uses of AI 6 uses 15
Depth distinct uses described as in production or at scale 4 uses 30
Measurement figures the company has disclosed about its AI, weighted by kind (below) 2 points 20
Governance, Responsible AI and people board oversight, an AI leader, AI policies, and Responsible AI and workforce practices in place (below) 6 points 30
Source quality share of uses backed by the company's own disclosures or SEC filings, not only vendors or press all 2.5
Recency share of uses with a source from the last 12 months all 2.5

Why these weights. Depth and Governance, Responsible AI and people carry the most, 30 each. Depth, because the index is about AI that is running. Governance, Responsible AI and people, because it is what makes AI accountable, and half of its points are for practices in place rather than statements. Measurement carries 20. Breadth carries 15: how many uses a company describes matters less than how many are running and how they are governed. Source quality and recency carry the least, because they separate companies least.

Measurement points. For each use of AI we take its single strongest figure:

  • 1 point: a cash figure the company credits directly to the AI initiative.
  • 0.6 points: hours or staff time.
  • 0.5 points: an operational result (speed, accuracy, a before-and-after) or work the AI did (calls or tasks handled, a share automated), or a cash figure where the company names AI as one of several causes.
  • Nothing: a count of people using a tool, such as "94% of staff use AI", with no result.
  • Nothing: a cash figure the company doesn't tie to AI. That includes correlations, a company-wide ratio mentioned beside AI, a level with no before-and-after, and savings credited to an unnamed party.
  • Half points: a forecast or target earns half the points of a result.

The size of a figure never matters.

Governance, Responsible AI and people. We count seven signals, each once, and only where the company names AI for the thing itself and says the policy or practice is in place, not planned. They add up to 6.5 points, and full marks are at 6, so a company can reach full marks without a C-suite AI leader.

  • Board oversight (1 point): the board, or a named board committee, oversees AI.
  • AI leader (1 point for a C-suite executive, half for anyone else): a named person has AI in their title, or the company describes them as leading its AI work.
  • AI policy (half a point): the company has an AI policy or framework, or a management committee for AI. An acceptable-use policy is enough.
  • Responsible AI policy (1 point): a policy, principles or program committing to the responsible use of AI, such as fairness, transparency, human oversight or consumer protection. That includes a program under the NAIC Model Bulletin on insurers' use of AI. It also earns the AI-policy half point.
  • Responsible AI control practice (1 point): a control on AI in place, such as an inventory of AI systems, reviewing or tiering AI by risk before it is used, validating models or monitoring them for drift, or checking AI vendors.
  • Responsible AI people practice (1 point): a practice that protects the people AI decisions affect: testing AI for bias or unfair discrimination, or a named human review, override or appeal of AI decisions about customers, applicants or staff, saying what is reviewed and by whom or at which step.
  • Workforce practice (1 point): from the company's own disclosures or SEC filings, one of:
    • an AI training, literacy or reskilling program for staff that is named, required or counted;
    • staff hours, roles or headcount the company credits to AI, with a figure, together with where the time or the people went;
    • a named way staff have a say in how AI is rolled out.

Three of the seven signals are practices in place, and they carry half of the points that count.

These don't count for any signal: - a director's AI experience, or board oversight of "technology" without AI named; - a heading that names AI, for the policy or practice under it, or AI named elsewhere in the document (see "How we read a source" for a sentence that refers back to AI).

These don't count as a practice (control, people or workforce): - commitments, expectations and plans; - statements that AI won't replace people; - general statements that people stay "in the loop"; - a customer's route to a person, such as "you can always reach a human"; - counts of staff with access to, or using, an AI tool; - staff acknowledging an AI policy; - for the workforce practice, a vendor's case study or a press account.

One passage, several signals. A sentence or paragraph can support more than one signal when each signal rests on different words in it, such as an inventory of AI systems for the control practice and bias testing of AI for the people practice. No words count toward two signals, and each signal must name AI for the thing itself on its own words. A list that only names topics, such as "a governance framework with monitoring, bias testing, disclosures", doesn't show a practice is in place unless it says what is done.

How we read a source. These apply to every edition:

  • A company's own statement that it uses AI counts wherever it appears, including in its risk factors. A sentence that names AI only as a risk, and no use of it, doesn't.
  • AI must be named for the thing itself. A sentence that doesn't name AI counts only when it clearly refers back to AI named in the sentence just before it: that sentence names the AI tool, model, agent, policy or committee itself, not AI or technology in general; the reference can only mean that AI; the two sentences are next to each other, with no change of topic; and the words we quote hold both. When it isn't clear, it doesn't count. A heading, a slide title or a label that names AI never carries over to what sits under it. A sentence that introduces a list does carry over to each item when it says AI is used for each ("machine learning is embedded in the following: ..."), or when it refers back, under the same tests, to the AI named in the sentence before; we quote it with the item. A lead-in that names AI beside something else, such as "data analytics and AI", doesn't. This applies to uses, figures and Governance, Responsible AI and people alike.
  • Trade press counts only for the company's own words, quoted directly.
  • A cash figure for a wider program counts as one where the company names AI as one of several causes when the company credits it to named levers that include AI.
  • When a company merged into another during an edition's research window, its disclosures about AI count toward the company it merged into, labeled as the predecessor's. Its governance doesn't carry over.

Notes for banks. The same method and weights apply to banks, with these readings:

  • Governance must name AI. Banks describe model risk management, model validation and model inventories under long-standing supervisory guidance. That counts toward Governance, Responsible AI and people only where the passage itself names AI or machine learning.
  • Fraud and anti-money-laundering vendors. A vendor's marketing that its platform uses AI doesn't make it the bank's use of AI. It counts when the bank says it uses AI or machine learning for that work, or a vendor case study names the bank and the use.
  • Fintech partners. AI run by a bank's fintech partner isn't the bank's own use unless the bank says it runs it or uses it.
  • Efficiency ratio. A bank's efficiency ratio mentioned beside AI is a company-wide ratio, and earns nothing.
  • Risk factors. AI named only as a risk, such as fraud by others using AI, isn't a use.

Notes for freight and logistics companies. The same method and weights apply, with these readings:

  • Telematics, dash cams and visibility platforms. These are often sold as AI. A vendor's marketing that its platform uses AI doesn't make it the company's use of AI. It counts when the company says it uses AI or machine learning for that work, or a vendor case study names the company and the use.
  • AI in the freight service. AI that a carrier or broker uses to run its own freight service, such as pricing, matching loads with carriers, routing or quoting, counts as a use even when shippers see it. AI software the company sells to others as a separate product doesn't.
  • Safety programs. An AI dash cam or driver-monitoring program counts as a use when the company names AI. A safety or compliance program isn't AI governance unless it names AI.
  • Operating ratio. A company-wide operating ratio mentioned beside AI earns nothing, as a bank's efficiency ratio doesn't.
  • Autonomous trucks. A company's own autonomous-truck pilot counts as a use, with the status the company gives it: a pilot unless the company says it is running.
  • Spin-offs. When a company was spun off during an edition's research window, its former parent's disclosures from before the spin-off count toward it only where they name the spun-off business, labeled as the predecessor's. The parent's governance doesn't carry over.

Notes for machinery makers. The same method and weights apply, with these readings:

  • AI in the product. AI in equipment or software a company sells, such as connected pumps, condition-monitoring platforms or vision in scanners, isn't a use of AI in its own operations. A service sold to customers on data from their equipment is part of the product. It counts as a use when the company says it uses that data, with AI, to run its own parts, service or production.
  • Selling to AI. Cooling, power, pumps and chip-making tools sold for AI data centers are demand for what a company makes, not a use of AI, and earn nothing.
  • Automation isn't AI by itself. Robotics, collaborative robots, automated cells, smart factories, Industry 4.0, connected sensors and analytics count only where the source names AI or machine learning for them.
  • Vendors. A vendor's marketing that its platform uses AI doesn't make it the company's use of AI. It counts when the company says it uses AI or machine learning for that work, or a vendor case study names the company and the use.
  • Company-wide figures. An operating margin, a lean or 80/20 program, or productivity savings mentioned beside AI earn nothing, as a bank's efficiency ratio doesn't.
  • Governance must name AI. A quality system, a product-safety program or a cybersecurity program isn't AI governance unless the passage names AI.

Editions

Each edition covers one defined group of companies, set by industry, size, listing and year, and is a snapshot as of a stated date:

  • Mid-Size Insurers 2026: the US SEC filers in insurance with calendar-2025 revenue of $50 million to $2 billion, as of November 9, 2026.
  • US Regional Banks 2026: the US bank and thrift holding companies that file 10-Ks, with total assets of $10 billion or more and calendar-2025 bank revenue of $50 million to $2 billion, as of November 12, 2026.
  • US Freight & Logistics 2026: the US freight transportation and logistics companies that file 10-Ks, with calendar-2025 revenue of $50 million to $10 billion, as of November 14, 2026. Its revenue band is wider than the other editions' because many US carriers and freight brokers have more than $2 billion in revenue. One leader rule covers the whole group, and the edition page reports the two revenue bands, $50 million to $2 billion and $2 billion to $10 billion, side by side.
  • US Industrial Machinery 2026: the US makers of industrial and commercial machinery (other than computer and office equipment) that file 10-Ks, with calendar-2025 revenue of $50 million to $10 billion, as of November 14, 2026. As in the freight edition, its band runs to $10 billion, one leader rule covers the whole group, and the edition page reports the two revenue bands side by side.

Companies in a deal. A company that merged into another company before an edition's as-of date is left out; the buyer counts only if it is in the same group. A deal that was agreed but hadn't closed by the as-of date leaves the company in, and if it closes before the edition is published, the company is left out. A take-private, where the company carries on under a new owner, never removes a company. When the company it merged into is in the same group, its disclosures about AI count toward that company, labeled as the predecessor's (see "How we read a source").

Who is ranked

AI-native companies, whose product is AI, are scored the same way as established companies but ranked separately. In 2026 we don't rank them or name them as leaders.

Who is named an AI in Practice Leader

A company is named an AI in Practice Leader when all of these are true:

  1. Its score is at least 55 out of 100.
  2. It has at least 1 point of Governance, Responsible AI and people on the record. A single half-point signal on its own, such as an AI policy or an AI leader who isn't a C-suite executive, isn't enough.
  3. It is in the top quarter of ranked companies. The number of places is a quarter of the ranked companies, and they go to the highest-scoring companies that meet rule 2, so a company that doesn't meet it leaves its place to the next one that does. Companies tied at the cut-off are all included.
  4. It has disclosed at least 3 distinct uses of AI, and at least 2 of them are in production or at scale.
  5. At least one of its own disclosures or SEC filings about AI is from the last 12 months.
  6. A Roiva staff member has checked every source behind its score.

Leaders are listed alphabetically. We don't publish scores or a ranking. We tell every company in the study its own score if it asks, after confirming the request with the company through a contact we find independently.

Profiles

Each leader gets a profile. Any other company in the study can claim a public profile; it is published only if the company opts in and its public evidence shows all of these:

  • at least 3 distinct uses of AI, at least 2 of them in production or at scale;
  • at least 2 of those uses backed by the company's own disclosures or SEC filings;
  • a company disclosure about AI from the last 12 months;
  • at least one disclosed figure that counts under "Measurement points", or one Governance, Responsible AI and people signal;
  • a Roiva staff member has checked every source the profile shows.

A profile lists the uses of AI, their status and use-case category, the governance, Responsible AI and people practices the company has on the record, and the date and link of every source. Every figure is labeled as the company's own claim, and a forecast is labeled as a forecast. Figures the index gives no points to never appear on a profile. A profile shows a figure only in the company's own words, from its filings, releases, calls, presentations or website. A use described only by a vendor or the press is listed with that source, but a figure that appears only there isn't shown, though it can still earn points.

Any company with a public profile, leader or not, can add a short section of its own under our terms for company-provided sections. It is labeled "Company-provided", shown apart from the sourced evidence, holds no figures, and never counts toward a score.

Who did the research

Under the direction of Roiva staff, AI research agents (Claude, made by Anthropic) found the sources for each company and copied the exact words used, 40 words or fewer per source. A script, and in some cases a separate checking agent, confirmed each quote on its source page. Roiva staff decided the method. A Roiva staff member checks every source behind a leader or a public profile before it is published.

Corrections and changes

  • To correct us, write to data@roiva.ai with a link to a public source. We add or change evidence only when anyone can check it in public.
  • Timing. We reply within one week, saying whether we've made the change or when to expect an outcome. We reach an outcome within 30 days, or say why not and give a new date.
  • A second look. If you disagree with the outcome, tell us why. We'll look again and give you our reasons in writing.
  • When a correction changes a result (the leader list or a profile), we recompute the score under the same method, update the edition within one week (or tell you why not and give a new date), and record the change, dated and with the reason, on the edition's public change log. A correction that changes only a score we don't publish is counted there without naming the company. The log starts at publication: changes made while an edition is drafted are revisions of a draft, not corrections, and aren't listed.
  • A correction can add a leader, never bump one. If a correction to a company's public evidence brings it within the leader rule, it is named a leader and the change is recorded on the change log. A leader is never removed because another company's score changed: it stays named, and keeps its badge, unless a correction to its own evidence means it no longer meets the leader rule. So an edition can name more leaders after a correction than at publication. The next edition starts afresh.
  • Removal. A company can ask, free, for its profile page to be taken down; its name stays on the dated leader list. A leader can also ask to be taken off the public list. Its score doesn't change and nobody moves up in its place. The list says how many leaders asked not to be listed, and a company off the list may not show the badge. It can ask to be listed again at any time.
  • The method only changes between editions. We don't change an edition's method after publishing it. A change applies to the next edition, and is recorded, dated and with the reason, under "Method history" below. Before each edition we raise any limit (for example "full marks at 6 uses") that more than a quarter of ranked companies have reached, and we show the previous edition re-scored under the new method, so a company's movement reflects its own disclosures.
  • Every badge is dated ("Mid-Size Insurers 2026"). A later method change doesn't take it back.

Limits

  • Public disclosure varies with company size, listing and style. Mutual insurers and other companies that don't file with the SEC disclose less, and they aren't in the 2026 insurer edition. Banks that don't file 10-Ks, including privately held and mutual banks, aren't in the 2026 bank edition. Freight companies that don't file 10-Ks, including privately held carriers and foreign shipping companies that file on Form 20-F, aren't in the 2026 freight edition. Machinery makers that don't file 10-Ks or have their business address outside the US aren't in the 2026 machinery edition, and neither are makers the SEC classes in another industry.
  • Vendor case studies are written by the vendor. They count toward breadth and depth, but not toward source quality.
  • Disclosure is rising fast: most of the 2026 evidence is from 2026 itself. Each edition is a snapshot as of its date.

Method history

  • November 10, 2026: the method first set out, for Mid-Size Insurers 2026.
  • November 13, 2026: before either edition was published, we set out how we read a source (risk factors, trade press, program-wide cash figures and merged companies) and applied it to both editions, and added the notes for banks for US Regional Banks 2026. The weights, limits and leader rule didn't change.
  • November 14, 2026: before either edition was published, we set the index to measure AI that is running, measured and accountable, and applied the method set out on this page to both editions. Governance & Responsible AI became Governance, Responsible AI and people, weighted 25 rather than 15 and with full marks at 6 points rather than 5. Breadth went from 25 to 20, and source quality and recency from 5 to 2.5 each. An AI policy now earns half a point. The Responsible AI practice signal became two: a control practice and a people practice. A new workforce practice counts AI training, redeployment and staff voice, from the company's own sources. The limits, the measurement points and the leader rule didn't change.
  • November 15, 2026: before US Freight & Logistics 2026 was published, we added the notes for freight and logistics companies, including the rule for spin-offs, and scored that edition under the method set out on this page. The weights, limits, measurement points and leader rule didn't change.
  • November 16, 2026: before Mid-Size Insurers 2026, US Regional Banks 2026, US Freight & Logistics 2026 and US Industrial Machinery 2026 were published, we raised Governance, Responsible AI and people from 25 to 30 and lowered breadth from 20 to 15, lowered the leader threshold from 60 to 55, and added the leader rule that a leader has at least 1 point of Governance, Responsible AI and people. We scored all four editions under the changed method. We also set out when a sentence that refers back to AI named just before it counts, and applied that to all four editions, and we added the notes for machinery makers. The signals, limits and measurement points didn't change.
  • November 17, 2026: we set out on this page when a sentence that introduces a list carries over to its items. That reading has applied to all four editions since November 16, so no score or result changed. We also set out that one passage can give several signals on different words. Before, we had counted one signal per paragraph, which the method never stated; the signals left out under that shorthand were read again, and the results that changed are on each edition's change log. Separately, we set out how specific a use of AI must be to count, for every edition after the four published on November 16, 2026. Those four editions keep the earlier reading, and their scores and leaders don't change.
  • November 18, 2026: we corrected this page, the independence policy, the overview and each edition's page, which said every figure in the index is the company's own claim, put on the record by the company. Under the source rules above, a vendor's case study or a press account can also give a figure. The method, the policy and every result are unchanged.