Agentic claims capture: same 64 companies, same instrument as the Outcomes study

Calidescope LLC. Capture run 2026-09-30. Counts describe the 64 companies and the pages read, not the market.

Summary

• Denominator. 64 companies checked out of the 140-company frame: all 20 tagged "AI-native", all 20 "Outcome sellers (deep read)" (Liftoff and Moloco sit in both lists, so 38 unique), plus 26 from other segments.
• Any agentic claim at all. 35 of 64 (55%) had at least one codable agentic claim on the pages read. 6 (9%) showed only a label or headline ("Agentic Advertising" as a nav item or blog title). 23 (36%) showed nothing we could code. Only 25 of the 35 had the claim on the home page or the frame URL; the other 10 came from a press release, blog or product page.
• By group. AI-native 15 of 20. Outcome sellers 6 of 20. Other segments 14 of 26. Among the 20 outcome sellers, only Kochava uses agent vocabulary on its home page.
• Autonomy split, 80 coded claims. ASSISTS 46 (57%) / ACTS 34 (42%). 25 of the 46 ASSISTS codes are "ambiguous, so ASSISTS" under the no-guessing-upward rule. If the 7 "you set the guardrails, the agent acts" claims are recoded ASSISTS, the split is 53 / 27. ACTS claims sit mostly in the AI-native group (23 ACTS to 16 ASSISTS). The other-segment group is 22 ASSISTS to 3 ACTS.
• Evidence rung, 80 coded claims. Says it 65 (81%) / prints a figure 7 (9%) / names the advertiser 6 (8%) / shows the method 2 (2%) / cites an outsider 0 (0%). Outcomes study: 221 / 159 / 290 / 42 / 28 of 740 (30% / 21% / 39% / 6% / 4%). The agentic set is thinner at every rung above "says it".
• Commitment. Of the 80 coded claims, 0 mention a guarantee, a remedy, pricing tied to a result, or paying when a number is missed. We did not find, on the pages we read, any company promising an agent that commits to a number with a remedy. The one contract we read (PubMatic AgenticOS agreement) says the opposite.
• Negotiation. One company in the frame lists negotiating among agent activities: Apostra (formerly Scope3). The word is one item in a list of 20 verbs. Beside it: "You set the price and rules." Two more negotiation claims sit outside the frame (Universal Ads, R2B2). All three keep a human or a rule-set upstream.
• Agents on outcomes. We did not find an agent priced on a result. We found agents whose pitch is a result (Taboola Realize+, Tatari Planning Engine, Quantcast Q+, MNTN), and a holding company (Omnicom) that says outcome pricing is coming for people-based work. The two lists do not touch on any page we read.
• Surprises. (1) PubMatic's homepage says AI agents "plan, execute, and optimize media"; its own blog says execution and approval should be separate; its contract puts all risk of agent actions on the customer. (2) The only method-rung agentic claims come from a performance-TV company (Tatari), which describes its own four-week randomized A/B test. (3) Scope3 no longer exists under that name: scope3.com redirects to apostra.com after a September 2026 rebrand.

How the instrument was applied

• Claim. A sentence or phrase asserting that an AI system acts, decides, executes, recommends or drafts for the user. Section headings and nav labels ("Agentic AI", "Agents at Adobe") were counted as label only, not as claims. 6 companies fall in that bucket.
• Vocabulary flag. 56 of the 80 claims use agent / agentic / autonomous / "human in the loop" language ("vocab = Y"). The other 24 say AI "automatically" finds, optimises or decides. Both are shown; totals use all 80 unless stated.
• ACTS vs ASSISTS. ACTS = the sentence says the agent does the thing (executes, optimises, buys, decides). ASSISTS = surfaces, recommends, drafts, summarises, or a human approves first. Operational rule for guardrails: "you set goals and guardrails, then the agent runs" is coded ACTS and flagged GRD (7 claims). "Agent proposes, you approve" is ASSISTS. Ambiguous claims are coded ASSISTS and flagged AMB (25 claims).
• Rung. Judged from what sits beside the claim on the same page or module. Homepage numbers with no source are rung 2. A named customer beside the claim is rung 3. A stated test design is rung 4. An outsider is someone who does not sell the thing; none sat beside a company-owned claim.
• Register is our coding, not the page's words.
• Verbatim status. Pages were read with WebFetch, which passes each page through a small extraction model. Every quote in the table is a string that model returned as verbatim. "B" in the verify column means a second, differently-worded fetch returned the same string. "A" means one fetch only. Treat "A" quotes as needing a spot check against the live page before publication. Page text can also differ between fetches (for example, some pages were mid-redirect).
• MADDB. news searches on agentic advertising, agents in media buying, agentic commerce, autonomous campaigns, agent-to-agent, negotiation, outcome pricing, guarantees and AdCP (about 20 queries; 5 hit a rate limit and were rerun; queries phrased as long sentences often returned zero). We did not pull company briefs. MADDB gave leads and context; every quoted claim in the table comes from a company or press URL.

Claims table

Flags: AMB = ambiguous, coded ASSISTS. GRD = agent acts inside guardrails the user sets, coded ACTS. Commitment column: "no" means the claim and its surrounding module mention no guarantee, remedy, results-tied price, or payment on a missed number. Group: AIN = AI-native, OUT = outcome seller, SPR = other segment. "Page" is home / product / press (a press release, blog post or insights page on the company's own site or wire).

Notes on specific rows:

• Row for Apostra "Negotiates": the string is one item in an agent-activity list under the heading "Human control, agent scale". The list as returned by the fetcher: "Finds opportunities / Checks audience fit / Scores potential / Compares pricing / Applies brand rules / Builds a plan / Requests terms / Reads proposals / Negotiates / Allocates budget / Books CTV / Books audio / Books OOH / Books social / Checks safety / Paces spend / Reallocates / Tracks delivery / Flags exceptions / Reports back". The item labels are verbatim by the second fetch; the "/" separators are probably the fetcher's rendering of a list or chip layout. Not certain of exact punctuation.
• Apostra homepage also says (second fetch, verbatim): "Buyers can browse or send a brief. You set the price and rules. Your agent handles the response." This is the seller-side agent.
• PubMatic row 3 uses a non-breaking hyphen in "natural‑language" as returned.
• Klaviyo "resolves 65% of questions autonomously": the homepage gives no source. A separate ROI-calculator page says "The 65% default resolution rate reflects Klaviyo's cross-company aggregate resolution rate across paying customers." and "ROI figures are based on a subset of Klaviyo Customer Agent customers that self-reported resolution rates. These customers participated in case studies and may not necessarily be representative." We coded the homepage claim rung 2 (source not beside it). A method note exists one click away, from the vendor, self-reported.
• Tatari rungs: the "17 of 18" and "without a human in the loop" sentences sit in an article that describes "a four-week randomized A/B test" run by Tatari on its own clients. No external auditor. Rung 4, method shown, not rung 5. Fetcher gave the figures "17 of 18 valid cases" and "12 of the 18" (the second is for clients under about $19.5K a week).
• Salesforce homepage prints figures ("32% increase in overall marketing ROI" etc.) attributed to "Salesforce Customer Success Metrics". They are not attached to the agent claims, so not coded against them.
• Mutinex: "Autonomous agentic build, human oversight" is coded ASSISTS because the same phrase includes human oversight.

Companies checked and result

Liftoff and Moloco appear in both the AI-native and outcome-seller lists and are counted once (64 unique).

Counter-evidence read from contracts and third parties (not counted in the 80)

Contract

Third-party and off-frame press (not company-owned claims for companies in the frame, or companies outside the frame)

Answers to the six questions

1. How many companies make an agentic claim at all?

Denominator: 64 companies, read on their home page (or the frame URL) plus, where the home page was thin, one or more press, blog or product pages found by search.

Notes. Moloco and Liftoff are in both lists (the AI-native and outcome-seller rows overlap by two; the "All" row de-duplicates). Only 25 of the 35 have the claim on the home page or frame URL. Nexxen, Taboola and Tatari have nothing agentic on their home pages; the claims came from a press release or blog. Under the stricter test (a coded sentence uses agent / agentic / autonomous / "human in the loop" language) the count is 29 of 64 (45%). Five of the 20 companies positioned as "AI-native" gave us no codable agentic sentence on their own pages (Liftoff, StackAdapt, Brandtech, Zefr) or a label only (Moloco). Zefr and MiQ have agent products reported in the trade press (see third-party table) that their own home pages did not mention; our absence count for them describes the page, not the product.

2. ASSISTS / ACTS split

80 coded claims: ASSISTS 46, ACTS 34 (57.5% / 42.5%).

Companies with at least one ACTS claim: 17 of the 35 that made a claim (6sense, Apostra, Equativ, Google Ads, Jasper, Klaviyo, Kochava, MNTN, PubMatic, Publicis, Quantcast, Salesforce, Smartly, Taboola, Tatari, Typeface, Writer). 25 of the 46 ASSISTS codes are ambiguity calls; a reader who codes those upward would see roughly 59 ACTS. The rule was to not guess upward, so this read is a floor for ACTS in the wording.

Pattern. The AI-native group writes in the ACTS voice ("agents that act", "An agent you delegate to", "Autonomous campaign performance is here"). Big platforms and incumbents write ASSISTS with an approval step in the sentence ("Campaigns only launch after you review and approve", "approval-based execution with clear guardrails", "Customize, override, or adopt recommendations"). The same company can do both: PubMatic's home page is ACTS ("lets AI agents plan, execute, and optimize"), its blog is ASSISTS ("a meaningful separation between the execution of a workflow and the approval of that workflow").

3. Evidence-rung distribution against the outcomes pattern

Read. 81% of agentic claims have nothing beside them, against 30% for outcomes. The modal outcomes rung (names the advertiser, 39%) is 8% here. The 6 rung-3 claims are vendor-owned: 6sense (a Reltio exec), Jasper (Anthropologie; Adidas), PubMatic (Amnet/Interbev; Klever/Rouge Care) and Tatari (Winona). Two claims reach rung 4, both Tatari, describing its own four-week randomized A/B test: "In 17 of 18 valid cases, the unmodified AI plan performed as well or better than the human-adjusted version." No company-owned claim reached rung 5. The nearest outsider data is DataBeat's 86%/13.4% bid analysis, which measures agentic buyers as a class (not a vendor's claim) and draws on partner SSPs including PubMatic.

Caution on the comparison. The two sets are not the same kind of sentence. Outcome claims say "this happened"; agentic claims mostly say "this is what the agent does", which is a capability description and may not need evidence in the way a result does. Also 80 claims from 35 companies is a small n. The comparison is directional. The fact that stands regardless: when agentic claims do carry a number, it is usually a speed or efficiency figure (40%, 50%, 87% faster setup, "1,098 hours") from the vendor, with the exception of Tatari.

4. Does any company promise an agent that negotiates terms or price, or commits to a number with a remedy?

Negotiation: one company in the frame lists it; two more sit outside the frame.

• Apostra (formerly Scope3), https://apostra.com/. The agent-activity list under "Human control, agent scale" contains "Requests terms", "Reads proposals" and "Negotiates". The same page says: "Humans set the objectives, context and rules. Agents handle the work required to act on them." and, on the seller side, "You set the price and rules. Your agent handles the response." So the negotiation claim is one word in a list, bounded by rules the human sets. No fee, price example or outcome is given on that page. The rebrand was reported 2026-09-23 to 2026-09-28 (MADDB: Business Insider, Adweek, MediaPost).
• Universal Ads (Comcast Advertising; not in the frame), 2026-09-02: "Buyer and seller agents can iterate in real time to finalize pricing and packaging." Beside it: "Guardrails, along with human oversight, ensure quality execution and alignment with brand standards." A pilot with one named advertiser (Crexi) and FOX Advertising as publishing partner.
• R2B2 pilot (not in the frame), with Omnicom Media as buyer, 2026-06-24: "Inventory was matched, terms were negotiated, and the purchase was executed end-to-end by autonomous AI agents." The publisher manually approved every campaign and creative. No budget or results disclosed.
• Trade-press framing, not a company claim: AdExchanger (MADDB snippet, 2026-01-29) says the industry "is racing toward a not-too-distant future where AI agents negotiate programmatic deals on their own". The Measure (CIMM, 2026-09-24; one fetch): "The market is now feeding signals into agent robots and letting them take the wheel on buying, planning and negotiating deals."
• Not found: on Magnite, PubMatic, Adform, The Trade Desk, Index Exchange, Equativ or Mediaocean pages we read, no claim that an agent negotiates price or terms. Magnite's orchestration blog and Adform's release describe discovery, evaluation and activation. PubMatic's agreement says nothing about agents negotiating terms.

Commitment to a number with a remedy: not found on any page we read.

• 0 of 80 coded claims mention a guarantee, a remedy, results-tied pricing, or payment on a missed number. Every page-level check for "guarantee, refund, pricing tied to results" returned none.
• The only agent contract read reverses the promise. PubMatic's AgenticOS agreement, Section 3.2: "PubMatic does not guarantee the outcome of any given campaign and does not guarantee any specific results." Section 3.3 puts on the customer "any and all risk" from "automated actions generated by Agentic Buying systems". Liability is capped at 12 months of fees (fetcher summary of the cap clause; wording not double-checked).
• Historic contrast for the study: the 8 guarantees in the outcomes set all sat upstream of the sale (fraud, attention, delivery, completion). We looked for an agentic version at IAS, DoubleVerify and Zefr. IAS Agent's page says only recommendations, and adds: "Please note certain features described on this page reflect the future state of IAS Agent and may evolve". We found none.
• Two adjacent items that are commitments about something else: Microsoft Advertising's agentic-commerce page says "Today, Microsoft does not take a commission or affiliate fee." Mutinex offers "No commitment required" for early access. Neither is tied to a result.

5. Is there an intersection: agents transacting on outcomes?

Not on the pages we read. What we found are two adjacent things that have not yet met.

• Agents whose pitch is a result, priced by something else. Quantcast: "let Q+ find and convert your next best customer, faster and more efficiently than manual optimization ever could." Taboola Realize+: "continuously makes and executes campaign decisions, helping drive incremental results" and "automatically moves budget in real time to the highest-performing campaigns". MNTN: "makes over 4 million bid decisions per second". Tatari: "trusting the AI to do a good job without a human in the loop for every decision". Klaviyo's Customer Agent "resolves 65% of questions autonomously", but the one price we could find (third-party site) is per AI conversation, not per resolution. None of these pages states a price tied to the result.
• Outcome-priced work, not agents. Omnicom's CFO: "Between 80% and 90% of Flywheel's work is priced according to results rather than hours." That is a holding company on agency labour, said at an investor conference and reported by ExchangeWire; it did not appear on omc.com in our read. Omnicom Media APAC's Harradine put it as a "may move" ("But we may move more towards business outcome-based metrics").
• The outcome sellers with the strongest outcome positioning (AppLovin, Attain, Cint, Keen, Rokt, Teads, Tinuiti, Vibe, MiQ, Recast) had no agentic claim on the pages we read. The six that did (Haus, Kochava, MNTN, Nexxen, Taboola, Tatari) framed the agent as speed, planning or automation, not as the party that takes the risk.
• Agentic commerce (agents that buy for shoppers) appears as a label or product page at Microsoft (Copilot Checkout), Moloco (nav item), Circana (blog title) and Criteo. It is a different transaction from advertisers paying on outcomes. Microsoft says it takes no commission.

6. Emotional registers

Our coding, primary register per claim, grouped:

Read. The dominant register is relief and delegation ("Hand over the outcome to bespoke agents", "Not a tool you prompt. An agent you delegate to.", "AI does the heavy lifting"), then speed and efficiency. Together they are 47 of 80 (59%). The second theme is control and safety (12%): "You make the final call", "Review and edit every recommendation before launch", "Campaigns only launch after you review and approve". That register hardly exists in the outcomes language as you describe it, which skews to growth, efficiency and waste. Growth / advantage is 6% here (5 claims). Waste does not appear in any coded agentic claim. Fear of falling behind is present but implicit, in words like "is here", "Ready" and "the first operating system", and in trade-press headlines; it appears in 8 claims as readiness / arrival (10%). It is louder in the surrounding press than in the claims. Mastery and power ("Multiply every marketer", "at a scale humans never could") are 9%.

One more contrast. The ACTS claims are the most relief-shaped (15 of 34 in relief / delegation / ease). The ASSISTS claims are where the safety language lives (all 10 control / safety claims are ASSISTS). Vendors that sell the agent as the actor speak in the language of relief; vendors that sell it as a helper speak in the language of control. Proof / certainty (3 claims) appears only in Haus's causal-recommendation line and Tatari's test write-up.

Limits: what we could not see

• Coverage. 64 of 140 companies. The 76 not read are not shown to be lower or higher on any measure. Selection favoured AI-native and outcome-sellers by design, so the 55% is not a market rate.
• Depth. Mostly one home page per company, plus 1 to 3 press, product or blog pages found by search where the home page was thin. Companies with thin home pages (AppLovin, Omnicom, Meta, Amazon, TTD, Viant, Yahoo, Samsung, Walmart) may run agentic products on deeper pages that we did not open. "None found" means none found on the pages we read.
• Rendering. WebFetch returns text that a small model extracts. JavaScript-heavy pages, carousels, tabs, videos, gated demos and PDFs may be under-read. Several home pages returned only nav labels. Two fetch targets returned 403 or 404 (The Drum; Samba blog).
• Verbatim risk. All quotes are strings the extraction model returned as exact. 7 rows (plus the contract, R2B2, Universal Ads and Omnicom CFO quotes) are confirmed by two differently-worded fetches (marked B); the rest rest on one fetch. Some strings arrived with the fetcher's formatting (list separators, ellipses, a non-breaking hyphen). Spot-check every "A" quote against the live page before any of it is published. Salesforce sentence in row 27 is ungrammatical in the source as returned ("let campaign agent handle generating content...").
• Time. Pages read 2026-09-30. Several claims come from press releases dated 2025-11 to 2026-09; a home page may have changed since. Scope3 became Apostra, reported 2026-09-23 to 2026-09-28; the frame still lists Scope3.
• Coding. Autonomy and register are one coder's calls. 25 of the 80 claims are ambiguity calls; 7 depend on the guardrail rule. Rung for a claim is judged from the same page or module; a case study on another page could raise it. Claim counts per company vary from 1 to 6, so companies with long pages weigh more.
• Not seen. Sales decks, order forms, IOs, RFP responses, private contracts, MSAs for other vendors, pilots with SLAs, and any pricing behind a login. A guarantee could sit in an order form and never on a web page. We read one contract (PubMatic). We did not read the terms pages of the other 63.
• Third-party items in the counter-evidence tables come from one or two fetches, some of which were partly paywalled (Adweek), and were not coded into the 80-claim counts. DataBeat's dates differ across the two write-ups (June 22 vs August 16); we did not resolve which report edition the 86% / 13.4% figures come from.
• MADDB. Company briefs were not pulled. Five news searches hit a rate limit and were rerun. Long-phrase queries such as "agent-to-agent negotiation advertising", "outcome guarantee agentic pay for performance agent" and "guaranteed outcomes AI agent advertising" returned nothing from MADDB; that is a statement about MADDB's index and query matching, not about the market.