State of AI in Programmatic 2027: Survey Methodology

The State of AI in Programmatic 2027 is an annual survey of 400 buy-side, sell-side and ad tech practitioners in the US, UK and EU, fielded in January and February 2027 and published in April 2027 with open data. This page is the pre-registered plan: questions, sample, timeline, analysis and release.

Publishing the plan before fieldwork is the point. Anyone can check in April that the questions asked are the questions listed here, that the sample met its quotas, and that the headline numbers come from the analysis plan rather than from whatever looked best.

1. Adoption. Which AI capabilities (bidding optimization, audience modeling, creative generation, campaign set-up agents, reporting and insight agents, supply-path curation) are in production, in pilot, or not used, by side and region?
2. Control. Who runs the models: the platform vendor, an agency, an in-house team, or a mix; and where does the data that trains or prompts them sit?
3. Money. What share of programmatic spend is placed by automated or agentic buying with no human in the loop per campaign, and how is that expected to change in 2027?
4. Trust and governance. Which controls exist (audit logs, explainability, spend caps, brand safety on generated creative, human sign-off), and which incidents have occurred?
5. Standards. Awareness and intended adoption of agentic advertising protocols (the AAMP vs AdCP tracker covers the two), and of explainability requirements.
6. Outlook. What practitioners expect to change in 2027 in budget, headcount, vendor relationships and in-house capability.

The questionnaire is 4 screeners and 22 questions, median completion 11 minutes in pretest (target). Exact wording is frozen on 15 December 2026; the wording below is the pre-registered draft. Fifteen questions (marked core) will be repeated unchanged in 2028 to build a trend.

Screeners

• S1. Which best describes your organization? (Advertiser or brand; Agency or trading desk; Retailer or commerce media network; Publisher or broadcaster; SSP, exchange or ad server; DSP or bidding platform; Data, identity or measurement vendor; Other.) Routes to buy side, sell side or vendor.
• S2. In which country are you based? (List; US, UK, EU member states, Other. "Other" terminates.)
• S3. Which best describes your role? (C-level; VP or Director; Manager or Lead; Analyst, Trader or Specialist; Other.)
• S4. In the past 12 months, have you been directly involved in buying, selling or operating programmatic advertising? (Yes; No. "No" terminates.)

Section A: adoption (core)

• Q1 (core). For each of the following, is your organization using AI in production, piloting, evaluating, or not using? Grid: bid optimization; audience building or lookalikes; creative generation (text, image, video); campaign set-up or trafficking agents; reporting and insight agents; supply-path or inventory curation; fraud or brand-safety classification; forecasting and pacing. (4-point status scale plus "don't know".)
• Q2 (core). Roughly what share of your programmatic spend in the past quarter was placed by systems that choose placements and bids with no human approval per campaign? (0%; 1 to 10%; 11 to 25%; 26 to 50%; 51 to 75%; over 75%; don't know.)
• Q3. Which single AI use case delivered the most measurable value in the past 12 months? (Single choice from the Q1 list plus "none".)
• Q4 (core). How long has your organization used AI in production for any of the Q1 use cases? (Not yet; under 1 year; 1 to 2 years; 2 to 4 years; over 4 years.)

Section B: control and data (core)

• Q5 (core). Who operates the AI models you rely on for programmatic? (Multi-select: DSP or SSP vendor; agency; in-house team; specialist AI vendor; open-source models run in-house; don't know.)
• Q6 (core). Where does the first-party data used to train or prompt those models reside? (Multi-select: vendor cloud; agency environment; our own cloud tenancy; our own data center; a clean room; don't know.)
• Q7. Can your team see the features and weights, or an equivalent explanation, behind an automated bid or placement decision? (Always; sometimes; never; don't know.)
• Q8 (core). In the past 12 months, has your organization moved any AI or decisioning capability in-house, or out to a vendor? (In-house; out to a vendor; both; neither.)

Section C: governance and incidents

• Q9 (core). Which controls are in place for AI-driven buying or selling? (Multi-select: spend caps; human sign-off thresholds; audit logs of decisions; model documentation; brand-safety review of generated creative; vendor contractual limits on data use; none.)
• Q10. In the past 12 months, has an AI-driven system caused an incident you had to correct? (Multi-select: overspend; wrong audience; brand-unsafe creative; placement on made-for-advertising sites; measurement error; data leakage; none; prefer not to say.)
• Q11. Who is accountable for AI decisions in programmatic at your organization? (Single: media or programmatic lead; marketing leadership; data or engineering leadership; legal or compliance; vendor; no one assigned.)
• Q12 (core). How much do you trust fully automated buying to act in your interest? (5-point scale from "not at all" to "completely".)

Section D: money and resources (core)

• Q13 (core). How did your programmatic budget change in 2026, and how do you expect it to change in 2027? (Two 5-point scales: down over 10%; down 1 to 10%; flat; up 1 to 10%; up over 10%.)
• Q14 (core). How did headcount in programmatic and ad operations change in 2026, and how do you expect it to change in 2027 as a result of AI? (Same two scales.)
• Q15. Which costs rose most because of AI in 2026? (Multi-select: platform fees; data fees; cloud or compute; specialist staff; measurement; none.)
• Q16 (core). Do you pay, or charge, an AI-specific fee or uplift on any programmatic platform? (Yes, as a share of spend; yes, as a flat fee; no; don't know.)

Section E: standards and outlook

• Q17. Before today, had you heard of the following? (Grid: AAMP; AdCP; IAB Tech Lab agentic advertising work; none.)
• Q18 (core). Does your organization plan to buy or sell inventory through an AI agent protocol in 2027? (Yes, already do; yes, planned; evaluating; no; don't know.)
• Q19 (core). Which of the following do you expect to be true by the end of 2027? (Multi-select: most of our spend is agent-placed; we run at least one model in-house; we require explainability from vendors contractually; we have reduced the number of vendors; we have moved data out of vendor environments; none.)
• Q20. What is the one thing a platform could change that would make you trust its AI more? (Open text, 300 characters.)
• Q21. What AI capability do you most want to build in-house that you currently rent? (Open text, 300 characters.)
• Q22. May we quote your open-text answers anonymously, and may we contact you for the 2028 wave? (Two yes/no.)

An attention check ("select 'piloting' for this row") sits in Q1. Q13 and Q14 are asked as paired items to measure expectation against experience.

• Population. People directly involved in buying, selling or operating programmatic advertising in the US, UK and EU in the past 12 months.
• Target. 400 completes. Quotas: region US 220, UK and EU 180; side buy 200, sell 120, vendor 80; seniority Director or above at least 240. Quotas are soft (±10%) to avoid closing cells early; the achieved sample is reported against them.
• Floor for reporting. No subgroup is reported below 50 completes (±13.9 points at 95%, computed), and differences between subgroups are reported only when a chi-square test is significant at 95%.
• Recruitment, in order of expected share. (1) Invitations to this site's newsletter list and visitors to its tools pages (intercept on the calculators, once per visitor). (2) Direct outreach on LinkedIn to practitioners whose titles match the screener, with a public post and a tracked link. (3) Partner newsletters from trade bodies and publications that agree to carry the link, each with its own tracked source. (4) A B2B research panel to fill cells that are short after week five, with the panel share capped at 30% of completes and reported. Every complete carries its source so that source effects can be tested.
• Incentive. A donation of $10 per complete to a charity chosen from three (editable), plus early access to the data file. No cash.
• Sample type. Non-probability. Margins of error are reported "as if" from a simple random sample with the design effect applied, and the report states that they are indicative, which is the AAPOR practice for non-probability samples (AAPOR, April 2021). Participation rate is computed per source.

Our calculation except the last row; formula z × √(p(1−p)/n), z = 1.96, before design effect.

• Cleaning. Remove completes under 40% of the median duration, straight-liners on Q1 and Q9 (same answer on every row), failed attention checks, and duplicates by hashed email or IP plus user agent. Report the number removed by rule.
• Weighting. Rim weighting to the target region × side distribution; weights trimmed at 0.3 and 3.0; design effect reported. Headline figures shown weighted and unweighted.
• Indices. AI Adoption Index (0 to 100): mean over the eight Q1 use cases of (production = 100, pilot = 50, evaluating = 25, not using = 0). Control Index: share of Q5 operators that are in-house or open-source run in-house. Governance Score: count of Q9 controls out of six. All three defined here so they cannot be tuned after the fact.
• Tests. Chi-square for categorical differences by side, region and seniority; t-tests on the indices; 95% level; Bonferroni correction within each question family. Only pre-specified comparisons (side, region, seniority, in-house vs vendor-operated) are reported as findings; anything else is labelled exploratory.
• Open ends. Coded by two coders against a codebook built from the first 50 responses; Cohen's kappa reported; disagreements resolved by a third reader; kappa below 0.7 triggers a codebook revision.
• Trend. The 15 core questions are locked for 2028; year-on-year differences are reported with the pooled margin of error.
• Pre-registration. This page is archived at publication (version 1.0.0); the April report cites the archived version and lists every deviation.

1. Headline numbers (five figures with margins of error).
2. Method in one page (the eleven AAPOR elements).
3. Adoption: the AI Adoption Index by side, region and seniority; the production-versus-pilot gap by use case.
4. Control: who operates the models and where the data sits; the in-house shift (Q8) and its direction.
5. Money: agent-placed share of spend (Q2); budget and headcount, 2026 actual against 2027 expected.
6. Governance: controls in place; incidents; accountability.
7. Standards: awareness and intent for agent protocols.
8. Outlook: the 2027 expectations list, ranked.
9. What buyers and sellers disagree on (the largest significant gaps).
10. Open-text themes with anonymized quotes.
11. Limitations and deviations from this plan.
12. Appendix: full tables, codebook, data dictionary, "cite as".

• Embargo. Briefings and the embargoed PDF to ten trade outlets (AdExchanger, Digiday, ExchangeWire, Adweek, Marketing Dive, The Drum, PPC Land, MediaPost, Campaign, Ad Age) from 15 April 2027; embargo lifts 22 April 2027 at 06:00 ET. Outlets listed as targets, not partners.
• Assets. Five headline charts as SVG and PNG under CC BY 4.0 with the attribution baked in; a one-page method summary; three anonymized respondent quotes cleared under Q22; a short video explainer.
• Analysts. The data file and codebook sent to analyst firms that cover programmatic on request from publication day; no exclusivity.
• Owned channels. Newsletter, a LinkedIn post series (one chart a day for a week), the webinar on 29 April with the research lead and two external practitioners.
• Follow-up. A mid-year note in September 2027 re-cutting the data against the ANA's Q2 2027 benchmark; a call for 2028 respondents in October 2027.

• Files. responses.csv (one row per complete; all closed questions as coded values; open-text answers only where Q22 permits and after removal of names, companies and locations; country collapsed to US, UK, DE, FR, other EU; company size in bands), codebook.csv (variable, label, type, value labels, question wording), weights.csv (respondent id, weight, source), questionnaire.pdf (as fielded), methodology.md (this page, archived).
• Licence. CC BY 4.0 (Creative Commons). Attribution text supplied.
• Identifier. A DOI minted through Zenodo for each edition (Zenodo); the DOI resolves to the archived files, the page URL to the latest edition.
• Schema. Dataset JSON-LD on the page: name, description, license, creator, datePublished, version, identifier (DOI), distribution (one DataDownload per file with contentUrl and encodingFormat), variableMeasured (the three indices and Q2), temporalCoverage (2027-01-12/2027-02-27), spatialCoverage (US, UK, EU).
• Privacy. No personal data in the release; the mapping from respondent id to contact details is kept separately, encrypted, and deleted after the 2028 wave invitation.

A stable URL that does not change by year, with each edition archived at its own path and DOI. The question wording fixed and public before fieldwork. Sample sizes and margins printed on every chart. Open data that lets a journalist or analyst recompute any number. A named research lead and a changelog. A trend design that makes the 2028 numbers comparable. And a limitations section that says what the study cannot claim, which is the element vendor surveys most often omit and the one AAPOR requires (AAPOR, April 2021). The explainable AI RTB simulator and programmatic strategy for 2027 will cite the results once published; DSP, DMP and external data in explainable RTB and dynamic creative optimization strategies are the practitioner reading behind questions Q7 and Q1.

Who can take the survey? Anyone directly involved in buying, selling or operating programmatic advertising in the US, UK or EU in the past 12 months, at any seniority. The screeners route respondents by side and country; vendors are included and reported separately so they do not color the buy-side and sell-side figures.

Is the survey anonymous? Responses are collected without names. Email is optional and used only for the Q22 follow-up and the 2028 invitation; it is stored apart from the responses and never released. Open-text answers are released only where the respondent agreed, and after removal of anything identifying.

How big does a sample need to be to be credible? There is no magic number; precision scales with the square root of n. At 400 the headline margin is about ±4.9 points at 95% (our calculation); Pew's example of 1,067 gives about ±3 (Pew, 8 September 2016). What matters more for a practitioner survey is that the sample is described, the quotas met, and subgroups below 50 not reported.

Will the raw data be available? Yes, respondent-level, on publication day, under CC BY 4.0 with a DOI, with a codebook and the weights. The only redactions are those needed to protect respondents.

How is this different from vendor surveys? The questions are published before fieldwork, the sample plan and achieved sample are disclosed per AAPOR, the data is open, and the sponsor is named. Vendor surveys usually publish results without the questionnaire, the base or the data.

Use the agentic advertising guide to compare the related operating models, definitions and evidence before making this decision.

1. AAPOR, Disclosure Standards, Code of Professional Ethics and Practices, Section III.A (revised April 2021). https://aapor.org/standards-and-ethics/disclosure-standards
2. Pew Research Center, "Understanding the margin of error in election polls" (8 September 2016). https://www.pewresearch.org/short-reads/2016/09/08/understanding-the-margin-of-error-in-election-polls/
3. ANA, "ANA, TAG TrustNet Launch Programmatic Transparency Benchmark" (17 June 2024). https://www.ana.net/content/show/id/pr-2024-06-tag
4. ANA, Programmatic Transparency Benchmark library (lists Q2 2026 findings, 26 August 2026; Q3 2026 webinar 4 November 2026). https://www.ana.net/micpackage/show/id/97114
5. Skai, "Ad spend rises across every channel in Q2 2026" (28 July 2026). https://skai.io/press-releases/ad-spend-rises-across-every-channel-in-q2-2026-as-ai-reshapes-commerce-media-skai-data-shows/
6. Creative Commons, Attribution 4.0 International license. https://creativecommons.org/licenses/by/4.0/
7. Zenodo, open research repository (DOI minting). https://zenodo.org/

• 18 October 2026: aligned the title and heading and applied recorded source-URL corrections where needed. Factual research cutoff remains 17 October 2026.

• 1.0.0, 17 October 2026. Methodology pre-registered: 4 screeners, 22 questions (15 core), sample plan, timeline, analysis plan, release plan. Next update 15 December 2026 (questionnaire freeze).