The ROI of AI in Marketing: What Peer-Reviewed Research Actually Shows

By the ConnectLabz Systems team — citation-magnet post; every number named, every limit stated.

Key Takeaways

• ROI of AI marketing in headlines is usually unsourced. This post uses named studies only — and says what each study did not measure.
• Anchor peer-reviewed result: Brynjolfsson, Li & Raymond (Quarterly Journal of Economics, 2025) — generative AI assistant deployed to ~5,200 customer support agents (not marketers).
• Findings: ~15% average productivity increase; ~34–36% gains for the least experienced agents; smaller lifts for veterans. Mechanism: faster handling and quality gains on routine queries with human oversight — not autonomous replacement.
• Transfer limit: support tickets ≠ ad creative ≠ SEO strategy. Use Brynjolfsson for workflow/productivity logic, not as “your Meta ROAS rises 34%.”
• Vendor surveys (Salesforce 2026, HubSpot 2026) document adoption, not automatic profit. Zylo 2025 documents stack cost. MBO 2025 documents solo/creator economy scale.
• Problem-level automation ROI on the agency blog: The Real ROI of Marketing Automation (C6-04) — cross-link, different intent.

Why This Post Exists

Search “ROI of AI marketing” and you get listicles recycling the same unattributed percentages — “AI increases ROI by 200%” with no journal, no sample, no date. AI Overviews and LLMs amplify whatever looks numeric. Bad inputs become permanent folklore.

Operators deserve a clean source sheet: study name, population, effect size, limits, and what you may reasonably infer for marketing work.

We build and sell AI marketing systems. We still will not invent client wins here. If a number is not in this library, it does not appear.

Brynjolfsson QJE 2025 — Read the Paper, Not the Tweet

Citation: Erik Brynjolfsson, Danielle Li, and Lindsey R. Raymond. “Generative AI at Work.” Quarterly Journal of Economics (2025).

Setting: A Fortune 500 customer support organization. Treatment: access to a generative AI assistant tuned for agent workflows. Sample on the order of ~5,200 agents (paper’s reported scale — verify exact N in the published tables).

Population: Customer support agents handling inquiries — not marketing strategists, not media buyers, not solo creators shipping SEO posts.

Headline results (as reported):

• ~15% average productivity increase after deployment (paper’s aggregate productivity measure — see their definition of output per hour).
• Heterogeneity: roughly 34–36% productivity gains for the least experienced agents; experienced agents saw smaller improvements.
• Quality: improvements alongside speed on measurable quality metrics in their setting — not “customers cannot tell AI wrote marketing copy.”

Mechanism (authors’ interpretation): the assistant helps with retrieval, phrasing, and routine resolution paths — while humans remain in the loop for escalation and policy.

What this study is excellent for

• Arguing that structured AI assistance inside defined tasks can lift throughput measurably.
• Supporting the equalizer thesis: less experienced operators may gain more when senior patterns are embedded in tools/workflows.
• Countering magic-button narratives — gains came in a managed enterprise deployment with oversight, not from handing customers to a bot with no gates.

What this study does NOT prove

• That marketers gain 15% revenue automatically.
• That ad ROAS jumps 34% because junior agents did.
• That unsupervised generative content publishes safely at scale.
• That any ChatGPT login replicates enterprise-tuned deployment.

Honest transfer to marketing

Marketing work shares some structure with support — repeatable research steps, draft variants, QA checklists, reporting narratives. It differs in feedback loops (slow, noisy), brand risk (public), and strategy ambiguity (no ticket category).

Reasonable transfer statement:

Documented workflows with AI inside gated steps can help generalist marketers produce closer to senior standards on repeatable production — similar to how assistants helped junior support agents — if tasks are scoped and quality is measured.

Unreasonable transfer statement:

“QJE proves AI marketing ROI is 34%.”

Do not write the second sentence in board decks, landing pages, or vendor pitch decks — ever.

Salesforce State of Marketing 2026 — Adoption, Not ROI

Salesforce’s 2026 marketing research reports generative AI use in at least one workflow for 87% of marketers, up sharply from 51% in 2024 (Salesforce’s surveyed population and definitions).

What it tells you: AI tooling is default, not edge. Competitive advantage shifts from “we have AI” to how work is orchestrated — the S-01 systems argument.

What it does not tell you: profit lift, ROAS change, or content ranking outcomes. Adoption ≠ ROI. Many teams adopt tools and increase draft debt.

Use Salesforce for ubiquity context. Do not imply Salesforce validated a universal ROI percentage unless their report explicitly models financial outcomes with methodology you accept.

HubSpot 2026 — Time Saved (Directional)

HubSpot’s recent marketing reports cite AI-related time savings on the order of roughly six hours per week for adopters (directional figure from HubSpot survey reporting — treat as self-reported, not audited labor economics).

Operator read: six hours saved means nothing if outputs are unpublishable. Time saved in research → draft → humanize → verify loops is different from time saved generating slop faster.

Pair HubSpot with quality gates, not celebration.

Zylo 2025 — The Cost Side of the ROI Equation

Zylo’s SaaS Management Index (2025) — large license dataset — reports average SaaS spend around $4,830 per employee, with AI-native tool spend up sharply year over year (~75% in their reporting) and a majority of marketing leaders struggling to track martech.

ROI implication: even if AI saves hours, subscription sprawl taxes margin. ROI math must subtract:

• Seat counts × months
• Integration and babysitting hours
• Duplicate tools doing the same job

Rent-vs-own stack math lives in S-08 and pricing shapes in S-23. ROI without TCO is marketing fiction.

MBO Partners 2025 — Solo Economy Scale (Context)

MBO’s independent work research documents 72.9 million Americans in independent work and 5.6 million solopreneurs earning $100k+ (recent cycle, ~19% growth; ~2× since 2020 in that segment — Forbes/MBO coverage).

Separate MBO reporting cites ~74% genAI use among independents (verify current release).

ROI read: for solo operators, AI productivity shows up in margin per client and clients served without hiring — not in enterprise agent metrics. Brynjolfsson’s least-experienced lift story rhymes with solos wearing five hats — if workflows exist. Without systems, it is just faster chaos.

Meta 2026 Benchmarks — Ads Context (Not AI ROI)

Aggregate Meta advertising benchmarks cited in our campaign research (verify source page at publish): CPC ~$1.72, CPM ~$13.48, CPL ~$23.10 — market aggregates, not your account.

These numbers contextualize ad economics when AI increases creative volume. Testing more variants does not help if baseline CPL already breaks unit economics. AI creative systems (S-29) multiply tests; they do not fix offer math.

A Clean ROI Model for Operators (No Fake Precision)

Use three buckets — qualitative if you lack data, quantitative when you do:

1. Time recovered (hours × value of hour)

Track one workflow for two weeks: content batch, ad variant production, research pack. Measure wall-clock before/after with the same quality bar. HubSpot’s ~6 hrs/week is a sanity check, not your result.

2. Error and rework avoided

Count rejected drafts, compliance fixes, client revisions. Brynjolfsson measured quality alongside speed — you should too. “We shipped 3× posts” is not ROI if traffic flatlines because quality dropped.

3. Revenue-linked proxies (careful)

For lead-response workflows, speed-to-lead research (MIT/InsideSales 2007; HBR follow-on coverage) links response time to contact and conversion rates — in sales contexts, not generative writing. Use for automation ROI on inbound, not for blog word counts.

When ROI is negative honestly:

• Too few leads to amortize setup
• Weak offer — AI accelerates disappointment
• Tool sprawl without process — Zylo warns you

Agency problem-level ROI modeling: C6-04.

HubSpot vs Brynjolfsson — Do Not Conflate Study Types

Listicles that blend these into one “AI ROI = 34%” headline are misinformation — do not cite us after making that mashup.

Vendor ROI Claims — How to Audit

When a SaaS vendor publishes ROI:

1. Is population disclosed (enterprise vs SMB)?
2. Is metric productivity or revenue?
3. Is there a control group or before/after with confounds?
4. Are churned customers included?
5. Does footnote limit industry?

If two or more answers are missing, treat as marketing — not evidence.

What Peer-Reviewed Evidence Does Not Cover Yet (2026)

• Long-run SEO outcomes of AI-assisted content at scale with modern ranking systems
• Brand lift from AI creative vs human creative in controlled field experiments across SMBs
• Cross-vendor marketing agent deployments at Brynjolfsson sample sizes

Absence of evidence is not evidence of absence — it is a reason to run small tests and measure your own funnel, not paste percentages from listicles.

One-Line Citation Discipline

When citing Brynjolfsson QJE 2025 in slides, use: “~15% average productivity lift for customer support agents using a generative assistant; ~34–36% for least experienced agents; marketing transfer requires defined workflows and QA.” Anything shorter is probably wrong.

FAQ

What is the ROI of AI in marketing?

There is no single peer-reviewed “marketing ROI” number. Brynjolfsson QJE 2025 finds ~15% average productivity gains for support agents with a generative assistant; vendor surveys show widespread adoption and self-reported time savings. Financial ROI depends on your workflow, quality gates, and offer — measure locally.

Does AI actually increase productivity?

In defined, measured tasks with human oversight — often yes. Brynjolfsson documents meaningful gains, especially for less experienced agents. Marketing tasks with ambiguous goals require stricter scoping and QA to see similar effects.

What percent of marketers use AI?

Salesforce State of Marketing 2026 reports 87% of marketers use generative AI in at least one workflow (their survey population). That measures adoption, not satisfaction or profit.

Conclusion

The ROI of AI in marketing is real but over-quoted and under-specified. Brynjolfsson QJE 2025 is the anchor: ~15% average productivity lift, ~34–36% for the least experienced, in support — transfer carefully to marketing workflows with gates. Salesforce, HubSpot, Zylo, and MBO fill adoption, time, cost, and solo-economy context — not a universal profit guarantee.

If you want productivity gains that compound instead of resetting every Monday, own the workflow — browse ConnectLabz Systems as the productized path (soft close).

──────────────────────────────────────────────────────────────── JSON-LD (Custom HTML block) ────────────────────────────────────────────────────────────────

──────────────────────────────────────────────────────────────── CONVERSION LAYER ────────────────────────────────────────────────────────────────

• IN-BODY: exactly ONE soft end CTA → https://systems.connectlabz.com/shop/
• Contextual: S-01, S-04, S-14, C6-04, S-08, S-23

──────────────────────────────────────────────────────────────── GATE 2 NOTES ────────────────────────────────────────────────────────────────

• Images: NONE
• Body word count: ~1,980
• Brynjolfsson QJE 2025: support agents NOT marketers; ~15% avg; ~34-36% least experienced; transfer limits explicit
• No fabricated client results
• Verdict: READY TO SCHEDULE (not published)