How Agencies Are Actually Using AI Agents With Clients Right Now
Marketing, creative, dev, and consulting agencies are past the pilot stage with agentic AI. Here is what is actually running in client work today, what is still mostly talk, and the billing, disclosure, and liability questions nobody has settled.
In June, with the advertising industry's dealmakers gathered in the south of France for their annual pilgrimage, agencies spent the week outdoing each other's AI agent announcements. WPP said it was testing a media buying and planning agent for premium video inventory. Dentsu struck a partnership with an AI agent developer called Newton Research to speed up campaign setup and analytics. Independent shop Butler/Till, working with DoubleVerify, had already built Claude-based agents to steer ad dollars away from low-quality inventory. None of this was a demo. It was agencies describing tools running inside live client accounts, according to reporting from Digiday.
That week captured something true about where agentic AI stands right now: it is no longer a slide in a pitch deck. It is showing up in campaign dashboards, audit workpapers, and codebases — faster than the billing models, disclosure norms, and liability frameworks built for a world without it. Here is what agencies are actually running today, and where the real friction is.
The work that has moved past pilot
The clearest data point on adoption comes from a January 2025 survey of 547 U.S. marketing agency decision-makers by NinjaCat: 91% were already using AI in some form, and 59% said they were actively using AI agents specifically — autonomous systems that execute tasks rather than just generate suggestions. The two applications agencies cited most were data and insights work (74%) and campaign strategy and optimization (72%). In practice, that means agents unifying metrics across ad platforms, catching anomalies before they waste spend, and delivering reporting straight to a client's inbox without an analyst assembling the deck by hand.
The IAB's State of Data 2025 report puts a number on how far that has actually gone: only 30% of agencies, brands, and publishers have fully integrated AI across the media campaign lifecycle so far, though the agencies and publishers in that group are ahead of the brands. Where it has been adopted, the IAB found AI meeting or exceeding expectations in more than 70% of cases, lifting productivity by 47% and cutting roughly 12 hours a week of low-value manual work per person.
The agentic layer specifically — media-buying agents that execute rather than just recommend — is still in the early-scale phase but moving fast. Digiday reported in September that 58% of ad executives expect agentic buying to hit meaningful scale within a year, citing IAB Europe research. Dentsu's Caitlin Gelles, evp of data technology and measurement, described the appeal to Digiday in blunt operational terms: "It takes hours to set up campaigns across systems… when you can store those nuances in a RAG database, or take how clients like to buy, and let AI find those efficiencies, that's where the value comes in."
Dev and digital-transformation agencies are running a parallel version of the same play. Dept, the Dutch agency group, built an e-commerce platform redesign for the Swiss fitness brand Blackroll using AI agents on Google's Gemini and Antigravity stack — Dept says the project finished 3.8 times faster than its own pre-AI estimate. Blackroll's CEO, Scott Zalaznik, told Digiday the company now uses AI agents for the majority of its e-commerce feature testing and development. Dept has since productized the approach into a paid offering called Agent Studio, effectively renting out its own agent-building workflow to clients who want the capability without building it themselves.
Consulting firms are further along, and still arguing about the payoff
If marketing agencies are mid-transition, the big consulting firms have already gone all in on volume — and are now stuck on the harder question of what any of it is worth. McKinsey CEO Bob Sternfels said in January that the firm has deployed tens of thousands of internal AI agents and eventually wants one for every one of its 40,000 employees, according to Business Insider. But as Mina Alaghband, a former McKinsey partner now chief customer officer at Writer, put it: "I think we are now in the age of confusion." A year ago the metric was usage. Now firms are trying to measure something harder — value actually created.
PwC's chief AI officer Dan Priest told Business Insider the firm has stopped counting agents and started counting how many humans actually use each one. EY tracks productivity, quality, and cost-efficiency KPIs on its agents monthly. BCG has the most concrete number: partner Scott Wilder said employees now spend about 15% less time on low-value work like building slideshows, reinvest roughly 70% of that recovered time into deeper analysis — and keep the remaining 30% as, in his words, more sleep or a yoga class. Even inside firms that bill by the hour, AI-recovered time isn't automatically converting into more billed hours.
"AI might finally be the death knell for the billable hour," said Clio CEO Jack Newton — a claim professional-services firms have made and walked back for thirty years, except this time the arithmetic behind it is public.
The billing model is the first thing actually breaking
Of the frictions agencies are hitting, billing is the most immediate because the math is inescapable. If AI cuts the time a task takes from five hours to one, and you still bill by the hour, you have handed your client an 80% discount without meaning to. A Forbes analysis of a 2026 Gartner survey of 743 audit professionals and separate research from legal software company Clio found the pattern showing up across adjacent professional-services industries: 93% of audit leaders report some AI use but only 38% have a strategy for it, and Clio estimates roughly three-quarters of billable legal tasks could now be automated by generative AI.
Pricing has not caught up. Clio found 86% of solo law firms and 78% of small firms have made no pricing changes at all in response, compared with 51% of mid-market and 46% of enterprise firms — the smaller the shop, the more it is quietly eating the discount. Meanwhile 71% of clients told researchers they actually prefer fixed or flat fees over hourly billing in the first place, which suggests hourly billing was always a stand-in for something clients actually wanted: a predictable outcome, not a timesheet. Salesforce's chief legal officer, Sabastian Niles, made the client-side expectation explicit in a Harvard Law School governance forum post cited by Forbes: "The gains must also of course be shared with clients through savings, cost-efficiencies, and new business models." The Wall Street Journal covered the same shift under the headline "Inside Consultants' Messy Shift From Hourly Billing," describing professional-services firms wrestling with reinventing how they charge as AI threatens to make the billable hour obsolete.
Marketing agencies are already testing the alternative NinjaCat's survey found: 28% have shifted away from standard billable hours toward value- or outcome-based pricing, and 45% now sell AI-based solutions and automations as a packaged service rather than folding them into existing retainers — turning the efficiency gain into a new product instead of a smaller invoice.
Should clients be told an agent touched their work?
The disclosure question is less settled than the billing one, mostly because there is no single rule to point to. The IAB's State of Data 2025 research found roughly half of brands are already concerned about how agencies and publishers use AI on their behalf — a trust gap that exists even before anyone has decided what disclosure should look like.
The clearest guidance so far comes from adjacent professional-services fields grappling with the same question for their own clients. In a piece syndicated from the Journal of Accountancy, insurer CNA advised CPA firms that in the absence of one controlling law — the legal landscape is a patchwork of the Gramm-Leach-Bliley Act, scattered state disclosure laws in places like California and Utah, and the EU's AI Act for firms with European clients — voluntary, specific disclosure is still the safer path, because clients who later learn AI touched their work without being told may read it as deception, whether or not anything went wrong. The piece pointed to the American Bar Association's 2024 Formal Opinion on generative AI, which draws a sharp line: boilerplate language buried in an engagement letter does not count as informed consent. A disclosure has to actually explain what the AI did and what a human checked afterward.
Every professional-services field hitting this question — accounting, law, and now agencies — is arriving at the same rough answer: disclosure is not legally required in most cases yet, but withholding it is a reputational bet that gets riskier the more clients start assuming AI is already involved.
Who is on the hook when the agent gets it wrong
Liability is the friction agencies have thought through the least, because it has not forced the issue yet — nobody has had a headline-grabbing agency AI failure the way the legal profession has had lawyers sanctioned for citing AI-fabricated case law in real filings. Law.com's David Partida has been running a six-part series through 2026 tracking exactly this problem as it moves from theoretical to routine, including installments titled "What Happens When AI Acts Alone — And Who Pays For It?" and "Who's Liable When AI Gets It Wrong?" A companion piece on the same site, "Accountability For AI Errors By Attorneys And Judges," notes that attorneys are increasingly filing briefs containing fake case citations generated by AI — usually traced back to user error rather than the tool itself — and that courts are already correcting the resulting rulings through motions and appeals. The professional guidance converging around this, from the ABA's opinion to CNA's advice to CPA firms, is consistent: a firm that relies on an AI agent's output without independent human review is on the hook exactly as if a junior staffer had made the mistake unsupervised. The agent does not carry the liability. The firm that shipped its output to a client does.
Agencies do not yet have their own version of this reckoning in the public record — which is arguably the friction point most worth watching, since agency deliverables move fast and often get less scrutiny before they reach a client's inbox.
One idea from The AI Organization is worth borrowing here: legibility, or being able to show — not just claim — who or what touched a piece of work and when. An agency that can produce an actual record of what an agent did, what a human reviewed, and when they signed off is answering the disclosure and liability questions with evidence instead of a policy statement. FabricLoop's own AI product, Loop Agent, is built on MCP with explicit, audited grants for exactly this reason: so client-facing AI use is something a firm can show, not just assert.
What this means if you run an agency
Strip away the announcements and the pattern is narrow but real. Agentic AI is earning its keep today on work that is repeatable, measurable, and reviewable before it reaches a client: campaign reporting, media-buying execution inside defined guardrails, first-pass research and competitive analysis, and — per Dept's numbers — a meaningful chunk of e-commerce build and test cycles. That is not a small list, and for agencies still doing that work by hand, it is a real cost advantage sitting on the table.
It is still mostly talk on work that requires judgment nobody has figured out how to verify at scale: creative decisions that carry brand risk, strategic recommendations clients pay for precisely because they trust the person making them, and anything where a mistake could reach a client before a human sees it. The agencies quoted across this reporting — Dentsu, PwC, BCG — describe the same model underneath their different applications: agents doing the compressible work, a human checking the parts that carry risk, and the time freed up getting reinvested rather than pocketed as margin. That is a more boring story than "AI replaces the agency." It is also the one actually happening in client accounts right now.