Contents10
- 01How has AI transformed digital marketing since 2023?
- 02Why is everything speeding up in 2026?
- 03What the agency market is going through
- 04Which AI tools matter for an agency in 2026?
- 05What does not work, or not yet
- 06What an agency brings that AI does not replace
- 07How we help agencies bring in AI
- 08What we run ourselves
- 09Where to start?
- 10Frequently asked questions
At the end of 2022, AI in digital marketing mostly meant ChatGPT writing draft posts. In October 2026, it chooses the queries your ads appear on, writes the copy of those ads, summarises Google’s results before anyone has clicked and produces a video ad for the price of a coffee. AI and digital marketing are no longer two separate subjects: they are the same trade, and it has changed in nature.
For agencies, the shock is twofold. The platforms now do part of the work that agencies used to be paid for, and their clients have access to the same tools. The trade does not disappear for all that: it shifts towards what machines cannot do. This article takes stock of what has changed, of the tools that matter and of what an agency can do with AI in its own processes. It is a subject we cover in our Digital marketing section, based on what we run ourselves.
In short
Google and Microsoft have made their search campaigns largely automatic: the choice of queries, copy and landing pages is handed to AI, often switched on by default.
Google’s AI summaries arrived in France on 22 July 2026; nearly two thirds of French searches already ended without a click.
Producing a text, an image or a video now costs a few cents; the time is lost in review, not in production.
The big agency groups are consolidating, and billing for time spent is being challenged.
An agency keeps what AI does not replace: trade-offs, knowledge of the client, the idea, independent measurement and responsibility for what goes out. We help it automate the rest, at the right cost, and sell new services.
How has AI transformed digital marketing since 2023?
The transformation came in waves, and each one hit a different part of the agency’s business.
| Period | What changed | Area affected |
|---|---|---|
| 2023 | Language models write draft texts and emails | copywriting, community management |
| May 2025 | Google launches AI Max for Search campaigns: queries without keywords, generated copy, landing page chosen by AI | media buying |
| July 2025 | Pew Research measures that AI summaries almost halve clicks to websites | SEO |
| 2025 and 2026 | Generated video becomes cheap and of advertising quality | creative, production |
| July 2026 | Google’s AI Overviews and AI Mode arrive in France | SEO, content |
| August 2026 | Microsoft switches AI Max on by default for new Search campaigns; the AI Act requires the labelling of certain generated content | media buying, legal |
| October 2026 | Google Ads extracts promotions from the advertiser’s site on its own | media buying, e-commerce |
This table does not say that everything happened at once. It says that every building block of the trade has been hit at least once, and that the platforms have moved faster than most agencies.
Why is everything speeding up in 2026?
Ad platforms are automating media buying
When Google introduced AI Max for Search campaigns in May 2025, it described it as a suite that can be switched on in one click, which extends queries beyond keywords, generates the copy and chooses the landing page. It announced 14% more conversions at a comparable cost per acquisition, according to its own data. Microsoft followed: since 27 August 2026, AI Max has been open to all advertisers and switched on by default on new Search campaigns, with 13.6% more conversions measured over 44 A/B tests, again by the vendor itself.
The latest step: from 12 October 2026, Google Ads automatically extracts promotions visible on an advertiser’s site (promo codes, offers) to insert them into its Search and Performance Max ads, unless the advertiser switches it off in its settings. An expired promotion, or one limited to part of the catalogue, can therefore end up in an ad without anyone having approved it.
The media buyer’s job is changing in nature. They no longer adjust bids keyword by keyword: they feed the machine and they monitor it.
The media buyer’s new job
- Feed
Real conversions, sales sent back from the CRM, exclusions, visuals.
- Let the platform optimise
Bids, queries, copy and landing pages.
- Monitor
Broadened queries, generated copy, promotions taken from the site.
We detail the first of these tasks in Send your CRM sales back to Google Ads.
Search ends more and more without a click
Pew Research tracked 900 American adults and nearly 69,000 searches in March 2025: when an AI summary appears, a traditional link is clicked in only 8% of visits, compared with 15% without a summary. Ahrefs, on 300,000 keywords, estimates in February 2026 that an AI Overview reduces the click-through rate of the top position by about 58%, a correlation and not proof of cause.
In France, both features have been available since 22 July 2026, with a new Search Console setting to opt a site out. The ground was already a minefield: according to SparkToro, 65.3% of Google searches in France ended without a click in the first half of 2026. What this does to traffic, and how to protect yourself, is covered in our article on AI Overviews and traffic. The new trade that comes out of it is called GEO: being cited in AI answers, and measuring it, which we explain in Measuring your visibility in AI.
Producing content costs almost nothing
On the Gemini API price list, here is what the computing for a piece of creative costs, before any human work:
per second of Veo 3.1 Lite video at 720p
of computing for a 30-second spot
US cents per generated image with a single model, depending on the resolution
Production cost is no longer what limits a campaign.
Clients know it. In the HubSpot State of Marketing 2026 survey, 83.5% of marketers say they are expected to produce more content thanks to AI, and 62.7% think that more human, distinctive content is needed to stand out. Both are true at the same time: more is produced, and what looks like everything else is no longer any use.
AI itself becomes an advertising channel
OpenAI opened its ads in ChatGPT in 31 European markets, including France, on 24 August 2026, sold at first by its sales team and partner agencies, and without personalised advertising in the European Economic Area to begin with. For an agency, this is one more channel to test, with its own rules and no performance history.
What the agency market is going through
The big groups are reorganising. Omnicom completed its acquisition of Interpublic on 26 November 2025, and WPP reported for the first half of 2026 net revenue down 4.7% on a like-for-like basis, with a savings plan and a platform presented as “agentic”. WPP’s chief executive believes that billing for time spent is probably not sustainable in the long term, but that outcome-based pay will still take a few years.
In France, the market, for its part, is growing. The e-advertising observatory of the SRI (France’s trade body for online advertising sales houses) measures €6.7 billion of digital advertising in the first half of 2026, up 12%, of which 41% for search. The same report counts the rise of AI agents among the three developments that could speed up the transformation of the market, by opening “a new phase of automation of the advertising chain”, and points out that European players account for only 17% of the market.
On the advertiser side, AI is everywhere but still pays off little. The Banque de France (France’s central bank) measures this gap in companies with 20 or more employees, where generative AI is used mainly in support functions, including marketing. HubSpot finds the same among marketers:
Use of AI and the productivity gain reported
Companies with 20 or more employees (Banque de France)
Marketers (HubSpot)
The time saved producing is lost again reviewing, correcting and putting the context back.
That is where an agency fits in: between a client who has the tools and results that do not follow.
Which AI tools matter for an agency in 2026?
We sort them by task, with what a simple rule already does and what must stay in human hands. Prices change every quarter: check them with the vendor on the day you choose.
| Task | Typical tools | Without AI, a rule is enough | What stays human |
|---|---|---|---|
| Media buying | AI Max (Google, Microsoft), Performance Max, Advantage+ (Meta) | budget alerts, exclusions, sending CRM sales back | the strategy, the signals given to the machine, what is switched off |
| SEO and GEO | Search Console, Ahrefs, Semrush (acquired by Adobe in April 2026), Profound, Peec AI | sitemap, tags, rank tracking | the choice of topics, the evidence, what gets cited |
| Content and creative | assistants (Claude, ChatGPT, Gemini), Canva, Veo, Gamma | document templates, format variations | the idea, the brand’s voice, the review |
| Reporting | Looker Studio, n8n, an assistant to comment on variances | the Monday report, pulled automatically from the platforms | the interpretation and the recommendation |
| Customer relations and CRM | HubSpot, Twenty, Odoo, with their agents or MCP server | follow-ups, lead assignment, updates | what is promised to the client |
| Combining models | OpenRouter, a gateway such as LiteLLM or Cloudflare’s | choosing the model by type of task | spending caps, the data you send |
Two signs that these tools are no longer gadgets: Profound, which measures a brand’s presence in AI answers, raised $96 million at a $1 billion valuation in February 2026, and Adobe completed its acquisition of Semrush in April 2026 to integrate it into its marketing suite.
OpenRouter, the missing piece for most agencies
An agency that uses AI often pays for several subscriptions per person, plus APIs from two or three providers, without knowing what each client costs. OpenRouter solves part of the problem: a single key and a single bill for hundreds of models, from Claude and GPT to open models such as DeepSeek or Qwen, which you switch by changing one line. According to its documentation, it passes on the providers’ prices with no markup on inference and charges 5.5% when you buy credits, and it does not log the content of requests unless you switch that on.
The benefit for an agency is concrete, and it has its limits:
What the agency gains
Each task goes to the cheapest model that does it well: a small model to classify or summarise, a large one to write what commits the brand
One key per client, with a cap, gives the real cost of AI per client, and therefore a line item that can be re-billed
If a provider changes its prices or goes down, you switch to another without rebuilding anything
The limits
The fee taken when buying credits has to be paid
The free variants of some models may be used for training
With us, no client data goes through a free model. The full method for keeping an AI bill under control is in The cost of AI in business: tokens, routing and forecasting.
What does not work, or not yet
Platform figures come from the platforms. Google’s 14%, Microsoft’s 13.6% or the 30% rise in conversions announced by Google for Demand Gen are internal measurements. They are not false for all that, but they do not replace a test on your own accounts, with a control group.
Default automation is expensive when nobody monitors it. An expanded query that is off topic, an out-of-date promotion picked up in an ad, a landing page chosen by the machine that does not convert: these are silent losses.
Mass content no longer protects you. Google penalises the production of pages at scale that have no value for the reader, whatever tool writes them. Producing more is not a strategy.
Legal obligations are arriving. Since 2 August 2026, Article 50 of the AI Act requires labelling deepfakes and certain generated texts published on matters of public interest, and telling the user that they are talking to an assistant; the European Commission has published a code of practice for doing so. What a company that uses AI must do is in AI Act and GDPR: what an SME using AI must do.
AI everywhere multiplies costs. A model called to send a reminder that a rule would send at no cost is a bill that climbs with no gain. Many marketing tasks do not need AI, as we show in Automation without AI: when a rule is enough.
What an agency brings that AI does not replace
AI takes over part of production and buying. It does not take over the trade. Platforms optimise a goal they are given, on the data they are given, and they grade their own work. Everything that happens before and after remains to be done, and that is precisely what clients find hard to do alone.
- Choosing the objective and making trade-offs. An automatic campaign maximises what it is asked to. Deciding which product to push, which margin to protect, how to split the budget between search, social networks and content, and what to switch off: that is a business judgement, not a calculation.
- Knowing the client better than the machine does. Its margins, its stock, its season, what its salespeople say, the customers it no longer wants. The ad platforms do not know this. The agency that translates it into signals (real sales sent back from the CRM, exclusions, priorities) gets better results than the one that adjusts bids.
- Having the idea and holding the brand. When everyone produces quickly and cheaply, what looks like everything else is no longer any use. The idea that catches attention and the recognisable voice are worth more than before, not less.
- Measuring as a referee. Between falling traffic, citations in AI answers and the conversions each platform claims for itself, the client needs someone independent to tell them what really pays off, and new indicators such as visibility in AI answers.
- Answering for what goes out. A false text, an out-of-date promotion, a piece of content to be labelled under the AI Act: someone has to review it and take responsibility. This guarantor role becomes a service in its own right.
- Making the client’s team succeed. Companies have the tools, but most of those that use generative AI report no productivity gain from it. Training, building good habits and closing that gap is an agency’s job.
Key takeaway
What changes most is the way of selling
A task that took a day now takes an hour: billing the day will not hold, and billing the hour would ruin the agency. A fixed fee per deliverable or a subscription, with a known AI cost per client, pays for these six roles rather than production time. And new services are opening up, such as GEO, measurement or automation, to be offered before the client discovers them elsewhere.
How we help agencies bring in AI
We are not an advertising agency and we are not after your clients. We are the technical team you often lack: the one that automates your production chain, combines AI models at the right cost and connects tools together, so that your teams spend their time on the six roles above. The work always follows the same order.
- An audit first.
We look at your real processes (brief, keyword research, creative variations, going live, monthly report, client meeting notes), your tools and what they allow you to connect, what holds the team back and what it wants to try, and what will need training. For each task: the time it takes, its frequency, what it would cost to automate and what that would bring in. This is our AI Audit Sprint, which also says what is not worth automating.
- Automating without AI whatever follows a rule.
The Monday report pulled from the platforms, budget alerts, sending CRM sales back to the ad platforms, creating tasks after a client meeting: with n8n, these flows run on their own and cost almost nothing each time. The detail, task by task, is in our guide to marketing automation for SMEs, with or without AI. This is our n8n process automation.
- Putting AI where it pays off, at the right cost.
We combine several models through a gateway such as OpenRouter: a small model to sort and summarise, a large one for whatever commits the brand, a model hosted on your own infrastructure for sensitive data. For multi-step tasks, an agent works in a loop (it drafts, checks the result against the brief, corrects) then hands over to a person. With a meter and a cap per client, you know the AI cost of each account and can re-bill it: this is our AI at the right cost offer.
- Opening new services to you.
Measuring visibility in AI and working to be cited there, ongoing SEO and GEO work, connectors between your clients’ CRM, shop and ad platforms: so many services you can sell without recruiting a technical team. This is what our digital marketing in one block offer covers, where ads, SEO and CRM work together. We can also work behind the scenes, under your brand, for your own clients.
- Setting guardrails and training.
Human review before anything goes out in a client’s name, data kept separate by client, the labelling required by the AI Act, a written list of what is switched off in the platforms. Then a trained team that knows what each automation does, without which it will be bypassed within three months. The time saved and the cost per deliverable are measured before and after.
What you gain: teams that spend less time producing and more time advising, a known AI cost for each client, and new services to sell on a technical base you did not have to build.
What we run ourselves
We apply these rules on the sites we operate before offering them.
From the field
On Palmora Property, a property listings site in six languages that we operate, switching from DeepL to Gemini cut the cost of the translation pipeline by a factor of 60, and 95% of segments come from the cache after the first pass, without calling the AI again. These figures apply to that pipeline only. Our calls to models go through OpenRouter at the start, then through a direct account when the volume justifies it, and through a model hosted on our own infrastructure for anything sensitive.
A client learned this the hard way: a model once published made-up content on their site, and it had to be removed. Since then, nothing that an AI writes is published without human review on our side, this blog included.
Where to start?
AI does not make agencies useless: it removes the part of the work that clients paid for least willingly, and makes the rest more valuable. An agency does not need to change everything at once. It needs to know where its time goes, what the AI it already uses costs it, and which two or three tasks will pay off most once automated. Our AI Audit Sprint answers these three questions in two to four weeks, at a fixed price, before any investment.
Frequently asked questions
Will AI replace digital marketing agencies?
It replaces part of the production and buying tasks, not the trade. Platforms automate bidding and creative, but they optimise on what they are given and do not know what matters to a brand. Agencies that mostly sold production time are suffering; those that sell strategy, data and measurement keep their place.
Which AI tools should a marketing agency use in 2026?
A general-purpose assistant for the team (Claude, ChatGPT or Gemini), the AI features of the ad platforms it already manages, a tool for tracking visibility in AI for GEO, and an automation platform such as n8n. For model calls, a gateway such as OpenRouter avoids multiplying accounts and gives the cost per client.
Does an agency need to hire technical profiles to bring in AI?
Not necessarily. An agency needs its automations to run, its AI costs to be tracked per client and its data to be protected, not a full-time team of engineers. It can entrust this technical base to a partner, including behind the scenes under its own brand, and keep its hiring for consulting, creative work and client relations.
Should you tell your clients that you use AI?
Yes, out of commercial honesty, and sometimes out of obligation. Since 2 August 2026, the AI Act requires labelling deepfakes and certain generated texts on matters of public interest, and flagging a conversational assistant. The simplest approach is to write it into the contract: what is produced with AI, what is reviewed, and by whom.
What is OpenRouter and why use it?
OpenRouter is a service that gives access to hundreds of AI models with a single key and a single bill. It passes on providers’ prices with no markup on inference and charges 5.5% when you buy credits. For an agency, it makes it possible to choose the cheapest model that is enough for each task, and to track spending client by client.
How do you bill clients for AI?
Not by the hour, since AI precisely reduces the time spent. A fixed fee per deliverable or a monthly subscription holds up better, with the real cost of AI tracked per client thanks to a dedicated key and cap. The client sees what they pay, and the agency does not silently fund its clients’ consumption.
Sources
- AI Max for Search campaigns (new tab)blog.google
- AI Max has been open to all advertisers and switched on by default (new tab)about.ads.microsoft.com
- Automatically extracts promotions (new tab)seroundtable.com
- A traditional link is clicked in only 8% of visits, compared with 15% without a summary (new tab)pewresearch.org
- Reduces the click-through rate of the top position by about 58% (new tab)ahrefs.com
- Available since 22 July 2026 (new tab)blog.google
- 65.3% of Google searches in France ended without a click (new tab)sparktoro.com
- Gemini API price list (new tab)ai.google.dev
- HubSpot State of Marketing 2026 (new tab)blog.hubspot.com
- Opened its ads in ChatGPT (new tab)openai.com
- Completed its acquisition of Interpublic on 26 November 2025 (new tab)omc.com
- Net revenue down 4.7% on a like-for-like basis (new tab)wpp.com
- Will still take a few years (new tab)digiday.com
- E-advertising observatory of the SRI (new tab)sri-france.org
- Banque de France (new tab)banque-france.fr
- Raised $96 million at a $1 billion valuation (new tab)tryprofound.com
- Completed its acquisition of Semrush (new tab)news.adobe.com
- Documentation (new tab)openrouter.ai
- 30% rise in conversions announced by Google for Demand Gen (new tab)blog.google
- Penalises the production of pages at scale (new tab)developers.google.com
- Code of practice (new tab)digital-strategy.ec.europa.eu
AI Max has been open to all advertisers and switched on by default
Read on about.ads.microsoft.com(new tab)A traditional link is clicked in only 8% of visits, compared with 15% without a summary
Read on pewresearch.org(new tab)Jim · founder of Brain Plus AI
Also in digital marketing
Practical guide11 min The topic guide
Marketing automation for SMEs: with or without AI
Emails, social media, SEO in several languages, support, reporting, catalogue: most of an SME's marketing tasks can be automated with simple rules. AI is justified at a few precise steps, and a person reviews whatever commits the brand.
Understand8 min
Google AI Overviews: what they do to your traffic
Google's generated summaries reduce clicks on some queries, especially questions. What the public data measures, and how to read the effect in your own Search Console before deciding.
Practical guide10 min
The cost of AI in business: tokens, routing and forecasting
What you really pay when you use an AI through an API: input and output tokens, the context sent back with every call, an agent's turns. And a method to forecast the bill before you receive it.