The Agency After Automation.

Ask in public whether AI will replace agencies and the answer comes back reassuring, because the question was built to invite one. The narrower question is the useful one, and it is less comfortable: which parts of what an agency invoices for are now produced at near-zero marginal cost by a system the client can buy directly.

AI has not automated marketing. It has absorbed one layer of it: the codified, procedural, volume layer, which is the layer agencies staffed most heavily, trained juniors on, and billed at the highest margin. Configuration, audit production, content generation, scheduling and reporting assembly were never the valuable part of the work. They were the visible part, the part that justified headcount, and the part through which judgment was manufactured as a by-product. Removing that layer does not shrink an agency proportionally. It changes its shape.

What survives is narrow: deciding what an automated system should be told to do, verifying independently whether it did it, and carrying accountability for the call. None of the three is something a platform can do to itself, which is why they are defensible. All are senior. That is why the transition is hard. The industry is cutting the entry tier faster than anything else, and that tier is how it made seniors. That problem matters more than any single platform capability.

What automation has absorbed

Paid media configuration

Configuration is collapsing as a billable craft.

• Performance Max and AI Max for Search handle keywordless query expansion, placement selection, bidding, audience determination, asset generation and final-URL expansion end to end. Google's 2026 roadmap adds steering and reporting controls; it does not return manual configuration (Search Engine Land, Google Ads Help).
• Meta has stated the destination plainly: the advertiser supplies an objective, a product image and a bank account, and Meta generates creative and targeting. Advantage+ is the default path, Andromeda ranking is live, and full end-to-end generation is announced for 2026 rather than confirmed shipped (Marketing Dive).
• AI-powered ad spend is projected at 57 billion dollars in 2026, roughly 12 per cent of the US ad market, growing 63 per cent year over year while the other 88 per cent grows at 5 per cent (eMarketer).

The SEO deliverable, and the market beneath it

• Full-site crawls producing prioritised, task-level technical and content audits are increasingly a tool output rather than a consulting engagement. The deliverable that carried a four- to five-figure fee is reproducible for a subscription.
• Pew found click-through to external links falls from 15 per cent to 8 per cent when an AI summary is present, and that only 1 per cent of users click links inside the summary.
• US zero-click searches reached 68 per cent in January to April 2026, up from 60 per cent in 2024 (Search Engine Land). Publisher Google referrals fell roughly a third year on year through November 2025, with US organic referrals down 38 per cent (Press Gazette, Reuters Institute).

Social media management, hollowed from both ends

• Listening is genuinely automated: sentiment and emotion classification, sarcasm and slang handling, visual and logo recognition, anomaly alerting and cross-source synthesis ship across Brandwatch, Talkwalker, Meltwater, Hootsuite and Sprout.
• The step from listening to ideation, which is what agencies charged strategy fees for, is where tooling is now pointed, though automatic idea generation is more marketed than demonstrated in the vendor documentation reviewed. Arriving, not arrived.
• Generation, scheduling and publishing are solved mechanically. Unsupervised posting is not: no credible trade evidence shows brands running fully autonomous social at scale, and the IAB's disclosure guidance assumes human oversight (Digiday).
• Community management is partly automated: spam filtering, intent classification, routing. The reported failure mode is that keyword-driven systems miss a meaningful share of sarcasm, memes and multimodal content, which is where brand risk sits.

What survives, and why

Three functions survive, and none can be performed by the system being optimised.

The independent check on a system that grades its own homework. The platform reports the conversions. It is also paid for them. Practitioner analysis cited in Search Engine Land's Performance Max guide reports GA4-to-Google Ads data gaps as high as 85 per cent on PMax-heavy accounts. That is a third-party figure rather than verified research, directional at best. Performance Max will still bid on branded terms a client already ranks for unless brand exclusions are configured, and that is opt-in. Automation optimises toward the objective it was given, with the data it was given, and reports against its own definition of success. Each is a judgment the machine cannot make about itself. Incrementality testing, holdout design, media mix modelling, and telling a client a channel is claiming credit for demand it did not create, are work the platform cannot sell.

Deciding what goes into the machine. Control relocates upstream: the offer, the creative, feed and data quality, the objective definition, the exclusion and constraint set, the first-party data. These are harder than configuration, because they require knowing the client's business, margin structure and customer rather than a platform interface. Configuration shops die; input shops compound.

What the market is short of because AI made it abundant. Consumer preference for AI-generated creator content fell from 60 per cent in 2023 to 26 per cent in 2025 (Billion Dollar Boy, MUSE V2), and over 20 per cent of videos shown to new YouTube users in late 2025 were flagged as AI slop (Digiday). Only 6 per cent of B2B marketers say AI tools significantly improved content performance, while 67 per cent increased output (eMarketer). More output, no better outcomes: that gap is the agency's market.

Add one that is commercial rather than technical. As execution costs fall toward zero, what the client is buying is someone who will be wrong on the record, someone who commits to a call and carries the consequence.

The role of the agency after the shift

• Measurement Authority. Independent verification of what marketing caused: incrementality, geo-holdouts, media mix modelling, unified reporting across platforms that each claim the same conversion. Highest defensibility, highest fee tolerance, hardest to build.
• Input Architect. Offer strategy, creative systems, feed and data quality, first-party data plumbing, and the constraint set fed into automated buying. Where paid media expertise migrates rather than dies.
• Demand Creator. Brand, distinctive assets, original research, digital PR, creator and community strategy. Creator economy spend is growing 26 per cent year on year, roughly four times faster than the media industry overall (Digiday, IAB).
• Capability Partner. Building the client's own AI-enabled marketing operation. With 82 per cent of brands already running an in-house agency (ANA, via eMarketer), being paid to build in-housing beats fighting it.

What stops being viable: channel-specialist execution shops, audit-and-recommend consultancies, content-volume vendors, scheduling and community retainers, and any agency whose primary claim is access to expertise in a platform interface.

The shape of the organisation

• UK, all agencies (IPA Census 2025): total employees fell to 24,963 from 26,787, down 6.8 per cent. Creative agencies fell 14.3 per cent; media agencies grew 2.4 per cent.
• The base is falling roughly three times faster than the whole: under-25 headcount down 19.2 per cent. Graduate and apprentice programmes fell from 56 per cent to 43.4 per cent of agencies.
• Holding companies: Omnicom and IPG combined fell from 128,200 to 120,000 across 2025, about 8,200 roles or 6.4 per cent, against a 1.5 billion dollar synergy target by 2028 of which 1 billion is labour (eMarketer, Campaign). Dentsu is cutting 8 per cent, about 3,400 roles, by end-2026 (MARKETECH APAC). WPP is running a 676 million dollar cost programme (Adweek).
• Intent: 8 per cent of UK agencies have already cut headcount because of AI and 24 per cent expect to within twelve months. In the US, 91 per cent of senior agency leaders expect AI to reduce headcount and 57 per cent have slowed or paused entry-level hiring (IPA, eMarketer).
• Beyond marketing: Stanford's Digital Economy Lab finds employment for 22 to 25 year-olds in highly AI-exposed occupations 19 per cent below trend as of June 2026, widened from 15 per cent a year earlier, driven by reduced hiring rather than layoffs.

Two caveats. Holdco cuts are confounded by merger synergies and the decline of the holdco model: Big Six share of US ad spend fell from 44.6 per cent in 2019 to 29.6 per cent by the first quarter of 2024. And no benchmarking body we could find publishes a junior, mid, senior, leadership and owner headcount ratio: not IPA, 4A's, Promethean, Parakeeto, AMI or SoDA. The table below is a model built from adjacent evidence. Argue with it rather than cite it.

The pyramid does not shrink proportionally; it inverts into a diamond. The pattern is documented across industries: entry-level roles fell from 6.8 per cent to 4.6 per cent of the workforce in two years across more than 8,700 companies, and 77 per cent of postings in leading sectors target two to nine years of experience against 12 per cent at entry level (Pave). Agency-specific confirmation is not published, though the IPA under-25 figure points the same way.

Elasticity is the obvious objection, and it is real: when execution price collapses, clients who could never afford an agency become servable. Three things stop it restoring the old org chart. Headcount is demand divided by productivity per head, and productivity is winning. Revenue holds while headcount falls across every major group. The demand that opens up at the bottom of the market is being captured by small AI-native shops whose cost structure a hundred-person agency cannot match. And price was never the only constraint there: trust, and the client's capacity to act on advice, were larger barriers cheap execution does not fix. The realistic path is contraction to roughly 65 to 70 heads on today's revenue, then 30 to 50 per cent revenue growth on higher-value work, returning to roughly 85 to 95 heads. Near where you started, and not the same organisation.

What this means for agency leaders

• Audit revenue against automation, line by line. Name the fee income a client could replicate with a subscription. That number is your exposure, and most agencies have never calculated it.
• Build the measurement practice first. Most defensible, hardest to copy, and it converts into new business: a discrepancy analysis on a prospect's own account is unarguable and it is about them.
• Reprice before the client asks. The source material carries a figure of roughly 60 per cent of US senior marketing leaders already spending less on agencies because of AI; it is unattributed there, so treat it as indicative, though the direction is not in doubt. The agency that admits an audit is now a tool output, and stops charging for it, earns the right to be believed about everything else.
• Never sell effort. Once fee is anchored to labour, AI-driven cost reduction becomes a demand you must concede. Retire "we manage your campaigns", "full-service digital marketing", "we'll audit your site", "X posts per month", "our team of specialists", and anything priced as time.
• Sequence the commercial model. Unbundle execution, judgment and verification into separate line items; price execution near cost, price judgment and verification at senior rates per decision, and take outcome exposure only where you control the variables. Expect slow movement: WPP's chief executive calls outcome-based pay a few years away with one client signed, Omnicom is negotiating performance-linked terms, and a 174-agency survey found no consensus, with a third having launched AI-linked revenue streams and impact still small (Digiday, Productive.io).
• Raise rigour rather than lowering it. A system spending budget autonomously across surfaces you cannot see compounds errors at machine speed and reveals them late. Input QA becomes the primary control, under named ownership: feed accuracy, conversion-tracking integrity, audience-signal hygiene, exclusion lists, brand-term protection. Constraint architecture becomes documented discipline, since platform protections are opt-in, and independent measurement becomes a standing test calendar with defined anomaly thresholds and escalation paths.
• Govern AI output explicitly. Name who approves machine-generated work, against what standard, with what disclosure position. The source material notes a gap between executive optimism and consumer sentiment on AI-made advertising widening from 32 to 37 points between 2024 and 2026, though it carries no citation for that figure.
• Withdraw rigour with equal discipline. Automate reporting assembly, and stop paying people for routine keyword research, bid management, individual placement exclusions, manual audit compilation, first-draft copy and scheduling. The test for any process and any meeting: does it produce a decision, or evidence of activity. Agencies that automate execution but keep the rituals get the old cost base and the new fee pressure.
• Redesign the org chart intentionally. Attrition produces a smaller version of the wrong structure. Organise around client problems rather than channels, run delivery on decision points rather than deliverable schedules, and move quality control from output review to input and judgment review.
• Keep hiring juniors, and change what you hire them for. Not execution capacity: judgment apprentices attached to senior decision-makers, adjudicating machine output rather than producing it. The industry is defunding its own senior supply.

What this means for the people inside agencies

The instinct is to train people on tools. That instinct built the org chart now being dismantled, and tool proficiency is the one capability that reliably depreciates on a schedule. Juniors used to acquire judgment as a by-product of volume: a thousand keyword decisions, a hundred campaign builds, endless reports nobody remembers writing. The apprenticeship ran on waste. It is gone, so judgment now has to be built deliberately or not at all.

• Paid media specialists carry the highest displacement risk and the clearest route out: from configuration to input design: offer construction, creative testing systems, feed and data quality, constraint architecture, objective definition. Measurement literacy is the core skill and the first thing to fund: incrementality, holdout design, geo-testing, significance, and reading platform-reported conversions with justified suspicion. Business and margin literacy follows, because an objective cannot be set without knowing what the business earns.
• SEO specialists move from audit production to audit adjudication: the tool produces two hundred recommendations, and the value is knowing which twelve matter, which are wrong, and what they cost. Entity, brand and citation strategy across AI answer surfaces is the growth area, with one honesty requirement. AEO and GEO have no independently verified evidence of efficacy yet, and belong in a proposal as informed experimentation. Measurement in a zero-click world, original research and digital PR are the durable inputs.
• Social and content teams move from production to editorial and brand judgment: what the brand should say, what it must never say, what is worth saying. Creator and community strategy is where spend is moving, with influencer spend growing more than three times faster than social ad spend (eMarketer). Brand risk judgment matters more as automated community management misses the sarcasm, memes and multimodal signals that cause incidents.
• Everybody needs two things. The first is client conversation capability, since explaining a judgment and defending it under challenge is revenue skill, not soft skill. The second is AI direction rather than AI operation: briefing, constraining, evaluating and rejecting machine output. That is quality judgment, which requires a standard, which requires domain competence.

Development structure changes with the content. Training becomes supervised judgment under real stakes: decision drills, post-mortems on live accounts, structured critique of AI output, rotation across client problems rather than channel silos. Assessment measures judgment rather than knowledge, because tool certifications say nothing about whether someone decides well under ambiguity. The apprenticeship gets compressed through manufactured decision repetitions, since the volume that produced a senior in eight years no longer exists. None of this is cheap, and the firms that need it most are the ones currently cutting the budget that would pay for it. The alternative is to buy senior people from a market that has stopped manufacturing them, at whatever price that market decides to set. The warning is already sitting in their own census data, and they have kept cutting anyway.

AFactor's programmes are built on this premise: role-based, assessed on reasoning under uncertainty rather than tool familiarity, with portfolio proof rather than attendance as the condition of graduation. Placement follows capability; it is an outcome, not a promise.

Where we could be wrong

The strongest counter-argument is that this brief has mistaken a hiring pause for a structural change. Almost all measured contraction sits at holding companies, where merger synergies and the decline of the holdco model confound the AI signal, and independent-agency data is thin. If the elasticity offset is larger than current data suggests, what looks like a permanent inversion of the pyramid is a cyclical trough in a market that will hire juniors again once demand catches up with capability.

Three further ways the position could fail. Automation may plateau below its promise: Meta's fully autonomous generation is announced rather than shipped, and Google keeps returning controls under advertiser pressure. If it stalls at the level of a very good assistant, execution skill retains value longer than argued here. The measurement opportunity may be captured by software rather than services: if independent incrementality becomes a cheap product rather than an expert practice, the most defensible position here erodes with it. And consumer reaction against AI content may prove cyclical rather than structural, compressing the premium on human distinctiveness.

What would change our position: independent-agency data showing entry-tier hiring recovering while revenue per head holds; a published benchmark of agency tier ratios, which does not currently exist, contradicting the model above; two consecutive years of rising graduate and apprentice participation in the IPA census; verified efficacy evidence for automated incrementality products.

The recommendations here were built to hold under all four scenarios. Each increases the seniority, judgment density and measurement independence of the firm, and none depends on a prediction about how fast the platforms ship.

The closing argument

Agencies were paid for two forms of scarcity: access to platforms and expertise, and capacity to execute at volume. Both are now abundant. Abundant things do not carry fees. What remains scarce is judgment exercised under accountability. That was always the real product, bundled inside execution and subsidised by a labour pyramid that no longer makes economic sense.

Sources

• IPA Agency Census 2025; Stanford Digital Economy Lab, Canaries August 2026; Pave on diamond-shaped org structures; Pew Research Center on AI summaries; Press Gazette and Reuters Institute on publisher traffic
• Search Engine Land, zero-click study 2026 and Performance Max guide; Google Ads Help, AI Max for Search; Google, PMax updates 2026; Marketing Dive on Meta AI-automated ads
• eMarketer: AI ad spend 2026; Omnicom and IPG cuts; AI adoption and agency hiring; consolidation FAQ; content and zero-click FAQ; influencer versus social ad spend
• Digiday: WPP on outcome-based pay; AI's branding problem; authenticity backlash; creator economy spend. Adweek on agency business models; MARKETECH APAC on Dentsu; Campaign on Omnicom and IPG headcount
• Productive.io, Agencies in the AI Era; Piscari, Agency Reset 2026; Promethean Research, agency growth guide; Billion Dollar Boy, MUSE V2; Hootsuite, AI social listening; Meltwater, social listening tools