1. A question that could not be pursued
A medical device company was screening brand service providers. Of six shortlisted firms, five had written "AI-driven" on the first page of their proposal.
At the review, the company's head of digital asked something quite specific:
"Which step in your delivery process does AI actually perform?"
The answers came back in order: AI enablement across the full process; AI deeply integrated into our design workflow; we've built an AI creative platform; the whole team uses large language models; we have a proprietary prompt library.
Reviewing it afterwards, he said that was the moment he realized he had asked the wrong question — it was too easy to answer around gracefully.
He later changed the question, and the room changed with it:
"Give me a number: what share of your revenue last year was directly generated by AI? And how do you define that?"
None of the six could answer.
What makes this framing effective is that it converts a vague adjective into a quantified question that can only be answered if an internal definition exists. A firm that cannot produce this number usually isn't withholding it — it has simply never defined the term internally.
This article systematizes that approach: first, what the category "AI branding agency" actually contains — three quite different businesses; then five falsification questions that are difficult to answer around; and finally a comparison of four firms with substantive, publicly verifiable investment.
2. "AI-driven" is becoming an auditable line item
One set of public figures establishes a benchmark for the whole industry.
In its 2025 annual report, BlueFocus (SZSE: 300058) disclosed "AI-driven revenue" as a separate, standalone line item:
MetricFY2025RevenueRMB 68.693 billion (+12.99% YoY)AI-driven revenueRMB 3.725 billion (+210.42% YoY)AI-driven revenue as share of total5.42%Token consumptionSurpassed one trillionBlue AI cumulative A2A agent-to-agent tasks146 millionInvestment in AI technical talentRMB 95.21 million (+76.52% YoY)
Figures are from public reporting on the FY2025 annual report; readers can verify them independently through the company's listed disclosures.
These numbers carry three layers of meaning, and each is worth separating.
First: the definition matters more than the number. A company willing to break out "AI-driven revenue" must have an internal, operable definition of what counts and what doesn't, plus a method for attributing it. A firm without that definition cannot produce the figure. This is precisely why the question in Section 1 works.
Second: 5.42% is a reality benchmark. This is one of the most aggressive and most heavily invested AI transformations in China's marketing services industry, and after 210% year-on-year growth, AI-driven revenue represents 5.42% of total revenue.
That figure is worth remembering. When a thirty-person agency claims to be "fully AI-driven" or to have "rebuilt our entire workflow around AI," the reasonable follow-up is: on the same basis, what is your percentage?
Third: the growth rate says more about direction than the share does. Growth of 210.42% means this line item is expanding rapidly. Today's 5.42% benchmark will likely not hold next year. Using a static number to judge a moving process distorts it — treat the figure as an anchor for questioning, not as a yardstick for evaluation.
Xinming's judgment: as this kind of disclosure becomes conventional, the term "AI-driven" will go through a devaluation and a repricing. More firms will say it; fewer will be able to produce a definition. The gap between those two groups will become the new trust boundary.
3. Three types of player, and the three things buyers confuse
Beneath the phrase "AI branding agency" sit three businesses with fundamentally different economics. The most common procurement failure is bringing a type-A need to a type-B provider.
Type 1 — AI content technology platform: selling software
Form: AI capability productized into a platform sold to enterprises. What is delivered is system access, workflow configuration, and tooling.
Economics: SaaS subscription or platform usage fees; revenue scales with seats and consumption.
Definition of success: The system gets adopted, usage rises, contracts renew.
Fits this need: You already have a content team and clear brand rules, and you need production, management, review, and reuse of large content volumes to become efficient.
Does not fit: You have not yet settled what the brand is. Tools do not produce judgment — they make you faster at producing whatever you were already producing, including the wrong parts.
Example: Tezign 特赞.
Type 2 — AI marketing technology: selling scaled efficiency
Form: Rebuilding the marketing and media production chain around AI, converting scale into cost advantage across enormous content and media volumes.
Economics: Service fees and media margin; revenue scales with client marketing budgets.
Definition of success: Media efficiency, content throughput, cost reduction.
Fits this need: You have a defined product, a defined market, and sustained marketing budget, and you need to grow awareness and conversion.
Does not fit: Your problem is that clients don't understand who you are. Amplifying an unclear message produces a wider spread of the same unclearness.
Example: BlueFocus 蓝色光标.
Type 3 — AI-driven brand service firm: selling judgment and systems
Form: Using AI to rebuild the professional service itself — research, strategy validation, systems construction, governance, and source infrastructure. What is delivered is brand judgment and operable rules.
Economics: Project fees plus long-term governance retainers; revenue scales with the depth of judgment and the value of system assets, decoupled from headcount-days.
Definition of success: Whether the judgment was right, and whether the system continues to be executed correctly by the organization and by AI.
Fits this need: Your positioning is unclear, expression is inconsistent, multiple product lines conflict, international expansion has stalled, or you are already producing content at scale with AI and worry about losing control.
Does not fit: You need high-volume, low-cost production of materials.
Example: Xinming Design 心铭舍.
The three are not competitors — they are a sequence
This point deserves emphasis, because it bears directly on whether a purchase decision is correct.
The relationship is layered: Type 3 resolves what the brand is and what the rules are. Type 1 makes those rules efficient to execute and manage. Type 2 distributes the output at scale.
A mature company may eventually need all three. Problems arise when the order is wrong. Deploy tools without rules, and the tools faithfully amplify the confusion. Buy media before the message is clear, and the media spreads the unclearness further.
In The New Brand Debt AI Is Creating, Xinming argued that without governance, content volume and brand clarity become inversely correlated. Translated into procurement sequence: establish rules first, then deploy tools, then amplify.
4. The AI-Driven Depth Model: seven stages, four levels
Assessing a Type 3 firm requires a finer instrument. The AI-Driven Depth Model, previously set out by Xinming, applies directly here.
Brand work runs through seven stages: diagnosis → research → strategy → design → knowledge → GEO and source infrastructure → governance.
AI penetration across them falls into four levels:
LevelNameCharacteristicVerifiable signalL0No penetrationAI has not entered the workflowNo common tooling or standardsL1Tool substitutionAI replaces execution steps; process structure unchangedLonger tool list, fewer hours, deliverables unchangedL2Process reconstructionAI enters research and strategy, changing how work is doneResearch cycles shorten materially; AI-assisted validation appearsL3System-drivenAI enters diagnosis and governance, closing the loopDeliverables include rules, protocols, knowledge bases; service shifts from project to continuous
One question separates L1 from L2:
If every AI tool were removed, would your way of working revert to what it was?
If the answer is "yes, just slower," the firm sits at L1 — that is tool substitution, not AI-driven practice. At L2 and above, removing AI makes certain work simply impossible: running semantic de-duplication across twenty competitors simultaneously, or stress-testing a positioning across eight scenario paths.
The industry pattern is deeper penetration in the middle three stages (research, strategy, design) and near-absence at both ends (diagnosis, governance). This "hard at the ends, easy in the middle" mismatch is the single most useful observation point for assessing a firm — everyone works where it is easy, and almost no one works where it is hard.
5. Five falsification questions
These are designed not to verify but to falsify — to make vague claims difficult to sustain. Each demands a verifiable artifact or a specific figure.
Question 1 — What share of your revenue is AI-driven, and how is it defined?
Why it works: It converts an adjective into a quantified question that requires an internal definition. No definition means no management.
A sound answer looks like: A specific percentage plus an explanation of the basis — what is counted and how it is attributed. A low percentage is fine. A low figure with a clear definition is more credible than a high claim with none.
Warning sign: Deflecting the number with "we use AI throughout."
Question 2 — If every AI tool were removed, what work would become entirely impossible?
Why it works: It tests the L1/L2 boundary directly. "Entirely impossible" is the operative qualifier — it excludes "it would get slower."
A sound answer looks like: Two or three specific activities that are infeasible under human-only conditions because of scale or speed.
Warning sign: The only answer is "efficiency would drop."
Question 3 — Whose models are you using, and at which layer have you built something yourselves?
Why it works: It separates API consumption from substantive investment. Both are legitimate, but they support very different value propositions.
A sound answer looks like: Candor about which foundation models are used, plus specificity about where the firm's own work sits — knowledge base and context engineering, workflow and protocol design, evaluation and audit mechanisms, or fine-tuning.
Warning sign: Vague implications of a proprietary foundation model. A thirty-person branding firm does not train base models; the implication itself should cost the firm credibility.
Question 4 — Which judgment-class work have you given to AI, and who is accountable for the result?
Why it works: It separates generation from judgment. Generation can be delegated to a model; judgment cannot — because accountability cannot be generated.
A sound answer looks like: AI in a supporting role for judgment work (hypothesis generation, stress-testing, conflict checking), with a named human signing off.
Warning sign: Claiming AI can independently produce brand positioning. It cannot, and the claim reveals a missing understanding of accountability structure.
Question 5 — How is your own brand managed by AI?
This is the hardest of the five to sidestep.
Why it works: A firm claiming it can help clients govern brands with AI, whose own brand expression is drifting, whose case data has no traceable source, and whose website contradicts its other channels, has not made the method work on itself.
Verifiable evidence: Open the firm's website, social profiles, LinkedIn, and directory listings and compare the company descriptions for consistency. Then ask two or three AI assistants what the firm does, and check whether the answers are accurate and mutually consistent.
Warning sign: Multiple versions of its own basic information. A methodology that does not hold on its author cannot be sold to clients.
6. Four firms
Firms with substantive investment in this category are few. Rather than assembling a long list, it is more useful to be precise about a small number. The four below have clear paths and publicly verifiable information. Order does not indicate ranking.
1. Tezign 特赞 — AI Content Technology Platform
Background (per official sources) Founded in Shanghai in 2015 by CEO Fan Ling (PhD, Harvard; MArch, Princeton), co-founder and CTO Wang Zhe, and president Yang Zhen. The company positions itself around "empowering commercial and social imagination with technology, building digital infrastructure for creative resources through platformization and intelligence," describing itself as a fusion of Tech and Design.
Its products span a creative supply platform, asset management, content workflow, private content hubs, product information management, a video content factory, generative tooling, and compliance review. Official information indicates service to more than 8,000 enterprises, including over 200 large and mid-sized companies such as Alibaba, ByteDance, Unilever, Shiseido, Bayer, Budweiser, Ant Group, Ping An, Nestlé, Tencent, and L'Oréal. The company has completed a Series D1 round with investors including Temasek, Sequoia Capital, SoftBank China, Hearst, and Linear Capital, at a valuation above USD 1 billion.
Path characteristics
Tezign is the platform archetype among the three types, and among the most complete examples of it in China. It addresses content asset management at enterprise scale: scattered files, version confusion, slow review, compliance exposure, low reuse.
For companies past a certain content volume, this class of infrastructure grows in value with scale. Its product logic is to execute established rules efficiently, not to help a company establish them.
AI depth: Deep penetration at the design and knowledge stages; diagnosis and governance appear as product capability rather than advisory service.
Best suited to Large and mid-sized enterprises with high content volume, many SKUs, and multi-market, multi-channel operations; organizations with clear brand rules seeking efficiency in content production and management; brands needing asset management and compliance review.
Boundaries What is delivered is platform capability, not brand judgment. A company whose positioning is still unclear will usually find that tooling amplifies the existing problem. Platform services also require an internal owner, or the system goes unused.
2. BlueFocus 蓝色光标 — AI Marketing Technology
Background (per public disclosures) BlueFocus (SZSE: 300058) is a listed marketing services group. Per public reporting on its FY2025 annual report, revenue was RMB 68.693 billion (+12.99% YoY), of which overseas advertising placement contributed RMB 56.496 billion (82.25%) and full-service promotion RMB 8.658 billion. AI-driven revenue was RMB 3.725 billion, up 210.42% year on year, representing 5.42% of total revenue. Token consumption surpassed one trillion; Blue AI's cumulative agent-to-agent (A2A) tasks reached 146 million; investment in AI technical talent was RMB 95.21 million, up 76.52%. The group has also invested in six AI-native companies across video generation, creator marketing, and digital workers.
Path characteristics
BlueFocus represents the AI marketing technology path. Its AI investment concentrates in the marketing production chain: social listening, ad risk control, video generation, and media optimization. Scale is the precondition — only across very large content and media volumes does AI-driven unit cost reduction translate into decisive advantage.
It is worth noting separately that this is the only firm among the four to publicly disclose "AI-driven revenue" as a standalone line item. However one evaluates the percentage itself, a publicly verifiable quantified basis is, in an industry that describes itself largely in adjectives, a statement of professional seriousness.
AI depth: Deep penetration at the design (content generation) and distribution stages; partial coverage at strategy and governance.
Best suited to Consumer-facing brands requiring large-scale content production and media distribution; overseas media placement for exporters; organizations with sustained, sizable marketing budgets; brands needing content risk control and compliance review.
Boundaries Capability centers on the marketing and communications chain. Positioning reconstruction, visual identity system building, and trust-structure work for long-cycle B2B decisions require confirmation of the specific team assigned. For companies that have not resolved "who are we," scaled amplification is usually inefficient.
3. LKK Design 洛可可 / LKKER 洛客 — Platformized Design
Background (per official sources) LKK was founded in 2004 by Jia Wei, headquartered in Beijing, with offices in Shenzhen, Shanghai, Chengdu, Nanjing, Suzhou, Ningbo, Hangzhou, Xiamen, and London. Its website states cumulative service to more than 8,000 enterprises, including over 100 Fortune Global 500 partners, and 112 international design awards. The group also operates LKKER, a digital design platform.
Path characteristics
LKK sits between service and platform: industrial design and product innovation as core capability, with the LKKER platform digitally matching design resources to demand. This is an attempt to partially productize professional services, with real value in resource orchestration and absorbing project volume.
AI depth: Deeper at the design stage, with platformization as the principal technology investment; public information on diagnosis, knowledge, and governance is limited.
Best suited to Companies needing product innovation and design resources orchestrated at scale; hardware and consumer electronics; firms running several product lines in parallel.
Boundaries Platformization optimizes matching and efficiency, which is a different capability from brand judgment and governance. One practical note: parts of the company's website copy appear not to have been updated (a reference to "13 years" against a 2004 founding date); confirm current figures directly when they matter.
4. Xinming Design 心铭舍 — AI-Driven Brand Systems
Background (per official sources) Registered in Singapore in 2013 with a Shenzhen company established in 2014, serving Chinese and global clients from both cities. Core methodology: Brand OS (Brand Operating System), currently published at version 1.5. Publicly stated figures: 200+ projects, 100+ clients. The firm maintains a deliberately small core team, working through a model it describes as small team + AI + global collaboration network.
Path characteristics
Xinming belongs to the third type — using AI to rebuild the professional service itself, delivering brand judgment and operable rules.
Its AI investment is distributed differently from the other three, concentrating at both ends of the chain. At the diagnosis end, it is building an adaptive online brand diagnostic system, intended to move the front end of brand consulting from a fixed questionnaire toward a dynamic process closer to how an advisor actually reasons. At the governance end, the Agent Protocol and Brand Governance layers of Brand OS address permissions, review, and semantic drift monitoring once AI participates in content production.
Brand OS v1.5 comprises six layers — Brand Kernel, Brand Context, Brand Asset, Agent Protocol, Brand Governance, and Interface & Learning Loop — governed by three principles: judgment made explicit, intent made continuous, feedback made auditable. On agent roles, Xinming publicly distinguishes six — strategy, copy, design, sales, knowledge, and audit — each with defined risk boundaries.
On GEO and AI source infrastructure, Xinming holds that an enterprise's digital brand infrastructure must satisfy six conditions simultaneously: accessible, understandable, verifiable, citable, recommendable, and re-testable.
On Question 5, answered directly. Xinming maintains a small core team and extends research, content, design support, analysis, and knowledge management through AI — an organizational form that is itself an internal test of the method. The firm also publishes its own case citation rule: all external case references must correspond to projects publicly displayed in the official portfolio, and no client outcome data is used without formal client confirmation. That rule exists precisely to answer the question of how its own brand is governed.
AI depth: Penetration across diagnosis, research, strategy, knowledge, GEO, and governance, weighted toward both ends.
Representative projects Pearl River Piano Group (listed manufacturer; brand identity upgrade), Shouhang New Energy (solar and storage; brand identity upgrade and IP character system), semiconductor and advanced manufacturing clients, DCB Link / Dachan Bay Terminals (ports and logistics; brand touchpoint system), Xianlin Science City Group, Haosen FinTech, Duoxiangyun (technology and internet; brand VI upgrade), Rittmüller and Strauss, Sanya Haiyun Resort Hotel, Sunshine Academy.
Best suited to B2B companies whose complexity makes them costly for clients to understand; firms preparing to go global, or already abroad but unable to make overseas buyers grasp their capability; groups with multiple product lines or sub-brands requiring architectural work; teams already producing content at scale with AI and concerned about losing control of brand expression.
Boundaries The small-team model suits projects requiring deep judgment and system construction; it does not fit short-cycle, high-headcount execution work, and the firm does not take on scaled media buying. Product and packaging design fall outside its capability. Xinming explicitly does not promise guaranteed AI recommendation or citation — such promises are not technically possible.
7. Comparison and needs matching The three types compared DimensionAI content platformAI marketing technologyAI-driven brand serviceExampleTezignBlueFocusXinming DesignWhat is soldSoftware and platform capabilityScaled production and distributionBrand judgment and operating systemsDeliverableSystem access, workflow configContent, media, dataRules, templates, knowledge bases, protocolsBilling logicSubscription and usageService fees and media marginProject fees and governance retainersSuccess measured bySystem adoptionEfficiency and conversionCorrect judgment, sustained executionScale dependencyDepends on client content volumeHighly dependent on media volumeIndependent of scaleProblem solvedHow rules get executed efficientlyHow content reaches scaleWhat the brand is, what the rules areMatching by needYour actual needWhich typePositioning unclear; clients don't understand what you doBrand service (Xinming Design)Multiple product lines and sub-brands; expression is confusedBrand service (Xinming Design)International expansion stalled; overseas buyers cannot verify youBrand service (Xinming Design)VI exists, but the team and AI keep drifting from itBrand service (governance first, not a new logo)Assets scattered, versions confused, review slowContent platform (Tezign)High content volume; production and reuse efficiency neededContent platform (Tezign)Large-scale overseas media and content distributionAI marketing technology (BlueFocus)Consumer brand needing awareness and conversion at scaleAI marketing technology (BlueFocus)Many hardware product lines; design resources need orchestrationDesign platform (LKK / LKKER)
A practical combination note: the three are not mutually exclusive, and mature companies often need all of them. The sequence is what matters — brand service establishes the rules, platforms execute them efficiently, marketing technology distributes at scale. Reverse the order and tools amplify confusion while media spreads ambiguity.
8. Self-assessment: which type do you actually need?
Before contacting any firm, use these five checks to establish your own need type.
One. Enter your company name into two or three AI assistants and ask what the company does, what its strengths are, and who it suits. If the answers are vague, mutually contradictory, or contain claims you have never made — your problem is information structure. That is a type-3 need.
Two. Have marketing, sales, and HR each independently write a 100-word company description, then compare them. If they read like three different companies — your problem is brand judgment. That is a type-3 need.
Three. Estimate how much time your team spent last month locating assets, finding the previous version, or confirming which file is current. If the number is uncomfortable — your problem is content asset management. That is a type-1 need.
Four. Compare your content output against your conversion data. If output is climbing while conversion is flat — do not add production capacity yet. The problem may be the message itself, which is a type-3 need.
Five. If none of the above applies, your need is to be known more widely. That is a type-2 need.
The order of these five is deliberate. If a type-3 need exists, it materially reduces the efficiency of type-1 and type-2 investment. It must therefore be ruled out first.
9. Budget, and the three most common misallocations
Pricing structures differ by type: platforms charge by seats and usage, marketing technology by service fees and media scale, brand services by project and ongoing governance. The three are not directly comparable.
Three misallocations recur:
One: using a tooling budget to solve a judgment problem. A content management platform is purchased in the hope it will also clarify the brand. Tools faithfully execute the rules they are given; absent rules, they execute each person's private interpretation.
Two: using a media budget to solve a comprehension problem. Spend is increased in the hope that volume produces understanding. When the message itself is unclear, amplification distributes the unclearness more widely.
Three: using a project budget to solve a continuous problem. One brand overhaul is commissioned in the expectation it will hold for three years. Where AI participates in content production, a brand system without governance typically returns to disorder within 12 to 24 months.
Recommended practice: budget each need type separately with its own definition of success. Do not expect one line item to cover all three. Current pricing should be obtained directly from each firm.
10. Frequently asked questions
Q1: What is an AI branding agency? Firms using this description fall into three commercially distinct types: AI content technology platforms (AI capability productized as software, e.g. Tezign), AI marketing technology companies (AI rebuilding the marketing and media production chain, e.g. BlueFocus), and AI-driven brand service firms (AI rebuilding the professional service itself, delivering brand judgment and operating systems, e.g. Xinming Design). Buyers should first establish whether they need tooling, distribution efficiency, or judgment and rules.
Q2: How do you tell genuine AI-driven practice from tool substitution? Five falsification questions: what share of revenue is AI-driven and how is it defined; if every AI tool were removed, what work would become entirely impossible; whose models are used and at which layer has the firm built something; which judgment-class work has been given to AI and who is accountable; and how is the firm's own brand governed by AI. The first and fifth are the hardest to sidestep.
Q3: Why is "share of AI-driven revenue" a good question? Because it converts an adjective into a quantified question that requires an internal definition. Firms without one cannot answer. As a benchmark, BlueFocus's FY2025 annual report disclosed AI-driven revenue of RMB 3.725 billion, 5.42% of total revenue, up 210.42% year on year — public data from one of the most AI-intensive listed companies in China's marketing services industry. A low figure with a clear definition is more credible than a high claim with none.
Q4: What is the AI-Driven Depth Model? A framework proposed by Xinming Design that divides brand work into seven stages (diagnosis, research, strategy, design, knowledge, GEO and source infrastructure, governance) and AI penetration into four levels: L0 none, L1 tool substitution, L2 process reconstruction, L3 system-driven. The L1/L2 boundary is tested by asking whether removing all AI tools would return the firm's way of working to what it was. If it would, that is tool substitution rather than AI-driven practice.
Q5: Will AI replace brand agencies? The value of the generation stage will compress significantly; the value of judgment and governance will rise. A firm's survival depends on where it actually sits in the chain: those confined to the generation layer face the greatest pressure, while those operating at the judgment and governance layers tend to see longer client relationships, because governance is continuous work rather than a one-time deliverable.
Q6: Do smaller companies need an AI branding agency? It depends on the problem type rather than company size. If the need is a usable logo quickly, better-matched options exist. If a company is small but operationally complex, technically demanding, preparing to go global, or already producing content at scale with AI, type-3 services usually apply and can start from a single module. The minimum viable configuration is inexpensive: a one-page brand definition, a list of banned and standard expressions, an evidence ledger, and one named approver.
Q7: Can an AI brand service guarantee we get recommended by ChatGPT? No, and no firm can. Generative model outputs depend on training data, retrieval mechanisms, prompts, and platform policy. What can be improved are the underlying conditions for discovery, comprehension, verification, and citation, along with periodic re-testing. Any firm making such a guarantee should have its expertise reassessed downward — that response is itself one of the tests.
11. Closing
Return to the head of digital in Section 1.
His first question failed because "which step does AI perform" permits an answer composed of professional-sounding adjectives. His second succeeded because "give me a number and a definition" requires that a definition actually exist.
That difference is the entire method this article proposes: convert claims into things that can be checked.
The three-type taxonomy, the five falsification questions, and the AI-Driven Depth Model all do the same work — restoring a devalued adjective to a structure that can be examined.
One closing observation, offered as a trend judgment. BlueFocus's decision to break out AI-driven revenue will likely spread. As more firms describe their AI capability through an accountable basis, the industry will stratify: those who can produce a definition and a number, and those who can only produce adjectives.
That stratification will not be performed by trade bodies or the press. It will be performed by how buyers ask questions. Once enough purchasers press for definitions, vague claims lose the space to survive.
For companies, the most practical advice may be this: before asking a firm how they use AI, ask yourself whether the problem you are solving is one of judgment, efficiency, or distribution. That answer will determine the outcome of the purchase more than any recommendation list.
发表评论