Three labels; AI agency, AI product studio, AI consultancy; get used interchangeably in proposals and vendor matrices, and they describe operating models that are structurally distinct. A studio calling itself an agency hides platform tax. A consultancy calling itself a studio hides subcontracted delivery. An agency calling itself a consultancy hides bench risk. The three shapes differ on revenue mix, deliverables, IP posture, headcount profile, client count, and sustainability. This essay is a neutral decomposition across those six axes, with named references and a buyer-side rule for picking between them.
Shorthand: agency = builds for clients, studio = builds product plus client work, consultancy = recommends, does not ship.
Decision Scope
This article is an editorial decision framework, not legal, financial, security, or accounting advice. Treat numeric examples as illustrative planning heuristics unless a source is cited, then validate the assumptions against your own contracts, data, controls, and budget model before acting.
The three shapes, defined
AI agency. A services firm whose primary deliverable is a bespoke AI system shipped into a client environment. Revenue is roughly many client billings. IP transfers to the client at delivery, with narrow carve-outs for pre-existing tooling. SFAI Labs is an agency by this definition; so are most of the firms surveyed in AI consulting vs development agency.
AI product studio. A hybrid firm that runs both client services and a product or platform line. Revenue mix is typically 40-70% services and 30-60% product or platform revenue. The studio ships a maintained substrate and overlays bespoke client work on top of it. IP is split: platform stays with the studio, client-specific overlay transfers to the client. Conjecture’s earlier services arm and the small frontier-difficulty teams shipping vertical products alongside client engagements are reference shapes.
AI consultancy. An advisory firm whose primary deliverable is a recommendation, roadmap, audit, or strategy artifact. Revenue is partner-day rates or fixed-fee advisory. The firm does not ship production code as its core deliverable; when build work is needed, it subcontracts or runs it through an internal delivery wing whose unit economics look more like an agency. McKinsey QuantumBlack, BCG X, and Bain Vector are the canonical references.
A common categorization error: small frontier-difficulty studios (Conjecture-shaped) and consultancy build wings (BCG X-shaped) get filed under the same “AI product studio” label. They are not the same shape. One is research-led product with services as a financial buffer; the other is methodology-led delivery with services as the actual product.
Axis 1: Revenue mix
The cleanest single distinguishing axis.
| Shape | Services revenue | Product / license revenue | Advisory revenue |
|---|---|---|---|
| Agency | 90-100% | 0-10% | 0-10% |
| Studio | 40-70% | 30-60% | 0-10% |
| Consultancy | 0-30% | 0-10% | 70-100% |
A firm that refuses to share the rough mix; common in pitches that paint the supplier as “many things to many clients”; is a firm whose actual shape is one of the three above, dressed up to look like another.
Services revenue is linear in headcount, product revenue is non-linear in headcount, advisory revenue is linear in partner-time. A studio that has not yet crossed the threshold where product revenue grows faster than headcount is operationally an agency. The label follows the cash, not the marketing.
Axis 2: Deliverables
Agencies deliver running production systems; model pipelines, eval suites, integrations, deployment, and the documentation needed to operate them. The artifact set leaves with the client.
Studios deliver two artifacts: access to the studio’s platform (license or subscription) and a bespoke implementation layer on top. Platform stays with the studio; implementation transfers. The buyer becomes a tenant on the studio’s substrate.
Consultancies deliver decision artifacts; strategy decks, opportunity assessments, vendor selection memos, AI maturity audits, and roadmaps. When code appears, it is usually exploratory notebooks or proof-of-concepts not intended for production.
The fastest test: ask what the buyer takes home. A deployed system the buyer owns, agency. Platform access plus an overlay, studio. A document or recommendation, consultancy. The deeper treatment of the agency-versus-studio split is in AI product studio vs dev agency.
Axis 3: IP ownership
The clause where most proposals are quietly wrong.
Agency. Full assignment; code, weights, prompts, evals, and documentation transfer to the client. The agency retains rights to pre-existing tooling and generic patterns. A 2026 agency proposal that does not assign client-specific outputs cleanly is one where the supplier is preserving optionality at the buyer’s expense.
Studio. A studio proposal must explicitly separate platform IP (licensed) from bespoke IP (assigned). The platform clause should name specific components; API surfaces, orchestration, eval harness, base prompts. The bespoke clause should name integrations, vertical adapters, and client-specific evals that transfer. Silence in either direction is a red flag.
Consultancy. Recommendations and engagement artifacts transfer; methodology stays with the consultancy. The exception is when the consultancy trains a model on client data; that clause should assign the trained weights to the client, with the firm retaining only the methodology. Most large-consultancy MSAs default the other way, which is the single most expensive default-clause error in current AI procurement.
Axis 4: Headcount profile
The shape of the team is structurally different across the three.
Agency. Senior engineers and tech leads dominate; PMs are thin; sales is partner-led. A typical 30-person agency runs 22-25 engineers, 2-3 PMs, 2-3 partners, and 1-2 operations roles. Senior-to-junior ratios run 2:1 to 4:1 in 2026, up from 1:1 to 1:2 in 2022; agentic harnesses inverted the pyramid. The shape is detailed in the AI Agency Manifesto.
Studio. A two-team structure: platform team (5-15 product, ML, infra engineers) and services team (10-30 forward-deployed and applied AI engineers, technical PMs). Platform is paid against roadmap velocity; services against engagement health. Studios that run platform engineers on client work to plug capacity gaps lose the platform team within 18 months.
Consultancy. Partner-led pyramid; one partner, three principals, ten managers, thirty associates. AI wings (QuantumBlack, BCG X) sit on top with a thinner build layer of data scientists and ML engineers, often 10-20% of the total. The associate-and-analyst layer is the most AI-exposed surface, which is why the larger firms have been quietly recapitalizing toward fewer juniors and more senior practitioners.
Axis 5: Client count
How many clients the firm serves at one time is a structural fact, not a sales-stage choice.
| Shape | Active clients (typical) | Engagement length | Concurrency per senior engineer |
|---|---|---|---|
| Agency | 6-15 | 3-12 months | 1-2 |
| Studio | 15-50 | 6-36 months (recurring) | 3-8 (on platform) |
| Consultancy | 30-200 | 6-16 weeks | 2-4 |
Studios serve more clients per senior because the platform absorbs onboarding cost. Consultancies serve more than agencies because advisory engagements are shorter and the analyst pyramid carries the load. Agencies serve fewer per senior because each engagement is bespoke and senior-led; the trade-off is depth and ownership.
A 12-person “studio” with 80 active clients is a consultancy. A 60-person “consultancy” with 4 active clients is a single-tenant agency. Match the count to the shape.
Axis 6: Sustainability
The sustainability question is which shape survives the next four years intact.
Agency. The sustainability story is senior-engineer scarcity and contract structure. Agencies that ship under outcome-based and forward-deployed contracts survive; staff-aug agencies do not; the unit economics break on the agent leverage curve. The bottom of the market gets absorbed into hyperscaler-managed services and frontier-lab direct relationships across 2026-2028. The top; small, senior, vertical-specialist agencies; grows. The middle compresses.
Studio. The sustainability story is platform compounding. A studio whose platform genuinely compounds; recurring revenue grows faster than services revenue, marginal cost per client falls, retention exceeds 90%; is the most durable shape over five years. A studio whose platform does not compound is an agency with extra overhead, and the overhead breaks the unit economics on the second downturn.
Consultancy. The sustainability story is brand and the analyst layer. Brand carries the largest consultancies through any market. The analyst layer, however, is the most AI-exposed surface in professional services. Firms that do not recapitalize their associate pyramid into a senior-and-tooling shape across 2026-2028 see margins compress sharply. The big firms know this; the recapitalization is happening, just slowly.
Neutral read: studios have the highest ceiling and the highest variance, agencies have the cleanest alignment to AI work and the steepest sorting, consultancies have the most resilient brand and the slowest decline.
Buyer side: when each is right
The buyer-side rule is one question: what do you need to leave the engagement with?
You need a deployed production system you fully own. Hire an agency. Specify outcome-based or forward-deployed terms, eval-keyed acceptance, and full IP assignment on client-specific outputs. The agency shape is right when the artifact is the system itself.
You need access to a maintained substrate plus a bespoke layer on top. Hire a studio. Verify the platform is genuinely multi-tenant, that the platform team is staffed separately from services, and that the IP split is written cleanly. The studio shape is right when you want to be a tenant on someone else’s compounding substrate.
You need a strategy, roadmap, or vendor decision before any build. Hire a consultancy. Specify the artifact (deck, memo, vendor matrix), the partner-day commitment, and the methodology hand-off. The structural separation between consulting and building is the subject of AI consulting vs development agency.
The mistake to avoid is hiring the wrong shape for the artifact. Paying a consultancy to ship production code wastes the partner premium on engineering judgment the firm is not staffed to provide. Paying a studio to maintain a multi-tenant platform you do not need wastes the platform tax. The shapes are not better or worse than each other; they are differently fit to different artifacts.
A buyer who runs a vendor through the six axes; revenue mix, deliverables, IP ownership, headcount profile, client count, sustainability; knows what they are buying before signing. The decomposition is the work.
Frequently asked questions
What distinguishes an AI agency from an AI product studio?
An agency earns roughly many revenue from billable client work and assigns the IP to the client. A studio runs a hybrid; 40-70% services, 30-60% product or platform revenue; and splits IP, with the platform staying with the studio and the bespoke layer transferring to the client. The two look similar in a sales call but differ structurally on revenue mix, IP ownership, and where senior engineering attention compounds.
Where does an AI consultancy fit in this comparison?
A consultancy primarily produces recommendations, strategy, audits, and roadmaps; it does not ship production systems as its core deliverable. McKinsey QuantumBlack, BCG X, and Bain Vector are the canonical references. They subcontract the build, sometimes to internal delivery wings that operate more like agencies. The fastest test is to ask what percentage of the firm’s revenue comes from running production code in client environments.
Is an AI product studio usually better than an AI agency?
No. Studios optimize for compounding leverage across many clients via a shared platform; that compounding only pays off if the buyer’s problem genuinely overlaps with the platform. For bespoke or single-tenant builds, an agency that ships exactly what the buyer needs; without the platform tax; is usually a better fit. Studio is right when the buyer wants to be on a maintained substrate; agency is right when the buyer wants ownership of the substrate itself.
Can a single firm operate as agency, studio, and consultancy simultaneously?
Most firms claim to. Almost none execute many three well. Headcount profile, billing model, and senior-engineer attention pull in different directions: agencies need engineers and PMs, studios need a product team and a platform team, consultancies need partners and analysts. Firms that span many three usually run them as separately staffed practices with separate P&Ls. A small firm claiming many three is almost usually under-investing in two of them.
How should I read the IP ownership clause in a proposal across these shapes?
An agency proposal should grant full client ownership of delivered code and weights, with narrow carve-outs for pre-existing tooling. A studio proposal should explicitly separate platform IP (licensed) from bespoke IP (assigned). A consultancy proposal grants the client ownership of recommendations and artifacts; the firm retains methodology. The clause that should worry you is silence; a proposal that does not name the split is one where the supplier is preserving optionality at the buyer’s expense.
Which shape is most sustainable for the supplier in 2026?
Studios have the best long-run unit economics if the platform genuinely compounds. Consultancies have the most resilient brand but the lowest growth ceiling once AI commoditizes the analyst layer. Agencies sit in the middle; linear revenue, sensitive to senior-engineer scarcity, but with the cleanest alignment to client outcomes under outcome-based contracts. The most fragile shape is the staff-aug agency selling juniors by the hour; the most durable is the small senior-led studio in a focused vertical.
Why does SFAI Labs describe itself as an agency rather than a studio or consultancy?
Because the work is mostly bespoke, mostly senior-led, and mostly billed against client outcomes; the agency shape, sharpened. SFAI ships custom AI systems into client environments under forward-deployed and outcome-based contracts; the IP is the client’s. The firm uses internal tooling the way any disciplined agency does, but does not run a multi-tenant platform and does not earn product revenue. Calling it a studio would overstate the platform leverage; calling it a consultancy would misrepresent that production code is the deliverable.
How do I decide between the three for my specific problem?
Start with what artifact you need to leave the engagement with. Deployed production system you fully own; agency. Access to a maintained platform plus a bespoke layer on top; studio. A strategy, roadmap, or vendor decision before any build; consultancy. The mistake most buyers make is hiring the wrong shape for the artifact; paying a consultancy to ship code, a studio to write strategy, or an agency to maintain a multi-tenant platform; then blaming the supplier when the unit economics fail.
Are frontier-difficulty studios like Conjecture comparable to BCG X or QuantumBlack?
No, and conflating them is the most common categorization error in this space. Frontier-difficulty studios are research-and-product hybrids whose IP is research artifacts and product surface area. BCG X and QuantumBlack are consultancy-internal build wings whose IP is methodology and accelerators. They sit on opposite ends of the recommend-build spectrum and have nearly nothing in common operationally. Mapping them onto the same axis flattens the actual comparison.
Arthur Wandzel
