Stop Selling AI Trinkets: The Five Postures of an AI-Native Consultancy

Consultancies fail at AI not for lack of ideas but for selling AI through the old model. Here is the ladder, the foundry, and the organisational move that actually produces new AI revenue.

Scott Farrell · LeverageAI · August 2026

I keep watching the same workshop. Partners gather the practice leads. Someone asks, in good faith, for AI ideas. The room produces copilots: a chatbot for the service desk, a summariser for workshops, a governance advisory deck, a “data quality assistant” bolted onto last year’s delivery pattern. The list is long. The commercial unit is unchanged. Three months later the firm still cannot point to new AI revenue that is not just the old engagement with an AI slide in the appendix.

That pattern is not a failure of imagination in the people. It is a failure of organisational design. The firm is trying to sell AI through the existing consulting model. The larger opportunity is to rebuild consulting through AI — and that requires a different organ than the one that sells days.

They are asking how to sell AI through consulting. The work that matters is showing them how to rebuild consulting through AI.

This piece is a firm diagnostic and an organisational design — a design, not a surveyed industry standard. The specimen I will use is real and limited: the Data Readiness Review built on FDE BI, sold and delivered as a fixed-price evidence product. Treat that as n=1. Where I make economic claims about utilisation P&Ls and product invention, treat them as hypotheses you can check against your own proposal cost, win rate, gross margin by offer type, and reuse across engagements — not as a verdict about any named firm.

The five postures — where most firms actually sit

Here is the ladder I use when a mid-sized data consultancy with a large installed client base asks what is going wrong with its AI programme. The postures are cumulative. Each rung can look busy. Only the top two change the commercial catalogue in a way that cheap cognition cannot casually erase.

Most firms I meet hover in the first three. That is not a moral failure. Advisory is familiar. Trinkets are fundable. AI-enabled delivery protects utilisation while making production cheaper — exactly what a labour-priced P&L is selected to do. The diagnostic is sharper than “are you doing AI?” It is: which rung are you actually on, and what would have to be true for the next one?

Rung 1 — AI advisory

Standalone AI advisory is still sold because clients ask for it and partners know how to staff it. The product is thinking: workshops, frameworks, policy language, board packs. The problem is not that judgment is worthless. The problem is that first-pass advice and polished narrative are becoming cheap. A field experiment with BCG consultants found that on tasks inside AI’s capability frontier, people using GPT-4 completed 12.2% more tasks and finished them 25.1% faster — with higher quality — than those without AI access.¹ That is good news for productivity. It is bad news for firms whose primary AI product is “we will think about AI for you” billed as senior days. Advice got cheap. Verification did not. When the deliverable is narrative without an evidence chain into implementation, the client can increasingly get a first draft elsewhere.

Advisory does not disappear on the higher rungs. It moves. Frameworks shape what the system notices. Judgment selects outcomes. Governance gates what may proceed. The senior person stops being the factory that produces the deck and becomes the control plane inside a product that carries evidence and a path to action. That is the preservation of what was always valuable — trust, accountability, domain context — without defending expensive prose as the main commercial unit.

Rung 2 — AI trinkets

Trinkets are the workshop’s natural output. Put AI on the current process. Show a chatbot. Attach a recommender. Ship a feature that demos well and leaves the statement of work, the pricing model and the client obligation unchanged. McKinsey’s 2025 State of AI survey still finds that nearly two-thirds of organisations have not begun scaling AI across the enterprise — most remain in experimentation or piloting.² Trinkets fit that regime perfectly: visible, marketable, easy to copy, easy to abandon.

The same HBS/BCG study that showed within-frontier gains also showed the jagged edge: on a complex managerial task deliberately placed outside the frontier, subjects with AI were 19% less likely to produce correct solutions.¹ Trinkets ignore the jagged edge. They assume “add AI” is uniformly good. An AI-native posture assumes the opposite: placement is the product. Some work belongs to models, some to deterministic code, some only to accountable humans — and the commercial promise is engineered around that placement, not around a widget.

Rung 3 — AI-enabled consulting

This is where serious firms often stop and congratulate themselves. Delivery staff use models. Proposals draft faster. Code and analysis accelerate. Margin improves on the same broad, bespoke offers. That is real value. It is also still the old game played faster.

In Great Reset language: an AI feature makes the old game slightly better; an AI factory gives the organisation the capacity to play a different game. AI-enabled consulting is the feature inside professional services. It does not create a new line item a buyer can order. It does not change what is economically impossible. It compresses labour inside a commercial structure still built for selling teams and days. When clients start expecting that compression as the baseline price, the firm has trained the market to pay less for the same shape of work.

There is a second trap inside rung three. Because AI-enabled delivery improves utilisation optics — fewer hours for the same SOW, happier internal dashboards — partners can sincerely believe the firm is “AI mature.” The scorecard says yes. The catalogue still sells the same broad, bespoke uncertainty. Terminal value has not migrated. The firm is faster at producing artefacts that are becoming less scarce. That is a good operational year and a dangerous strategic decade if nothing else changes.

Rung 4 — AI-constituted services

Here the commercial object changes. The test is simple and falsifiable — it is owned in full by the AI-constituted services taxonomy, so I will not re-derive it: remove the AI. Does the same offer still exist, only more slowly? Then it is AI-enabled. Does the fixed price, coverage, specificity or evidence promise collapse? Then the service is AI-constituted.

The Data Readiness Review is the specimen I have actually built and can stand behind as n=1. The client does not buy “an AI spreadsheet-analysis experience.” The client buys certainty about a data estate: observed evidence, mapped gaps, human disposition of findings, and an evidence-backed scope for what to do next — at a fixed price that would have been reckless under pure human sampling economics. The AI is often invisible to the buyer and indispensable to the producer. That is constitutive embedding, not surface embedding. The commercial protocol that sells certainty first, before the implementation SOW, is the sibling argument. The biography of how the deck became software is another.

One offer does not make a firm AI-native. It makes rung four real. The strategic error is to treat the first product as the destination and try to copy its screens five times by hand.

Rung 5 — AI-native consultancy

An AI-native consultancy is not a consulting firm with a lot of copilots. It is a firm whose product catalogue, delivery model and learning system are designed around cheap cognition. Its question is not “what AI can we add to our projects?” It is:

What valuable client outcome can we now package, price and deliver that no traditional consulting operation could reliably offer before?

That question cannot be answered once. It has to be answered on a cadence — which is why the organisational move is not “hire three AI people into the existing service lines.” It is to install an offer foundry.

The search object is not an “AI use case”

If you inventory current processes and ask where AI can help, you will get rungs one through three. The hunting ground for rungs four and five is different: economically suppressed services — valuable work absent from the catalogue not because clients would not pay, but because cognition, customisation, uncertainty or coordination cost made them commercially irrational under human labour economics.

That is the service-market face of a cognition scarcity audit: fund the analysis you never do, not only the process you already staff.

Examples of suppressed services in a data consultancy’s world include: an assessment that would require reading hundreds of workbooks; line-by-line reconciliation of requirements against an existing estate; continuous assurance over documents, policies and actual evidence; a defensible implementation scope before a major commitment; a tailored decision package for every significant account rather than the top three. These are missing because they were impossible to price and deliver repeatedly — not because they lacked value.

Take the bespoke statement of work as a worked surface of the same problem — not because SOW automation is the product, but because it makes the economics of suppressed cognition concrete. In many firms the SOW process consumes expensive senior attention, is usually unpaid, recurs constantly, reuses prior proposals poorly, varies wildly by author, and is written often for deals that never convert. At a low win rate the winning job absorbs the pre-sales effort of many failed attempts; then the won project can lose more money through omitted assumptions and change. The firm calls this business development. Economically, a large portion is unpriced cognition spent reconstructing the same commercial object from scratch.

The senior SOW author is the join algorithm: client request, account history, firm capability, available people, prior architectures, commercial risk and personal scar tissue, manually assembled into a number. One person lands at two hundred thousand dollars; another, same brief, six hundred. There is no explicit substrate against which either answer can be tested. AI will not primarily attack this by writing proposal prose faster. It will attack it by allowing an alternative firm to retain engagement knowledge, measure the estate before pricing, expose assumptions, reuse tested work packages and learn from quoted, won, lost and delivered work. The native output of that system is not a prettier SOW. It is a defensible commercial decision — from which the SOW becomes a generated receipt. That is why the Data Readiness shape matters: observed evidence → declared requirement → mapped gap → explicit questions → human decisions → bounded work packages. The suppressed service was “certainty before commitment.” It was not “write RFPs faster.”

AI makes the missing product economically reachable. That is the whole point of climbing past trinkets.

You cannot ask the horse division for the car

Here is the structural trap, stated as a hypothesis you can test inside your own firm rather than as an accusation about someone else’s.

Existing service lines are selected and rewarded to make the current model work: sell consulting days, staff projects, prepare bespoke statements of work, manage utilisation, protect account relationships, deliver the agreed technology, absorb enough ambiguity to keep the client happy. Asking those leaders to invent the model that reduces custom scoping, compresses billable labour and changes what consultants sell is like asking the horse division to design the car. They are not stupid. They are doing the job the organisation designed.

You cannot ask the AI-advisory horse division to invent the product that makes standalone AI advisory less central.

That sentence is Terminal Value’s Self-Disintermediation Doctrine applied to professional services: if an AI-native competitor could destroy part of your business, build that competitor inside your own company first — as harvest, migrate and construct running in parallel, not as a slogan on a strategy offsite.

Harvest the cash-generating advisory, implementation and managed-service book. Do not vandalise today’s revenue. Be honest internally that this is the model that still funds the firm while value migrates.

Migrate the durable assets — client relationships, delivery history, domain expertise, consultant judgment, methodology, reusable code, governance knowledge, account context — into firm memory, patterns, sensors and evaluation machinery rather than leaving them trapped in people’s heads and last year’s proposals.

Construct a separate but connected portfolio of AI-constituted service products. Separate enough that its metrics are not utilisation of the old pyramid. Connected enough that it can launch through the installed base the firm already has.

The foundry is the construct motion given a name, a budget and a loop. Without structural independence, every proposal for a labour-compressing product loses to the next billable project. Without installed-base access, the foundry invents products nobody can sell. Both failures are organisational, not technical.

The organisational move: an AI-native offer foundry

The Practice OS already describes how a consultancy scales scarce forward-deployed judgment: a capability kernel, firm and client kernels, a client-contained vessel, field operators with recognition-and-routing authority, escalation that leaves fossils, and engagement two as the proof that transfer happened. This piece does not re-teach that machinery. It extends it with the organ that invents the next commercial unit: the offer foundry.

The foundry’s job is not generic AI advisory and not building whatever widget a client requested last Tuesday. Its job is to repeatedly produce commercial offers. The loop:

I organise that foundry around six functions. Walk them as mechanisms, not as a slide title.

1. Sense

The installed base is the highest-value dataset a mature consultancy owns. It already knows — for many accounts — the systems estate, sponsors, delivery history, governance friction, stranded projects, architecture constraints, and where the relationship is losing energy. Business development here is matching, not brainstorming.

Sensing means using that knowledge deliberately. Delivery staff, proposals, projects and account history surface: repeated client friction; unpaid or weakly paid consulting labour; difficult decisions that recur; context reconstructed from scratch every time; sampling caused by human limits; services clients ask for that the firm cannot safely price. The output of sense is not a backlog of forty-seven AI ideas. It is a short list of friction nominations with evidence — “we rebuild this SOW shape twelve times a year and lose margin on half the wins,” not “maybe a chatbot for finance.”

Operationally, sense looks like a standing cadence, not a brainstorm. Account leads submit signal cards: what friction they saw, which buyer felt it, how often it recurs, what the firm currently does instead, and why pricing a better answer has failed. Delivery retrospectives tag unpaid senior hours by cause class — reconstruction, reconciliation, exception hunting, political translation — rather than only by project code. Lost proposals are sampled for the same classes: did we lose on price, or because we could not make the unknown known cheaply enough to commit? The foundry ranks nominations by recurrence × economic pain × evidence availability. Cold “AI use case” workshops skip all of that and invent features detached from the firm’s actual friction rent.

A mid-sized data consultancy with a large installed client base has an advantage startups do not: account access and trust. Sensing without that base is cold ideation. Sensing with it is pattern recognition over lived delivery. The installed base is not only a sales channel for launch (function five). It is the sensing instrument itself.

2. Select

Selection is where most innovation programmes fail by refusing to choose. Candidates pass through filters: horse versus car (is this optimising the old process or constructing a new economic unit?); terminal-value effect (does this matter when generic analysis is cheap?); friction removed (client and firm); willingness to buy (named buyer, budget shape); latency and governance fit; availability of evidence; potential for repeated delivery without heroic staff.

A useful selection board forces a single paragraph commercial hypothesis before any architecture discussion: who buys, what decision they make, what they stop doing, what the firm stops inventing per engagement, and what would falsify the hypothesis in ninety days. If the paragraph cannot be written, the candidate is not ready for build. If twenty candidates all pass because the bar is “sounds AI-ish,” the bar is wrong. Prefer one constituted offer with a named buyer over five trinkets with enthusiastic sponsors.

The output should be one selected product candidate for the next cycle — not a portfolio theatre of half-funded experiments. Kill, fix or double-down is the discipline; celebration of every pilot is the anti-discipline. The cognition scarcity audit’s portfolio test is exact: tolerate horse optimisation as operations spend; fund car construction as strategy.

3. Compose the commercial transaction

Before anyone builds software, define the purchase. Buyer. Decision the product enables. Fixed price or measurable pricing band. Timebox. Eligibility. Evidence inputs. Client-owned outputs. Exclusions. Next-step options. A legitimate stop outcome — including “build nothing” as a successful paid result when the evidence says so.

This is the inversion most technical teams resist: the commercial product comes before the AI architecture. If you cannot name the decision the client is buying and the stop states the engagement can land in, you do not yet have a product. You have a project shape waiting to leak margin. The certainty-first protocol is the worked form of this for readiness-style offers.

Composition also forces pricing honesty. A fixed price without measured complexity is either reckless or padded. Volume bands, human disposition load and typed-surprise reserves are design choices in the commercial unit, not afterthoughts when the first engagement overruns. Exclusions are not legal boilerplate; they are the product boundary that keeps the promise deliverable by ordinary staff. Next-step options are not upsell theatre; they are the natural outputs of an evidence process — including the option to stop. If “build nothing” cannot be a successful paid outcome, you have not sold certainty. You have sold a wedge into implementation regardless of evidence.

4. Build the delivery machine

Only then determine what deterministic sensors inspect; what safe representation the model receives; what institutional knowledge grounds judgment; what AI proposes; what humans approve; what the interface must expose; what code compiles into the final artefact. The architecture that makes AI-constituted delivery governable — sensors, models, human disposition, privilege separation — is sibling territory. The foundry’s job is to insist that build serves the commercial unit already composed, not the other way around.

FDE BI is useful here as a generative pattern, not a Power BI template: evidence in, comparison and uncertainty surfaced, human decisions required, scope compiled from accepted findings. Copying the screens without the commercial composition and the disposition discipline produces a trinket with a nicer UI.

Build also decides what must never be automated. Accountable disposition of material mappings, acceptance of scope that binds the firm, and exceptions that create new claim classes stay human. The machine prepares and proposes; people own consequence. That placement is why the product can carry a fixed-price promise without pretending the model is the partner. When build starts from “what can the model do?” instead of “what must the commercial unit guarantee?”, you get impressive demos and uninsurable commitments.

5. Launch through the installed base

Cold-selling a new AI product to strangers is the hard path. The foundry launches through accounts with an obvious fit — where the firm already knows the estate, the sponsors and the friction. Delivery consultants act as account sensors: they recognise and route. They do not invent claims, prices or unusual commitments. The central foundry controls qualification, offer definitions, pricing envelope and exceptions.

That sales membrane is load-bearing. Without it you get two hundred consultants making improvised AI promises. With it you get recognition at the edge and product integrity at the centre — the same recognition-and-routing authority Practice OS gives field operators, applied to offers rather than only to escalations.

Train the membrane as a skill, not a memo. Delivery consultants learn eligibility criteria, the one-sentence promise, the exclusions, and the escalation path for “almost fits.” They do not learn to freestyle price or invent features in the room. When a sponsor asks for a customisation that would break the product boundary, the consultant routes rather than negotiates a special. That feels slower in the moment. It is how the offer remains a product instead of re-becoming a bespoke project with an AI label.

Internal deployment work already says: measure the funnel, not the vibes — recognition, qualification, conversion, delivery, reusable learning. Run the productised lane as a separate commercial lane from legacy RFQ statistics. Blending them recreates the red-bead mistake: you rank people on a mixture of unrelated processes and call it sales performance.

In the legacy RFQ lane you still track unpaid sales hours, time to submission, win rate, discounting, forecast versus actual margin, estimate variance and change-request frequency — with uncertainty acknowledged. In the productised readiness lane you track eligible accounts identified, percentage offered the product, conversation-to-purchase conversion, days from offer to decision, salesperson and technical hours before purchase, price integrity, delivery cost and cycle time, consultant review effort, exception rate, review-to-build conversion, estimated versus actual downstream effort, downstream gross margin, and change requests attributable to missed readiness findings. The comparison is not which lane has the prettier close rate. It is which lane produces more total economic value per eligible opportunity, with less client effort, less speculative labour, better downstream margin and more reusable learning. That is company-building measurement. It only becomes possible when the commercial unit is stable enough to compare.

6. Compound

Every engagement produces a field-pattern record: what was invariant; what was client-specific; what worked; what failed; what required escalation; what should become configuration; what should remain local; what must be rejected. Recurrence nominates. Humans promote. Copying is not compounding. Compounding is when engagement two is easier because engagement one improved shared machinery — the engagement-two test applied to offers, not only to practice transfer.

This is how one product becomes a portfolio without degenerating into custom patches. An offer that looks reusable but creates excessive support or exception handling can be demoted. A client-specific solution stays client-specific. A recurring delivery tool may become an internal primitive rather than a saleable promise. That discipline is what prevents the foundry from turning into a workshop full of half-supported trinkets.

Compounding also clarifies the partnership problem many firms fear: “if we learn the first product, why would we need the capability that built it?” The healthy answer is not permanent dependence on a hero to run yesterday’s offer. It is that the firm becomes competent at delivering the current product while the foundry remains the fastest, safest way to evolve it and invent the next five. Screens reveal what was promoted. They do not automatically transfer the rejection history and evaluation function that made promotion meaningful. The durable business is the repeated verb — friction to product to transfer to next product — not the first noun on the menu.

Organise the portfolio by problem shape, not department

“Now build one for CFOs, one for sales and one for operations” is a useful way to imagine scale and a dangerous way to design products. Department labels invite solution-first thinking. The reusable unit is the problem shape. Same compiler, different fitted products:

Margin Integrity Review. A finance leader may buy a fixed-price evidence product that reconciles ERP structures, management spreadsheets, contract terms, margin assumptions, reporting gaps and unresolved finance decisions — not “AI for CFOs,” but a concrete question: what is currently knowable, unsupported, contradictory or missing before management relies on reported margin?

Account Growth Decision Pack. A sales or account leader may buy an evidence-backed next-opportunity package compiled from account history, prior projects, proposals, delivery outcomes, client communications, firm capability and current public signals. The client does not buy “AI-powered sales intelligence.” They buy a decision pack for a named account.

Operational Control Readiness. An operations or risk buyer may want policies, SOPs, incidents, evidence and actual execution reconciled into line-item findings for human disposition — a readiness review for control, not a chatbot for the policy wiki.

These are sketches of shape, not claims of a multi-product track record. The point is portfolio discipline: problem shapes that share a compiler amortise the foundry; department skins that each reinvent the stack do not.

Notice what is shared across the three sketches: a messy private reality is lowered into evidence; judgment is applied under known frameworks; humans dispose line items; the output is a decision-ready product with a legitimate stop. That is why problem-shape organisation beats department organisation. The foundry amortises the compiler. The market-facing name changes. The organs do not. When every practice builds its own AI widget stack, you get trinket sprawl and no learning system. When the foundry owns the compiler and practices own domain fit and account access, you get a portfolio.

Manage offers as products — the lifecycle

Innovation projects have kick-offs and demos. Products have stages and disposition. Each candidate offer should move through:

The specimen stage matters. A working specimen proves the method exists; it is not a prescription that every client needs this exact product. Discovery still decides fit. Showing a concrete object early raises the resolution of discovery — objections become evidence about constraints, not status fights about abstract strategy.

First client proves someone will pay. Second engagement with greater ordinary-staff leadership proves transfer. Supported offer means claim boundaries, pricing, training and support burden are known. Then scale — or repair — or demote. Demotion is success when it prevents a zombie product from consuming senior time forever.

Measure each offer, not just each project:

• client effort removed;
• time to paid entry;
• sales effort before purchase;
• delivery gross margin;
• machine versus senior-consultant effort;
• decision usefulness (did a human make a consequential call from the artefact?);
• downstream conversion and estimate accuracy;
• reuse without code forks;
• escalation frequency by class;
• whether engagement two was easier because engagement one improved shared machinery.

Those metrics make the portfolio learn. Utilisation of the old pyramid does not measure whether a new offer should live. If you only have utilisation, you will only invent what utilisation rewards.

Two measurement failures are worth naming because they feel rigorous while destroying the foundry. The first is ranking salespeople on small-sample product close rates without accounting for account mix — the red-bead error with a new dashboard. The second is declaring product success because engagement one was impressive when engagement one was still hero-led. Engagement one can be theatre with good people. Engagement two, with ordinary staff leading more because shared machinery improved, is the honest test. Without that transfer metric, every “product” is a project that happens to have a name.

Machine versus senior effort deserves special attention. If an offer still requires the same senior hours as the old discovery, you have not constituted a new service; you have rebranded consulting with software. Track the ratio over successive engagements. It should move. If it does not, demote or repair before you scale the promise into the installed base.

A self-assessment instrument

Before the Monday list, a short instrument you can run in a partners’ meeting without a consultant present. Answer honestly; the point is location, not branding.

1. Catalogue test. Open the public or private offer list. How many items are AI advisory, how many are features on existing services, how many are named products with fixed or banded price, eligibility and stop states? If you cannot find any of the last category, you are not on rung four.
2. Workshop test. Pull the last AI ideation session’s outputs. What fraction would still exist if the word “AI” were banned and you only allowed commercially new promises? If nearly all fail that filter, you are inventing trinkets.
3. Utilisation test. Who owns the P&L for inventing labour-compressing offers? If the answer is “the same leaders measured on utilisation of the current pyramid,” the horse division is still in charge of the car.
4. Membrane test. Can any delivery consultant invent an AI claim or price in front of a client? If yes, you have enthusiasm without a product system.
5. Transfer test. For your best AI-shaped engagement, was engagement two materially easier for ordinary staff because shared machinery improved — or did the same experts reappear under a new project code?
6. Lifecycle test. Is there a board where offers move through nomination, hypothesis, specimen, first client, transfer, support, and scale/repair/demote — with demotion allowed? Or only a graveyard of pilots and a showcase of demos?

Six honest answers locate the firm. They also tell you which foundry function is missing. Sensing without selection is a backlog. Selection without composition is a wish list. Composition without build is a deck. Build without installed-base launch is a lab. Launch without compounding is a one-off. Compounding without lifecycle discipline is trinket accumulation under a better name.

What this piece is not — and what to do Monday

This piece does not re-teach the Practice OS kernel, vessel and escalation design. It does not re-derive what an AI-constituted service is — that taxonomy is live. It does not settle vendor-side commercial terms for suppliers installing foundries inside consultancies; that is a later problem. Adjacent sensing networks, category-transition diagnostics and knowledge-base pre-mortems live in their own siblings.

What you can do with this design on Monday:

1. Locate the firm on the ladder. List the last ten AI initiatives. Classify each as advisory, trinket, enabled, constituted, or native-system. If everything sits in the first three, you have a discovery problem, not a model problem.
2. Name the foundry even if it is three people. Separate metrics from utilisation. Give it installed-base access and claim control. Without both, it will either starve or freestyle.
3. Pick one suppressed service, not ten use cases. Compose the commercial transaction before architecture. Build a specimen. Sell to a known account. Instrument engagement one and design engagement two as the transfer test.
4. Put every offer on a lifecycle board. Friction → hypothesis → specimen → first client → transfer → supported → scale/repair/demote. Kill zombies. Promote fossils. Refuse to score product success on the old pyramid’s utilisation alone.

The compact line is the one I would put in front of a partnership that is still hunting trinkets:

Stop selling AI as the subject of the engagement. Embed AI deeply enough that a better engagement can exist. FDE BI is one working example of that embedding — n=1, not a survey. The next move is not to copy that product five times. It is to install the organisational capability that repeatedly discovers and builds the next one.

The first product is the car. The durable business is the foundry that decides which car should exist next, builds it, proves it, and makes each subsequent one cheaper and safer to launch — using the firm’s own installed base before someone else uses cheap cognition to compress the labour model from outside.