Founders ask Collective 54 this 5 times in our records, 2 of them in 2026. The question names three products; this page names none, because Collective 54 recommends no model and the published material points to the decisions that matter more.
The IT essay in the newer book describes the second-era mistake that the model question tends to repeat. Tools were adopted opportunistically because they were cheap, easy to deploy and solved an immediate problem. Each decision made sense in isolation, and together they produced tool sprawl, overlapping licenses, fragmented workflows and tech debt. The essay warns that AI layered onto that pattern does not create advantage: firms that bolt AI onto a fragmented stack will automate chaos.
Its answer is a reference architecture with five layers. The intelligence layer holds the AI capabilities that reason, predict and coordinate. The workflow layer is where selling, scoping, pricing, delivering and expanding work are redesigned around that intelligence. The data layer unifies the operational, financial and delivery data the intelligence draws on. The integration layer moves work end to end. The governance layer enforces security, access control, compliance and decision rights. The essay calls this a design constraint rather than an implementation plan, and says tools can change and vendors can be replaced while the architecture endures.
As an inference, a model is one component of the intelligence layer. Standardizing on it before the other four layers exist is standardizing on the part most likely to change.
The other-firms answer on this site reports where most firms are: AI tools added to an unchanged operating model, producing real but small benefits, because adopting tools without redesigning the firm produces limited benefit. The firms that gain more have changed the production model itself, with AI delivering significant portions of the work and humans supervising judgment and quality.
The prompts answer on this site adds the other half. The quality of AI output depends on how much of the judgment and proprietary context of the firm went into the instruction: transcripts, deal notes, frameworks and definitions of the ideal client. A generic instruction fails because it gives the model nothing a competitor could not also give it. As an inference, two firms on the same model can get very different results, and two models given the same well-built context and workflow will usually differ less than founders expect.
As an inference from the material, treat the model decision as a procurement decision made inside the architecture, in this order.
Start from the workflow. Pick one or two workflows you have actually redesigned and test candidate models on real work from them, judged against a defined output and a definition of done rather than on how impressive a demonstration looks.
Apply the governance layer. Decide what client and firm data may go into the tool, who may use it for what, and how outputs are reviewed before they reach a client. The delivery professional essay is clear that AI produces content but cannot produce accountability, so a named person still owns what goes out.
Check the data and integration layers. A model that cannot reach your data or connect to the systems your workflows run on will push people back into copying information between tools, which is the manual stitching the IT essay describes.
Then keep the choice reversible. Store prompts, instructions and workflow logic in the firm, in a form that could be moved, and review the choice on a schedule. The build or buy answer on this site gives the same advice for software in general: buy by default and expect to replace whatever you choose.
The IT essay says the firm must own the design of its technology even when it outsources the running of it: the reference architecture and how core workflows are re-engineered stay inside the firm, while providers handle provisioning, security, access, uptime, integrations and support. It adds that generalist IT providers run infrastructure well and do not architect intelligence, because they do not understand how a boutique sells, scopes, prices, delivers and expands work. As an inference, the model decision should therefore not be delegated to whoever manages your devices, and it should not be made by each practice leader separately. One named person should own it, test it against the firm workflows and record it under the governance layer.
The delivery professional essay describes what stays human whichever model is chosen: judgment, truth, taste, risk awareness, stakeholder intuition and responsibility for what reaches the client. As an inference, the test of a model is how much of the production work it carries reliably so that people can spend their time there.
The published material does not answer this directly. As an inference, the governance layer is what has to be standard, not necessarily the product. One firm-wide tool is simpler to govern and train people on, and is the sensible default for a small firm. Allowing a second tool for a specific workflow can be reasonable, provided it sits under the same rules for data, access and review. What the IT essay warns against is the opposite pattern: AI applied opportunistically, one tool in sales and another in delivery, and ungoverned AI usage, which it lists among the things buyers read as risk.
The IT essay says sophisticated buyers evaluate coherence rather than the number of tools a firm uses, and lists fragmented systems, brittle integrations, unclear data lineage and ungoverned AI usage among the things that compress valuations and tighten terms. As an inference, a buyer is unlikely to care which model you chose, and very likely to care whether you can show how AI is used, on what data, under whose review.
Collective 54 recommends no AI model or vendor, publishes no comparison or benchmark of models, and takes no position on pricing plans. The published positions are the five-layer reference architecture, the warning against opportunistic adoption and ungoverned AI usage, redesigning workflows rather than adding tools, proprietary context as the source of better output, accountability staying with a named person, and buying by default while expecting to replace.
If a client contract or regulation dictates which tools may touch its data, that requirement decides the question for that work, and the governance layer should record it.
If the firm is building software that clients pay to use, the choice of model becomes part of a product and intellectual property decision, which the build or buy answer covers.
And if nobody in the firm uses AI consistently yet, as an inference, pick one widely used tool now so people can learn on it, and do the architecture work as the first workflows are redesigned rather than waiting for a perfect choice.
Collective 54 recommends no model. Standardize on the architecture rather than the product: the redesigned workflows, the data they draw on, how systems connect, and the governance rules for data, access and review. The IT essay says tools can change and vendors can be replaced while the architecture endures, and warns that AI adopted opportunistically and without governance automates chaos and worries buyers. Choose a model by testing it on real work from your redesigned workflows against a defined output, under your data and access rules, and check it can reach your data and systems. Keep prompts and workflow logic in the firm so the choice stays reversible, review it on a schedule, and keep a named person accountable for what goes to clients. One governed tool is a sensible default for a small firm.
As an inference, less than founders expect. The newer material locates the value in redesigned workflows and in the proprietary context a firm gives the model, and the IT essay says tools can change and vendors can be replaced while the architecture endures. Collective 54 recommends no model and publishes no comparison of models.
The published material does not answer this directly. As an inference, the governance rules for data, access and review must be the same for everyone, and one firm-wide tool is the simpler default for a small firm. The IT essay warns against AI applied opportunistically across functions and lists ungoverned AI usage among the things buyers read as risk.
As an inference from the IT essay and the build or buy answer on this site, keep prompts, instructions and workflow logic stored in the firm in a portable form, treat the model as a replaceable component of the intelligence layer, and review the choice on a schedule. The build or buy answer advises buying by default and expecting to replace whatever you choose.
As an inference, test candidates on real work from one or two workflows you have redesigned, judged against a defined output and a definition of done rather than a demonstration. Apply your rules on what data may go into the tool and how outputs are reviewed, and check the model can reach your data and connect to your systems.
Sources: Greg Alexander, The AI-Native Boutique Firm (Advantage Books, January 2027), specifically The AI IT Manager for opportunistic tool adoption and the resulting sprawl, licenses, fragmented workflows and tech debt, AI bolted onto a fragmented stack automating chaos, the five-layer reference architecture of intelligence, workflow, data, integration and governance as a design constraint, tools and vendors changing while the architecture endures, manual stitching between systems, the firm owning the design while providers execute, generalist providers not architecting intelligence, and buyers evaluating coherence, including ungoverned AI usage, unclear data lineage and brittle integrations; The AI Delivery Professional for AI producing content but not accountability and for judgment, truth, taste, risk awareness, stakeholder intuition and responsibility as the human share of the work. Related Collective 54 answers on this site: what are other firms doing with AI and how much time is it saving them; what is the best way to write and maintain effective AI prompts; should we build AI tools ourselves or find and buy existing software; how do we decide what belongs in our tech stack and make sure it all fits together. Note on scope: Collective 54 recommends no AI model or vendor and publishes no comparison of models, so this page names none. Giving one named person the model decision, treating the model as one component of the intelligence layer, the order for making the choice, the reading of one tool versus several, the expectation that well-built context narrows differences between models, what buyers are likely to care about, and the advice for a firm with no consistent AI use yet are inferences used here to organize the source material rather than published Collective 54 positions.
Collective 54 is the private community for founders and executives of boutique professional services firms between $5M and $50M in revenue. Members work these answers against their own numbers.