Decision as a Service: The Architecture That Changes Everything
The insights industry has been building toward this business model without naming it. It is time to name it.
This post is the next installment in our ongoing structural deep-dive series, following The Future of Professional Services Is a Warning Shot for Market Research, The Data Infrastructure Layer Is Being Rebuilt, Synthetic Sample Is Not the Market. Decision-Grade Data Is., The Industry Is Being Redrawn. Here’s the Map., and Self-Serve Is Not a Category Anymore. It Is a Capability.
Where This Started
A few days ago I received a text from Jeff Krentz, a senior capital markets professional who has spent his career at the intersection of financial markets and the insights industry. The message was casual, but the idea was not:
“I don’t think it’s insane to consider a future ‘Decision as a Service.’ Would be interesting to discuss and debate.”
He was riffing off a conversation about Data as a Service — a model that has been gaining significant traction in our coverage. His instinct was that DaaS model was important but incomplete. The end state the industry is building toward is not delivering better data. It is delivering better decisions.
Jeff is right. And the convergence of three concepts — Domain-Specific Language Models, Service as Software, and Decision as a Service — describes an architecture that is already being built, at pieces and at scale, across multiple parts of the market. This post names that architecture, maps how the pieces connect, and identifies what it means for every participant in the value chain.
Credit where it belongs: the label “Decision as a Service” is Jeff’s. The architecture, as we’ll show, was already taking shape. He just gave it the right name.
Three Concepts. One Architecture.
Let’s start with the definitions, because precision matters here.
Domain-Specific Language Models (DSLMs) are AI models trained or fine-tuned on specialized industry data rather than general internet-scale text. Gartner named DSLMs one of its Top Strategic Technology Trends for 2026 and predicts DSLM and DSLM-underpinned application market revenue will reach $131 billion by 2035. By 2028, Gartner further predicts, most enterprise GenAI models will be domain-specific rather than general-purpose. The core insight: a general LLM asked to simulate consumer behavior is an educated guesser. A model trained on 60 years of consumer attitudinal data, 900,000 retail scanner locations, 250,000 household purchase histories, and 60,000 psychographic and media variables per US consumer is a calibrated expert operating from verified ground truth. Those are not the same thing, and they will never produce the same quality of decision.
Service as Software is the structural inversion of SaaS. Software as a Service sells you access to a tool that you then operate to produce an outcome. Service as Software deploys AI agents that autonomously plan, execute, and deliver the outcome itself — the service is the software. You do not buy the tool; you buy the result. As The New Stack describes it: “Service as Software is where AI agents deliver complete outcomes rather than simply the assistance that Software as a Service provides.” The pricing model shifts from subscriptions and seats to tasks completed and outcomes delivered. HFS Research projects this model growing into a $1.5 trillion market by 2035.
Decision as a Service is the logical terminus of this trajectory. Rather than providing data, insight, or even a recommendation, the provider delivers an optimized, confidence-scored decision — tested against synthetic market conditions before commitment, grounded in verified human data, and embedded directly in the client’s operational workflow. The output is not “here is what consumers think.” The output is “here is what you should do, and here is the confidence level behind it.”
The mash-up of these three concepts produces a single coherent architecture: a domain-specific AI system that autonomously operates as an expert service, with the output being not a report, not a recommendation, but a decision — delivered as infrastructure.
The Cisco Analogy Is the Right Frame
Jeff’s analogy was Cisco, and it is the right one.
Twenty years ago, Cisco sold networking pipes — hardware that moved data. The decisive value creation came when Cisco embedded intelligence into the pipes themselves: routing logic, security decisions, traffic prioritization, protocol optimization. The pipe became a thinking system. The infrastructure became the decision-maker.
The AI parallel is direct. NIQ Optiq, Palantir AIP, and the emerging decision orchestration layers in the insights industry are doing to enterprise decision-making what Cisco did to network routing — embedding intelligence into the infrastructure itself. The client no longer calls the expert to ask what to do. The infrastructure knows what to do, has tested the options in simulation, and surfaces a confidence-scored recommendation before the meeting starts.
Microsoft’s just-announced $2.5 billion Frontier Company — embedding engineers inside clients to build and run AI systems against proprietary internal data — is institutional confirmation of this architecture arriving at commercial scale. Palantir pioneered this model. Microsoft is industrializing it. The difference between this and traditional decision support is the same difference between Cisco’s dumb pipes and intelligent network infrastructure: the decision capability is no longer episodic and requested — it is continuous and embedded.
And there is one more consequence of recent market dynamics that opens the field: the gradual loosening of Amazon/OpenAI and Microsoft/OpenAI model exclusivity arrangements creates significant oxygen for application-layer and vertical decision-service specialists to build on top of foundation models without needing to own them. The infrastructure moment has arrived. The application layer is now open to players who could not have entered two years ago.
The Four-Layer Architecture
Here is how the three concepts map onto a complete Decision as a Service architecture. Each layer is distinct. Each is necessary. And — critically — the layers are more separable than they first appear, which matters enormously for understanding who can compete where.
The key insight from this architecture — one that changes the competitive landscape significantly — is that Layers 3 and 4 are more separable than they appear. Composable decision services — discrete, reusable, invocable decision logic units — mean vertical specialists can build at the ontology and delivery layer without owning the data foundation, provided they can access trusted data via API. This is the open competitive space the new foundation model accessibility has created. The firms in the transition cohort (Dig, Suzy, Material, Escalent) are de facto building at Layers 3 and 4. That is both their opening — the SME market — and their ceiling — bounded by the quality of purchased rather than owned data.
The Expert Network Question
Jeff floated expert networks as one possible architecture for Decision as a Service. It is worth addressing directly.
Expert networks (GLG, Tegus, AlphaSense) provide access to human expertise on demand. Decision as a Service is the inversion: the expertise is pre-trained into a domain-specific model, continuously available, and scalable without marginal cost per query. Expert networks are episodic (you access them when you know you need them, not continuously) and they do not compound (each engagement is discrete).
But expert networks are not irrelevant to this architecture. The right framing is: expert networks are the training source and governance layer for DSLMs in high-judgment domains. The human expert validates the model’s outputs, flags edge cases, and provides the calibration that transforms a sophisticated language model into a trustworthy decision infrastructure. This is what AlphaSense has done in financial intelligence — proprietary content, AI-native retrieval, workflow-embedded delivery, expert-validated governance. The expert network is not the service delivery layer; it is the continuous quality assurance and edge-case adjudication layer that keeps the Decision Ontology trustworthy as it compounds.
This Is a Business Model Revolution, Not a Product Feature
Here is where the architecture connects to money — and to the structural implication that the insights industry has not yet fully absorbed.
The $153 billion global insights industry (ESOMAR 2025) is priced almost entirely on project fees and subscription access. You pay for the process: the study, the analysis, the report. Decision as a Service changes the pricing unit entirely:
From: “We will conduct this study for $X” → To: “We will deliver this decision for $X, at Y confidence level”
From: Annual platform subscription → To: Per-decision or per-decision-cycle pricing
From: CMO/insights buyer → To: CEO/CFO buyer — because the accountability for the decision now attaches to the output
This is the commercial expression of what we have been describing throughout this series as the valuation gap. As we wrote in our investment thesis work: the same underlying human data asset commands 6–8x EBITDA as a point solution and 12–18x as an integrated Decision Intelligence platform. Decision as a Service is why. The unit of value is no longer data volume or survey throughput. It is position in the decision stack, and ultimately, accountability for the decision itself.
The firms that shift from selling data and research to selling decisions will reprice at fundamentally higher multiples. But they will also take on a fundamentally different accountability posture. A research firm that sells a report has always had a clear exit from responsibility: the client makes the decision, the firm provides inputs. Decision as a Service removes that insulation. That is the trade. And it is why this model can only be credibly built by firms with earned, board-level institutional trust — the kind that cannot be purchased or accelerated, only accumulated through decades of methodological rigor and an unbroken track record of defensible answers.
A Decision as a Service system nobody trusts is simply a faster way to make the wrong decision.
The Accountability Problem Nobody Is Resolving Yet
Every structurally honest post in this series includes the uncomfortable counterargument. Here it is.
Decision as a Service changes the liability posture in ways the industry is not yet equipped to address. Three specific challenges:
Governance and auditability. The Decision Ontology layer must be versioned, documented, and explainable at a level that satisfies enterprise risk management and, as of August 2, 2026, the EU AI Act’s requirements for relevant, representative, accurate, consistent, and complete training data. For Decision as a Service providers, this is not a compliance checkbox — it is the legal expression of the same trust architecture that has always governed consequential decision-making. Firms that cannot document the provenance and integrity of their decision-intelligence training data will not be able to operate this model in regulated markets.
Model drift. DSLMs trained on historical consumer data optimize for patterns that produced the training corpus. A model trained on a stable behavioral world will give confident answers calibrated to that world — even as the world changes. Continuous retraining with fresh real-human data is not optional. It is the ongoing operating cost of trustworthy Decision as a Service.
The trust gap for new entrants. A vertical application specialist can build excellent composable decision services on top of foundation models. “Excellent” is not sufficient when a CMO’s career depends on the output. The trust infrastructure — earned through decades of rigorous methodology and board-level client relationships — cannot be accelerated. It must be inherited. This is not sentiment. It is the structural argument that has run through everything we have written in this series: the most defensible moat in the Decision Intelligence stack is not the data. It is the trust that makes someone willing to act on it.
Who Is Building This Right Now
NIQ has the strongest declared architecture in Brand Management. NIQ Cadence — 19 specialized AI agents coordinated by Optiq, built on scanner data across 900,000 stores, the Homescan panel of 250,000 households, and MRI-Simmons’ 60,000 attitudinal, psychographic, and media elements per US consumer — is the closest existing production system to a Decision as a Service architecture in the insights industry. The data moat and the model moat are the same moat here. The gap: commercial availability below enterprise CPG scale. The SME market — brands with $10M–$500M in revenue — remains structurally unaddressed by both NIQ Cadence and Walmart’s architecture.
Walmart is the complementary execution layer, not the competition. Walmart Connect’s June 2026 expansion of Scintilla first-party purchase data into YouTube streaming inventory via Google’s Display Video 360 — connecting verified purchase behavior to media activation and closing the loop to measured sales outcomes — is Decision as a Service made visible in a live market transaction. Verified behavioral data, not modeled signals. Measurable outcomes, not inferred ones. That distinction is the entire argument.
Palantir built the ontology model. The Palantir Ontology — data, logic, and actions modeled together so AI agents can reason and act within a governed enterprise context — is the architectural template for what “decisions as invocable services” means in practice. Decisions are not one-off outputs in this model. They are versioned, governed, auditable services that agents can call repeatedly and consistently. The limitation for the insights industry: Palantir operates on whatever structured operational data the client brings. It does not own the verified consumer human data layer.
Aera Technology — now part of ServiceNow — is the earliest production DaaS example in operational enterprise. Aera has been delivering AI-powered decision intelligence for supply chain and operations since before the term “agentic AI” became common. Its architecture — autonomous agents operating against a live operational digital twin to surface and execute operational decisions — is the industrial proof of concept for what the insights industry is now building for Brand Management decisions.
And then there is HelloTwin.
HelloTwin, launched from Snowflake’s AI Hub in Menlo Park in late June 2026 and backed by BMP Ventures, introduced what it calls the “Digital Authority” — an accountable AI that runs on top of a semantic digital twin of the business, owns specific business outcomes, directs AI agents, and holds the mandate to align both humans and machines around measurable results.
The architectural principle HelloTwin is building on is the same one Palantir pioneered at enterprise scale — and it is the right one. Their patent-pending compiler architecture resolves answers against a governed semantic model rather than generating them probabilistically from an LLM. The same business question asked twice produces the same answer. Deterministic, not probabilistic. Governed, not guessed. As CPTO Hemant Kumar of design partner CloudKompass states: “HelloTwin doesn’t replace rigor with AI, it combines them. The AI reasons on top of a governed digital twin, working from numbers that were computed correctly, not guessed.”
HelloTwin is not a domain-specific language model play — it does not contain a proprietary training corpus on consumer behavior or industry intelligence. What it is building is the Decision Ontology layer for the SME market: the governed semantic foundation that makes any domain’s decisions trustworthy, auditable, and continuously owned — without requiring a data warehouse, a data science team, or a consulting project. One customer reports: “We built a digital twin from the applications we already use — no warehouse and no data science team. That’s $100,000 we never spent.”
That is Layer 3 of the Decision as a Service architecture, made accessible to the 99% of companies that cannot afford a data science organization. The SME market that NIQ Cadence and Walmart’s architecture structurally leaves open is exactly the market HelloTwin is entering. Watch this space.
What This Means for the Industry
For suppliers: The commercial model transition from project fees to decision pricing is not a future scenario — it is an active architectural shift happening at the production level. NIQ Cadence is the enterprise expression. HelloTwin is the SME expression. Palantir is the operational enterprise expression. The firms that build the accountability infrastructure, the governance documentation, and the trust architecture to sell decisions rather than data will compound. The firms that wait will find the market has moved — and that they have become a supplier to someone else’s decision system, not the owner of their own.
For brand-side buyers: The question for your function is no longer “how do we use AI better?” As we argued in our buyer-facing piece, it is: “Are we the function that owns decision architecture in this organization, or are we being reduced to a data supplier inside someone else’s decision system?” Decision as a Service is being built at the supplier level. If your internal function is not building toward the same architecture — continuous intelligence, decision-embedded workflows, governance and provenance documentation — you will eventually be purchasing a service that replaces what you currently provide.
For investors: The valuation framework crystallizes around this. The unit of value is position in the decision stack. Firms priced as single-layer data suppliers but actively building toward proprietary human data ownership, DSLM capability, and a decision layer represent the primary mid-transition opportunity. The gap between their current multiple and their forward architecture multiple is structural, not cyclical. As we wrote in our PE analysis: the same underlying human data asset commands 6–8x EBITDA as a point solution and 12–18x as an integrated Decision Intelligence platform. Decision as a Service is how the 12–18x gets earned.
The Closing Provocation
The insights industry has spent 115 years being paid for the process of understanding consumers. Data collection. Analysis. Report delivery. All of it in service of one moment: the executive decision.
Decision as a Service collapses the distance between the process and that moment — and in doing so, it changes what the industry is selling, what it is accountable for, and who will survive the transition.
The architecture to build it already exists. The data moats are established. The domain intelligence layer is in formation. The delivery infrastructure is open. The composable decision-service opportunity is real for new entrants. The industry is being redrawn — on top of everyone in it, in real time.
What does not yet exist — at scale, with full accountability — is the commercial model and the trust architecture that allows a CEO to stake a board-level decision on the output of an AI-powered decision service.
Building that is the work of the next three years.
The question is who is going to build it first.
Thanks to Jeff Krentz for the spark that became this piece. Insight Innovation Ventures invests in AI and analytics startups across the USA and UK, with a focus on data infrastructure, research technology, and AI-native insight platforms.
Related reading from the IIV series:
The most interesting piece to me is this: "Decision as a Service collapses the distance between the process and that moment — and in doing so, it changes what the industry is selling, what it is accountable for, and who will survive the transition."
Makes me wonder if the reticence of the research industry to commit to decisions is an artifact of the safety inherent in historically having access to a single input in the decision process. This architecture greatly expands the purview of available outputs.
That said, depending on the decision being made, there are many other factors (capital expenditures, staffing, etc.) that still need to be weighed.
This is a very compelling future. Question - are there types of decisions you see as early use cases and then complexity may build with experience? Or how do you see that unfolding?
Appreciate this very thought-provoking piece.
Kay here, co-founder of HelloTwin - thank you for the precise placement of our work in this architecture. One thought to add to your accountability section: in the SME market, there’s no risk team or analyst sitting between the AI and the owner making the call. So governance can’t be wrapped around the product, it has to be the product. Deterministic resolution, audit trails, reproducibility: not features, the foundation. Rare to see the counterarguments taken as seriously as the thesis. That’s why this series stands out.