Every commercial leader of a mid-to-large business now has a version of the same conversation. The data exists. The CRM exists. The analytics team exists. The dashboards have been built. And yet the commercial decisions being made on the ground bear only a loose relationship to the data being produced upstream. The sales rep is improvising. The pricing team is approximating. The marketing function is targeting against last year's segmentation. The intelligence the company has invested in is not arriving where decisions are being made, in a form that decisions can be made from.
This is the practical problem of commercial intelligence, and it is a structural problem rather than a software one. The fix is not another tool. The fix is an architecture: a specific way of connecting data sources, analytics, CRM, and commercial workflow that produces a closed loop. Signals come in. Decisions get made. Actions get taken. Outcomes feed back. The system learns. Each cycle, the next decision is better-informed than the last one. This is what a commercial intelligence engine actually does. The technologies that enable it are widely available. The architecture that connects them is not, which is why so few companies have one.
This piece describes that architecture in the form that has produced durable outcomes across roughly thirty implementations the author has been involved in or observed. It is not the only architecture that works. It is one that is durable, that does not require unusual technology investment, and that can be built incrementally by a commercial leader without rebuilding the company's entire data stack.
- Companies with mature commercial intelligence capabilities show 2.0 to 2.5x higher revenue per commercial FTE than peers (McKinsey B2B Pulse, 2024).
- The single largest investment area for top-quartile commercial organizations is no longer headcount; it is data integration and decision tooling (Gartner Sales Investment Survey).
- Among companies attempting commercial intelligence builds, roughly 60 percent deliver below business case, almost always for reasons of architecture rather than technology selection (Bain, 2023).
The four layers
A working commercial intelligence engine has four functional layers. Each is necessary; none is sufficient on its own. Companies that succeed build all four; companies that fail typically build one or two well and assume the others will follow.
Layer 1: Signal
The signal layer is the set of data sources that feed the engine. There are five categories that matter, and most companies have all five somewhere in their organization. The work is not acquiring them. It is connecting them.
The first is transactional data: orders, invoices, returns, prices paid. This is the highest-fidelity data the company has, and it is usually trapped inside the ERP. Liberating it for commercial use requires nothing more than an integration that most companies have not bothered to build.
The second is interaction data: meetings, emails, calls, portal events. This data exists across half a dozen systems (CRM, email, calendar, conferencing, portal) and is rarely consolidated. The integrations exist as standard products. The decision to use them does not.
The third is third-party intent: signals from external sources about what the buyer is researching, what they are reading, what jobs they are posting, what their competitors are doing. Bombora, ZoomInfo, LinkedIn, and a dozen others sell this data. The data on its own is noisy. The data combined with a company's own first-party signals is informative.
The fourth is support and service data: what is the customer asking, where are they having trouble, what is the trajectory of their usage. This data lives in the support platform. It is rarely visible to the commercial team. The disconnect is one of the most expensive in commercial operations: the company has a clear signal that an account is in trouble, and the people responsible for retaining that account are the last to see it.
The fifth is external context: industry signals, news, regulatory changes, supply chain events. This data is necessary for sophisticated account engagement and is the layer companies most often skip because it requires human curation.
The architectural decision here is the consolidation point. A modern company should consolidate these five signal categories into a single customer data layer (typically a data warehouse with a customer 360 model on top). The technology is well-understood. Snowflake, BigQuery, and Redshift are all viable. The discipline is in defining the customer master and the entity resolution that connects records across systems.
Layer 2: Record
The record layer is what is conventionally called the CRM. Its function is to be the canonical, current account of every customer relationship. This is where the signals get consolidated into a story that humans can act on.
The mistake here is treating the CRM as the place where data is gathered. It is not. It is the place where data is consolidated and made actionable. The gathering happens automatically from Layer 1. The CRM should reflect the signal data, not duplicate the effort of capturing it. Companies that ask their sales reps to type information into the CRM that already exists in another system are losing the most expensive commercial time the company has, and producing data that will be incomplete and stale.
The discipline at this layer is twofold. First, the data model has to be defined: what is an account, what is an opportunity, what is a contact, what is the relationship between them. Most companies have inherited a data model from their CRM's defaults, which were designed for a generic case that may not match the business. Re-modeling at this layer is unglamorous and high-leverage.
Second, the integration has to be automated. The CRM should be receiving from Layer 1 (signals), passing to Layer 3 (intelligence), and consuming from Layer 4 (actions taken). Every manual data entry step that survives this design is a tax on the commercial team and a source of data corruption.
Layer 3: Intelligence
The intelligence layer is the analytical work that transforms the consolidated data into commercial signals. This is what most companies call "commercial analytics" or "revenue intelligence." It is the layer that produces the predictions, scores, and recommendations that commercial teams can act on.
The key categories are: account scoring (which accounts to prioritize), churn prediction (which customers are at risk), expansion signals (which customers are ready to grow), pricing intelligence (where the company is leaving money on the table), competitive intelligence (where the company is winning and losing), and win-loss analysis (why deals are turning out the way they are).
Each of these can be built with reasonable accuracy using the data from Layers 1 and 2. The technology is widely available. The work is in defining what counts as a useful prediction for the specific business. A churn model that flags 80 percent of accounts as "at risk" is not a useful model, even if its accuracy is high. A model that flags 5 percent of accounts and is correct 60 percent of the time is a useful model.
The architectural decision at Layer 3 is whether to build, buy, or hybrid. The market has matured enough that buying for most of these capabilities is now viable. Building still makes sense where the company has a genuinely differentiated approach to the signal or where the integration with proprietary data is the source of advantage. Hybrid (buy the engine, customize the configuration) is usually the right starting point.
Layer 4: Action
The action layer is where the engine produces commercial outcomes. Without it, the first three layers are a very expensive reporting system. With it, the system compounds.
The action layer is the smallest layer in terms of technology and the largest in terms of organizational discipline. The technology is mostly orchestration: workflow tools that translate signals from Layer 3 into specific outreach sequences, alerts, or workflow triggers. Salesloft, Outreach, the native automation in modern CRMs, custom workflow engines. The choice matters less than the design.
The design has three properties that distinguish working systems from failing ones.
The first is that actions are deterministic, not advisory. The system does not produce a recommendation that a human evaluates and then ignores. It produces a workflow that runs. The human can override, but the default is action. The number of commercial actions per signal goes from approximately 0.2 (advisory model) to approximately 1.1 (deterministic model) when the design is changed. The five-times improvement is the work the engine is doing.
The second is that the playbooks are owned by the commercial team, not by the analytics team. The analytics team has produced the signal. The commercial team has produced the playbook that responds to the signal. The ownership division is essential. When analytics owns the playbook, the commercial team treats the output as something done to them. When commercial owns the playbook, they treat it as their own work.
The third is that the loop closes. Every action taken produces a measurable outcome (response, meeting, opportunity, closed deal, churn averted). The outcome feeds back to Layer 3, which improves the model. The improved model produces better signals. The better signals produce better actions. This is the compounding mechanism. Companies that have built the first three layers but failed to close the loop have a system that performs at its initial level forever. Companies that have closed the loop have a system that gets better month over month.
The intelligence is not the destination. It is the input to the action. The action is what produces the next signal, and the signal is what produces the next round of intelligence.
The build sequence
The hardest decision for an operator is the build sequence. The temptation is to build the visible layer (the dashboards, the AI, the recommendations) first. This usually fails, because the data underneath is not yet good enough to produce useful outputs. The discipline is to build bottom-up: Signal, then Record, then Intelligence, then Action. Each layer requires the layer below it to be functional.
A realistic timeline for a $500 million to $5 billion revenue B2B company is roughly 18 to 24 months for the full architecture, with visible commercial impact starting around month 9 (when Layer 3 begins producing usable signals) and compounding from month 15 (when Layer 4 closes the loop). The mistake to avoid is committing to outcomes in the first 6 months. The first 6 months are infrastructure work, and the temptation to declare premature victory at month 4 will destroy the architecture by encouraging shortcuts that produce visible results without the foundation to sustain them.
The cost is significant but not extraordinary. A mid-sized B2B company should expect to invest $3 to $8 million in the architecture and operations over two years, plus ongoing run-rate. The return, measured rigorously, is typically 8 to 15 percent revenue uplift sustained, plus material reductions in commercial cost-to-serve. The payback period is roughly 18 to 30 months. The compounding starts then.
What separates the successes
The companies that have built durable commercial intelligence engines share three traits beyond the architecture itself.
The first is that the CRO owns it. Not the CIO, not the CDO, not the analytics lead. The Chief Revenue Officer is the customer of the system and the decision-maker on the design. Without this ownership, the system gets built as an IT project, optimized for technical elegance rather than commercial impact, and used inconsistently by the commercial team.
The second is that they have invested heavily in commercial team enablement. The architecture is useless if the field cannot use it. Successful companies have invested roughly as much in training, change management, and playbook design as they have in the technology itself. The companies that have under-invested in this layer have produced engines that nobody operates.
The third is that they have measured rigorously. Every layer has its own metrics: signal completeness for Layer 1, data quality for Layer 2, model accuracy and uplift for Layer 3, action completion and outcome rates for Layer 4. The metrics are reviewed monthly. The architecture is treated as a product that has performance, and the performance is managed actively.
The engine is not a project that ends. It is a capability that compounds. Companies that build it well are five years into a curve that competitors who start today cannot catch by 2030. The work begins with the data the company already has, and the architecture that almost nobody has built.