Who your customer is changes what the AI has to survive
A copilot for fifty enterprise accounts and a support system for two million consumers use overlapping technology and almost nothing else. Volume, latency budget, tolerance for a wrong answer and who carries the liability all differ, and each one pulls the architecture in a different direction.
The same technology, four completely different problems
Who your customer is decides what the AI has to survive. Query volume, tolerance for a wrong answer, latency budget, who carries the liability. These four shapes cover almost every engagement we take.
B2B and enterprise software
Low query volume, high stakes per query. A wrong answer reaches a named account and lands in a QBR. The work is accuracy, permissions and per-tenant isolation, not scale.
What we build: internal and customer-facing copilots, revenue intelligence layers, multi-tenant LLM features inside your existing product, usage metering that makes margin visible per customer.
Consumer brands and B2C
Enormous volume, thin margin per interaction, and a public failure mode. Cost per request and latency matter more than benchmark scores, and brand safety is a hard requirement rather than a preference.
What we build: autonomous support across channels, real-time personalisation and recommendations, demand forecasting, content operations, sentiment-based escalation to humans.
Marketplaces and platforms
Two-sided, with strangers transacting. The AI problems are matching, trust and abuse, and they are adversarial: someone is actively trying to defeat the system, which is a different engineering problem from serving cooperative users.
What we build: matching and ranking, listing quality and enrichment, fraud and trust-and-safety classifiers, dispute triage, seller and buyer support agents, moderation queues with human review.
Internal operations
No external customer at all. The measure is hours removed and errors avoided, and the hardest constraint is usually the data: undocumented legacy systems, inconsistent schemas and processes that live in somebody's head.
What we build: knowledge search over institutional documents, document processing, compliance and audit automation, engineering and delivery copilots, agents that move work between systems.
What changes in the architecture, model by model
This table is the short version of a conversation we have on almost every scoping call.
| Model | Dominant constraint | What it forces | Proof |
|---|---|---|---|
| B2B and enterprise | A wrong answer reaches a named account | Citation grounding, refusal behaviour, tenant isolation, per-customer audit trail | Knowledge search |
| Consumer and B2C | Volume and cost per interaction | Small-model routing, aggressive caching, sub-100ms paths, sentiment escalation | Support platform |
| Marketplaces | Adversarial users | Classical ML classifiers, drift monitoring, human review queues, rapid retraining | ML systems |
| Internal operations | Messy, undocumented data | Longer discovery, layout-aware parsing, integration work, change management | DevOps platform |
Marketplaces are where we most often say no to generative AI
Fraud scoring, listing quality and abuse detection are adversarial classification problems on structured features. A gradient boosted model retrains in minutes when attackers adapt, explains its decision to a compliance team and costs a fraction of an LLM call per event. We reach for a language model on the parts that need language, and not on the parts that do not.
Every number on this site is traceable to a system
We have a standing rule: no figure appears in a proposal, a deck or on this website unless it came out of a deployed system with a client attached. Each one below carries the engagement it was measured in.
Citation accuracy on legal queries, with zero hallucinated citations after DPO alignment
measured · legal-ai · prodSupport queries resolved without human escalation across seven channels
measured · support-platform · prodQualified pipeline generated in the first six months by the autonomous revenue layer
measured · revenue-platform · prodAccuracy classifying regulatory compliance gaps across eight jurisdictions
measured · regtech · prodReduction in debugging and incident resolution time inside an existing DevOps toolchain
measured · devops-platform · prodData residency inside the client VPC on the private deployment, with no external API calls
measured · private-llm · prodNot sure which of these you are?
Most companies are two of them at once. Bring the workflow and the customer, and we will tell you which constraint dominates and what that means for the build.