MarketSage: building our own AI sales intelligence platform

How Cognis Group designed and shipped MarketSage — an AI sales intelligence product — as a reference build for the agent engineering practice we take into client engagements.

The problem

Sales and revenue teams across African and emerging-market enterprises were under-served by existing sales intelligence platforms. Tools built for US and European markets rarely covered the firmographic, regulatory, and market-structure data that B2B sellers on the ground actually needed — and the ones that did sat behind licensing models that locked out everyone but the largest incumbents.

We wanted a product that solved the real problem: help a seller walk into a meeting with a credible, grounded view of the buyer's business, the market around it, and the trigger events that make this week different from last week.

Building it ourselves also gave the Cognis engineering team a reference system — one where every agent engineering decision we recommend to clients is one we have made, shipped, and operated on our own product.

The approach

MarketSage was scoped using the same six-phase methodology we use with clients — discovery, assessment, strategy, build, deploy, govern. The team spent the first four weeks understanding the actual seller workflow, not the tool wish-list. That changed the architecture materially: the product had to be composable into existing CRM and outreach flows, not a destination application.

Architecture at a glance

• Data layer. Multi-source firmographic and market data with lineage metadata on every field, so every agent-generated claim can be traced to a primary source.
• Agent layer. A small set of focused agents — company research, contact enrichment, trigger detection, briefing generation — each with its own evaluation harness and its own graded task set.
• Policy layer. Structured guardrails for data usage (NDPA-aligned), output citation requirements, and hallucination detection on factual claims before they reach the seller.
• Integration surface. API-first from day one. Native integrations with the CRMs and outreach tools customers already run, not a proprietary workflow trap.
• Observability. Every agent call logged, attributed, and tied to the model version and prompt template that produced it.

What we shipped

Version 1 reached internal production at week 14 and entered private beta at week 16. The same evaluation harness the build team used to promote each agent into production now runs continuously against live traffic — so regressions surface in days rather than in support tickets.

Results

What changed for the engineering practice

MarketSage now functions as Cognis Group's reference implementation. When we recommend an evaluation harness, a policy layer, or a phased rollout pattern to a client, we are recommending the one running on our own product — with the scars to prove it works. That is the standard we hold AI Agent & Automation Engineering engagements to.

Further reading

• AI Agent & Automation Engineering — the practice this product was built to reference.
• Building AI Agents That Actually Ship — the engineering pattern behind MarketSage.
• Claims processing: 14 days to under 48 hours — the same agent discipline applied in financial services.