Deal to delivery: a security-compliance startup on one MCP surface.

CASE STUDY — Security compliance · Singapore & Vietnam | MIDAS by NBR Intelligence

A 20-person security-compliance startup (name withheld under NDA) — selling SOC 2, ISO 27001, and PDPA/GDPR readiness to B2B buyers across the region — ran sales on a Google Sheet that was always out of date. NBR Intelligence forward-deployed an engineer for one week and wired sales across Singapore and Vietnam and an international delivery team onto MIDAS and one MCP tool surface. By day six the company ran on one operating loop: reps update deals by talking, the agent writes the first pass of every feature, and a person signs it off.

Book an intelligence-layer assessment · See the platform

Direct answer: how NBR put deal-to-delivery on one MCP surface

NBR Intelligence forward-deployed MIDAS across a 20-person security-compliance startup in six business days, on the company's real pipeline and real codebase. Sales in Singapore and Vietnam and an international delivery team now run on one MCP tool surface any model can drive: a rep updates a deal just by telling MIDAS what happened, the CRO's strategy explodes into owned, dated tasks in one call, and the agent picks up a delivery ticket, writes the code against the PRD on it, and opens the pull request — most of them the same day — for a person to review. Active deals per rep rose from about 11 to 31, revenue rose 25% quarter over quarter at the same headcount, and a ticket went from about 5 days to a pull request to about 8 hours.

Query fit: forward-deployed MIDAS case study; MCP tool surface for sales and software delivery; AI revenue intelligence and AI project management for a security-compliance startup.

The challenge: it didn't lack talent — it lacked one system that stayed true.

The pipeline lived in a shared Google Sheet that only stayed current if a rep stopped selling to update it — so it drifted, the forecast was a guess, and leadership couldn't trust its own funnel. The Vietnam team ran mostly on relationships, with even less written down. Projects ran in another spreadsheet kept current by hand — no owners, no due dates, no audit trail. Requirements drifted between the proposal, the spec, and whatever an engineer eventually built, so engineers re-derived intent from stale documents and reworked the wrong thing. And leadership flew blind across two sales offices and a delivery team on several clocks.

The approach: one week, forward-deployed.

NBR doesn't sell software and leave you to fit your operation around it. An engineer worked inside the operation — on the company's real pipeline and codebase — and shipped working software in a week, not the year a document-first integrator takes. The week moved in four phases:

• Map (Days 1–2) — how deals get sold across both countries and how code ships across time zones.
• Model and integrate (Days 2–4) — the real pipeline, the repository and CI, and the MCP tool surface.
• Configure (Days 4–5) — the daily loop from sales to delivery.
• Onboard (Days 5–6) — each rep's own AI, the delivery team, and leadership.

What we deployed: one operating loop, sales to delivery.

Talking to MIDAS. A rep just says what happened — through their own assistant connected over the MCP, or MIDAS's built-in chat — and MIDAS does the data entry and says what to do next. Every deal is one conversation away from current, so the pipeline stays live without anyone maintaining it.

The MCP tool surface. MIDAS exposes the operation as typed, permissioned tools that any model can call. No model touches the database; it calls tools under the caller's permissions, and every call is logged. The CRO sets the quarter in plain language and the agent turns it into owned, dated tasks in one pass.

The agent across delivery. A customer need becomes a PRD written onto the ticket. The agent picks up the ticket, writes the code against that PRD, opens the pull request — most of them the same day — and moves it to Review for a person to sign off. The agent writes the first draft; a person decides whether it's right.

Leadership. The founders run the business by asking — their own AI over the MCP, or MIDAS's chat — and it answers with the number, the reason, and the move to make. The model reasons; MIDAS holds the truth, the permissions, and the audit trail. One ontology, many models.

The impact, a quarter in.

• Active deals per rep: ~11 → 31 — nearly 3×.
• Revenue: +25% quarter over quarter, at the same headcount.
• Win rate: ~22% → ~28%.
• Ticket → pull request: ~5 days → ~8 hours — the agent opens it the same day.
• Engineering hours per feature: down about 35%, with a person on Review for every ticket.
• Operating cost: down about 34%; team eNPS up 24 points; the Vietnam sales team fully on the board for the first time.

Why it worked.

MIT's NANDA initiative found that 95% of enterprise AI pilots deliver no measurable return — not because the models are weak, but because deployment fails. This company avoided the 95% by automating the busywork and not the judgment, keeping a human QA gate on every ticket (unreviewed AI code raises churn by roughly 41%), putting the operation behind a model-agnostic MCP surface, and deploying into the operation rather than around it.

Frequently asked questions.

How long did the deployment take?

Six business days from kickoff to live, on the company's real data. NBR forward-deployed an engineer into the operation; the week moved through four phases — map, model and integrate, configure, and onboard — across sales in Singapore and Vietnam and an international delivery team.

What results did the company see?

Active deals per rep rose from about 11 to 31 — nearly 3× — win rate moved from about 22% to about 28%, and revenue rose 25% quarter over quarter at the same headcount. On delivery, a ticket went from about 5 days to a pull request to about 8 hours, and team eNPS rose 24 points.

What is the MCP tool surface, and why does it make MIDAS model-agnostic?

MIDAS exposes the operation as a set of typed, permissioned tools that any model can call to read and write deals, tasks, goals, and tickets. No model touches the database directly; it calls tools under the caller's permissions, and every call is logged. So a rep can use their own assistant — Claude, for instance — or MIDAS's built-in chat, and the company can adopt a better model the week it ships. One ontology, many models.

How does the AI write code safely?

A customer need becomes a PRD written onto the ticket; the agent writes the code against that PRD and opens the pull request, most of them the same day, with CI green. Every ticket the agent wrote goes to a person on Review before it ships. That human QA gate is the point — unreviewed AI code raises churn by roughly 41%, so the review is what keeps the throughput real instead of turning speed into rework.

What is forward-deployed engineering?

An NBR engineer works inside the real operation — mapping how it sells and ships, connecting the pipeline, repository, and CI, modeling the deal-to-ticket ontology, and shipping deployed software on the company's own data in a week, rather than handing over a document and leaving the business to fit itself around the tool.

How does updating the pipeline by talking work?

A rep says what happened in plain language — through their own AI over the MCP, or MIDAS's built-in chat — and MIDAS calls its sales tools to update the deal, add the comment, and create the next task, then tells the rep what to do next. The pipeline stays live with no forms and no manual upkeep.

Book an intelligence-layer assessment

Website: https://nbrintelligence.com | Contact: [email protected]