TLDR

• Built an AI-powered due diligence system to automate approximately 40% of a high-volume due diligence workflow, from discovery through to production deployment on the client's own infrastructure

• 25% of submissions are now fully reviewed by the AI system without requiring any human involvement

• Saved approximately 23 hours of manual review time per month, freeing skilled reviewers to focus on cases requiring human judgment

• SLA commitments now comfortably met, with operational capacity freed up to absorb volume surges and support other areas

The starting point

CoinGecko is one of the world's largest cryptocurrency data platforms. Their token listings team reviews dozens of applications per day, each requiring a rigorous due diligence process to verify legitimacy, quality, and compliance before a token appears on the platform.

Before the engagement, this process was entirely manual. Each application went through a multi-step due diligence process, requiring lookups, cross-referencing, and judgment calls from experienced operations staff at every stage.

CoinGecko's team was capable and their process was sound. But with dozens of submissions arriving daily, volume was outpacing capacity. The review backlog was growing, meeting SLA commitments on processing times required the team's full attention, and the growing queue left no room to absorb surges or focus on other priorities.

CoinGecko recognized that AI could help, but they didn't have the capacity to build a solution in-house. They made a deliberate choice: bring in an external specialist to validate the approach and deliver a production system.

The problem

The core tension was familiar to any operations leader: rising volume, fixed team.

CoinGecko's operations staff spent the bulk of their time on repetitive verification work. The same checks, applied manually to every submission. Many of these were routine: clear-cut rejections that didn't require human judgment, or straightforward validations that could be reliably automated. But without an automation layer, every submission received the same level of manual attention regardless of complexity.

The effects cascaded. The growing backlog meant the team had to work harder to maintain processing commitments. Skilled reviewers spent hours on routine cases instead of focusing on the ambiguous ones that genuinely needed their expertise. And as submission volume continued to grow, hiring proportionally wasn't a sustainable answer. CoinGecko needed to scale capacity without scaling headcount.

Automation also offered a consistency benefit. Some verification criteria involved subjective judgment, particularly ambiguous cases where standardized AI analysis could support more uniform decision-making across the team.

The approach

Hivekind's first step was not building. It was understanding.

We started with a structured discovery phase: requesting examples of past verification decisions, complete with the reasoning behind each outcome. We mapped CoinGecko's full review process in detail, documenting how the operations team applied their criteria and made decisions day to day.

This assessment identified which steps were most amenable to automation. Some involved clear, structured criteria that AI could handle reliably. Others required nuanced judgment or access to information that wasn't suited to automation at this stage. Rather than attempting to automate everything at once, we scoped the initial engagement to the steps with the highest automation potential, covering approximately 40% of the overall workflow.

The system was designed around a principle that mattered to CoinGecko: human oversight stays in the loop. The AI would analyze, score, and recommend, but humans would retain the ability to review, override, and make final decisions. For clear-cut cases where the AI had high confidence, CoinGecko could configure rules to complete the review automatically without human involvement. For everything else, the AI's analysis would assist the reviewer, not replace them.

This approach (discovery first, scope conservatively, keep humans in control) meant CoinGecko could adopt AI automation incrementally, building confidence with real results before expanding further.

The work

We built an AI-powered analysis system with specialized components for different verification tasks. Each component focused on a specific area of assessment, running checks in parallel and producing an overall evaluation with confidence scores and detailed reasoning for each finding.

The system integrated with multiple external data sources for comprehensive verification, with fallback capabilities built in for reliability. End-to-end tracing and error monitoring were included from the start. This was a production system handling real operational decisions, and it needed to be observable and debuggable.

Development followed a structured sequence. Once core features were built and validated, the system was migrated to CoinGecko's own infrastructure, a deliberate design choice. CoinGecko would own and operate the system on their servers, with their own credentials and accounts, retaining full control over the technology. Finally, CoinGecko's engineering team completed the integration with their internal admin tool, and the system went live.

Once deployed, the workflow changed meaningfully. A reviewer opens a submission, clicks a single button, and the AI system runs its automated checks in seconds. Results appear directly in the admin tool: pass/fail indicators with confidence scores, and specific findings with reasoning for each check. For submissions where the AI reaches a high-confidence determination, the review completes automatically without human involvement. For the remaining submissions, the reviewer uses the AI analysis to inform their assessment of the automated steps, then completes the remaining manual steps.

Adoption was immediate. CoinGecko's operations team started using the system on day one, with no resistance and no extended ramp-up. They could see the AI's reasoning, verify its conclusions, and override when they disagreed. Trust was built into the design.

The outcome

The results were measurable from launch.

25% of submissions are now reviewed end-to-end by the AI system, requiring no human involvement at all. These are clear-cut cases where the system reaches a high-confidence determination and completes the review automatically. For the remaining submissions, reviewers spend significantly less time on the automated steps, focusing their expertise on the checks that genuinely require human judgment.

The team saves approximately 23 hours of manual review work per month. Each fully automated review removes around 90 seconds of manual evaluation time, and across the remaining submissions, the AI's analysis helps reviewers assess cases significantly faster by surfacing the signals that matter most.

The backlog that had been growing steadily was cleared. CoinGecko now comfortably meets its processing commitments. The additional operational capacity means the team is better prepared to handle unexpected surges in submission volume, which in crypto markets can fluctuate sharply with market conditions. The freed capacity also allows CoinGecko to deploy resources to other operational areas, a bottom-line benefit beyond the listings workflow itself.

Critically, this was achieved without adding headcount. The same team now handles the same (and growing) volume of submissions, with capacity to spare.

Accuracy was maintained throughout. CoinGecko and Hivekind jointly established evaluation frameworks, ran system-testing workshops, and built measurement processes to ensure that efficiency gains didn't come at the cost of review quality.

The return: a permanent reduction in manual workload, comfortable SLA performance, and a scalable operational foundation, all owned and operated by CoinGecko on their own infrastructure.

What changed

CoinGecko now has a production AI system processing every listing application, every day. Their team operates differently now, reviewing AI-generated analysis instead of working through every step manually for each submission. The repetitive verification work that consumed skilled reviewers' time is handled automatically, freeing the team to focus on the decisions that require human expertise and judgment.

The engagement also shifted how CoinGecko thinks about AI in operations. What started as a specific automation project demonstrated a broader principle: that AI can be adopted incrementally, with measurable results at each stage, rather than requiring a large upfront bet. The team now has both the technology and the confidence to consider where else this approach could apply.

Key results

• 25% of submissions fully reviewed automatically, with clear-cut cases requiring no human involvement

• ~23 hours of manual review time saved per month

• SLA commitments comfortably met, with operational capacity freed for volume surges and redeployment to other areas