Speedinvest had already seen the AI shift from the front row. Its teams were watching founders, portfolio companies, and board conversations move from experimentation toward AI-native ways of working. The question inside the firm was more practical: how to bring the same shift into the operating model of a regulated, operationally complex venture business.
The firm already had access. Company-wide ChatGPT usage existed; some teams were also using Claude and Perplexity. Access was not the bottleneck. The bottleneck was the distance between a chat window and the work itself: dealflow, compliance, Affinity CRM, the fund's ERP system, email, notes, and internal knowledge.
The problem class
AI access is not operational AI. In a venture firm, the valuable work crosses systems and teams. A new deal arrives by email, needs to be understood, routed, and entered into the CRM. Compliance work depends on external research, evidence capture, and a fileable result. Finance and operations depend on the fund's ERP system. Notes and knowledge sit around those systems, not neatly inside one of them.
That is why a simple prompt rollout was never enough. The first idea was to identify five to ten processes and automate them quickly. Once the work began, the more important foundation became clear: teams needed a shared understanding of what modern AI systems could do, how human review should stay in the loop, and how repeated work could become reviewable skills and agents rather than isolated prompts.
What we did
Specialty Tokens started with company-wide enablement and then worked team by team. The sessions were practical by design: not abstract AI demos, but examples from each team's own work. Employees learned where AI could help with routine tasks, where judgment still mattered, and how workflows could interact with external systems rather than only draft text beside them.
That training layer mattered because the technical layer would otherwise have had no owner. We worked closely with Speedinvest's internal AI owner, so the capability could keep spreading after the initial engagement. The limiting factor was not only whether a workflow could be built. It was whether each team had someone close enough to the work to translate a repeated process into a working AI skill.
The technical work moved in parallel. Speedinvest's ERP system sat at the center of many workflows, so AI had to begin reaching systems of record, not just documents and chat history. We helped start that integration work and began extending the same approach to Affinity, the firm's CRM. The available connector layer for Affinity was still early, so the work had to account for those limits instead of pretending every system was already easy to reach.
From there, ERP, CRM, email, notes, and internal knowledge workflows started moving from isolated prompts toward connected systems. That is the core of both CRM AI integration and ERP AI integration: the model is only useful when it can work with the systems where the business state actually lives.
Where the workflows landed first
One early workflow focused on dealflow intake. Speedinvest had an inbox where potential new deals arrived before being entered into the CRM. The process was repetitive and high-volume, but not mindless: the workflow had to associate opportunities with the right investment manager, route them to the right team, and define the right information at that stage.
Together, Speedinvest and Specialty Tokens built an initial skill or agent to work through that inbox automatically. The workflow reduced an estimated 10–20 hours per week of manual work. More importantly, it gave teams a concrete pattern for the next process: repeated manual handling could become a structured, reviewable automation tied to the systems they already used.
Another workflow began on the compliance side: LP adverse-media checks. The existing process involved searching for a person or company, using relevant keywords, capturing evidence such as screenshots, and filing the result. The AI-supported version under development was designed to support that work with skills and external data sources. The compliance team had been spending an estimated 10–20 hours per week on this work, and the new workflow had the potential to reduce that load while expanding multilingual coverage.
That matters in a regulated operating context. AI is useful here only when it improves coverage, consistency, and speed without removing the review discipline that the process requires.
Why this matters
By the end of the engagement, Speedinvest had moved beyond basic AI access. More employees understood that AI could work inside operational systems, not only answer questions in a separate window. Teams had seen how skills and agents could attach to their own processes. Early workflow wins made the change concrete: dealflow intake and LP checks were no longer theoretical use cases, but examples of where repeated work could be redesigned.
That is the shape of serious AI for venture capital. The fund's edge is still judgment. The machinery around that judgment should be lighter: less copying, less manual routing, less repeated search, and better use of the systems where the firm already records its work.
It is also why AI workshops and training belong next to implementation. Workshops without connected workflows fade. Connected workflows without team understanding go unused. Speedinvest needed both: a shared operating language for AI, and working processes that showed the language was real.
Ownership
The important ownership point was internal. Speedinvest left the engagement with a clearer understanding of where AI could support real work, an internal AI owner helping teams continue, and early workflows proving that meaningful operational time could be reduced.
For a venture firm, that is the foundation for responsible AI adoption: leadership attention, connected systems, team-by-team enablement, and concrete workflows where the time savings can be reviewed against the underlying process.
