Agentic VC funds are coming. Are you willing to dream?

Over the past year, I’ve been obsessed with exploring the paths that might be taken to create an agentic capital allocator in our asset class (venture capital), and to think about how far it is really useful to go.

The most natural path was to start with experimenting on the Platform side of the fund due to my own proximity to the function. However, the observations and its application naturally extend to the operations and investing side of the fund.

VC Platform work timelines are collapsing, and its scope is (or could be) expanding significantly with increasing AI capabilities. Many brilliant people across the ecosystem are capturing opportunities that come with collapsing timelines by building automations and workflows, and finding insights in data in no time.

However, it’s worth paying attention to what happens when we succeed in collapsing these timelines to almost nothing. I believe this will unlock exciting opportunities to expand the scope. Applying the Jevons paradox, the cheaper it is to serve portfolio companies, the more demand there will be - making platform more impactful than ever.

So my question is, what would you build if you could do a decade of platform related work in one day?

‍

A framework for understanding your VC Platform AI maturity level

To answer this question, we first have to rethink the scope. To be able to do that though, it’s useful to review the current levels of AI maturity across firms, so we can better understand what collapsing timelines might look like for your fund.

This framework draws on my own observations across the ecosystem, as well as my experience moving BACKED through these levels.

I learned a lot about where many funds are on their AI journey through hundreds of conversations, and I’ve also had the chance to run a fun initiative—Backathons—AI hackathons for VCs in any role (investing, platform, ops, finance, etc.). There’s no better way to experience a Cambrian explosion of ideas for collapsing work timelines than bringing together people from across funds, roles, and seniority, all working on the same problem but approaching it differently (if you want to join the next Backathon, my DMs are open!)

It became clear to me that everyone is aware of “timelines are collapsing - we need to capture the benefit”. But there is striking variance in how far, and how seriously, funds take that observation.

‍

To place your fund along the AI adoption levels (with a focus on Platform for now), start by considering these questions:

• How visible is the output people produce? Access to tools doesn’t necessarily translate into results.
• Can the whole firm (eventually) benefit? Many firms have individuals who can build something impressive with AI, but the question is whether those gains are shared.
• Can the improving AI competence survive personnel change? If the biggest AI champion left tomorrow, would the rest of the firm know how to keep progressing?
• Who notices the problem first? AI or humans?
• What happens when agents get something wrong? Have you addressed who bears responsibility?

Answering yes to all of these questions suggests you’re well along in your AI adoption journey. If you can answer yes to the first few, you’ve begun, and you’re somewhere on the spectrum (more on that below). If you answered no, or couldn’t answer all of them, you’re at the beginning.

L0: Some talk, no action

At this level, some individuals are experimenting with AI tools privately, but there is no firm’s shared access to data, nobody is building automations or workflows. All “work” remains in LLM chats. Some team members might be reading newsletters, talking to peers in the sector, and are genuinely curious about what’s possible.

Who triggers work: nobody

What AI can access: no access granted

Who extends the building capability: nobody builds

Where human sits for judgement and review: n/a

Adoption & behavioural change: just talk, zero behavioural change

Ownership & accountability: nobody is accountable

Weekly active users: some team members might use chat function LLMs

Automations in production: 0

Agents on triggers: 0

L1: Personal productivity gains

What’s going on here is that the outputs are real and visible, but they happen on the level of individuals, so the fund’s collective output looks exactly the same.

Who triggers work: individuals, per task

What AI can access: whatever one person pastes in

Who extends the building capability: enthusiasts hack for themselves

Where human sits for judgement and review: self-checking, informal

Adoption & behavioural change: a few individuals, roles untouched

Ownership & accountability: leaves with the person

Weekly active users: a ****few individuals

Automations in production: 0 (personal scripts don't count)

Agents on triggers: 0

L2: Workflows that benefit teams

The fund has adopted some simple processes. The median team member benefits without having to be a builder. Team member(s) are fully in control, they initiate everything.

Who triggers work: humans trigger shared workflows

What AI can access: shared drives, connected tools, common templates

Who extends the building capability: a few designated builders create for the team

Where human sits for judgement and review: peer-visible output, informal norms

Adoption & behavioural change: median team member on shared workflows, roles unchanged

Ownership & accountability: named workflow owners, documented

Weekly active users: majority of team

Automations in production: several, each with a named owner

Agents on triggers: 0-1 pilot

L3: Agents do part of the work, humans direct them

AI notices first, and work arrives partly done. The job shifts from doing to directing and reviewing. This is where the scope question, and the point about ambition, starts to matter: if the busy, manual work becomes less busy, what do you do with the newly gained time? What happens when agents get it wrong also needs attention here.

Who triggers work: agents initiate on triggers; humans direct

What AI can access: systems of record (portfolio data, CRM, meeting transcripts/notes, tasks, intro history, etc) structured and readable

Who extends the building capability: platform/ops team builds agents; power users contribute

Where human sits for judgement and review: human reviews most output; escalation paths defined

Adoption & behavioural change: universal use; job content shifts toward directing/reviewing

Ownership & accountability: named agent owner who tunes, monitors, can kill

Weekly active users: effectively everyone

Automations in production: 10+

Agents on triggers: a handful, in at least one function

L4: Agents do majority of the work, humans supply judgement

The fund is structurally different. Humans set policy and handle exceptions, and Platform becomes a portfolio of agents with a manager. Weekly active users may fall because the system is designed so that fewer humans need to actively operate the tools for the organization to produce the same (or more) output.

This is the stage we’re aiming for. The work here is about governance and continuously improving the quality of judgment. A Monday at L4 might look like reviewing what the intro agent flagged as unusual overnight, adjusting the policy for one of those cases, and spending the rest of the day with founders.

Who triggers work: agents run the process; humans set policy

What AI can access: read and write across systems, plus external signals (hiring, traffic, filings)

Who extends the building capability: anyone composes agents within established governance guardrails

Where human sits for judgement and review: review by exception; error rates measured; external-output policy explicit

Adoption & behavioural change: roles formally redefined; ratios change; structure redesigned around agents

Ownership & accountability: owner runs a portfolio of agents, reviewed like team performance

Weekly active users: may plateau or fall

Automations in production: absorbed into infrastructure

Agents on triggers: portfolio of agents, majority of function output

L5: Fully autonomous

I haven’t experienced this level, so this is fully hypothetical. Here, there is no human input anywhere, this is fully autonomous operation. Venture is a trust and judgement business, and accountability to founders and LPs should not be delegated to an agent, so I think that’s the reason to stop at L4 (and I’m hoping that this statement will age well :)).

Who triggers work: no human, including exceptions

What AI can access: everything, self-extending access

Who extends the building capability: agents build and extend agents

Where human sits for judgement and review: None (the reason to stop at L4)

Adoption & behavioural change: Function has no human org

Ownership & accountability: Accountability unresolved (boundary, not target)

What makes  progression between levels possible?

To be able to traverse between these levels and run your platform (and eventually the entire fund) on abundant abundant intelligence, four things need solving: permissioned access, context, provenance and trust. And, of course, a human champion that is in the loop at every level to guide the progress.

Moving from L1 to L2

Progressing from L1 to L2 requires building shared context that’s accessible to the entire team.

We did this at BACKED by creating a shared context layer, composed of multiple databases - vector, graph, and relational - that let us ingest information across all the internal tools we use (G Drive, Notion, Attio, Granola, etc.). It also includes an ontology that helps agents traverse the graph to find the right information quickly. The hard part, but also the most magical once it’s solved well, is entity resolution: the ability to pull information on a person or a company (or any other entity) across all your tools.

Moving from L2 to L3

Moving from L2 to L3 requires solving for provenance.

This became clear when we noticed how we relate to information. When team members share information with each other, the context is often implicit (you remember who said it and how much to trust it). But if we want to act on information surfaced by agents, provenance needs to be explicit—source, timestamp, confidence, and confidentiality tier—so we can judge and use it appropriately. So we built provenance into the the shared context layer.

And how do you go from L3 to L4?

This is the transition we’re currently making. Here, we’re solving for permissioning, which equals trust.

Trust is an asymmetric asset: it can be built over years and destroyed in a single incident (and once it’s broken, you de-level instantly). At L4, trust is earned by solving L3 well. It also depends on having enough data to act on, paired with a permission layer so the right people can access the right information.

We invested a lot of time, energy, and creativity to building a context layer with provenance and timestamps on every record, working entity resolution, a relationship graph, and permissioning - kept continuously up to date. It’s queryable by humans and machines through the same layer, independent of any one tool.

Essential core of AI adoption: the role of a human champion.

Majority of funds have teams with someone occupying the AI outlier skills, and some to many people on the long tail of AI newbies. The role of the human AI champion is to recognise that, build for the benefit of all and bring onboard the whole team (which includes up-skilling - internal hackathons are a great tool for that). Some funds have Head of Product or CTOs to fulfil this role, some have enthusiastic champions. What’s useful to remember is that you can run only as fast as your slowest team member.

The value of becoming an agentic VC fund

This change is profound. At the beginning of this piece I was asking what would you build if you could do a decade of platform related work in one day? I’m still working through the answer, myself, or as our friend Claude would say - cerebrating - so in the meantime, let’s picture a few of the examples of what compressing the timeliness look like for us at BACKED:

Talent

Old world: the talent partner's personal Rolodex run one search at a time.

Present: AI-sourced longlists, drafted outreach, enriched candidate records; the talent lead curates rather than trawls and spends their hours on live candidates.

New world: agents run the top of the funnel end to end: they auto-recognise when a portfolio company is hiring and for which roles, handle sourcing (including the fund’s internal network), run screening, manage scheduling, and produce a structured first-pass assessment across every open search simultaneously. The human focuses on the close (e.g., the reference call that requires trust, persuading someone who wasn’t looking, and judging whether a strong CV is actually a fit for the company).

Customer and GTM introductions

Old world: "does anyone know someone at Hugging Face?" in the partner meeting. The matching is done by memory, and massive part of the network's value is unused because even the best person can only track so many opportunities, process so much information, and make so many high-quality decisions.

Present: the network mapped and queryable; AI drafts the blurbs and tracks the funnel; platform runs matching as a workflows

Future: an agent continuously reads portfolio needs against the full relationship graph (LP base, exec network, portfolio-to-portfolio and so on) and proposes matches with the context of why now. The human vets the fit and makes the warm handoff.

Next-round fundraising support

Old world: a spreadsheet of Series A investors built by hand, pulled from the top of investors’ heads. Intro requests surfaced in a meeting. Prep started when the founder said, “I’m raising.”

Present: AI-drafted investor maps enriched from databases and the fund’s own intro history; narrative and data-room prep accelerated by shared workflows; humans still decide timing, sequencing, and who makes each call.

Future: an agent watches runway, metrics, and market appetite across the whole portfolio and opens a graduation file on each company 12 months before the raise. The target list is scored against what those funds actually underwrite, and gaps in the story are flagged while there’s still time to fix them. The human focuses on what was always the part they loved: fundraising strategy and making the personal introductions.

What will future look like for VC firms?

If AI automates everything - meaning compresses timelines on majority of our current exciting tasks - I suspect the only differentiators for funds will be their honed judgement and built relationships. People will have more time to build and maintain relationships, strengthen them, and expand them - instead of spending their talents shuffling bits back and forth.

Now is the time to apply creativity to what an expanded scope for Platform (and, eventually, the entire fund) could look like. We already have the tools; now we need ambition to dream bigger.

And I’m keen to hear your thoughts - my DMs are open.

Lucia

Lucia Cerchlan - BACKED’s AI champion and Head of Portfolio Management and Product