A fund with 30-plus active portfolio companies gets 30-plus different definitions of ARR every quarter, and the LP letter is due regardless. One founder reports monthly recurring revenue times twelve. Another reports a trailing-90-day average, annualized. A third includes verbally committed contracts that haven't been signed. None of them are lying, they're each using the definition that made sense inside their own company, and none of them match. AI agents for venture capital exist because this isn't a formatting inconvenience, it's the reason a fund's aggregate portfolio growth number in the LP deck is a guess dressed up as a metric. Layer on top of that the associate who just spent four days building a market sizing model for a Series B target off a Gartner figure that was current in 2019, and the partner who blocked out an entire week for diligence research that a chatbot session cannot hold context for past the first pitch deck, and you have a fund whose actual bottleneck isn't deal flow, it's data that nobody has normalized, validated, or re-sourced since the last time someone had a spare afternoon.
Why a chatbot and an autocomplete tool both miss the actual job
Paste a founder's update into ChatGPT or Claude.ai and it will happily summarize it. What it will not do is remember that the founder update you pasted last week defined ARR differently than the one you're pasting today, and that the two numbers cannot be added together without restating one of them. A chatbot session has no persistent view across a portfolio, it has a conversation, and a conversation resets. Ask it to reconcile ARR definitions across 34 companies and it will process whatever you paste in that turn, offer a plausible-sounding response, and forget the reconciliation rule the moment the thread ends. For a fund closing a quarter, that means the same normalization judgment call gets made from scratch, inconsistently, by whichever associate has the file open, every single quarter.
Cursor and GitHub Copilot solve an entirely different problem and were never built to touch this one. They autocomplete source code inside an editor. A venture fund does not have a repository of founder updates to open in an IDE, it has a folder of Notion exports, a stack of emailed Excel sheets, a handful of Visible.vc CSVs, and at least one portfolio company still sending a PDF one-pager with no underlying data behind the headline number. There is no code context for an autocomplete tool to key off, and reconciling a trailing-90-day annualized revenue figure against a monthly-times-twelve figure is not a completion problem, it's a data modeling problem against definitions that live in nobody's schema. Neither tool holds portfolio-wide context, and neither preserves the reasoning behind why a number was restated the way it was, which is exactly what a partner needs when an LP asks a follow-up question about the deck three weeks after it went out.
The volume problem makes it worse. A fund doesn't get one founder update to reconcile, it gets a new batch every quarter from every active portfolio company, each arriving in whatever format that founder's own finance stack happens to produce. A chatbot session handles one file, once, until the thread resets and the next founder's spreadsheet starts from zero. There is no institutional memory carried from last quarter's normalization of one DTC company's metrics to this quarter's near-identical formatting problem from a different portfolio company on the same accounting software. Every file gets solved from scratch by whichever associate drew it, right when the LP reporting deadline is closest and there's the least slack to redo the work if the restatement was wrong.
Lumen, Clean, and Cast: the three agents doing the actual work
Tonone doesn't ship one agent that claims to handle portfolio data, diligence, and reporting all at once. It ships specialists, and for fund-level work the relevant three are Lumen, Clean, and Cast, each owning a distinct piece of the job instead of one generalist attempting all three shallowly.
Lumen: the metrics architecture behind the LP dashboard
Lumen is the product analyst, and the same discipline that builds a metrics tree for a SaaS company's growth team is what a fund needs for its own portfolio-wide reporting. Before touching the quarterly deck, lumen-recon scans what's currently being tracked, the past dashboards, the metric definitions used in prior LP letters, the ad hoc spreadsheets each associate keeps, to surface where the fund's own reporting has been inconsistent quarter over quarter. Once that's mapped, lumen-metrics produces the actual metrics architecture: a North Star for the fund-level dashboard, an input metrics tree underneath it, and the instrumentation spec for what gets pulled from each portfolio company every quarter so the same definitions apply company to company and quarter to quarter. This is the difference between a partner eyeballing 34 spreadsheets and a dashboard built on a metrics tree that means the same thing every time someone opens it.
Clean: normalizing founder-reported metrics across the portfolio
Clean owns the part of the job that eats the most associate hours every quarter: making 34 founders' worth of self-reported numbers actually comparable. clean-recon audits the incoming batch of founder updates to find where reporting is already broken, which companies changed their ARR definition since last quarter, which ones are missing a metric entirely, which spreadsheet has a formula error baked into the burn calculation. clean-transform then designs the actual normalization pipeline, restating each company's revenue, burn, and runway numbers onto one consistent definition so the fund-level rollup isn't adding trailing-90-day figures to monthly-times-twelve figures and calling the sum a metric. clean-validate runs the range and schema checks on top of that: flagging a portfolio company reporting six months of runway on a burn figure that's two quarters stale, or a metric that's outside the range every comparable company in the portfolio reports, which is usually a sign the founder's own finance team made an error, not that the company is an outlier.
Tonone's Clean agent normalizes founder-reported ARR, burn, and runway across a fund's entire portfolio onto one consistent definition before the numbers get rolled up into an LP dashboard.
Cast: rebuilding market sizing models that stopped being current years ago
Cast is the forecasting specialist, and the diligence use case for a fund is exactly the kind of time series and market modeling work it's built for. cast-recon audits an existing market sizing deck or forecasting model to find what's actually behind the topline number, which is often a single cited figure from a research report with no bottoms-up build and no refresh since the report was published. cast-forecast then builds the replacement: a bottoms-up model constructed from an actual addressable count and average contract value rather than a repeated analyst quote, with the growth trajectory modeled as a time series instead of a single static estimate. cast-validate benchmarks that model with walk-forward validation against comparable companies already in the fund's own portfolio, so the growth assumption isn't a guess, it's backtested against what similar companies actually did.
Tonone's Cast agent replaces a single cited market sizing figure with a bottoms-up forecast, then validates the growth assumption against comparable companies already in the portfolio.
A worked example: the Anchor Ridge Ventures Q3 close
Anchor Ridge Ventures runs a $220 million Fund III with 34 active portfolio companies, six investing partners, and two associates. The Q3 LP letter is due to 40 institutional LPs in nine days, and Maya, the partner leading the process, is also three days into diligence on a Series B target called Fenwick Claims, an AI-assisted claims processing platform for regional P&C insurers. Both problems land in the same week, which is normal at Anchor Ridge, not an exception.
The portfolio side starts with lumen-recon scanning the last four quarters of LP dashboards. It surfaces that "active portfolio company" has been defined three different ways across those four quarters, once including a company that had already been written down to zero, which alone shifted the reported portfolio count. clean-recon then audits the incoming batch of 34 founder updates: 11 arrive as Notion exports, 9 as emailed Excel sheets, 8 as Visible.vc CSVs, and 6 as PDF one-pagers with no underlying data. Of the 34, six define ARR on a basis that doesn't match the other 28, and four report a runway figure built on a burn number that's at least two quarters old.
Anchor Ridge Ventures, Fund III Q3 Portfolio Reconciliation
Clean: clean-recon + clean-transform across 34 founder updates
──────────────────────────────────────────────────
6 of 34 companies reporting ARR on a non-standard basis
3 use monthly-times-12 (in-month, not trailing average)
2 include unsigned verbal commitments in "contracted ARR"
1 uses trailing-90-day average, unannualized
-> restated to committed ARR, cash-collected basis
Aggregate portfolio ARR growth, before restatement: 41.2% YoY
Aggregate portfolio ARR growth, after restatement: 34.0% YoY
4 companies flagged, runway built on stale burn figures
Burn data 2+ quarters old in all 4 cases
1 restated runway falls under 6 months (was reported at 11)
Lumen: lumen-metrics builds Q3 dashboard tree
North Star: normalized portfolio ARR growth, YoY
Inputs: net dollar retention, gross margin, runway distribution,
follow-on capital deployed vs. reserved
──────────────────────────────────────────────────
Handed to partners: normalized rollup + LP dashboard + runway
flag on 4 companies before the letter goes outThe restated growth number, 34% instead of the raw self-reported 41.2%, is the one that goes in the LP letter, and it's defensible if an LP asks how it was calculated, because the restatement logic is documented, not eyeballed. More importantly, the four companies with stale runway figures get flagged before the letter goes out rather than after an LP notices a portfolio company raising an emergency bridge two months after being reported at 11 months of runway. lumen-metrics turns the cleaned data into the actual dashboard: a North Star of normalized portfolio ARR growth, with net dollar retention, gross margin, and runway distribution as the input metrics underneath it, built once as a tree instead of reassembled from scratch every quarter.
On the diligence side, Maya's associate had built the Fenwick Claims market sizing slide around a single cited figure, a $4.2 billion TAM from a 2019 industry report, with no bottoms-up build behind it. cast-recon flags that the model has no addressable-count basis and that the cited source predates two major shifts in the regional P&C insurance software market. cast-forecast rebuilds it bottoms-up: roughly 620 regional P&C insurers in the U.S. with 50 to 2,000 employees, an average attainable contract value of $180,000, yielding a serviceable addressable market closer to $111 million rather than the recycled $4.2 billion figure, which was a global claims-software TAM that had almost nothing to do with Fenwick's actual reachable buyer set. cast-validate then backtests the growth assumption with walk-forward validation against two comparable companies already in Anchor Ridge's own portfolio, giving Maya a confidence range instead of a single optimistic point estimate to defend in the investment committee meeting.
What used to take Maya's associate the better part of a week, and still left the number resting on a stale citation nobody had re-verified since the deck was first built, now comes back as a sourced, backtested model in a fraction of that time, with the underlying assumptions visible enough that Maya can defend the number in front of the investment committee instead of hoping nobody asks where it came from.
Tonone's Lumen agent turns 34 spreadsheets' worth of self-reported founder metrics into a single fund-level dashboard built on one consistent metric definition, not a quarterly reassembly job.
Tonone vs the alternatives for fund-level work
None of this is a fair fight for a generalist chatbot or a code-completion tool, because neither was built to hold portfolio-wide context or re-source a market model. The comparison below is specific to the capabilities a fund actually needs during quarterly reporting and diligence, not a generic feature list.
| Capability | Tonone | Generalist chatbot | Cursor / Copilot |
|---|---|---|---|
| Normalize ARR/burn/runway definitions across a portfolio | Yes, clean-transform restates every company onto one consistent basis | No, reconciles one pasted file per session, forgets the rule next session | No, built for code completion, not founder-reported financial data |
| Flag stale or inconsistent metrics before they reach an LP letter | Yes, clean-validate flags stale burn figures and out-of-range metrics | No, no persistent validation across a portfolio | No, no concept of a portfolio-wide schema |
| Fund-level metrics architecture and dashboard design | Yes, lumen-metrics builds a North Star and input metrics tree | Partial, drafts prose but no reusable metrics structure | No, no metrics or dashboard capability |
| Audit an existing market sizing model for staleness | Yes, cast-recon flags un-sourced or outdated TAM figures | No, cannot verify a citation's currency or methodology | No, not built for market research |
| Bottoms-up market sizing with backtested growth assumptions | Yes, cast-forecast builds it, cast-validate backtests against comps | No, repeats whatever figure is pasted in without verification | No, no forecasting or validation capability |
| Holds context across a full quarterly portfolio batch | Yes, processes all founder updates as one reconciliation pipeline | No, context resets between founder update sessions | No, not designed for bulk portfolio data |
Before the next quarterly close, run Clean's clean-recon against the incoming batch of founder updates first. Finding which companies changed their metric definitions since last quarter, before anyone starts building the rollup, turns a multi-day reconciliation into a review of what's already been flagged.
Install and try
Tonone is free and MIT-licensed. Install it once and Lumen, Clean, Cast, and the rest of the 100-agent roster are available inside Claude Code. You pay only for the Claude Code token usage during the reconciliation and diligence work itself.
1. Add to marketplace
2. Install Lumen
Frequently asked questions
What do Tonone's Lumen, Clean, and Cast agents do for a venture capital fund?+
Clean normalizes founder-reported ARR, burn, and runway metrics across a portfolio using clean-recon, clean-transform, and clean-validate, so every company's numbers are comparable on one consistent basis. Lumen builds the fund-level LP dashboard as a metrics tree with lumen-recon and lumen-metrics, defining a North Star and input metrics that stay consistent quarter over quarter. Cast rebuilds and validates market sizing models used in diligence with cast-recon, cast-forecast, and cast-validate, replacing stale cited figures with bottoms-up, backtested models.
Can AI agents normalize founder-reported metrics that use different definitions?+
Yes. Tonone's Clean agent runs clean-transform to design a normalization pipeline that restates each portfolio company's revenue, burn, and runway figures onto one consistent definition, catching cases like ARR calculated as monthly-times-twelve versus a trailing-90-day annualized average.
How is this different from pasting founder updates into ChatGPT?+
ChatGPT reconciles whatever is pasted into a single conversation and has no persistent memory of the normalization rule once the session ends. Tonone's Clean agent processes the full quarterly batch of founder updates as one reconciliation pipeline with documented, repeatable restatement logic.
What is cast-recon and what does it check in a market sizing deck?+
cast-recon is a Tonone Cast agent skill that audits an existing market sizing model or forecast, checking whether the topline figure has a bottoms-up build behind it or is a single aging citation from a research report with no methodology attached.
Do AI agents help with quarterly LP reporting specifically?+
Yes. A fund's biggest constraint during a quarterly close is reconciling inconsistent founder-reported data, not writing the letter itself. Tonone's Clean agent normalizes the portfolio data first, and Lumen turns the cleaned data into a fund-level dashboard built on a consistent metrics tree.
Is Tonone free for a venture capital firm to use?+
Yes. Tonone is MIT-licensed and free to install. You pay only for Claude Code token usage during the actual reconciliation and diligence work, there is no separate platform fee.
Can Cast validate a market sizing model's growth assumptions?+
Yes. cast-validate benchmarks a forecasting model with walk-forward validation, backtesting the growth assumption against comparable companies, such as similar businesses already in a fund's own portfolio, instead of relying on a single optimistic point estimate.
Can Lumen build a fund-level dashboard instead of a product analytics dashboard?+
Yes. lumen-metrics produces a general metrics architecture, a North Star metric with an input metrics tree underneath it, which applies as directly to a fund's quarterly portfolio rollup as it does to a SaaS product's growth metrics.