Published 29 June 2026 · 14 min read
Why extraction is the bottleneck
Alternative investment data does not arrive in APIs. It arrives in PDFs. Capital call notices, quarterly portfolio company reports, GP performance packs, capital account statements, financial statements, limited partnership agreements. The volume is significant, the formats are inconsistent, and the stakes are high: a missed data point in a NAV reconciliation or an investor report can trigger regulatory questions.
For years, the industry solved this with people. Analysts manually keying numbers from documents into spreadsheets and accounting systems. That approach does not scale, and the error rates increase as document volumes grow and reporting deadlines compress.
Data extraction software for alternative assets is designed to replace that manual process: ingest a document, identify the relevant fields, extract the values, and deliver structured data to a downstream system. But the products on the market differ significantly in how they approach this, what they do with the data after extraction, and which part of the market they serve.
Before evaluating any product, it is worth understanding what a good extraction workflow actually requires:
- Document classification to identify what type of document has been received (capital call, NAV statement, financial report) without manual tagging.
- Field-level extraction that handles tables, multi-page layouts, and inconsistent formatting across different fund managers and administrators.
- Confidence scoring so the system can flag uncertain extractions for human review rather than silently guessing.
- Human review workflow because no extraction model is perfect on every document, and regulated firms need a clear process for exceptions.
- Audit trail linking every extracted value back to its source page and position in the original document.
- Downstream delivery so extracted data flows into the systems that actually use it, whether that is an accounting platform, a data warehouse, or an investor portal.
The last point is where the market is splitting. Some products treat extraction as a standalone step. Others embed extraction into a broader data platform. That architectural difference has significant implications for how much manual work remains after the AI has done its part.
The landscape in 2026
The extraction market for alternative assets has matured considerably. Canoe Intelligence has built the largest LP-side operation. Accelex was acquired by Carta in late 2025, signalling that extraction is becoming a feature of larger platforms rather than a standalone category. 73 Strings raised a significant Series B and positions extraction as the intake layer for its valuation engines. FactSet launched Cobalt AI Doc Ingest as an add-on to its GP monitoring platform. Alkymi has carved out a niche in private credit extraction. And daappa has built Extractor AI as part of a broader operations stack.
What follows is an honest assessment of six products. We build one of them, so apply appropriate scepticism to our view of our own product. But we have tried to be fair about where each competitor genuinely excels, because the right choice depends on your operating model, not on marketing.
How the main options compare
Canoe Intelligence
Canoe is the most established name in alternative investment document extraction. The platform processes 1.6M+ documents per month, extracting over 200M data points. That scale is not just a vanity metric: it means the AI models have been trained on a volume of real alternative investment documents that no competitor can currently match.
Canoe started with pattern recognition and has evolved through multiple generations of AI, now incorporating LLMs for contextual understanding. The models are trained exclusively on real alternatives data, and the company states that data never leaves their firewalls. The Bloomberg integration launched in April 2026 adds a distribution channel for extracted data that matters to institutional LPs.
The focus, however, is squarely on the LP side. Canoe excels at helping allocators and institutional investors process the documents they receive from GPs. If you are a fund administrator or GP looking to extract data from portfolio company reports for your own operational use, the fit is less direct. And while Canoe is excellent at extraction and collection, the analytics and downstream data governance sit with partners like Confluence and PrimePlus rather than being native to the platform.
Strengths
- Scale: 1.6M+ documents per month, a decade of alts-specific AI refinement
- Models trained exclusively on real alternative investment documents
- Bloomberg integration for institutional data distribution
- Clear LP-side market leadership with deep domain expertise
Limitations
- LP-focused: less suited to GP or fund administrator operational workflows
- Primarily extraction and collection; analytics via partners, not native
- Not a governed data platform that feeds downstream operations directly
Accelex (now part of Carta)
Accelex was acquired by Carta in October 2025, bringing AI-powered document extraction into Carta's broader alternative investment infrastructure. The platform focuses on institutional LPs and covers a wide range of document types: performance reports, capital notices, capital account statements, fund financial statements, LPAs, PPMs, and quarterly reports.
The AI classifies incoming documents into 30+ private market categories and auto-tags them with the relevant fund or vehicle. That classification step is genuinely useful at scale, reducing the manual sorting that typically precedes extraction. Post-extraction, Accelex provides analysis-ready dashboards covering IRR, MOIC, EBITDA, and portfolio look-throughs.
The accuracy claim of 99.5% warrants a note: this is post-extraction accuracy, meaning after human review has been applied. That is still a meaningful figure, but it is different from raw extraction accuracy. The Carta acquisition brings distribution advantages but also introduces integration uncertainty. Accelex's roadmap is now part of Carta's broader platform strategy, which may or may not align with what standalone extraction users need.
Strengths
- Strong document classification across 30+ private market categories
- Carta backing and distribution reach
- LP analytics built in: IRR, MOIC, portfolio look-throughs
- Broad document type coverage including LPAs and PPMs
Limitations
- Now part of Carta's roadmap, introducing integration uncertainty
- LP-focused: less suited to fund admin or GP operational extraction
- 99.5% accuracy claim is post-extraction (after human review), not raw
73 Strings (73 Extract)
73 Strings positions extraction as part of a broader intelligence platform. 73 Extract sits alongside 73 Monitor (portfolio dashboards) and 73 Value (valuation engines). The $55M Series B led by Goldman Sachs, with participation from Blackstone and Hamilton Lane, signals institutional credibility.
The extraction module claims up to 99% accuracy and a 90% reduction in manual work, with 10x workflow acceleration. These are broad figures and the "up to" qualifier is worth noting. Where 73 Strings is genuinely differentiated is in the connection between extraction and valuation. If the documents you are extracting feed directly into fair value calculations, the integrated flow from document to valuation model is a real advantage over tools that extract data and then leave you to figure out where it goes.
The trade-off is that extraction is in service of the valuation workflow. If you need extraction for operational purposes (feeding an accounting system, populating an investor portal, building a governed data warehouse) and do not need the valuation modules, you are adopting more platform than the problem requires.
Strengths
- Extraction integrated with valuation engines, a natural pipeline
- Strong investor backing signals enterprise credibility
- Handles complex financial documents across private markets
- 10x workflow acceleration for valuation-driven extraction
Limitations
- Valuation-led: if you do not need valuation, you are buying more than you need
- Accuracy claims are broad ("up to 99%")
- Less focus on the operational extraction workflow for fund administrators
Cobalt AI Doc Ingest (FactSet)
Cobalt AI Doc Ingest launched in February 2026 as an add-on to FactSet's Cobalt Portfolio Monitoring platform. The approach is distinctive: no model training, no configuration, no ramp-up time. The system uses agentic AI to extract, normalise, and map financial data from source documents automatically.
The GP focus is clear. This is built for private capital managers collecting portfolio company data, and clients report reducing multi-day processes to minutes. The integration with the broader FactSet and Hamilton Lane data ecosystem is a genuine advantage for GPs already in that world.
The limitations are equally clear. Cobalt AI Doc Ingest is an add-on, not a standalone product. It feeds Cobalt's monitoring module, not a broader data platform. It is GP-only, so LP-side or fund administrator extraction workflows are out of scope. And it is very new: the beta launched in North America in February 2026 with general release in March, which means the track record is measured in months, not years.
Strengths
- Zero-configuration approach: no model training or setup required
- FactSet and Hamilton Lane data ecosystem integration
- GP-focused portfolio monitoring integration
- Agentic AI handles normalisation and mapping automatically
Limitations
- Add-on to Cobalt, not available as a standalone product
- GP-only: not for LP or fund administrator extraction
- Very new (Feb 2026 beta), limited production track record
- Extraction feeds monitoring only, not a broader data platform
Alkymi
Alkymi takes a bundle-based approach to extraction. Rather than offering a general-purpose extraction engine that you configure for your document types, Alkymi ships pre-built bundles: Alkymi Alts (capital account statements, capital notices, schedules of investments) and Alkymi Private Credit (financial statements, compliance certificates).
This turnkey model has a clear advantage for firms that deal primarily with the document types covered by the bundles. There is less configuration, faster time to value, and less risk of the "it works on your demo data but not on ours" problem. Alkymi won the FTF Technology Innovation Award 2026, and the private credit focus is well timed given market growth projections (Morgan Stanley projects the private credit market at $5T by 2029).
The limitation of the bundle approach is coverage. If your firm encounters document types outside the standard bundles, the extraction may not cover them without customisation. And while the extraction itself is strong, Alkymi is less focused on what happens downstream: analytics, portal delivery, and governed data models are not core to the platform.
Strengths
- Turnkey bundles for specific document types, fast time to value
- Strong private credit capabilities, well positioned for market growth
- Enterprise-grade with FTF Technology Innovation Award 2026
- Pre-built for common alternative investment document types
Limitations
- Bundle-based approach may not cover all document types firms encounter
- Less focus on downstream analytics or portal delivery
- Newer entrant with less track record than Canoe
daappa Extractor AI (part of daappa Studio+)
Extractor AI is not a standalone extraction product. It is the intake layer of daappa Studio+, designed so that extraction feeds directly into a governed data model rather than stopping at a spreadsheet or a staging area that someone has to reconcile manually.
The extraction itself uses a three-tier model: automated AI extraction using computer vision, generative AI, and parallel LLM validation; confidence scoring on every extracted field; and managed human validation for low-confidence outputs. The result is ~99% extraction accuracy on private markets documents, with human review on exceptions. The operational impact is ~99% less time to process a fund's report pack, reducing what previously took days to under an hour. These figures are drawn from 20+ years of building private markets technology.
Document types covered include quarterly portfolio company reports, GP performance reports, LP statements, capital account statements, and financial statements. Every extracted value carries a full audit trail back to its source page and position in the original document.
Where Extractor AI is architecturally different is in what happens after extraction. The data flows into DataHub, a governed data model that normalises and structures everything. From there, it feeds Analytics, NAV Oversight, and Investor Portal. The advantage is that there is no gap between "data extracted" and "data usable in operations." The disadvantage is that Extractor AI requires the Studio+ stack. If you only need extraction as a standalone step, this is more platform than the problem requires. And daappa is newer to market with a smaller installed base than Canoe.
Strengths
- Embedded in an operational stack: extraction feeds governance, analytics, and portal in one flow
- Human review built into the workflow, not bolted on after the fact
- Full audit trail from any dashboard number to the source document page
- On-premise and sovereign cloud deployment options for regulated firms
- ~99% extraction accuracy; ~99% less processing time; 20+ years in private markets
Limitations
- Not standalone: requires the Studio+ stack
- Newer to market with a smaller installed base than Canoe
- Less suited if you only need extraction without the downstream data platform
Summary comparison
| Product | Approach | Document types | Key differentiator | Limitation | Best for |
|---|---|---|---|---|---|
| Canoe Intelligence | LP-side extraction at scale | GP reports, capital notices, NAV statements, financial statements | 1.6M+ docs/month, decade of alts AI, Bloomberg integration | LP-focused; analytics via partners, not native | LPs and allocators |
| Accelex (Carta) | Extraction + LP analytics | Performance reports, capital notices, LPAs, PPMs, quarterly reports | 30+ document categories, Carta ecosystem | LP-only; post-review accuracy; Carta roadmap uncertainty | Institutional LPs in the Carta ecosystem |
| 73 Strings | Valuation-led extraction | Financial documents feeding valuation models | Extraction integrated with valuation engines | Over-scoped if you do not need valuation | Firms where valuation is the primary use case |
| Cobalt AI Doc Ingest | Zero-config GP add-on | Portfolio company reports and financial data | No training/setup; FactSet data ecosystem | Add-on only; GP-only; very new | GPs using Cobalt for monitoring |
| Alkymi | Turnkey bundles | Capital account statements, capital notices, compliance certificates | Pre-built for common types; strong in private credit | Bundle gaps; less downstream focus | Private credit managers, wealth managers |
| daappa Extractor AI | Embedded operational stack | Portco reports, GP reports, LP statements, capital accounts, financials | Extraction to governed data to analytics in one flow | Requires Studio+; smaller installed base | Fund admins and GPs wanting end-to-end data flow |
How to evaluate: questions to ask during demos
Every vendor will demonstrate extraction on documents that work well with their models. The questions that reveal how a product will perform in your environment are the ones about edge cases, accuracy measurement, and what happens after extraction.
- How does accuracy get measured? Pre-review accuracy (what the AI gets right before any human touches it) and post-review accuracy (after human correction) are fundamentally different numbers. Ask which one the vendor is quoting, and ask to see both.
- What happens with documents outside the standard templates? Every firm receives documents that do not fit the usual patterns. Ask the vendor to show you what happens when the system encounters a document format it has not seen before. Does it fail silently, flag it for review, or attempt extraction with reduced confidence?
- Where does the extracted data go? This is the question that separates extraction tools from extraction platforms. Does the data land in a governed data model, a staging area, a CSV export, or a spreadsheet? What systems can it feed, and what integration work is required to make that happen?
- What is the human review workflow? No extraction model is perfect on every document. Ask to see the workflow for handling low-confidence extractions. Is it built into the platform, or does it happen in email and spreadsheets? Can reviewers see the source document alongside the extracted values?
- How do you handle multi-entity, multi-currency documents? Alternative investment documents frequently contain data for multiple entities, in multiple currencies, across multiple reporting periods in a single PDF. Ask the vendor to demonstrate extraction from a genuinely complex document, not a clean single-entity example.
- Can you show the audit trail from a reported number back to the source page? For regulated firms, the ability to trace any extracted value back to its exact position in the original document is not optional. Ask the vendor to click on a number in the output and show you where it came from.
- Where is the extraction processing done? Data residency matters for regulated firms. Ask whether extraction happens in the vendor's cloud, your cloud tenancy, or on-premise. Ask which jurisdictions the processing infrastructure sits in, and whether that can be changed.
What to watch: extraction is becoming table stakes
The most important trend in this market is not better extraction accuracy. Accuracy will continue to improve across all products as the underlying AI models get better. The more significant shift is the convergence of extraction and data governance.
Standalone extraction is becoming table stakes. The question that increasingly determines operational efficiency is not "can you extract data from this PDF?" but "what happens to the data after extraction?" Does it land in a governed model with field-level lineage, ready to feed analytics and investor reporting? Or does it land in a staging area that someone has to reconcile manually before it becomes useful?
The platforms that will define this category over the next few years are the ones where extraction is the first step in a governed data pipeline, not the last step before a manual process. The gap between "data extracted" and "data trusted and operational" is where most firms still lose time, and closing that gap matters more than shaving another tenth of a percentage point off extraction accuracy.
If you are evaluating extraction tools today, look beyond the extraction itself. Ask where the data goes, who governs it, and how many manual steps sit between "document received" and "data in production." That is where the real cost lives.
Next steps
If you want to see how an extraction-to-governance approach works in practice, we are happy to run Extractor AI against your actual documents and show you how the data flows through DataHub into analytics and reporting.
See how Extractor AI works with your actual documents→
Or explore the individual components: Extractor AI · DataHub · Studio+ overview