Review of Pecan AI, predictive analytics software vendor

By Léon Martin-Ménard

Last updated: November, 2025

Pecan AI is a Tel Aviv– and New York–based predictive analytics software company offering a low-code platform that automates feature engineering and model training for tabular data and time-series, then wraps it in a chat- and notebook-driven user experience built on large language models. The company has raised roughly $116M in venture capital, including a $66M Series C round led by Insight Partners in 2022, with earlier funding from Dell Technologies Capital, S-Capital, and others, and now positions itself as a “no-code” or “low-code” alternative to traditional data science stacks.¹²³⁴ Its core engine trains and selects among standard machine-learning models (tree-based methods, LSTMs, ARIMA, Prophet) via automated feature engineering and Bayesian hyper-parameter search, while newer “Predictive GenAI” capabilities use generative AI to help business and data users define predictive questions, assemble training datasets in SQL, and operationalize models through a guided notebook interface.⁵⁶⁷⁸⁹ In August 2025 Pecan introduced DemandForecast.ai, a SaaS product built on this platform and aimed specifically at medium-to-large supply chain organizations, with marketing that emphasizes faster deployment of explainable demand forecasts and a 2025 Gartner “Cool Vendor” badge in cross-functional supply chain technology.¹⁰¹¹¹² What Pecan demonstrably delivers today is an accessible, fairly modern AutoML environment with an LLM front-end; what remains less substantiated is any claim to fundamentally novel forecasting algorithms or to end-to-end supply chain optimization beyond demand-side predictive models.

Pecan AI overview

At heart, Pecan is a vertically agnostic predictive analytics platform: given historical tabular data (transactions, customers, events), it automates feature engineering, trains multiple candidate models, selects the best based on a chosen metric, and deploys the resulting model to score future events. The company markets this as a way for BI analysts and other “data-adjacent” professionals to build churn models, LTV predictions, fraud scores, demand forecasts, and similar use-cases without needing to code in Python or manage ML infrastructure.¹⁶⁹⁴

The current product stack has three layers:

Core AutoML and data science pipeline – a back-end that transforms raw tabular data into predictive features and trains models (tree-based ensembles, time-series models, certain deep-learning architectures) on top of that feature space.⁹
Predictive GenAI UX – an LLM-powered “Predictive Chat” and “Predictive Notebook” that prompt users to describe a predictive goal in natural language, then generate SQL that defines the training dataset, which can be edited before model training.⁶
Verticalized packaging such as DemandForecast.ai – domain-specific front-ends and messaging, in this case targeted at supply chain planners who want demand forecasts and explanations without writing code.¹⁰¹¹¹²

Public documentation and marketing material consistently indicate that Pecan’s underlying models are standard supervised learning techniques orchestrated into an automated pipeline rather than new forecasting or optimization algorithms; its main innovations are in automation, packaging, and user experience. The DemandForecast.ai product, introduced in 2025, repackages this stack for demand forecasting: Pecan emphasizes explainable forecasts, GenAI-driven guidance, and integration to ERP/planning systems, but does not provide detailed public evidence of prescriptive inventory or capacity optimization on top of those forecasts.¹⁰¹¹¹²

From a supply chain perspective, Pecan therefore sits closer to “predictive layer for many domains (including supply chain)” than to a specialized supply chain optimization suite. This distinction matters when comparing it with vendors whose core proposition is not just forecasting but optimization of replenishment, allocation, and production decisions.

Pecan AI vs Lokad

Pecan and Lokad both anchor their value propositions in “AI” and forecasting, but they occupy different positions in the stack and make meaningfully different technical choices.

Scope and focus
- Pecan is cross-vertical: the same platform supports churn prediction, LTV, lead scoring, fraud detection, campaign ROAS, and demand forecasting. DemandForecast.ai is a relatively recent packaging targeted at supply chain, where demand forecasting is one among several use-cases.⁶¹¹
- Lokad is specialized in quantitative supply chain optimization: its platform is designed from the outset to compute probabilistic demand distributions and then optimize decisions (orders, allocations, production, pricing) to maximize expected economic outcomes under uncertainty, across retail, manufacturing, and aerospace supply chains.
User experience and programmability
- Pecan emphasises no-code/low-code, using a chat interface and auto-generated SQL notebooks to help analysts design predictive projects and data transformations, with classical ML models trained under the hood.⁶⁹ Changes are expressed via configuration and SQL.
- Lokad exposes a domain-specific language (Envision) where all data transformations, probabilistic forecasts, and optimization logic are encoded as scripts. This makes the platform closer to a programmable analytics engine; learning curve is higher, but the resulting models can capture granular constraints (MOQs, lead-times, compatible parts, budget constraints, etc.) and complex optimization objectives.
Predictive vs prescriptive depth
- Pecan clearly documents time-series forecasting capabilities (LSTM, ARIMA, Prophet, tree-based models) and automated feature engineering for tabular data.⁹ DemandForecast.ai focuses on producing accurate, explainable forecasts and surfacing them to business users; public materials do not describe mathematical inventory or network optimization on top of those forecasts.¹⁰¹¹
- Lokad’s published work centres on full demand distributions plus specialized stochastic optimization algorithms (e.g., stochastic discrete descent) and differentiable programming to learn forecasts directly for decision quality. Its deliverables are prioritized lists of decisions (e.g., purchase order lines, transfers, maintenance tasks) with estimated financial impact, not just forecasts.
Supply chain maturity
- Pecan’s supply chain offering is recent. DemandForecast.ai was launched in August 2025, and the main evidence so far consists of a Product Hunt-style marketing site, a detailed press release, and a handful of third-party write-ups repeating the vendor’s description.¹⁰¹¹¹²¹³¹⁴ The product emphasises being live “in weeks,” explainability, and GenAI-guided UX; it does not yet have a deep public case-law of supply-chain-specific ROI, nor detailed descriptions of how forecasts are converted into order quantities or capacity plans.
- Lokad has more than a decade of publicly documented supply chain use-cases (large retailers, auto-parts distributors, aerospace MRO, etc.) where optimization and economic drivers are central, and its technology roadmap is tightly bound to those problems (quantile forecasting, probabilistic demand over lead time, custom optimization algorithms, etc.).

In short: Pecan is a generalized, ML-as-a-service environment with a strong usability layer and a new supply-chain-flavoured front-end; Lokad is a specialized probabilistic optimization engine for supply chains with a programming interface. For a company whose main problem is “we need forecasts and some predictive models, ideally without hiring many data scientists,” Pecan is a credible entrant among modern AutoML platforms. For an organisation whose core problem is “we must optimize replenishment, allocation, and capacity decisions under uncertainty,” Lokad’s architecture and decision-centric design remain more aligned with that goal.

Company history, funding, and commercial profile

Founding and locations

Pecan AI was founded by CEO Zohar Bronfman and CTO Noam Brezis in Israel and later expanded to New York. Early coverage in 2020 described Pecan as a four-year-old company founded in 2016, framing it as a platform that lets analysts build predictive models without in-house data science teams.² More recent materials, including Pecan’s own Predictive GenAI launch press release, state that the company was founded in 2018.⁵

This discrepancy—2016 vs 2018—likely reflects a difference between early R&D/incorporation versus the current corporate entity, but public sources do not clarify the exact legal timeline. What can be said with confidence is that Pecan emerged publicly around 2019–2020 with an AutoML-style predictive analytics platform and has since maintained offices in Tel Aviv (or Ramat Gan) and New York.²⁴

Funding rounds and investors

Pecan is a heavily venture-backed company. Publicly reported funding includes:

Seed / Series A (~$15M, 2020)
- In 2020, the company announced an $11M Series A led by Dell Technologies Capital and S-Capital, bringing total funding to $15M when combined with prior seed investment.² Coverage at the time emphasized Pecan’s mission to bring predictive analytics to companies lacking in-house data scientists, and positioned the product as an easier alternative to building custom models in Python.
Series B ($35M, 2021)
- Subsequent news and investor materials mention a $35M Series B round in 2021 (co-led by GGV Capital and others), though this round is less extensively documented in the sources reviewed. Aggregate funding totals cited after the Series C strongly imply a B-round of this magnitude when back-calculating from the total.¹³⁴
Series C ($66M, 2022)
- In February 2022, BusinessWire reported a $66M Series C round led by Insight Partners with participation from GV (Google’s venture arm), Dell Technologies Capital, S-Capital, and Vintage Investment Partners.¹ Pecan’s own statements and third-party coverage (e.g., The Jerusalem Post and Built In NYC) confirm this figure and cite a cumulative funding total of roughly $116–117M.¹³⁴

The investor syndicate (Insight, GV, Dell, etc.) is consistent with a mid-stage enterprise software company in growth mode. The Series C press materials claim that Pecan had “dozens of customers” across industries, generating “tens of millions of predictions every day,” but these metrics are self-reported and not independently audited.¹³

Headcount and recognitions

Precise headcount fluctuates and is hard to verify. Built In NYC’s 2022 piece cites around 80–90 employees across Israel and the U.S., while later press materials reference teams serving customers in fintech, insurance, retail, CPG, mobile gaming and other verticals.¹⁴

In the supply chain context, the August 2025 DemandForecast.ai launch press release highlights that Pecan was named a 2025 Gartner “Cool Vendor in Cross-Functional Supply Chain Technology,” with Gartner’s rationale focusing on the no-code platform, GenAI-driven guidance, and rapid, explainable model deployment.¹⁰ “Cool Vendor” status signals industry interest but does not in itself demonstrate depth of adoption or technical superiority; Gartner’s own disclaimer stresses that these designations reflect analyst opinion rather than fact.¹⁰

Taken together, Pecan appears to be a mid-stage, growth-phase SaaS vendor: well funded, with a credible investor list and some brand recognition, but not yet a dominant player in any single domain (including supply chain), at least based on publicly verifiable evidence.

Technology and product architecture

Core data science pipeline and AutoML engine

The clearest technical details about Pecan’s internals come from its own help-center article, “Pecan’s Data Science: A Peek Behind the Scenes,” which describes an automated pipeline for feature engineering, model training, and model selection.⁹ Key points:

Automated feature engineering
- For continuous numerical variables, Pecan automatically derives statistics such as mean, standard deviation, min/max, mode, and coefficients of linear trends over historical values.⁹
- For categorical variables, it identifies common categories and applies encodings such as one-hot, ordinal and target encoding depending on data distribution.⁹
- For dates, it extracts features like day of week, month, seasonal patterns and distances between dates.⁹
- It also uses denoising autoencoders and other unsupervised methods (e.g., clustering) to create higher-level features (e.g., “lookalike” groups), and applies feature selection methods such as variance thresholds, correlation checks, permutation importance and SHAP values.⁹
Model zoo and selection
- Pecan lists the following “best-in-class” algorithms as supported: time-series LSTMs, ARIMA, Prophet, and tree-based models such as LightGBM and CatBoost.⁹
- Hyperparameter search is done via Bayesian optimization, typically preferring tree-based models for tabular problems, with a validation split around 10% of training data.⁹
- Different loss functions (e.g., log-loss, Tweedie) are evaluated depending on task and data distribution.⁹
Optimization metrics
- Users can select the metric the AutoML engine optimizes (e.g., accuracy, F1, business-specific metrics), reinforcing that the platform is general purpose rather than specialized to a single vertical.⁹

There is no public discussion of bespoke forecasting algorithms beyond this model zoo; the stack is essentially a well-engineered orchestration of mainstream ML methods. From a technical standpoint, that is entirely reasonable—Sophisticated feature engineering plus strong tree-based models are state-of-practice for tabular problems—but it is not state-of-the-art in the sense of new algorithms.

Notably, Pecan’s public materials do not claim to implement probabilistic forecasts in the sense of full demand distributions (e.g., quantile grids over lead time) for supply chain. Loss functions and metrics may allow some probabilistic outputs (e.g., Tweedie, quantile regression), but this is not documented as a core supply-chain-specific feature.

Predictive GenAI: chat and notebook

In January 2024 Pecan announced “Predictive GenAI,” framing it as an infusion of generative AI into its existing predictive platform.⁵⁶ The launch blog and press coverage describe two primary components:

Predictive Chat – a chat interface where users describe a predictive goal (e.g., “which customers are at risk of churn?”) in natural language. The LLM then asks follow-up questions and formulates a “predictive question” suitable for modeling.⁶
Predictive Notebook – based on the chat, the system generates a notebook containing auto-generated SQL queries that define the training dataset. Users can inspect and edit the SQL before proceeding to model training.⁶

Third-party articles (TDWI, FutureOfWorkNews) essentially restate this: Predictive GenAI uses generative AI to guide users from business question to tabular dataset and model, combining the conversational affordances of LLMs with Pecan’s existing AutoML backend.⁷⁸

Technically, this is an LLM-powered UX layer on top of a classical ML engine:

The LLM interprets natural-language input and drafts SQL and configuration.
The underlying training, feature engineering, model selection, and scoring use the same supervised learning techniques described in Pecan’s data science article.⁹

Pecan’s materials present Predictive GenAI as “state-of-the-art,” but the novelty lies in the integration and productization, not in the underlying ML algorithms, which remain mainstream. There is also no public explanation of guardrails, evaluation of LLM-generated SQL, or failure modes of the chat (e.g., hallucinated columns); those aspects must be inferred.

DemandForecast.ai: supply-chain-specific packaging

DemandForecast.ai, launched in August 2025, is a vertically branded product built on Pecan’s Predictive GenAI platform and aimed at “supply chain leaders.”¹⁰¹¹ The PR and microsite describe:

Target users: medium-to-large enterprises with complex supply chains, especially in retail and consumer goods.¹⁰¹²
Key capabilities:
- “Live in weeks, not months” – accelerated onboarding and model deployment.¹⁰
- “Business-user-first design” – forecasts presented directly to planners with no coding required.¹⁰¹¹
- “Explainable AI” – GenAI-powered explanations and guidance around forecasts.¹⁰¹²
- “Enterprise-grade integration” – connectors or APIs to ERP and planning systems.¹⁰

Third-party write-ups (AI TechPark, TechIntelPro, and others) essentially echo the press release, highlighting the “trillion-dollar forecasting gap,” the Gartner Cool Vendor recognition, and the focus on explainability and GenAI-driven guidance.¹²¹³¹⁴

Crucially, public materials stop at the forecasting layer:

They detail that DemandForecast.ai produces accurate, explainable forecasts and surfaces them in a GenAI-guided UI.
They do not describe the mathematical translation of these forecasts into order quantities, safety stocks, network flows, production plans, or price recommendations.
There is no public mention of probabilistic forecasts over lead time, multi-echelon inventory optimization, or cost-based objective functions that would be typical in a supply-chain-specific optimization engine.

On the evidence available, DemandForecast.ai should be viewed as “forecasting plus explanations for planners”, not as a full prescriptive optimization suite.

AI claims and evidence

Pecan’s branding leans heavily on “Predictive AI” and “Predictive GenAI.” Evaluating these claims:

Classical ML: The AutoML pipeline is well documented and squarely in line with current industry practice: feature engineering over numeric/categorical/date types, autoencoder-based dimensionality reduction, SHAP-based feature importance, and a model zoo of LSTMs, ARIMA, Prophet, and tree-based methods with Bayesian hyperparameter optimization.⁹ This is solid and modern, though not particularly unique—similar combinations exist in many AutoML frameworks.
Generative AI: The Predictive GenAI launch and follow-up blog posts make clear that generative AI is used at the UX and orchestration layer, not as a forecasting model per se.⁵⁶ LLMs help express business problems in predictive form and generate SQL notebooks; the heavy lifting of forecasting remains with the classical models described above. Third-party coverage does not add independent evidence beyond reiterating Pecan’s claims.⁷⁸
Scale and impact: Pecan PRs claim the platform runs “over 30 million predictions per day” for customers across industries such as fintech, insurance, retail, consumer goods, and mobile gaming.⁵ These numbers are plausible for a SaaS predictive service but are self-reported; there is no external audit or benchmark comparable to, say, public forecasting competitions.

Overall, Pecan clearly has a credible, engineering-heavy implementation of mainstream supervised learning, and there is no reason to doubt that it can automate a lot of feature engineering and model-selection work. However, there is little public evidence of proprietary, state-of-the-art algorithms beyond the orchestration of known techniques, and the “GenAI” side is primarily a usability enhancement rather than a new learning paradigm.

Clients, case studies, and commercial maturity

Named customer references across Pecan’s materials include companies such as Johnson & Johnson, Creative Artists Agency (CAA), Ideal Image, and SciPlay, among others.¹⁵ These are presented as users of Pecan’s platform for various predictive tasks; however, the level of detail is limited and there are no deeply technical case studies in public domain that, for example, show uplift curves, hold-out validation, or post-deployment business impact in a rigorous way.

The Predictive GenAI press release includes a testimonial quote from a Kenvue executive (Global Consumer Supply Chain), praising Pecan’s approach to bringing advanced data science to more business teams.⁵ This confirms that at least some supply-chain-adjacent stakeholders are engaged, but again, the evidence is anecdotal and framed as marketing endorsement.

For DemandForecast.ai specifically:

The launch PR emphasises the magnitude of global forecasting errors (citing IHL Group and Harvard Business Review for $1.7T+ in annual stockout/overstock losses) and positions DemandForecast.ai as a remedy via better forecasts.¹⁰
It does not name specific supply chain clients for the product; nor do subsequent articles add concrete case examples beyond the Gartner Cool Vendor mention.¹⁰¹²¹³¹⁴

From this, a cautious assessment is that Pecan is commercially credible as a generic predictive analytics vendor, with recognizable brands among its users, but its supply-chain-specific track record is still emerging and not well documented in public sources.

Critical assessment for supply chain use

From a supply chain decision-maker’s perspective, Pecan’s strengths and limitations can be summarized as follows:

Strengths

Modern AutoML for tabular/time-series data: The documented pipeline (feature engineering, SHAP-based importance, LSTMs, ARIMA, Prophet, tree-based ensembles, Bayesian optimization) should be capable of producing strong point forecasts given sufficient data quality.⁹
Accessibility and UX: Predictive GenAI’s chat + notebook flow can meaningfully lower the barrier for BI teams who want to move from retrospective reporting to predictive models without hiring data scientists or building MLOps infrastructure.⁶⁷
Rapid experimentation across use-cases: Because the platform is cross-vertical, a company can apply Pecan to multiple predictive problems (churn, LTV, demand) with a single vendor, potentially speeding up experimentation if the data is already ingested.

Limitations / open questions

Forecasts vs decisions: Public information focuses on forecasting accuracy and explainability, not on end-to-end decision optimization (inventory policies, multi-echelon stock placement, capacity allocation, pricing, etc.). DemandForecast.ai appears to stop at the forecast and explanatory layer.¹⁰¹¹¹² Any prescriptive decisions would need to be implemented either manually by planners or via additional tools.
Probabilistic modelling depth: There is no explicit support documented for full probabilistic demand distributions over lead-time—or for directly optimizing service levels or expected costs under uncertainty—which are central to advanced supply chain optimization. Loss functions such as Tweedie can approximate some aspects, but that is not the same as a distribution-aware optimization engine.
Evidence for “state-of-the-art”: While Pecan’s pipeline is modern and thorough, it is built from standard components widely used in the ML community; public sources do not showcase novel forecasting algorithms or optimization techniques specific to supply chain. “State-of-the-art” in this context mainly refers to using current mainstream methods in a production-grade AutoML system, not to advancing the algorithmic frontier.
GenAI reliability and governance: Pecan does not disclose technical details about how Predictive GenAI prevents erroneous SQL generation, ensures schema consistency, or manages versioning of LLM-generated artifacts. For regulated or safety-critical supply chains, this lack of transparency may be a concern.

Given these points, Pecan’s current sweet spot appears to be organisations that:

Want to deploy predictive models quickly across several business domains.
Have BI or analytics teams comfortable with SQL but not with building full ML pipelines.
Are primarily looking for better forecasts and predictive scores, rather than deep, mathematically rigorous optimization of supply chain decisions.

For companies whose primary need is decision-centric supply chain optimization under uncertainty, Pecan could serve as a forecasting component or experimentation environment, but would likely need to be complemented (or replaced) by more specialized tools to reach best-in-class performance in inventory and network optimization.

Conclusion

Pecan AI offers a credible, modern predictive analytics platform with a strong emphasis on accessibility: automated feature engineering, a well-curated set of standard ML models, and a generative-AI-driven interface that guides users from business question to SQL notebook and trained model. Its funding history and investor base indicate that it is a serious, mid-stage SaaS vendor rather than an early experiment, and its cross-vertical focus makes it attractive to organisations seeking to apply predictive models across multiple business problems with limited data-science resources.¹²⁵⁴

However, from a strictly technical and supply-chain-specific perspective, the publicly documented evidence does not support viewing Pecan as a state-of-the-art supply chain optimization system. Its algorithms are standard supervised learning methods, not novel probabilistic or optimization techniques; its GenAI features primarily enhance UX; and its DemandForecast.ai product focuses on demand-side forecasting and explainability without clearly documented prescriptive decision logic.⁹¹⁰¹¹¹² This does not diminish the platform’s usefulness as an AutoML-plus-GenAI environment—especially for organisations early on their predictive analytics journey—but it places realistic bounds on what a supply chain organisation should expect.

In direct comparison with Lokad, Pecan is a generalist predictive tool newly entering the supply chain forecasting space, while Lokad is a specialist decision-optimization platform whose architecture, algorithms, and case-law are tightly coupled to supply chain economics and uncertainty. The choice between them should be driven by whether the primary need is “we need to build predictive models more easily” (where Pecan is competitive) or “we must optimize complex, uncertain supply chain decisions for economic performance” (where Lokad’s approach is structurally more aligned).