The DeepVero view
The thesis in one minute
Lark is building an operational agent layer for wholesale distributors: it connects to existing systems, learns a distributor's rules and historical decisions, and helps run purchasing, forecasting, inventory, sales, and analytical workflows without replacing the ERP.
The wedge is attractive because the ERP is usually the system of record, not the complete decision system. Employees still resolve mismatched records, inspect long reorder lists, interpret customer context, and carry operating knowledge in their heads. If Lark can encode those decisions safely, it can sell against labor, working-capital leakage, and missed revenue—not merely software convenience. The open questions are integration repeatability, measured ROI, permissioning, and whether incumbents or vertical AI competitors absorb the same control point.
“Your ERP records transactions. Your team still makes the decisions.”
Lark · See the operating-gap argument →
01 · Problem
The transaction is digitized; the judgment around it often is not
A wholesale distributor can have an ERP, warehouse system, CRM, spreadsheets, email, and supplier portals while still relying on people to reconcile what those systems mean. A buyer deciding whether to replenish a SKU may inspect on-hand units, customer backorders, open purchase orders, supplier constraints, recent demand, and exceptions before acting. Sales staff perform a similar synthesis around quotes, product availability, pricing, and follow-up.
This work is costly because it is repetitive but not perfectly deterministic. The obvious cases consume attention; the unusual cases depend on tacit knowledge. Delays can translate into stockouts, excess inventory, slow quotes, stale CRM records, and decisions that are difficult to audit later.
- Records live across systemsERP, warehouse, CRM, email, and spreadsheets each hold part of the answer.
- People reconcile contextOperators resolve inconsistent data and business exceptions manually.
- Actions finally happenPurchasing, inventory, and sales decisions wait for that reconciliation.
The scale of the underlying decisions is material. The 2022 Economic Census reports about $864 billion of year-end wholesale inventory across covered establishments. That does not measure Lark's addressable market, but it shows why small improvements in purchasing and inventory decisions can matter economically. Census inventory table →
02 · Product
A controlled agent that observes, recommends, acts, and learns
Lark presents a conversational interface above existing operational systems. Its published workflows cover purchasing, demand forecasting, inventory monitoring, sales operations, and data analysis. In purchasing, for example, the product prioritizes shortages, shows its reasoning, recommends a quantity, and prepares a purchase-order draft and supplier email for approval.
The implementation process is consultative: discover a manual workflow, connect the relevant systems and data, configure rules and exceptions, run a controlled human-in-the-loop pilot, then expand. Lark says most setups take one to two months and that customer data is isolated and not used to train models for other customers.
- Read operational dataBring ERP and distributor systems into a shared context and rules layer.
- Analyze and draftA specialized agent proposes a purchasing, inventory, or sales action.
- Approve or handle exceptionsA human remains accountable for controlled execution.
- Write back and learnRecord the action and use its outcome in later decisions.
Purchasing
Prioritize shortages, inspect stock and commitments, recommend quantities, and draft purchase orders and vendor messages.
Forecasting and inventory
Compare demand windows, flag anomalous records, surface stock risk, and explain why an item requires attention.
Sales operations
Monitor sales inboxes, match messages to opportunities, update CRM stages, and surface the next follow-up.
Data analysis
Answer plain-English questions against operational data while exposing the records and caveats behind the answer.
03 · Why now
Models can interpret messy context while the installed ERP remains hard to replace
Three conditions make the timing plausible. First, language models can now interpret emails, explanations, product text, and semi-structured records that conventional workflow automation handles poorly. Second, tool use allows a model to move from answering a question to preparing or executing a bounded action. Third, many distributors already have decades of transaction history but cannot justify a disruptive ERP replacement.
Data already exists
Sales, purchase, inventory, and customer records can ground recommendations without creating a new system of record.
Exceptions are legible
Modern models can combine structured records with email, notes, and local operating language.
Human control is deployable
Draft-and-approve workflows can create value before a customer is willing to allow autonomous write-back.
ERP coexistence lowers disruption
An overlay can enter through one workflow rather than asking the distributor to replace its operational backbone.
04 · Opportunity
A large firm base, narrowed by fit, adoption, and contract value
The 2022 Economic Census counted 232,225 merchant-wholesaler firms, excluding manufacturers' sales branches and offices. Not every firm has enough workflow volume, usable data, integration readiness, or budget for a specialized agent. The model below therefore applies explicit fit and adoption filters rather than treating every wholesaler as a customer.
| Scenario | Operational-fit share | Adoption in fit segment | Annual contract value | Implied customers | Annual pool |
|---|---|---|---|---|---|
| Conservative | 15% | 5% | $18,000 | ≈ 1,742 | ≈ $31M |
| Base | 30% | 15% | $48,000 | ≈ 10,450 | ≈ $502M |
| Upside | 50% | 30% | $120,000 | ≈ 34,834 | ≈ $4.2B |
DeepVero estimate Fit, adoption, contract values, and implied revenue are analytical assumptions—not Lark guidance, current market size, or a forecast. Public pricing is custom, and no customer, revenue, retention, or contract-size data were disclosed when reviewed. The Census denominator includes a wide range of firm sizes and verticals, so the model is most useful as a sensitivity table. Census firm count →
The upside case requires Lark to become a multi-workflow operating layer, not a single-task assistant. The lower cases can still support a meaningful company if initial workflows produce provable working-capital, labor, or revenue gains and expand inside the same account.
05 · Buyers and go-to-market
Land with one painful workflow; expand across the operating loop
| Buyer or user | Job to be done | Proof required |
|---|---|---|
| COO / general manager | Increase throughput without adding equivalent headcount | Cycle-time reduction, exception coverage, adoption |
| Purchasing leader | Review shortages and reorder decisions faster | Fewer stockouts, stable service level, no excess inventory |
| Inventory / supply-chain leader | Find mismatches and prioritize attention across many SKUs | Alert precision, working-capital impact, explainability |
| Sales operations | Keep quotes, inboxes, CRM stages, and follow-up synchronized | Faster response, cleaner CRM, conversion lift |
| IT / ERP owner | Add automation without destabilizing the system of record | Security, permissions, audit trail, integration reliability |
Company-reported Lark says it begins by mapping the workflow and controls, then runs a controlled live pilot with a human in the loop. Most setups reportedly take one to two months. This implies a founder-led, implementation-heavy motion today, with expansion after one workflow works as expected.
DeepVero estimate Purchasing is a strong initial wedge because the decisions are frequent, quantifiable, and connected to both customer service and inventory capital. Expansion can then move into forecasting, monitoring, sales, and analysis using the same integrations.
06 · Competition and moat
The contest is for the decision layer above distribution data
| Alternative | Strength | Opening for Lark |
|---|---|---|
| Distribution ERPs such as Epicor | System-of-record position, workflow breadth, installed trust, native data | A focused agent can iterate across systems and local exceptions without an ERP migration |
| Distribution AI platforms such as Proton.ai | Vertical data model, customer references, sales and commerce products | Lark can compete on custom operational workflows and cross-functional agency |
| Transaction automation such as Conexiom | Production scale in orders, RFQs, acknowledgements, and invoices | Lark targets judgment-rich decisions beyond document conversion |
| Generic copilots and automation platforms | Broad ecosystems, procurement leverage, flexible tooling | Less distributor-specific context, workflow packaging, and implementation knowledge |
| Internal operations teams | Deep local knowledge and trusted exception handling | Knowledge is difficult to scale, audit, and preserve when employees change |
The potential moat is encoded operating judgment—not model access
The foundation models, chat interface, and basic ERP connectors are replicable. A stronger advantage would accumulate from distributor-specific decision histories, reusable vertical schemas, integration mappings, approval policies, exception taxonomies, outcome feedback, and deployment playbooks. Each implementation should make the next one faster while making the current customer harder to displace.
The counter-risk is services gravity. If every ERP instance, product taxonomy, and local rule requires bespoke engineering, customer value may be real while software margins and deployment speed remain weak. Lark needs a growing reusable core beneath customer-specific configuration.
07 · Evidence and traction
The product story is specific; commercial proof is still absent
Detailed workflows
The public product shows concrete purchasing, forecasting, inventory, sales, and analysis flows rather than a generic “AI for operations” claim.
Controlled deployment
Published implementation includes rules, exceptions, approvals, historical testing, and a human-in-the-loop pilot.
Large operational base
Census data counts 232,225 merchant-wholesaler firms and roughly $864 billion of wholesale inventory at 2022 year-end.
Relevant founder experience
YC says CEO Jason Wang previously built agentic systems for industrial supply chain and a stream-processing control plane.
No named customer, paid-customer count, revenue, growth, retention, production action volume, independently measured time savings, inventory improvement, model-error rate, integration count, or security certification was found as of September 23, 2026. Lark's examples are labeled illustrative; they demonstrate product intent, not customer outcomes.
08 · Risks
Six failure modes define the investment case
- Integration drag: legacy ERPs, custom fields, inconsistent master data, and weak APIs can turn every deployment into a consulting project.
- Decision liability: a wrong purchase quantity, inventory interpretation, price, or customer commitment can create direct financial damage.
- Trust and adoption: experienced operators may reject recommendations that are hard to explain or that disregard local exceptions.
- Incumbent response: ERP vendors and established distribution AI platforms already own data, workflow, customer references, and procurement relationships.
- Weak data foundations: conflicting inventory records and undocumented rules are the opportunity, but they also degrade agent accuracy and evaluation.
- Services-heavy economics: custom discovery and configuration can slow growth and suppress gross margin unless patterns become reusable.
09 · Investment thesis
A valuable control point, if implementations compound into software
What to believe: wholesale distribution contains a broad class of recurring, high-consequence decisions that sit between systems. Lark has selected a coherent vertical and described product flows that keep people in control while allowing automation to deepen over time.
What remains unproven: that early pilots produce repeatable ROI, that deployments can be standardized across heterogeneous distributors, and that Lark can accumulate a proprietary decision layer before ERP and vertical-AI incumbents close the gap.
Signals that would strengthen the thesis
- Named production customers with before-and-after purchasing or inventory metrics
- Deployment time falling while the number of supported ERPs and workflows rises
- High recommendation acceptance with low costly-error and override rates
- Expansion from one workflow into multiple departments inside the same distributor
- Evidence that reusable schemas and policies reduce implementation labor
- Security controls, auditability, permissioning, and enterprise references
Signals that would weaken the thesis
- Pilots remain dashboards or drafts without trusted system write-back
- Each new customer requires extensive custom engineering and data cleanup
- Inventory or purchasing gains disappear after controlling for operator selection
- Users override recommendations frequently or cannot understand the reasoning
- ERP vendors bundle comparable agents into existing contracts
- Customer concentration and implementation staffing grow faster than recurring revenue
This profile is an analytical company teardown, not investment advice. Lark is private, company operating data is limited, and scenario values are illustrative.