AI Production Planning for Manufacturing: Cut Lead Times Without Adding Headcount

Your ERP has a scheduling module. You use it. And yet your production floor still runs hot — expedited orders, machines sitting idle while queues pile up elsewhere, and lead times that slip by days because a single constraint nobody modelled was never caught until it was too late.

This is not an ERP problem. It is a scheduling intelligence problem. And the distinction matters enormously before you spend money trying to solve it.

What ERP Scheduling Actually Does (And Where It Stops)

Every major ERP, including ERPNext, provides production scheduling functionality. Work Orders get planned start and end dates. Bill of Materials drives material requirements. Routing defines which work centres handle each operation. Capacity planning shows you total hours against total availability.

What ERP scheduling does not do: it does not reason about sequencing. It assigns dates within capacity envelopes, but it does not answer the question "given everything that needs to happen today across twelve work centres, what is the optimal order to run these jobs to minimise total flow time and protect my most constrained resource?"

That reasoning gap is where lead time bleeds away. A facility running 85% capacity utilisation with poor sequencing will consistently underperform a facility at 92% utilisation with intelligent sequencing. The difference is not headcount. It is the logic applied to the decision of what to build next.

Industry benchmark: manufacturers relying solely on ERP scheduling report average schedule adherence of 61–68%. Facilities using constraint-based AI scheduling report 84–91% adherence. That 20-point gap translates directly to lead time, customer promise accuracy, and revenue per machine hour.

How AI Production Planning Actually Works

The production planning agent Techseria builds on LangGraph.js does three things that ERP scheduling does not.

First, it models the constraint, not just the capacity. Theory of Constraints (TOC), developed by Eliyahu Goldratt, holds that every production system has exactly one binding constraint at any moment — the resource whose capacity limits overall system throughput. All scheduling logic should subordinate to protecting and exploiting that constraint. Standard ERP capacity planning treats all work centres as equally important. An AI agent identifies the current constraint dynamically, then builds sequences that ensure the constraint never starves.

Second, it sequences at the job-operation level, not the order level. ERP plans Work Orders. The AI agent plans individual operations within those orders across all work centres simultaneously, resolving sequencing conflicts before they reach the floor.

Third, it responds to real-time state. If a machine goes down at 10:00 AM, the agent re-sequences the remaining queue by 10:05 AM. ERP schedules are typically static for the day or the week.

The Data the Agent Needs from ERPNext

The agent connects to ERPNext via its REST API and reads six data sets at the start of each scheduling cycle:

Work Orders — open orders with item code, quantity, planned start, required by date, and current status. ERPNext API: `/api/resource/Work Order`.

Bill of Materials — routing operations, work centre assignments, standard times per operation. ERPNext API: `/api/resource/BOM`.

Work Centre Capacity — available hours per shift, current breakdown status, efficiency factors. ERPNext API: `/api/resource/Work Center`.

Stock Levels — raw material and WIP availability to confirm each Work Order can actually start. ERPNext API: `/api/resource/Bin`.

Shift Patterns — planned shift schedule by work centre, including planned maintenance windows. ERPNext API: `/api/resource/Shift Type`.

Open Purchase Orders — for materials not yet received, to understand when they will be available and whether any orders need to be held. ERPNext API: `/api/resource/Purchase Order`.

The agent pulls these on a configurable cadence — typically every 15 minutes during production hours — and re-solves the scheduling problem if the current sequence deviates materially from the previous solution.

The LangGraph.js State Machine

The scheduling agent is implemented as a LangGraph.js state machine with the following nodes:

1. Data Ingestion Node — pulls all six data sets from ERPNext APIs, validates completeness, flags missing data.
2. Constraint Identification Node — calculates workload-to-capacity ratio across all work centres for the planning horizon, identifies the binding constraint.
3. Sequence Optimisation Node — runs the constraint-based scheduling algorithm. For the constraint work centre, sequences jobs using the Critical Ratio (CR) method weighted by margin and due date. For non-constraint work centres, schedules to feed the constraint without creating unnecessary WIP buffers.
4. Conflict Resolution Node — detects scheduling conflicts (material not available, work centre double-booked) and resolves them by adjusting start times or flagging to the planner.
5. ERPNext Write Node — pushes the optimised schedule back to ERPNext as updated Work Order planned start dates and operation sequences.
6. Planner Review Node — for any order where the algorithm cannot find a feasible schedule within the required date, surfaces an exception to the production planner with the specific constraint causing the infeasibility.

The state machine runs autonomously for the majority of scheduling decisions. Planners see a dashboard showing schedule adherence, constraint utilisation, and the exception queue. They interact with exceptions rather than with the schedule itself.

Real-World Impact Metrics

Across Techseria's manufacturing clients implementing AI production planning:

34% lead time reduction — measured as average days from Work Order creation to completion, compared to the 90-day baseline period before implementation. The reduction comes primarily from eliminating queue time at the constraint and reducing WIP that accumulates when non-constraint work centres run ahead of the constraint.

22% OEE improvement — Overall Equipment Effectiveness gains split roughly equally between Availability (fewer expedited changeovers disrupting planned runs) and Performance (better sequencing reduces minor stoppages from material not being ready).

Schedule adherence from 64% to 87% — measured as percentage of Work Orders completed within one day of planned completion date.

These are not theoretical. They are measured outcomes from ERPNext-integrated deployments in discrete and process manufacturing environments with 20–200 work centres.

What Needs to Be True Before You Start

The agent needs reliable data in ERPNext. Three conditions matter most:

Routing accuracy — your BOM needs accurate operation times. If standard times are wrong by more than 20%, the schedule optimisation will be imprecise. Techseria's implementation process includes a routing audit in weeks 1–2 to identify and correct the most consequential inaccuracies.

Work centre status visibility — breakdowns need to be logged in ERPNext promptly. If maintenance events take hours to appear in the system, the agent is scheduling against phantom capacity. This often requires a small process change with the maintenance team.

Work Order discipline — orders need to be created and updated in ERPNext as they progress. Shops where planners keep shadow spreadsheets alongside ERPNext will need to consolidate before the agent can be effective.

Implementation Timeline: 8–10 Weeks

Weeks 1–2: Data audit and ERPNext configuration review. Assess routing accuracy, identify data gaps, configure API access, establish baseline metrics (current lead time, OEE, schedule adherence).

Weeks 3–5: Agent development. Build LangGraph.js state machine, develop constraint identification logic, implement sequencing algorithm calibrated to the facility's specific product mix and work centre configuration.

Weeks 6–7: Integration and testing. Connect agent to ERPNext APIs, run parallel scheduling (agent generates schedule alongside current process), compare outputs, tune parameters.

Week 8: Controlled rollout. Agent takes over scheduling for a defined subset of products or work centres. Production planning team monitors, provides feedback.

Weeks 9–10: Full deployment and handover. Full work centre coverage, exception handling protocols established, planner training complete, performance dashboard live.

Investment: £20,000–£40,000 Fixed Fee

The range reflects facility complexity:

• £20k–£28k: Up to 15 work centres, single production site, ERPNext already in good data shape.
• £28k–£40k: 16–50 work centres, multi-shift complexity, routing audit required, or process manufacturing with variable batch parameters.

Annual maintenance (software updates, algorithm retuning, ERPNext version compatibility): 15–18% of build cost.

At 34% lead time reduction for a facility currently at 18-day average lead times, the value recovered in customer promise accuracy, expedited freight cost reduction, and WIP capital release typically exceeds £200k–£800k annually for a mid-market manufacturer — depending on revenue, margin, and current inefficiency baseline.

The Planner's Role Changes, Not Disappears

A common concern: will the AI replace the production planner? No. The planner's role shifts from daily schedule construction — which is manual, time-consuming, and impossible to fully optimise by hand — to exception management and continuous improvement.

Planners interact with the agent's exception queue. When the agent cannot find a feasible schedule for an order, it explains why (material shortage, capacity infeasibility, conflicting customer priority) and presents options. The planner decides. The agent executes. This is faster, less stressful, and more accurate than the current state where planners are simultaneously firefighting and trying to build tomorrow's schedule.

Techseria designs every manufacturing AI engagement with human-in-the-loop architecture as a first-order requirement. The agent acts autonomously within confidence thresholds; consequential decisions require human approval.

Is This the Right Next Step for Your Facility?

If your production floor suffers from any of the following, AI production planning will deliver measurable ROI within the first quarter post-deployment:

• Lead times that are difficult to quote accurately because the schedule shifts too frequently
• Machine capacity that looks adequate on paper but still produces backlogs
• Planners spending the majority of their time rescheduling rather than planning
• Customer complaints tied to delivery date misses rather than quality

If your ERPNext data is substantially incomplete — routing times are guesses, Work Orders are not updated as jobs progress — the implementation will include a data remediation phase. Techseria's team has executed this for multiple clients; it adds 2–3 weeks and £3k–£8k but is necessary to ensure the agent has reliable inputs.

Book a production planning scoping call. Techseria will review your current ERPNext configuration, identify the data readiness gaps, and give you a fixed-fee proposal within 5 business days. No obligation. No vague "discovery" engagement — a specific cost and timeline for your facility.

[Book a Strategy Session] | [Get a Fixed-Fee Quote]

Three Questions to Ask Before Engaging Any AI Scheduling Vendor

The market for production scheduling AI includes a wide range of vendors — some built specifically for manufacturing, some repurposed from logistics or supply chain contexts. Before engaging any provider, three questions cut through the noise:

1. Does the scheduling logic model the constraint explicitly? If the vendor cannot explain how their system identifies the binding constraint and subordinates non-constraint scheduling to it, their system is doing capacity planning, not constraint-based scheduling. The distinction determines whether you get 10% lead time improvement or 34%.

2. Does it write back to your ERP, or is it a parallel system? Scheduling systems that generate their own separate schedule alongside ERP create a two-system problem: planners must manually translate recommendations into ERP, which introduces errors and delay. The agent should read from and write to ERPNext directly, with the ERP as the single source of truth.

3. What happens when data is missing or wrong? Any real production environment has data gaps. The question is whether the agent handles them gracefully (flags the gap, uses conservative defaults, alerts the planner) or fails silently (produces a schedule that looks correct but is based on phantom data). Ask to see the error handling and exception logic before you commit.

Techseria's scheduling agent is built to satisfy all three requirements. The answer to each question is documented in the technical specification delivered at project kickoff.