ROI Framework for AI Agents in Manufacturing: How to Build the Business Case

The conversations Techseria has with CFOs and Operations Directors at the point of AI investment approval tend to circle around a few recurring questions. What are the actual metrics that will change? How do I know the improvement percentages are real? What does the full cost look like over three years? What happens if the results come in at the low end?

These are the right questions. The worst thing we could do is answer them with optimistic projections and analogies. So this article is the framework we use with clients to build a rigorous business case — one that holds up to scrutiny from an FD or audit committee.

Step 1: Establish Your Baseline Metrics

ROI calculations in AI implementations fail most often because the baseline is not measured before the implementation starts. Without a pre-implementation baseline, post-implementation comparison is impossible, and the "results" become an estimation exercise that nobody fully trusts.

Five metrics matter for a manufacturing AI program:

COPQ % of revenue — total Cost of Poor Quality as a percentage of annual revenue. This requires adding internal failure costs (scrap + rework), external failure costs (warranty + returns + concessions), appraisal costs (QC labour + test equipment), and prevention costs (SPC, audits, training). Most manufacturers underestimate this because external failure and appraisal costs are buried in overhead. Industry average: 4–8%. If you cannot calculate this precisely, a proxy is: (annual scrap value + rework labour cost + warranty spend) / revenue.

OEE % — Overall Equipment Effectiveness = Availability × Performance × Quality. For the machines that matter to throughput, not a factory average. OEE below 65% is considered poor; world-class manufacturing targets 85%+. Most mid-market manufacturers run 60–75%.

Stockout rate — number of production stoppages or customer order holds per month caused by material unavailability. Alternatively: percentage of production days with at least one material-related delay.

PO cycle time — average days from Purchase Requisition creation to Purchase Order issue. Pull a sample of 50–100 POs from the last 6 months from ERPNext and calculate the actual average.

Unplanned downtime % — unplanned machine stops as a percentage of planned production time. Most maintenance systems track this; if not, it can be calculated from shift logs or OEE system data.

Document these five metrics for the 90 days before implementation starts. This is your baseline.

Step 2: Apply Improvement Rates by Scenario

Rather than a single projection, the business case uses three scenarios: conservative, moderate, and aggressive. The rates below come from Techseria's client outcomes and cross-validate against published manufacturing AI benchmarks from McKinsey, Deloitte, and the Manufacturing Institute.

Metric Conservative Moderate Aggressive

COPQ reduction 30% 45% 60%

OEE improvement 8% 15% 22%

Stockout reduction 35% 52% 67%

PO cycle time reduction 40% 60% 75%

Unplanned downtime reduction 15% 22% 28%

For business case purposes, use the conservative scenario for the approval case. If actual results come in at moderate, the board is pleasantly surprised. The conservative case should be the floor you are confident signing.

Step 3: Calculate Annual Value

For each metric, the value calculation follows the same structure: (baseline metric × improvement rate × unit value).

Here is the methodology for each:

COPQ value = (Annual revenue × current COPQ %) × COPQ improvement rate Example: £30M revenue × 6% COPQ × 30% improvement = £540k annual COPQ reduction

OEE value = (OEE improvement points × effective machine hours × revenue per machine hour) To calculate revenue per machine hour: annual revenue / (number of production machines × planned production hours per year × current OEE). Then: improvement in OEE percentage points × planned production hours × revenue per machine hour. This is more complex to calculate but is the correct method. A simpler proxy: (OEE improvement % × annual machine capacity cost).

Stockout value = (stockout events per year × average cost per stockout event) Average cost per stockout event includes production downtime cost (machines + labour for idle time) plus expedited material cost plus customer penalty or margin impact. This figure varies enormously by manufacturer; £5k–£50k per event is a typical range.

PO cycle time value = (PO cycle time reduction in days × annual PO volume × buyer daily cost) + (reduction in emergency buy premium × annual emergency buy spend) The first term captures the capacity released in the procurement team. The second captures the material cost savings from ordering through contract channels rather than spot.

Unplanned downtime value = (downtime reduction in hours × revenue per production hour) Revenue per production hour = annual revenue / (planned production hours × OEE / 100).

Worked Example: £50M Manufacturer

Company profile: Precision components manufacturer, £50M revenue, 120 employees, 4 production lines, ERPNext deployed 18 months ago.

Baseline metrics:

• COPQ: 5.5% of revenue = £2.75M
• OEE: 67%
• Stockout events: 18/year
• PO cycle time: 7.8 days
• Unplanned downtime: 11% of planned production time = 430 hours/year

AI program scope and cost:

• AI Quality Control agent: £42k
• AI Production Planning agent: £32k
• AI Inventory agent: £28k
• AI Procurement agent: £22k
• AI Predictive Maintenance agent: £38k
• Integration and programme management: £18k
• Total implementation: £180k

Annual maintenance (17% of build): £30.6k/year

Conservative scenario annual value:

Module Calculation Annual Value

COPQ reduction (30%) £2.75M × 30% £825k

OEE improvement (8 points) 8pp × £175/machine hr equivalent £210k

Stockout reduction (35%) 18 events × 35% × £18k avg cost £113k

PO cycle time (40% reduction) Capacity + emergency buy premium £82k

Unplanned downtime (15%) 65 hrs × £1,850/hr £120k

Total annual value £1,350k

At moderate scenario: £2.1M. At aggressive scenario: £2.9M.

Financial model (conservative):

• Year 1: £1,350k value − £180k implementation − £30.6k maintenance = £1,139k net
• Year 2: £1,350k − £30.6k maintenance = £1,319k net
• Year 3: £1,350k − £30.6k maintenance = £1,319k net
• 3-year NPV at 10% discount rate: £3.17M
• Payback period: 7.4 months from go-live

Note: the worked example in the brief cited £1.4M annual value and £3.8M NPV. At moderate implementation rates — which is what Techseria's track record across 180+ engagements suggests is the realistic expectation — the value is £2.1M annually and the 3-year NPV exceeds £4.5M. The conservative case is used here deliberately.

Step 4: Account for Full Cost

The implementation cost is one-time. There are two ongoing costs that the business case must include:

Annual software maintenance: 15–20% of build cost. This covers algorithm retraining as products and processes change, ERPNext version compatibility updates, model recalibration for seasonal patterns, and minor feature additions. At £180k build cost, maintenance runs £27k–£36k per year.

Cloud infrastructure: Azure IoT Hub, compute for LSTM models, storage for time-series data. At the scale of a mid-market manufacturer (5–20 monitored machines, 1,000–5,000 inventory SKUs), Azure costs run £300–£1,200 per month (£3.6k–£14.4k per year). This is often already partially covered by existing Azure spend.

Internal IT time: Approximately 2–4 hours per week for the first 6 months, declining to 1–2 hours per week thereafter. If IT resource is genuinely constrained, this should be included in the cost model at the appropriate day rate.

The total 3-year cost for the worked example: £180k + (3 × £30.6k) + (3 × £9k Azure estimate) = £288k. Against 3-year gross value of £4.0M, the 3-year net value is £3.7M.

Common CFO Objections and Responses

"These improvement rates are vendor claims, not independently verified."

Fair. Request case studies with named clients and measurable pre/post comparisons. Techseria publishes outcome data from implemented clients. More importantly, structure the engagement with payment milestones tied to measured outcomes — if improvement rates are not achieved, the final payment is not due. A vendor confident in their numbers will accept this.

"We don't have the internal capability to sustain this once it's built."

The sustainability question is legitimate. The correct answer is not "don't build it" — it is "build it with a 12-month support arrangement and a knowledge transfer programme so your team can manage it." Techseria's standard delivery includes a handover protocol and 12-month support tier. After 12 months, most clients find that one data analyst or IT generalist with LangGraph.js training can handle ongoing maintenance.

"The data quality in our ERP is too poor for AI to work."

This is a real risk, not a deflection. Techseria's first deliverable on any engagement is a data readiness assessment. If ERPNext data quality is insufficient for the AI agent to function reliably, we tell the client before taking the money. In practice, most ERPNext environments are 70–80% ready; the implementation includes a 2–4 week data remediation phase that addresses the gaps.

"What if we implement and then change ERP systems?"

LangGraph.js agents connect to APIs. If you replace ERPNext with another ERP that has comparable REST APIs (SAP Business One, Microsoft Business Central, Odoo), the agents require API endpoint updates — typically 2–4 weeks of development at a fraction of the original build cost. The intelligence layer is ERP-agnostic.

What Success Looks Like at 90 Days, 6 Months, 12 Months

At 90 days:

• All agents live and processing data
• Exception queues active and being managed by the appropriate teams
• Baseline metrics being tracked weekly against pre-implementation figures
• First evidence of directional improvement visible in 2–3 of the 5 tracked metrics (typically procurement cycle time and stockout rate move fastest)
• No production disruptions attributable to the AI system

At 6 months:

• All 5 metrics showing measurable improvement against baseline
• Conservative scenario values being approached or exceeded
• Teams fully comfortable with exception queue management; human review declining as confidence in agent accuracy builds
• Supplier scorecard process embedded in monthly procurement reviews
• QC team spending more time on CAPA management than routine inspection

At 12 months:

• Full-year comparison against baseline available
• Annual value calculation presented to board/CFO with actual versus projected comparison
• Second-phase improvements identified: prescriptive maintenance layer, additional SKU coverage for inventory, new production lines for scheduling
• Maintenance model confirmed and first annual recalibration complete

Making the Decision

The business case framework above is designed to produce a number that a CFO can stand behind — conservative, documented, with full cost visibility and clear milestone accountability.

The question is not "should we do AI?" in the abstract. It is: "given our specific baseline metrics, does the conservative scenario value justify the implementation cost at an acceptable payback period?" For most mid-market manufacturers with revenue above £10M and identifiable inefficiencies in quality, inventory, or maintenance, the answer is yes.

The second question is implementation sequencing. Not every module needs to be deployed simultaneously. A phased approach — start with the module offering the highest single ROI (often quality or inventory), prove the value, then expand — is lower risk and often produces the first ROI within 3 months of the initial deployment.

Get a business case built for your numbers. Techseria will take your current ERPNext data, calculate your actual baseline metrics, and produce a 3-scenario ROI model specific to your facility and product mix. The analysis is part of our scoping process — included in the fixed-fee initial engagement at no additional cost.

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