← All Blog Articles

Failure Mode and Effects Analysis (FMEA) for Software and Supply Chain Risk

The Half-Day Workshop That Turns Implicit Risk Into a Plan

Failure Mode and Effects Analysis (FMEA) for Software and Supply Chain Risk

This is one of RoadmapOne ’s articles on Objective Prioritisation frameworks .

On the night of 11 October 2026, hundreds of supermarkets across France had their fridges and freezers remotely switched to defrost or pushed to absurd temperatures. The attackers came in through internet-facing refrigeration controllers; equipment the retailers had bought, installed and largely forgotten about. A second wave hit Belgium, the Netherlands, Germany and others on 2 November. Staff arrived in the morning to find the stock ruined.

A month later, on the evening of 12 November, Microsoft 365 went down. Outlook and Teams were out across the UK into the following day, and conveyancers who confirm property completions by email were left phoning each other to get deals over the line. Some completions slipped to the following week. The mortgage money itself never touched Microsoft; the confirmations that release it did.

In neither case did the attack or outage start in the business’s own core systems. It came from further down the supply chain, and somebody could have written it on a whiteboard in advance.

Failure Mode and Effects Analysis (FMEA) is a structured method for listing the ways a process, product or supply chain can fail, then scoring each failure for Severity, Occurrence and Detection on 1-10 scales. Multiplying the three gives a Risk Priority Number (RPN), which ranks the failures so the team can fix the most dangerous ones first.

My Personal Experience

TL;DR: The value of FMEA is the workshop. It takes risk that lives implicitly in people’s heads and makes it explicit, scored and owned; then you have to do something with it.

I first used FMEA at GE Capital in the early 2000s, when Six Sigma was close to a religion there, and it has earned its keep in every business I’ve used it in since. Today I advise businesses with extreme reliability requirements, Trayport in energy trading and a core banking platform among them, where an inventory of failure modes is table stakes. Their supply chains are deep, and they run across several cloud services. Nobody can hold all of that in their head, which is exactly why you need to write it down.

What Is Failure Mode and Effects Analysis?

A failure mode is the way in which something fails to do its job. An effect is what the customer (or the regulator, or the business) experiences as a result. A cause is why the failure mode happens. A delivery van breaking down is a failure mode; a customer’s shopping not arriving at the agreed time is the effect; a skipped service is the cause.

The single most common mistake in an FMEA is confusing causes with failure modes. “The database server runs out of disk” is a cause. “Customers cannot register” is an effect. The failure mode sits between them: “the database cannot write data”. Keep the three separate and the scoring makes sense; blur them and the spreadsheet fills up with duplicates that score differently depending on who wrote them.

The US armed forces described the procedure in 1949. NASA used it heavily on Apollo (and is often wrongly credited with inventing it), the automotive industry adopted it from the 1970s, and it later became a staple of Six Sigma, which is where I met it.

Design, Process and Supply Chain FMEA

Manufacturing distinguishes Design FMEA (how could the product fail?) from Process FMEA (how could the way we make it fail?). Software and service businesses need a third flavour that the textbooks rarely cover: supply chain FMEA, where the components are the suppliers, platforms and services you depend on, and the failure modes are the ways they can let you down.

Why Digital Supply Chains Need FMEA Now

A typical SaaS business of 200 people depends on a cloud provider (often two), an identity provider, a payment processor, a transactional email service, a DNS and CDN provider, an observability vendor, a customer support platform and Microsoft 365 or Google Workspace. Each of those has its own suppliers. Your cloud provider has a power company; your email provider runs on someone else’s cloud. Some of the most painful outages are common-cause failures, where three of your “independent” suppliers turn out to sit in the same data centre region.

AI has made the chain longer. Building software is now cheap, so businesses stitch in more third-party services than ever (model APIs, vector databases, agent platforms), and everyone assumes each new link will just work.

Why It Usually Takes a Crisis

If the knowledge already exists in people’s heads, why doesn’t it get written down? In my experience, because everyone is too busy running the business. Resilience work is not sexy growth stuff, and most businesses simply aren’t set up to see it as important, so it never reaches the top of the agenda. It takes a crisis (a frozen supermarket, a week of missed completions) to bring risk-based thinking up the list; by which point you are writing the FMEA as a post-mortem.

The Risk & Audit Committee Question

Most boards have a dedicated Risk & Audit committee, and from the NED seat I’ve seen how much depends on its composition. A committee made up entirely of finance experts will tend to see financial and regulatory risk, usually presented as a red-amber-green heat map of Likelihood × Impact. A committee that includes people with an operational background will ask what happens when the payment provider has a bad afternoon. Most of all it depends on the chair; if the chair isn’t driving this thinking, the business gets round to it after the crisis.

The Three Numbers: Severity, Occurrence and Detection

Every failure mode gets three scores on 1-10 scales. The tables below come from the instruction manual I wrote many years ago, updated for software and supplier risk. Treat them as a starting point and change them to suit your business (in most software businesses, reserving 10 for loss of life seems somewhat extreme). The one rule that matters is consistency across the organisation: a 7 in the payments team has to mean the same as a 7 in customer support, or you can’t compare the risks.

Severity: How Bad Is the Effect?

Severity scores the worst effect of a failure mode. If a failure mode has several effects, record only the highest severity.

Score Severity of effect
10 Regulatory breach or safety risk, occurring without warning
9 Regulatory breach or safety risk, occurring with warning
8 Core service down for all customers; customers angry
7 Core service down for some customers, or degraded for all; customers angry
6 Core journey slow or partially failing; customers dissatisfied
5 Secondary feature unavailable; customers need a workaround
4 Minor defect noticed by most customers
3 Minor defect noticed by some customers
2 Minor defect noticed only by the most discerning customers
1 No effect on customers

Occurrence: How Often Will It Happen?

My original scale ran from “every second” down to “once in the lifetime of the system”, which suits transactional systems handling thousands of requests a minute. It is useless for supplier events, where a cloud region failing every few years is a meaningful frequency, so I now keep both columns and use whichever fits the analysis.

Score Transactional systems Supplier and third-party events
10 Every second Weekly or more often
9 Every 10 seconds Monthly
8 Every minute Quarterly
7 Every hour Twice a year
6 Every day Once a year
5 Every week Every 2 years
4 Every month Every 3-5 years
3 Every quarter Every 5-10 years
2 Every year Seen in the industry, never by us
1 Once in the system’s lifetime Never seen anywhere

Detection: Would You Even Know?

Detection estimates how well your current controls would catch the cause or the failure mode before the customer is affected. The scale is inverted, which catches everybody out at least once: 1 means you are certain to detect it, 10 means you have no chance.

Score Likelihood of detection before customers are affected
10 No control exists; we’d hear from customers, the press or the regulator
9 A control exists in theory, but nobody watches it
8 We’d find out from the supplier’s status page (eventually)
7 A manual check would catch it, but only at the next scheduled check
6 We alert on our own uptime, but not on the supplier’s degradation
5 We alert on hard failure of the supplier integration
4 We alert on error rates in the integration
3 Synthetic transactions exercise the journey end to end
2 Latency and error-rate alerting on the core journey, with an on-call response
1 An automated control is certain to detect and fail over before customers notice

Most risk-based prioritisation I see leaves this column out entirely. The typical risk register is a 5×5 grid of Likelihood and Impact, and it treats a failure you would spot in seconds exactly the same as one that festers silently for a week. It is also often the cheapest gap to close.

A client of mine is heavily dependent on a third-party platform for a core workflow. They built proper latency monitoring on that workflow, alerting on the p95 latency of user creation rather than waiting for errors. When the supplier’s platform started to degrade, the p95 crept up long before anything actually broke, and my client was able to call the supplier and tell them they had a problem; hours before the supplier would have noticed it themselves. The supplier was no more reliable than before. In FMEA terms my client’s Detection had moved from roughly a 6 to a 2, and what would have been an outage became a phone call.

Calculating the Risk Priority Number

Once every failure mode has its three scores, the arithmetic is simple:

$$ \text{RPN} = \text{Severity} \times \text{Occurrence} \times \text{Detection} $$

RPN runs from 1 to 1,000. I also like to calculate Criticality, which is Severity × Occurrence: the risk before you give yourself any credit for spotting it. Comparing the two tells you whether a failure mode is dangerous because it is bad and frequent, or because you are blind to it.

Take a failure mode with Severity 7, Occurrence 6 and Detection 6. Its RPN is 252 and its Criticality is 42. Improve the monitoring so that Detection drops to 2, and the RPN falls to 84 while Criticality stays at 42. The risk hasn’t gone away; you have simply made sure that you will see it coming.

Where RPN Misleads You

There is an argument that RPN is mathematically unsound, and it is a fair one. The three scales are ordinal, so a Severity of 8 is not “twice as bad” as a 4, and multiplying ordinal numbers produces figures that look more precise than they are. Very different risks can share a score: 9 × 2 × 2 and 3 × 3 × 4 both give an RPN of 36, yet the first is a potentially catastrophic failure that happens rarely and the second is a nuisance. And because the three factors carry equal weight, a catastrophic failure with excellent detection can rank below a cosmetic one that nobody monitors.

My answer is to use RPN to order the conversation, not to make the decision. Sort by RPN, then read down the list with Severity and Criticality alongside it, and pull any Severity 9 or 10 to the top regardless of its score.

AIAG-VDA Action Priority: The 2019 Update

The automotive industry reached the same conclusion. In 2019 the AIAG and VDA (the American and German automotive bodies) published a harmonised FMEA handbook that replaced RPN with Action Priority: a lookup table that combines the three scores into High, Medium or Low, with Severity weighted most heavily. The handbook also formalised seven steps: planning and preparation, structure analysis, function analysis, failure analysis, risk analysis, optimisation, and results documentation.

The Action Priority tables were built for cars, so I would not copy them cell for cell into a SaaS business. The principle ports perfectly, though: severity first, then occurrence, then detection, and never let a high-severity failure mode hide behind a modest number.

How to Run an FMEA Workshop in Half a Day

The scoring is the easy part. The hard part, and the valuable one, is getting the right people into a room to do the work. I fundamentally believe the conversation is ALWAYS more valuable than the spreadsheet it produces; the spreadsheet is just the evidence that the conversation happened.

Choose the Level: Company, Product or Component

FMEA works at several altitudes. You can run it for the whole company (which suppliers could stop us trading?), for a specific product (what could stop customers completing a mortgage application?), or for a collection of components (what happens when the identity provider slows down?). I’d encourage you to build the muscle of running these at every level of the organisation, because each altitude finds failure modes the others miss. The company-level session spots that Microsoft 365 is a critical supplier to the completions process; the component-level session spots that the observability vendor has no fallback.

Who Should Be in the Room

Who you invite depends on the level, but the recipe is the same: five or six people giving a cross-section of product, operations and engineering talent. You want the engineer who knows the DNS provider is a single point of failure, the operations lead who knows the supplier’s status page is always the last to admit anything, and the product manager who knows which customer journeys make the money.

The chair is the most important appointment. You want somebody who won’t accept half-baked answers (“I’m sure it’s fine, it’s AWS”) and who has the authority to drive the remediation afterwards. An FMEA chaired by somebody with no power to fund the fixes produces a beautifully scored list of problems that nobody will ever solve.

Map the Landscape Breadth First

Start wide. Spend the first hour on a block diagram of the landscape you are reviewing: every component, platform and supplier that touches the customer outcome, drawn on a whiteboard with arrows. Resist the urge to score anything yet. The breadth-first pass is where the room discovers that the payment flow depends on a fraud-screening service nobody remembered signing up for.

Creativity: Inventing Failure Modes Nobody Has Seen

Then go deep on the specific components that matter most. For each one, ask the team to call out as many failure modes as they can, and capture them straight into the spreadsheet. It’s creative work, and the most enjoyable part of the session. The best failure modes come from questions like “what if it doesn’t fail, but gets slow?”, “what if it succeeds but sends the wrong thing?” and “what if it fails at the same time as the thing that’s supposed to warn us?” For each failure mode, ask what the customer experiences, and score the worst answer.

Strawman Scores, Not Scoring Debates

The big watch-out is that people love to argue about scoring methodology. Should a Detection score of 2 mean three minutes of outage or six? Is this failure mode a 6 or a 7? Forty minutes later, nobody has listed a single new failure mode.

The cure is strawman scoring. The chair proposes a score, the room either nods or offers a quick counter, and you move on. Scores can always be tweaked later if people don’t think they’re useful; a missing failure mode can’t be scored at all. The breadth-first map, then a deep dive with strawman scores, will get most teams through two or three components in an afternoon.

A Worked Example: A SaaS Platform’s Supply Chain

Here is what the output might look like for a mid-sized SaaS business, using the supplier column of the Occurrence scale.

Component Failure mode Effect S O D RPN S × O
Transactional email Messages silently dropped Password resets and invitations never arrive 6 7 9 378 42
Observability vendor Alerting unavailable Another incident goes unnoticed 7 5 10 350 35
Identity provider Sign-up latency degrades New customers abandon registration 7 6 6 252 42
Payment processor API returns errors Renewals fail; revenue at risk 7 6 4 168 42
DNS provider Resolution fails Platform unreachable for everyone 8 3 3 72 24
Primary cloud region Region outage Platform unavailable for all customers 8 4 2 64 32
Microsoft 365 Email and Teams down Support can’t reach customers mid-incident 5 6 2 60 30

The cloud region outage, the one that dominates every board conversation about resilience, ranks near the bottom, because everybody would notice it within seconds. Email, identity and payments share an identical Criticality of 42, and only Detection separates them. And the observability vendor (the tool that is supposed to tell you something is wrong) is itself a supplier that can fail, and in most businesses nobody has ever asked what happens when it does!

That last row also illustrates one of FMEA’s classic limitations: it scores failures one at a time, and it is poor at multiple-failure interactions. The observability outage only really hurts when something else breaks at the same time. Ask the question explicitly in the workshop, because the method won’t ask it for you.

From Inventory to Roadmap: Prioritising the Remediation

An FMEA that ends with the spreadsheet is a waste of an afternoon. For each failure mode near the top of the list, the team agrees a recommended action, an owner and a target date. Actions come in three types: reduce the Severity (a fallback email provider, a degraded mode that still lets customers sign up), reduce the Occurrence (a better supplier, a redundant design), or improve the Detection (synthetic transactions, latency alerting, an independent heartbeat for the monitoring system itself). Detection improvements are often the quickest win; a seeded test inbox that checks delivery every five minutes takes the email row from a 9 to a 3 in a sprint.

Remediation Competes for the Same Squads

This is where good FMEAs go to die. Every remediation action needs engineering capacity, and it competes for the same squads as the growth roadmap. Unless it appears on the roadmap with real squad-sprints allocated to it , it loses that competition every quarter. There is always more work than capacity.

So treat the output as roadmap work. In Run, Grow, Transform terms it is Run work, and it belongs alongside the rest of your keep-the-lights-on allocation. Frame it as an Objective with a measurable outcome, such as “no supplier failure mode above RPN 150” or “Detection of 3 or better on our top ten supplier failure modes”, and allocate it to a squad. In RoadmapOne that Objective sits in the grid next to the growth work, so the board can see exactly how much capacity resilience is getting, and whether it is getting any at all. If you use BRICE or Cost of Delay for the rest of the roadmap, the FMEA gives you the evidence to score resilience work honestly rather than leaving it to a gut feeling. A firebreak sprint is also a good home for the smaller detection fixes.

Keeping It a Living Document

An FMEA performed once and left on the shelf for a year delivers very little. As actions complete, record the new scores so the business can see the risk falling. Revisit it when you add a supplier, launch a product or suffer an incident; every post-mortem should ask “was this failure mode on the FMEA, and if not, why not?” Quarterly at product level and annually at company level is a sensible rhythm, and the company-level version makes an excellent paper for the Risk & Audit committee.

Common FMEA Pitfalls

Scope is the first trap: try to analyse the whole company at component level in one afternoon and you will finish nothing. The scoring debate is the second. The third is built into the method, because the team can only score failure modes it foresees; a room full of engineers misses the commercial failures and a room full of finance people misses the technical ones. A mix of experience in the room is the only defence.

The most insidious pitfall is the green spreadsheet. If the scores drift gently downwards at each review without any remediation work being done, you have built yourself a watermelon project , green on the outside and red in the middle. The chair who won’t accept half-baked answers is the cure here too.

The level of detail also irritates some people, so choose the team carefully and give it enough time to do a thorough job. Done properly, the session is fun; there’s real satisfaction when the room realises it has found a failure mode nobody had ever written down.

FMEA vs Other Risk Tools

FMEA is one tool in the kit bag. It is excellent at listing and ranking single failures across a known estate, weaker at combinations of failures, and silent on whether you are building the right product at all.

Tool Best for Blind spot
5×5 risk register Board-level overview of enterprise risk No detection lens; rarely granular enough to act on
FMEA Systematically listing and ranking failure modes across processes and suppliers Single failures only; depends on the team foreseeing modes
Fault tree analysis Working backwards from one catastrophic event through combinations of causes Time-consuming; covers one top event at a time
SVPG product risks Value, usability, feasibility and viability risk in new products Says nothing about operational or supplier failure
Riskiest assumption test Testing the assumption most likely to sink a new idea Pre-launch only

For a business with extreme reliability requirements I’d use FMEA as the inventory, fault trees for the two or three catastrophic scenarios the board loses sleep over, and the risk register as the summary that goes to committee. Your crown jewels deserve the deepest analysis of all.

Frequently Asked Questions

What is the risk priority number in FMEA?

The risk priority number (RPN) is the score FMEA uses to rank failure modes. It is the product of three 1-10 ratings: Severity (how bad the effect is), Occurrence (how often it happens) and Detection (how likely current controls are to catch it before the customer is affected). RPN runs from 1 to 1,000; higher numbers indicate higher priority for remediation.

How do you calculate a risk priority number?

Multiply the three ratings: RPN = Severity × Occurrence × Detection. A failure mode with Severity 7, Occurrence 6 and Detection 6 has an RPN of 252. Improving monitoring so Detection falls to 2 cuts the RPN to 84. Always read RPN alongside Severity, because very different risks can share a score, and any Severity 9 or 10 deserves attention regardless.

What is the detection rating in FMEA?

The detection rating scores how likely your current controls are to detect a cause or failure mode before its effect reaches the customer. The scale is inverted: 1 means detection is almost certain, 10 means no control exists and you would hear about it from customers. It is the column most risk registers leave out, and improving it is often the cheapest remediation available.

How are severity, occurrence, and detection rated in an FMEA?

Each is rated on a 1-10 scale agreed by the team. Severity scores the worst effect on the customer, from no effect to a regulatory or safety breach. Occurrence scores frequency, from almost never to constantly. Detection scores how likely controls are to catch the failure in time, with 10 meaning not at all. Use published scales as a starting point, then adapt them consistently across the organisation.

What are the 7 steps of FMEA?

The 2019 AIAG-VDA handbook defines seven steps: planning and preparation, structure analysis, function analysis, failure analysis, risk analysis, optimisation, and results documentation. In practice, for a software or supply chain FMEA, that means scoping the level, mapping the components, listing failure modes and effects, scoring them, agreeing remediation actions with owners, and re-scoring as actions complete.

What is the latest version of FMEA?

The current reference is the AIAG-VDA FMEA Handbook, published in 2019, which harmonised the American and German automotive approaches. Its biggest change was replacing the risk priority number with Action Priority, a lookup table that rates each failure mode High, Medium or Low with Severity weighted most heavily. Outside automotive, many teams still use RPN, ideally with severity-first judgement.

Is FMEA part of Six Sigma?

Yes. FMEA is one of the standard tools in Six Sigma, typically used in the Analyse and Improve phases of DMAIC to identify and rank the ways a process can fail. I learnt it at GE Capital, where Six Sigma was embedded in every function, but FMEA predates Six Sigma by decades and works perfectly well on its own.

Conclusion

The frozen supermarkets and the stalled completions had one thing in common: the failure mode was knowable in advance, and nobody had written it down. FMEA is a deliberately unglamorous answer. Put five or six people from product, operations and engineering in a room with a chair who won’t accept half-baked answers, map the supply chain breadth first, invent as many failure modes as you can, strawman the scores and stop arguing about whether it’s a 6 or a 7. Then take the top of the list and fight for the capacity to fix it, because a perfectly scored spreadsheet changes nothing on its own.

Use the scales here as a starting point rather than a methodology to follow slavishly; adapt them, keep them consistent and revisit them. The businesses that ride out a supplier failure usually got there with half a day, a whiteboard and a squad with the remediation on its roadmap.

Book the afternoon before the crisis books it for you.