Blog

Your QMS Is a Melting Ice Cube

$1.2 trillion wiped from software valuations in six weeks. Per-seat SaaS pricing in freefall. LLM costs dropping 10x per year. The QMS you just renewed is already depreciating.

Mustaq Bijral | Mar 12, 2026 | 12 min read

I’ve spent over a decade building software for pharmaceutical manufacturing. I’ve sat across the table from VP Quality leaders at 50-site companies who spend more time navigating their quality management system than actually managing quality. I’ve watched teams burn 18 hours investigating a minor deviation — not because the investigation was complex, but because the system demanded it. And I’ve seen the same deviation recur three months later, because the CAPA generated inside the QMS was a compliance artefact, not a genuine corrective action.

If you’re running Veeva QualityOne, TrackWise, MasterControl, or any of the legacy QMS platforms that dominate pharma today, I need you to understand something: the software you’re paying $150–300 per seat per month for is a melting ice cube. Not because it’s poorly built. Not because the vendors aren’t trying. But because the entire category — the thing we’ve called “quality management software” for two decades — is structurally obsolete.

This isn’t a prediction. The market is already pricing it in. And the underlying technology shift is moving faster than most pharma leaders realise.

In February 2026, Anthropic launched Claude Cowork — an AI agent that can autonomously navigate enterprise software workflows. Within 48 hours, $285 billion in market capitalisation evaporated from software, financial services, and asset management stocks. The S&P 500 Software & Services Index fell 30% in six weeks. This wasn’t a correction. It was a repricing of an entire industry.


The SaaS Model Is Collapsing

This isn't a blog post thesis. It's priced into the market.

Let me start with the broader picture, because QMS doesn’t exist in isolation. The entire SaaS industry is undergoing a structural repricing that will reshape enterprise software over the next 24 months.

Satya Nadella put it bluntly: SaaS applications are “CRUD databases with business logic,” and that business logic layer will be absorbed by AI agents. Microsoft is already “aggressively collapsing” the backend of its Dynamics platform to reflect this reality.

PitchBook’s Q1 2026 analyst note — titled “SaaS Is Dead, Long Live SaS” — argues that public software valuations are being repriced for obsolescence. The economic logic is flipping: an annual charge of $1,200 per seat becomes an annual charge of $10,000 per automated workflow. Software spend is now competing directly with payroll spend.

Foundation Capital coined the term “Service as Software” — AI is flipping SaaS from selling tool access to selling outcomes. The addressable market expands from the ~$500 billion software market to the $4.6 trillion global services market.

Forrester identified four investor fears driving what they call the “SaaS-pocalypse”: (1) SaaS companies won’t be the provider of choice for AI agents, (2) per-seat pricing becomes obsolete when AI reduces headcount, (3) vibe coding lets startups replicate complex SaaS platforms in weeks, and (4) SaaS products are fundamentally too complex — hundreds of apps that don’t talk to each other.

And Bain & Company predicted that within three years, most routine digital tasks will shift from “human plus app” to “AI agent plus API.”

Traditional B2B SaaS funding dropped 60% year-over-year in Q4 2025. Meanwhile, AI-native enterprise companies raised record amounts. The capital markets have already made their bet.

$1.2T+

Software Market Cap Wiped

S&P 500 Software & Services Index fell ~30% in six weeks following AI agent product launches.

60%

Drop in SaaS Funding

Traditional B2B SaaS funding fell 60% YoY in Q4 2025, while AI-native companies raised records.

$4.6T

Service-as-Software TAM

AI expands the addressable market from $500B in software to $4.6T in global services (Foundation Capital).


Now Apply This to QMS

Quality management software is the textbook case of a CRUD database with routing logic.

Think about what your QMS actually does. Strip away the branding, the validation documentation, the sales deck about “connected quality.” At its core, every QMS on the market does the same thing:

  1. Someone creates a record — a deviation, a CAPA, a change control, a complaint
  2. The system routes it — to an investigator, a reviewer, an approver, based on predefined rules
  3. Humans do the actual work — investigating root causes, assessing impact, writing corrective actions, reviewing documents
  4. The system tracks status — open, in progress, pending approval, closed
  5. Someone reports on it — cycle times, open items, overdue CAPAs

That’s it. That’s the entire value proposition. A CRUD database with routing logic and a reporting layer.

Now re-read Nadella’s statement: “SaaS applications are CRUD databases with business logic, and that business logic layer will be absorbed by AI agents.”

QMS is not adjacent to this disruption. QMS is this disruption. It is the purest example of a software category where the human is doing all the work, and the system is merely tracking that the human did it.

Your QMS doesn’t manage quality. It manages the paperwork around quality. There is a fundamental difference — and that difference is exactly the gap that agentic AI fills.


The Numbers That Should Alarm You

The cost of operating a traditional QMS is staggering — and most of it is human labour that AI can perform.

McKinsey estimates that 25–30% of pharmaceutical manufacturing costs relate to quality — with QC lab costs and documentation being the primary drivers. Approximately 30% of staff time is spent on documentation-related activities: product dossiers, machine logs, batch records, deviation reports. A single biotech batch record can comprise 5,000 to 45,000 manual entries.

BioPhorum’s cross-industry study on deviation management is even more revealing:

  • A minor deviation requires 18.1 hours of activity time and takes 29 calendar days to close
  • 24% of all deviations are repeats — the same issue, recurring because the CAPA was a checkbox exercise, not a genuine fix
  • 60–70% of deviations are minor or low-risk but get treated with the same heavyweight process as critical ones
  • A risk-based approach saves an average of 22,200 work hours per site per year — roughly $888,000 per site

One company in the BioPhorum study reduced its QA team by 200 people simply by implementing risk-based deviation management — without any AI, without any new technology. Just by not forcing minor deviations through a process designed for critical ones.

Now imagine what happens when you add AI agents that can autonomously investigate, correlate, assess risk, and draft corrective actions. The labour component of quality management — which is the vast majority of the cost — doesn’t compress by 20%. It compresses by 80%.

18.1 hrs

Per Minor Deviation

Average activity time for a single minor deviation — most of it spent on investigation and documentation (BioPhorum).

24%

Repeat Deviations

Nearly a quarter of all deviations recur — indicating CAPAs generated inside traditional QMS are ineffective.

$888K

Wasted Per Site Per Year

22,200 hours of avoidable work per site annually, driven by one-size-fits-all QMS workflows (BioPhorum).

30%

Staff Time on Documentation

Nearly a third of pharma manufacturing staff time goes to documentation — the core workflow that AI agents automate (McKinsey).


The Cost Curve That Changes Everything

LLM capabilities are improving at a rate that makes traditional software economics untenable.

The reason this disruption is happening now — not in five years, not in ten — is the unprecedented rate at which large language models are improving in both capability and cost.

a16z calls it “LLMflation”: LLM inference costs are decreasing by 10x every year. What cost $60 per million tokens in 2021 costs $0.06 today — a 1,000x reduction in three years. This price decline is faster than compute costs during the PC revolution or bandwidth costs during the dotcom boom.

Epoch AI’s analysis is even more striking: the cost to achieve GPT-4-level performance on PhD-level science questions fell by 40x per year. Across benchmarks, the median rate of cost decline increased from 50x/year to 200x/year after January 2024.

Stanford’s 2025 AI Index Report shows capability improvements that are hard to overstate:

  • SWE-bench (autonomous coding): scores rose by 67.3 percentage points in one year
  • GPQA (PhD-level reasoning): scores rose by 48.9 percentage points in one year
  • HumanEval (code generation): top models now score 99%
  • In 2022, you needed 540 billion parameters to score 60% on MMLU. By 2024, 3.8 billion parameters achieved the same — a 142x efficiency gain in two years

Epoch AI also found that AI capability progress has accelerated — the best scores on their Capabilities Index grew almost twice as fast over the last two years as they did over the two years before that. A 90% acceleration starting April 2024.

Let me make this concrete for pharma quality. Today, an AI agent can:

  • Read a deviation report and correlate it against batch records, equipment logs, and environmental monitoring data in seconds
  • Identify whether a similar deviation has occurred before, across any facility in the network
  • Draft a root cause analysis with supporting evidence
  • Recommend corrective actions based on what actually worked for similar deviations
  • Assess risk and classify severity automatically

In 12 months, that same agent — running on models that are 4x more capable at half the cost — will do all of the above better than your most experienced QA investigator, and it will do it in minutes instead of days.

In 24 months, the agent will handle most minor deviations end-to-end without human intervention. A human will review and approve — the way a manager reviews a direct report’s work — but the investigation, the analysis, the corrective action, the documentation will all be agent-generated.

At that point, what is the QMS for?

When LLM inference costs drop 10x per year and capabilities double every few months, the crossover point where an AI agent handles a deviation better, faster, and cheaper than a human navigating Veeva is not five years away. It is months away. And it moves the same direction every quarter — permanently.


Why Traditional QMS Vendors Can't Adapt

The architecture is wrong. Bolting AI onto a CRUD database is like adding a turbocharger to a horse.

The obvious counter-argument is: “Veeva and TrackWise will just add AI features.” Every legacy QMS vendor is already announcing AI capabilities — Veeva has its AI assistant, TrackWise is integrating generative AI, MasterControl is adding AI-powered document review.

But here’s why these efforts will fail to save the category:

The architecture is fundamentally wrong.

Traditional QMS is built around forms, workflows, and approval chains. The data model assumes a human will fill in fields, route documents, and make decisions. AI-native platforms are built around signals, ontologies, and autonomous agents. The data model assumes an AI will detect anomalies, correlate data, and recommend actions. You cannot retrofit one architecture into the other — the same way you couldn't retrofit a mobile app onto a mainframe.


The business model is misaligned.

QMS vendors charge per seat. Their revenue grows when you add more users — more QA specialists, more investigators, more reviewers, more approvers. AI agents eliminate the need for most of those users. A QMS vendor that successfully deploys AI is actively destroying its own revenue base. This is the innovator's dilemma in its purest form.


The data is siloed by design.

In a traditional QMS, deviations live in one module, CAPAs in another, change controls in a third, batch records in an MES, equipment data in a CMMS. The system was designed for humans to navigate between modules. An AI agent needs all the data in a unified ontology — connected, contextualised, queryable. Legacy QMS vendors would need to re-architect their entire data layer to enable genuine AI reasoning across quality events.

Validation creates inertia, not a moat.

QMS vendors often argue that GxP validation requirements create switching costs that protect their installed base. This is true in the short term. But validation is a one-time cost. The ongoing cost of operating an obsolete system — $888K per site per year in wasted deviation management alone — accumulates every month. Within 18–24 months, the economic case for migration overwhelms the validation switching cost.


AI features ≠ AI-native.

Adding an AI chatbot to TrackWise is like adding spell-check to a typewriter. It makes the existing workflow slightly faster without changing the fundamental model. An AI-native platform doesn't have a 'deviation management module' that an AI assistant helps you fill out. It has an autonomous agent that detects the deviation, investigates it, correlates it, classifies it, drafts the CAPA, and presents a complete resolution package for human review. The human's role shifts from 'doer' to 'approver.'


What Replaces QMS

The shift from quality management software to quality management outcomes.

The successor to QMS isn’t a better QMS. It’s a platform that delivers quality outcomes directly — without requiring humans to operate a system as an intermediary.

Here’s what that looks like in practice:

Deviation Management

Traditional QMS

QA receives notification → opens deviation record → assigns investigator → investigator manually reviews batch records, equipment logs, environmental data → writes root cause analysis → drafts CAPA → routes for review → manager reviews and approves → closure verification

Typical timeline: 15–30 days

Agentic AI Platform

System detects anomaly in real-time → correlates against batch, equipment, and environmental data across all facilities → identifies root cause pattern → generates investigation report with evidence → drafts risk-appropriate CAPA → presents complete package for human approval

Typical timeline: 2–4 hours

CAPA Effectiveness

Traditional QMS

CAPA generated as a compliance requirement → generic corrective actions based on templates → effectiveness check scheduled 90 days out → check is a form completion exercise → 24% of deviations recur anyway

Effectiveness: 76% (BioPhorum benchmark)

Agentic AI Platform

AI analyses what actually worked for similar deviations across all sites → generates specific, evidence-based corrective actions → continuously monitors for recurrence → automatically escalates if pattern re-emerges → learns from outcomes across the network

Effectiveness: continuous, measurable, improving

Batch Record Review

Traditional QMS + MES

QA reviewer manually checks each entry against specifications → cross-references equipment calibration records → verifies environmental monitoring data → flags exceptions for investigation → signs off page by page

Typical timeline: 5–20 days per batch

Agentic AI Platform

AI agent reviews entire batch record against specifications, calibration data, and environmental monitoring in seconds → highlights genuine exceptions with context → provides a confidence-scored release recommendation → human reviewer approves or investigates flagged items only

Typical timeline: 1 day (Valent BioSciences achieved this)

Notice the pattern. In every case, the traditional QMS is orchestrating human labour. The agentic platform is performing the labour and presenting results for human judgement. The human doesn’t disappear — they become the decision-maker rather than the worker.

This is the “Service as Software” model that Foundation Capital describes. Instead of paying $300/seat/month for 50 QA specialists to operate a QMS, you pay for deviation resolutions, batch reviews, CAPA completions. The pricing model shifts from inputs (seats) to outputs (outcomes). And the outputs are better, faster, and cheaper.


The Counter-Arguments — and Why They're Temporary

Every structural objection to this shift is valid today and irrelevant within 24 months.

I hear the same objections from every quality leader I talk to. They’re reasonable. And they’re all time-limited.

“Regulated industries move slowly.” True. But the economic pressure is now extraordinary. When you can demonstrate $888,000 per site per year in savings from deviation management alone — savings that compound across 50 sites into $44 million annually — even conservative pharma companies find budget. The question isn’t whether to move. It’s how fast your competitors move first.

“You can’t trust AI for GMP decisions.” You don’t have to. The AI does 90% of the work — investigation, correlation, drafting, risk assessment. A human reviews and approves in minutes instead of days. The human is still accountable. The audit trail is still complete. 21 CFR Part 11 requirements are still met. But the human’s role has shifted from data entry clerk to quality executive.

“Our QMS is validated — switching costs are enormous.” Validation is a one-time project cost. The ongoing cost of operating an obsolete system is permanent. If your QMS costs $2 million per year in licensing and your quality team spends $44 million per year in labour operating it, the validation cost of a new platform pays for itself in the first quarter.

“Veeva/TrackWise will catch up.” This is the strongest objection and the most dangerous assumption. Legacy vendors will add AI features. But as I argued above, AI features bolted onto a CRUD architecture cannot deliver the same outcomes as an AI-native platform built around a unified data ontology. The gap will widen, not narrow, as models improve. Every quarter, the AI-native platform gets meaningfully better. The legacy platform gets incrementally better. The curves diverge.

“Our data is in the QMS — that’s the moat.” Your data is in the QMS the same way your photos were on Kodak film. It’s a data prison, not a data moat. A platform that can ingest, contextualise, and reason across your deviation, CAPA, batch, equipment, and environmental data is more valuable than a system that stores it in disconnected modules.


The Timeline

Where we are, where we're going, and what happens to QMS along the way.

Now — Q1 2026

AI Assists Inside Legacy QMS

AI copilots help draft deviation investigations, summarise documents, suggest root causes. The QMS is still the system of record. Humans still do most of the work. But the speed advantage of AI-assisted teams over manual teams is already 30–50%, creating competitive pressure.

The QMS workflow is the bottleneck. AI can draft faster, but the routing, approval, and review steps still take days.

Q3 2026 — Q1 2027

Agentic Platforms Handle Routine Quality Events

AI agents autonomously handle 60–70% of minor deviations, routine CAPAs, and batch record reviews end-to-end. Humans review and approve agent work. Quality teams shrink as the labour component compresses. Per-seat QMS licensing starts to look absurd when you need 15 reviewers instead of 50.

Complex, multi-system investigations still require human expertise. Regulatory agencies are still developing frameworks for AI-driven quality decisions.

2027 — 2028

QMS as a Category Dissolves

Quality management becomes a capability embedded in the manufacturing platform, not a standalone software category. AI agents manage quality events as part of a unified production-quality-compliance system. The 'quality management system' becomes the 'quality intelligence layer' — always on, always learning, always improving. Traditional QMS vendors either pivot to AI-native architectures or become legacy maintenance businesses.

Full autonomy for critical quality decisions remains human-supervised. But the system does the work; the human provides judgement.


What You Should Do Now

Practical steps for quality leaders who see the ice melting.

Audit Your QMS Economics

Calculate the true cost of your QMS — not just licensing, but the fully loaded cost of every person-hour spent operating it. Deviation investigation time, CAPA documentation, batch record review, change control routing. Most quality leaders are shocked to find the labour cost is 10–20x the software cost. That labour cost is what AI compresses.

Cost analysisBaseline metrics

Start Building Data Readiness

AI agents are only as effective as the data they can access. If your deviations are in Veeva, batch records in an MES, equipment data in a CMMS, and environmental monitoring in a separate system, no AI can reason across them. Start planning data unification now — even if you're not ready to switch platforms, having connected, contextualised data is the prerequisite for everything that follows.

Data unificationOntology design

Run a Parallel Pilot

Pick one quality workflow — minor deviation management is ideal — and run it on an AI-native platform alongside your existing QMS for 90 days. Measure cycle time, investigation quality, CAPA effectiveness, and reviewer satisfaction. Let the data make the case. When your minor deviations close in hours instead of weeks with better root cause analysis, the business case for migration writes itself.

Proof of concept90-day pilot

I recognise that telling a VP Quality to plan for replacing their QMS is a big statement. These are systems that took years to validate, that your entire quality organisation is trained on, that your SOPs reference by name.

But the ice is melting. Every quarter, LLMs get meaningfully more capable. Every quarter, inference costs drop. Every quarter, the gap between what an AI agent can do and what your QMS requires a human to do gets smaller. The $1.2 trillion repricing of software valuations in early 2026 wasn’t a market overreaction. It was investors calculating exactly this curve and pricing it in.

The question is not whether traditional QMS will be replaced. The capital markets, the technology curve, and the competitive economics all point in the same direction. The question is whether you will be the quality leader who navigated the transition — or the one who renewed a five-year contract on a melting ice cube.

The winner in pharmaceutical quality will not be the company with the best QMS. It will be the company with the best quality outcomes — delivered by AI agents that investigate, correlate, reason, and act at a speed and scale no human team can match. The QMS was never the goal. Quality was. The tool is about to change.

Exit