How Much Can AI Help With PQC Migration?

Introduction

Every few weeks, someone tells me that AI is about to solve the PQC migration timeline problem. The pitch varies. Sometimes it comes from a vendor demonstrating an AI-powered cryptographic discovery tool. Sometimes it surfaces in a LinkedIn thread where someone argues that frontier models will “automate away” the multi-year programs that organizations are staring at. Sometimes it appears in a conference panel where a speaker projects a slide showing “Traditional: 10+ years” on one side and “AI-Accelerated: 2-3 years” on the other, with an arrow between them.

The pitch is appealing. PQC migration at enterprise scale involves upwards of 120,000 discrete tasks spanning network infrastructure, application code, PKI hierarchies, key management, hardware security modules, embedded systems, operational technology, and vendor relationships. The deadlines are already arriving from multiple directions. Executive Order 14412, signed on June 22, 2026, directs federal agencies to migrate high-value and high-impact systems to PQC for key establishment by December 31, 2030, and for digital signatures by December 31, 2031, and requires federal contractors to comply with PQC FIPS by the end of 2030. NIST IR 8547 proposes deprecating quantum-vulnerable public-key algorithms (RSA, ECDSA, ECDH, DH) by 2030 and disallowing them by 2035. CNSA 2.0 imposes separate requirements for National Security Systems. Google and Cloudflare have set internal 2029 migration targets. These are not interchangeable commitments, but they all point the same direction: organizations no longer have an unlimited planning horizon. If AI could compress the timeline from a decade to a few years, the problem would be far more tractable.

AI does help with PQC migration. I use AI tools in my own practice. Applied Quantum builds AI-assisted triage into its migration methodology. Several commercial products do the same. The question is not whether AI helps. The question is how much. And on that question, I think the industry is developing a dangerous gap between expectation and reality.

Where AI Helps Today

Before I argue about where AI falls short, let me be specific about where it already delivers value. I am not an AI skeptic in this domain. I have built AI into my own workflow and recommended it to clients.

Cryptographic discovery triage. Once a passive discovery tool has captured network traffic and enumerated cryptographic instances across an enterprise, AI-assisted classification can process hundreds of thousands of results and categorize them by algorithm type, key length, protocol context, data sensitivity, and regulatory applicability. This triage would take a human team months. AI reduces it to days. At Applied Quantum, we use this capability routinely.

Migration strategy automation. For each discovered instance, the migration strategy decision (direct replacement, hybrid deployment, algorithm swap, system retirement, or residual risk acceptance) depends on a matrix of factors. AI can generate draft recommendations across the full inventory, which human analysts then review and validate. This accelerates the planning phase meaningfully.

Code change generation. Where the migration involves mechanical code changes (library substitution, buffer resizing, handshake-record-length adjustments), frontier models can generate the diffs faster than human developers. The changes still require peer review and testing, but the drafting step compresses from hours to minutes.

Test case generation. AI models can generate interoperability test scenarios, enumerate edge cases for hybrid deployments, and produce protocol-level test vectors that would take human teams considerable effort to construct manually.

These are real contributions. They are already reflected in current migration timeline estimates. When I say a large enterprise takes 10 to 15 years to complete PQC migration, that estimate already assumes AI-assisted tooling is part of the process.

The Most Ambitious Version of the Argument

One of the most detailed recent treatments of AI-accelerated PQC migration appeared last month: a hypothesis paper by Robert Campbell in MDPI’s Cryptography journal, titled “Mythos-Class Frontier Models and the Compression of Post-Quantum Cryptography Migration Timelines.” Campbell proposes a 2 to 4 year scenario for migrating the highest-exposure subset of systems under concentrated resources, redesigned governance, and overlapping migration phases, down from the traditional 5 to 15 year baselines for full-portfolio migration. The accelerant in his model is Anthropic’s Claude Mythos Preview.

I covered Campbell’s earlier work in December 2025, when he published a study in MDPI’s Computers journal projecting those same 5 to 15 year timelines. I agreed with those estimates then, and I still think they are broadly correct. His new paper retains those full-portfolio baselines but introduces an accelerated scenario whose relationship to them is not controlled enough to support a clean compression ratio. I think the original estimates remain the right planning assumption for any enterprise-wide program.

Campbell’s paper includes a useful dual-use framing. He models Mythos as both a defender accelerator and an adversary destabilizer through six feedback loops (three on each side), and that framing is sound. He correctly identifies that embedded systems, OT, and tactical RF have non-compressible timelines governed by certification cycles and hardware replacement schedules. He is transparent about the paper’s limitations: it is a hypothesis with no empirical data, no direct model evaluation, and a timeline estimate he describes as “a modeling choice about scenario bounds, not a number derivable from the cited capability evidence alone.”

The paper does include a critical-path sensitivity model, so my disagreement is with its inputs rather than its methodology. Campbell assigns eight years to AI-compressible software-analytical work and only two years to the non-compressible institutional, regulatory, and hardware floor. He also treats much of the traditional baseline as sequential overhead that can be removed through concurrency. Those assumptions are clearly labeled as analytical stipulations rather than empirical measurements. But in every large program I have observed, those proportions are close to reversed. The institutional, governance, and coordination work dwarfs the software-analytical work on the critical path.

The paper also changes several variables at once in its accelerated scenario: scope narrows from the full portfolio to the highest-exposure subset, staffing density increases, governance is restructured, phases run concurrently, and a frontier model is introduced. Without a comparison using the same systems, staffing, governance, and concurrency without Mythos, the model cannot isolate how much compression is caused by AI versus how much comes from narrowing scope and reorganizing the program.

Where the Years Actually Go

I have led and advised PQC migration programs across government and enterprise clients. The pattern is consistent regardless of organization size: the majority of calendar time is consumed by activities that AI cannot accelerate, because they run on human and institutional clocks.

Let me walk through the actual anatomy of a large enterprise migration.

Securing executive mandate and budget takes 3 to 12 months before any technical work begins. This involves educating the board on quantum risk, positioning PQC within the organization’s risk register, building a business case that a CFO will approve, and competing for budget against every other priority in the security portfolio. AI does not accelerate board decision-making.

Standing up program governance takes another 3 to 6 months. A program with 120,000 tasks needs a dedicated program office with an accountable executive, a PMO of 3 to 8 FTEs at peak, a steering committee, and workstream leads. If the “PQC lead” is one person juggling quantum readiness on Fridays between SOC escalations, you do not have a program. You have a checkbox. AI does not staff your program office.

Discovery, including access negotiations, takes 12 to 24 months across the full estate. And here I want to be specific about what that time is spent on, because this is where the “AI compresses discovery” claim fails most visibly.

The technical analysis portion of discovery is the part AI accelerates. Passive discovery tools scan network traffic, parse certificate stores, fingerprint cipher suites, flag quantum-vulnerable configurations. AI improves the classification and prioritization of those results. Fine. That portion might take two to four months of a human team’s effort, and AI compresses it to weeks.

But discovery in a real program is mostly not technical analysis. Discovery is a political negotiation across business units. The PQC program office asks the payments division for access to production network segments to run passive discovery. Payments says no because they are mid-audit. The program office asks the manufacturing unit for access to the OT network. Plant engineering says they need to validate that the discovery tools will not interfere with real-time control systems, and that validation takes four months because it requires a maintenance window on the test line. The program office asks the cloud platform team to enumerate cryptographic dependencies across 2,000 microservices. The platform team says they will get to it after their current container orchestration migration, six months out.

No AI model resolves any of this. The bottleneck in discovery is getting permission and access rather than analysis speed. It is convincing the plant manager that yes, you need to inventory the SCADA network. It is waiting for the maintenance window. It is working through the internal politics of a large organization where every team has its own priorities. I see this in every program I advise.

Strategy and planning, with AI-assisted triage already built in, takes 6 to 12 months. Even with AI generating draft recommendations across the full inventory, a human cryptographic authority validates every one. “Should we deploy ML-KEM-768 or ML-KEM-1024 for this payment channel?” depends on the counterparty’s readiness, the regulatory regime, the performance budget on the terminating hardware, and the organization’s risk appetite. AI informs the decision. A person owns it.

Pilot implementations on non-critical systems take 6 to 12 months.

Scaled migration execution across thousands of systems, gated by change management processes and vendor release schedules, takes 3 to 7 years. This is the dominant line item, and it is the one most resistant to AI compression. AI can generate a code diff in minutes. That diff then enters the organization’s change management process: peer review, change advisory board approval (the CAB meets weekly or biweekly), staging environment deployment, regression testing, production scheduling with rollback procedures and on-call assignments. For vendor-dependent systems, the timeline is entirely outside the organization’s control. The enterprise cannot migrate its HSMs until the vendor ships PQC-capable firmware. It cannot migrate its VPN concentrators until the vendor certifies a PQC-compatible release. As I noted in my cost analysis, vendor lead times are often the longest single dependency.

Testing consumes 25 to 40 percent of total migration budget. One organization I worked with found testing consumed 40 percent of their spend against an initial estimate of 15 percent. AI can generate test cases. It cannot eliminate the testing phase, because interoperability testing requires real protocol exchanges with real counterparties, and those counterparties operate on their own schedules.

Partner and cross-industry coordination adds latency at every integration point. Financial networks, payment processors, certificate authorities, standards bodies. Every cross-organizational handshake is a calendar dependency.

The redo multiplier. Every PQC program I have worked on has required at least one significant redo of a completed phase. This adds a 1.5 to 2x factor over the program’s lifetime.

These phases overlap; the total is not a simple sum. The 12 to 15 year range for a large enterprise reflects dependencies, resource contention, external gates, and rework on the critical path. But look at which activities AI compresses: discovery analysis (from months to weeks), strategy triage (already baked into estimates), code generation within execution (from hours to minutes per change). Everything else runs on institutional and human clocks.

Why Effort Compression Is Not Schedule Compression

This is the distinction that the “AI compresses PQC migration” narrative consistently misses: effort share is not schedule share.

AI materially reduces analyst and developer effort inside a PQC program. Discovery triage, strategy recommendation, code change generation, test scenario creation. These are real contributions. In the programs I have observed, AI-compressible technical work accounts for perhaps 15 to 20 percent of total program effort. Those are illustrative proportions, not published benchmarks, but the pattern holds across large, heterogeneous enterprises and government estates with regulated, embedded, OT, or vendor-controlled systems.

The reduction in elapsed time is smaller because much of that work either runs in parallel with other activities or waits behind dependencies that AI does not control. Twenty percent of effort could be performed by resources that are not on the schedule-constrained critical path. It could be distributed across all phases. It could be blocked behind access negotiations, vendor firmware releases, or CAB approval queues. Reducing it by 80 percent may save substantial cost without reducing the program finish date by anything close to that proportion.

A better question is: how much AI-compressible duration sits on the resource-constrained critical path after governance, access, vendor, certification, testing, hardware, and partner dependencies are modeled?” In my experience, the answer is: far less than Campbell’s model assumes. His critical-path model assigns eight years to AI-compressible software-analytical work. In the programs I have led, the institutional and coordination dependencies dominate the critical path, and accelerating the analytical work shifts the finish date by months rather than years.

The economic benefit of AI in PQC migration may therefore be larger than the calendar benefit: fewer analyst hours, faster production of candidate changes, greater inventory coverage, lower cost per system migrated. That is a different and more defensible claim than “AI compresses a 13-year program to 4 years.”

Building CBOM and Crypto-Agility Are Organizational Transformations

Two outcomes of PQC migration deserve separate attention because they are consistently mischaracterized as technology problems when they are organizational problems.

A live cryptographic bill of materials (CBOM) means every team that deploys software, provisions certificates, configures network devices, or procures hardware updates the cryptographic inventory as part of its standard workflow. That requires new processes, new tooling integrations, new responsibilities assigned to people who currently do not think about cryptography. It requires training and enforcement. It requires connecting the CBOM to the organization’s configuration management database, its CI/CD pipeline, its procurement approval process, and its vendor management system.

Crypto-agility means the organization can swap cryptographic primitives without launching a multi-year program every time. That requires abstraction layers in application code, algorithm-negotiation capabilities in protocols, certificate infrastructure that supports rapid re-issuance, and testing automation that validates new primitives against the full deployment surface.

Both are measured in organizational readiness, not computational speed. AI cannot implement them. People and process changes implement them.

The Dangerous Misreading

The reason I care about getting the proportions right is not academic. It is practical.

Campbell does not recommend delay; his policy conclusion calls for immediate action, parallelization, and governance restructuring. I want to be clear about that. The practical danger is that the 2 to 4 year headline will travel farther than the paper’s scope conditions and caveats. If a CISO reads “AI compresses PQC migration to four years” and mistakes a scenario for the highest-exposure subset for a full-enterprise migration plan, they could conclude that program standup can be deferred until 2028, that they can still meet a 2032 deadline, and that AI tools will close the gap. That would be a serious misreading, but it is a foreseeable one. They will discover around 2030 that the organizational work I have described takes a decade regardless of how fast the AI generates code diffs. By then, the first regulatory deadlines will have passed. CNSA 2.0 already requires quantum-resistant algorithms for new National Security System acquisitions starting January 2027. EO 14412 requires federal agencies to migrate high-value systems to PQC key establishment by the end of 2030.

The “AI will handle it” narrative is the newest version of an older pattern I have been pushing back against for years: finding reasons to defer action on PQC. First it was “quantum computers are 20 years away.” Then it was “we’ll wait for the standards to be finalized.” Now it is “AI will compress the timeline so we can start later.” Each version provides a plausible-sounding justification for inaction that dissolves on contact with how migration programs actually run.

The correct response to AI-accelerated adversary capability is not “compress the timeline from 15 years to 4 years with AI.” The correct response is: start the 15-year program now. Use AI within it. Use AI aggressively for discovery triage, strategy automation, code generation, test case creation, and protocol analysis. It is a real force multiplier on the technical work. Then spend the remaining ten years doing the governance, access negotiation, vendor coordination, change management, and organizational transformation that no AI model can do for you.

Campbell’s paper makes one observation I fully endorse: AI does accelerate offensive capability, which means the urgency of migration increases. Anthropic reported that Mythos Preview can autonomously discover and exploit vulnerabilities that survived decades of human review. As I analyzed in my OT security piece, that class of capability is especially concerning for critical infrastructure running legacy firmware with limited or poorly documented security review. The adversary clock is tightening.

But a tighter adversary clock makes the organizational work more urgent. It does not make it faster. Until we have autonomous organizations, AI compresses the fraction of PQC migration that is technical analysis. That fraction is smaller than most people who have never led these programs assume.

What compresses migration timelines is starting.