From Electricity to Autopilots: Why Hong Kong Corporate Service Firms Have 18 Months, Not Five Years

TL;DR: Sequoia says autopilots will eat the $6 of services for every $1 of software. Helpman–Trajtenberg say GPTs reorganise sectors in two-phased cycles. Tetration says AI compounding is faster than exponential. For Hong Kong corporate services, the window is ~18 months — and using AI like Excel is the most expensive mistake on the table.

Key Takeaways

• Productivity surges arrive with workflow redesign, not technology purchase
• Autopilots price the outcome and bypass the copilot+human stack entirely
• AI diffusion is compressed because rollout is an API call and adopters are software
• Hong Kong service firms’ hourly pricing model punishes them for getting faster
• Five concrete quarterly moves: rebuild one workflow, re-price one offering, instrument everything, reorganise around judgement, kill non-production pilots

From Electricity to Autopilots: Why Hong Kong Corporate Service Firms Have 18 Months, Not Five Years

A combined reading of Sequoia’s Services: The New Software, Helpman & Trajtenberg’s 1996 GPT diffusion paper, and Tim Dasey’s tetration framing — translated into a five-move playbook for Hong Kong corporate service firms.

From Electricity to Autopilots: Why Hong Kong Corporate Service Firms Have 18 Months, Not Five Years

Two recent pieces of writing — one from 1996, one from 2026 — describe the same approaching wave from opposite ends of the telescope. Read them together and the strategic question facing every Hong Kong corporate service firm becomes uncomfortably concrete: rebuild the workflow now, or be the cautionary case study in someone else’s deck in 2028.

1. The 1996 paper everyone in services should re-read

In Diffusion of General Purpose Technologies (NBER Working Paper 5773, 1996), Elhanan Helpman and Manuel Trajtenberg formalised what economic historians had long suspected: General Purpose Technologies (GPTs) — the steam engine, electricity, the transistor — do not lift productivity on contact. They diffuse sector by sector, in a sequence of two-phased cycles: first a slow, costly adaptation phase where complementary innovations are built, then a rapid growth phase once the technology and the surrounding workflows finally fit.

Productivity Time → Phase 1: Adaptation Complementary innovation, no productivity gain Phase 2: Diffusion Workflow redesign → output surge inflection Source: Helpman & Trajtenberg (1996), A Time to Sow and a Time to Reap: Growth Based on General Purpose Technologies, NBER Working Paper 5773. Curve illustrative of the paper’s two-phased cycle.

Their case study was the transistor. The pattern they described is the same one Paul David documented for the electric dynamo in The Dynamo and the Computer (1990, American Economic Review 80(2): 355–361): factories had electric motors for nearly 40 years before productivity surged, because the surge required redesigning the entire factory floor — not just swapping steam for volts.

The lesson is uncomfortable for incumbents: the productivity payoff arrives with the workflow redesign, not with the technology purchase. Firms that buy the GPT but keep the old process get the cost without the gain. Firms that redesign capture the surge.

2. AI is a GPT — but the diffusion clock is different

Electricity took ~40 years to reorganise manufacturing. The transistor took ~25 years to reorganise computing. Broadband took ~15 years to reorganise commerce.

Years from arrival to sector-wide reorganisation Electricity (1880s)~40 yrs Transistor (1947)~25 yrs Broadband (1990s)~15 yrs Cloud / SaaS (2006)~8 yrs AI / Autopilots (2022)~3 yrs (projected) Sources: Paul David (1990), The Dynamo and the Computer, AER 80(2); Helpman & Trajtenberg (1996), NBER WP 5773; Sequoia Capital (2026), Services: The New Software. AI line is the authors’ projection from current autopilot deployment cadence.

AI is compressing this further, for three structural reasons the 1996 model did not have to contend with:

1. Zero physical rollout. There is no grid to lay, no fab to build, no last-mile to wire. Distribution is an API call.
2. The complementary innovations are themselves AI. Tooling, evaluation, orchestration, agents, retrieval — the “factory redesign” is being built by the same technology, in public, every week.
3. The adopters are software. Each user sector in Helpman–Trajtenberg had to physically retool. A modern services firm retools by changing a prompt, a workflow, and a permission.

The two-phased cycle still applies. The phase lengths do not.

3. Exponential is the optimistic case. AI looks tetrational.

The empirical backdrop is the steepest scaling curve in the history of computing. Epoch AI’s dataset of notable training runs shows compute used to train frontier models doubling roughly every six months since 2010 — versus a ~21-month doubling time in the pre-deep-learning era. That step-change is what makes the next sections’ diffusion arithmetic so unforgiving.

Compute used to train notable AI models Total training FLOP, log scale Deep Learning Era 1⁰⁰ 1⁰⁶ 1⁰¹² 1⁰¹⁸ 1⁰²⁴ FLOP

<text x="100" y="316" text-anchor="middle">1950</text>
<text x="220" y="316" text-anchor="middle">1970</text>
<text x="340" y="316" text-anchor="middle">1990</text>
<text x="460" y="316" text-anchor="middle">2010</text>
<text x="580" y="316" text-anchor="middle">2020</text>
<text x="670" y="316" text-anchor="middle">2025</text>

Pre-DL: doubling ~21 months Deep-learning era: doubling ~6 months Recreated by the authors from data and analysis published by Epoch AI — Notable AI Models and GovAI, Computing Power and the Governance of AI (Anderljung et al., 2023). Schematic — not a reproduction of the original Epoch/GovAI chart.

Most strategy decks still draw AI progress as an exponential curve — a steepening hockey stick. Tim Dasey has argued that this understates what is actually happening. The better mental model, he suggests, is tetration: not repeated multiplication, but repeated exponentiation.

Linear vs Exponential vs Tetrational (log y-axis) 1⁰⁰¹⁰²¹⁰⁴¹⁰⁶¹⁰⁸ Step → Linear Exponential Tetrational Source: Tim Dasey, Tetration vs Exponentiation (LinkedIn, 2025). Curves are schematic representations of a(n)=n, 2ⁿ, and ⁿ2 plotted on a log scale.

A quick refresher:

• Linear: 2, 4, 6, 8, 10
• Exponential: 2, 4, 8, 16, 32 (each step multiplies)
• Tetrational: 2, 4, 16, 65 536, 2⁶⁵ 536 (each step exponentiates the previous result)

Why this matters for AI:

• Models get better → they train and evaluate the next generation of models.
• Models get better → they write the tools that orchestrate models.
• Models get better → they design the agents that use those tools to make models better.

Each of those loops is itself an exponential. Stack them, and the compound rate is no longer exponential — it is a tower. The visible lag between “interesting demo” and “production-grade autopilot” has gone from years to months to weeks. The lag between “production-grade autopilot” and “it has replaced a job function” is now measured in single quarters.

You do not need to believe the strict mathematical claim to take the operational point: planning for an exponential AI roadmap is already conservative. Planning for a linear one — which is what most corporate service firms are quietly doing — is malpractice.

4. Sequoia’s diagnosis: services are the new software

In March 2026, Julien Bek published *Services: The New Software* at Sequoia Capital. The thesis is short and uncomfortable: the next $1T company will be a software company masquerading as a services firm. Not a tool that makes accountants faster — a company that just closes the books.

Enterprise spend: software vs services Software ($1) $1 Services ($6) — the autopilot TAM $1 ×6 Source: Julien Bek, Services: The New Software, Sequoia Capital (March 2026). For every $1 spent on enterprise software, ~$6 is spent on the services that surround it.

Sequoia draws a sharp line between two business models:

• Copilot — you sell the tool to the professional. The professional uses it to do their job faster. Harvey for lawyers, Rogo for bankers, Cursor for engineers (in its early form).
• Autopilot — you sell the work to the customer. The customer skips the professional entirely. Crosby drafts the NDA. WithCoverage gets the insurance. Rillet closes the books.

The budget difference is staggering. In Sequoia’s framing: for every dollar spent on software, six are spent on services. Copilots compete for the $1. Autopilots eat the $6.

This is the operational counterpart to the Helpman–Trajtenberg cycle. The “complementary innovation” phase is not abstract anymore — it has a name (autopilot), a business model (price the outcome, not the hour), and a venture capital pipeline pointed directly at the work corporate service firms perform today.

5. Why this is existential for corporate service firms

Look at the verticals Sequoia explicitly names — and notice how many describe exactly what corporate service firms sell:

US outsourced services TAM by vertical (USD billions, midpoint) Insurance brokerage$140–200B IT managed services$100B+ Accounting & audit$50–80B Tax advisory$30–35B Legal transactional$20–25B Each line is a Sequoia-named autopilot wedge. Combined US TAM Sequoia is hunting exceeds $400B before HK / APAC equivalents. Source: Sequoia Capital (2026), Services: The New Software — vertical sizes are Sequoia’s published US estimates. Bar lengths use range midpoints.

• Accounting and audit ($50–80B outsourced in the US alone). Sequoia notes the US has lost roughly 340,000 accountants in five years while demand has grown, and 75% of CPAs are nearing retirement. Rillet is building the AI-native ERP that closes the books, not the tool that helps an accountant close them.
• Tax advisory ($30–35B). 80–90% of the underlying work is intelligence, even though CPA licensing creates a regulatory moat.
• Legal transactional work ($20–25B). Contract drafting, NDAs, regulatory filings — high intelligence, routinely outsourced, quality verifiable. Crosby and Lawhive are autopilot-native.
• Insurance brokerage ($140–200B). Largely shopping carriers and filling forms — pure intelligence.
• IT managed services ($100B+). Patching, monitoring, alert triage running across thousands of identical environments.

Now translate that to the corporate services world: annual returns, statutory filings, board minutes, register maintenance, sanctions screening, KYC refresh cycles, bank reconciliation, AP processing, payroll, expense approvals, license renewals, contract review, beneficial ownership reporting. Every single one of these is intelligence-heavy work with a clean outsourced budget line. They are exactly the wedge Sequoia is telling its founders to attack.

Hong Kong’s exposure is structurally worse than most:

• High share of routine, document-driven work. Exactly the workload AI agents now do end-to-end.
• Pricing tied to hours, not outcomes. A pricing model that punishes you for getting faster.
• Compliance moats that AI-native competitors will simply encode. HKICPA standards, SFC rules, IRD filing formats, Companies Registry forms — all expressible as machine-readable workflows.
• Client expectations being reset elsewhere. Your clients already use ChatGPT. They know what a 30-second answer looks like. A three-day turnaround on a board minute is no longer “thorough” — it is conspicuous.

6. The ‘AI as Excel’ trap

The most expensive mistake a Hong Kong firm can make right now is to treat AI as a productivity tool bolted onto the existing workflow — a “smarter Excel,” a “better search box,” a “drafting assistant.” This is exactly the mistake factory owners made in 1900 when they put electric motors on the same overhead shafts that had been driven by steam.

Margin trajectory: ‘AI-as-Excel’ incumbent vs autopilot challenger 0%50%100% Year 0Year 1.5Year 3 Hourly + AI-as-Excel Outcome-priced autopilot crossover Source: Authors’ analysis derived from Sequoia Capital (2026), Services: The New Software. Curves illustrate the pricing dynamic Sequoia describes — hourly margins compress as AI lowers the floor; outcome-priced autopilots capture the freed surplus.

It feels productive. It preserves the org chart. It does not require redesigning a single workflow. And it is precisely the posture Sequoia’s autopilot companies are pricing into their attack.

The innovator’s dilemma is brutal. A firm whose revenue depends on billable hours has every incentive to use AI as a speed multiplier — same workflow, fewer minutes per file, slightly better margins. But the autopilot competitor isn’t optimising the workflow. They are deleting it. They charge a flat fee for the outcome — the closed books, the filed return, the cleared sanctions check — at a price that assumes the intelligence work costs near zero.

When a CFO can buy ‘books closed monthly’ for a fixed fee instead of paying both QuickBooks and an accountant, the QuickBooks-plus-accountant stack does not get cheaper. It gets bypassed.

The productivity surge does not come from the tool. It comes from redesigning the workflow around what the tool can now do autonomously:

• Not “AI helps the secretary draft the AGM notice” — the workflow produces, files, and follows up on the AGM notice, with the secretary approving exceptions.
• Not “AI helps the accountant categorise transactions” — the ledger reconciles itself nightly, and the accountant reviews flagged anomalies.
• Not “AI helps the lawyer summarise the contract” — the contract is reviewed against playbook, redlined, and routed for signature, with the lawyer owning judgement calls.

7. Intelligence vs judgement: the only planning lens that matters

Sequoia’s framing — intelligence vs judgement — is the most useful planning lens a corporate services firm can adopt right now. Walk through every service line and ask: which parts are rules-based intelligence work, and which parts require judgement built on years of practice?

Intelligence vs judgement — share of work by service line Tax preparation~90% / 10% Annual return filing~94% / 6% KYC refresh~80% / 20% Contract review~70% / 30% Tax advisory~50% / 50% Intelligence (automatable) Judgement (stays human) Source: Sequoia Capital (2026), Services: The New Software — the tax-advisory intelligence share (80–90%) is Sequoia’s figure; other service-line splits are the authors’ estimates anchored to Sequoia’s framework and Hong Kong corporate-services workload composition.

Intelligence work that should be automated end-to-end (autopilot):

• Extracting data from BR certificates, passports, bank statements, invoices, contracts.
• Classifying documents into the right entity, matter, and workflow.
• Reconciling transactions, matching payments to invoices, flagging anomalies.
• Running sanctions, PEP, and adverse media screens with native-script variants.
• Drafting standard documents: NDAs, board resolutions, NAR1s, engagement letters, simple contracts.
• Monitoring expiry dates and triggering renewal workflows.
• Generating filing packages and routing them through approvals.

Judgement work that stays human (and where senior people should now spend most of their time):

• Edge-case interpretation of regulation across jurisdictions.
• Client relationships, scoping, and discovery.
• Risk decisions on borderline KYC cases or unusual structures.
• Strategic advisory: structuring, tax planning, restructuring, dispute strategy.
• Quality oversight and accountability for AI-produced work.

The shift is not ‘add AI to existing workflows.’ The shift is: the workflow itself becomes the AI, with human checkpoints inserted only where judgement is genuinely required. Junior-level intelligence work doesn’t get faster — it gets absorbed.

8. The playbook for this quarter (not this strategy cycle)

If you take both Sequoia and Helpman–Trajtenberg seriously, five moves become urgent — and the unit of time is the quarter, not the year:

Five-quarter autopilot playbook Q1Rebuild 1 workflowautopilot-first Q2Re-price to outcomesflat fee, not hourly Q3Capture evidencestructured workflow data Q4Reorganise talentjudgement > throughput Q5Kill theatre pilotsproduction or nothing Each step compounds: priced outcomes generate evidence; evidence retrains the autopilot; autopilot frees senior judgement. Source: Authors, synthesising Sequoia Capital (2026) Services: The New Software and Helpman & Trajtenberg (1996) NBER WP 5773.

1. Pick one workflow per service line and rebuild it autopilot-first. Annual return filing. Expense categorisation. KYC refresh. Contract review against a standard playbook. End-to-end, evidence-logged, human-approved at the exception path only. Charge a flat fee. Measure margin. Then move to the next workflow.
2. Re-price at least one offering to outcomes. Per filing, per onboarded entity, per reviewed contract — not per hour. If you cannot, your competitors will.
3. Capture operational evidence as the work happens. Sequoia’s argument is that today’s judgement becomes tomorrow’s intelligence as AI systems accumulate proprietary data about what good looks like. Every approval, correction, exception, and edge-case decision your team makes is training data for your future autopilot. If that data lives in email threads, Excel files, and people’s heads, you have nothing to compound. If it lives in a structured workflow with linked documents, decisions, and outcomes, you have a moat.
4. Reorganise talent around judgement, not throughput. The firms that win will not be the ones with the largest paraprofessional bench. They will be the ones with the deepest judgement layer — senior people whose time is spent on the 20% of cases where AI cannot decide, on client relationships, and on continuously raising the quality bar of the autopilot.
5. Stop running pilots that cannot go to production. A 12-week pilot in an environment with no audit trail, no permissions model, and no client-facing surface is theatre. Run pilots inside the real platform your team already uses, or do not run them. Assume your competitor is doing all of the above. They are.

9. The bottom line

Helpman and Trajtenberg told us, three decades ago, that GPTs reshape sectors in two-phased cycles, and that the firms which redesign around the technology capture the surge while the rest get selected out. Paul David told us, even earlier, that the lag between having the technology and benefiting from it is determined by how fast you are willing to rebuild your own factory. Sequoia is now telling its founders, in plain language, to build the companies that replace the service firm — not the service firm’s software.

The adaptation window has collapsed Electricity~40 years Transistor~25 years Broadband~15 years AI in HK corporate services~18 months Late adopters of electricity did not become slower factories. They became closed factories. Sources: Paul David (1990); Helpman & Trajtenberg (1996); Sequoia Capital (2026). The 18-month estimate is the authors’ for the Hong Kong corporate-services market.

Two stories that prove the thesis is already cashing in

The shift from “billable services” to “autopilot software” is no longer hypothetical. Two of the most-cited corporate case studies of the last 24 months come from industries adjacent to corporate services — and both compress decades of theory into hard numbers a managing partner can read in one minute.

Story 1 — Allen & Overy + Harvey: a magic-circle firm bets the workflow

In February 2023, Allen & Overy — one of the world’s most prestigious “magic circle” law firms — announced an exclusive launch partnership with Harvey, a GPT-4-based legal AI built on OpenAI models. Within months, more than 3,500 of the firm’s lawyers across 43 offices were using Harvey on live client matters — drafting, research, contract analysis, due diligence, regulatory work. By April 2025, A&O Shearman (the post-merger entity) extended the deployment with agentic, multi-step AI agents that execute complete legal workflows — not just answer questions. Partner David Wakeling, who runs the firm’s Markets Innovation Group, has been quoted publicly saying he had “never seen anything like Harvey” in 15 years of legal tech.

Read that again with a Hong Kong corporate-services lens. The most reputationally cautious firm on the planet — one that bills in tenths of an hour and earns its margins from human judgement — concluded that the riskier move was not deploying AI. They are now openly engineering away the very billable hours that funded the firm. They did not “explore.” They re-platformed.

The lesson is not “lawyers use ChatGPT.” It is that the prestige players moved first because they understood the Sequoia point earlier than everyone else: when autopilots arrive, the $6 of services collapses into the $1 of software — and whoever owns the software keeps the margin.

Sources: A&O announcement, 16 Feb 2023; A&O Shearman + Harvey agentic rollout, 6 Apr 2025; The Times, 24 Mar 2023.

Story 2 — Klarna: one AI agent, 700 humans, USD 40 million

In February 2024, Swedish fintech Klarna disclosed the operating numbers from its OpenAI-powered customer-service assistant after just one month live: 2.3 million customer conversations handled, equivalent to the workload of 700 full-time agents, with an estimated USD 40 million profit improvement in 2024 alone. Average resolution time fell from 11 minutes to under 2. Repeat inquiries dropped 25%. CEO Sebastian Siemiatkowski stated publicly that the AI was already performing on par with — or above — human agents on customer-satisfaction scores. By later updates, Klarna characterized the agent as doing the work of roughly 853 employees.

Klarna is not a law firm or a CoSec. But the parallel is exact. Customer service, like company secretarial work, KYC review, AR reconciliation, statutory filings, and most of the recurring deliverables of a Hong Kong corporate-services firm, is a structured, evidence-bounded, high-volume task with regulated outputs. It is precisely the work that Sequoia’s “services-are-the-new-software” thesis predicts will be absorbed by autopilots first.

Klarna’s number — one agent doing the work of 700 — is what 700:1 leverage looks like in practice. It is also what your competitor’s price list looks like in 18 months if they deploy and you do not.

Sources: Klarna press release, 27 Feb 2024; Business Insider, 28 Feb 2024; Customer Experience Dive — 853 employees update.

Why these two stories matter, together

Allen & Overy proves the prestige and judgement end of professional services is willingly cannibalising its own billable hours rather than wait. Klarna proves the volume and operations end can be compressed by a factor of several hundred. Hong Kong corporate-services firms sit squarely between those two poles — and therefore inherit both pressures at once. Helpman & Trajtenberg’s two-phase GPT cycle predicts exactly this: the leading sector (legal AI) re-platforms first, then the productivity wave sweeps the connected sectors (compliance, accounting, secretarial). The 18-month window is not a guess pulled from the air. It is the gap that is already visible in the data.

AI is the GPT of this generation. Its diffusion clock is faster than electricity’s, faster than the transistor’s, faster than the internet’s, and — if the tetration framing is even directionally right — faster than exponential. Capital is flowing aggressively into autopilots aimed at exactly the work corporate service firms perform today.

A firm that responds with ‘we use ChatGPT now’ is not adapting. It is decorating. The firms that survive the next five years will treat AI not as a productivity feature but as an operating system rewrite — where the workflow, the evidence, the handoffs, and the pricing model are all designed around the assumption that intelligence work is essentially free, and judgement work is the entire business.

Late adopters of electricity did not become slower factories. They became closed factories. The same selection pressure is now arriving — at compounded speed — for professional services in Hong Kong. On current trajectory, the window between “we should probably look at this” and “our clients have left” is not five years. It is closer to 18 months.

The choice Sequoia is forcing onto this industry is not ‘should we adopt AI.’ It is: do you want to be the autopilot, or do you want to be replaced by one?

The tsunami is not coming. It is already in the harbour.

Sources

1. Services: The New Software — Sequoia Capital (Julien Bek, March 2026)
2. 2026: This is AGI — Sequoia Capital (Pat Grady & Sonya Huang)
3. Diffusion of General Purpose Technologies (Helpman & Trajtenberg, NBER WP 5773, 1996) — NBER
4. Diffusion of General Purpose Technologies — full PDF — NBER
5. The Dynamo and the Computer — Paul A. David (1990, AER 80(2): 355–361) — JSTOR / American Economic Association
6. Tim Dasey — Tetration vs Exponentiation — LinkedIn

Tags: ai strategy, corporate services, hong kong, workflow automation, autopilot, gpt diffusion, tetration