For most of my career, the scarce resource in a strategy team was synthesis. Data was abundant — Factiva alerts, Refinitiv terminals, analyst reports, earnings calls — but the ability to compress it into a defensible point of view by Thursday's leadership review was rare. We hired for it, trained analysts toward it, and built entire operating rhythms around it.

Large language models attack exactly that scarcity. A capable model can now produce a competent first-pass competitor profile, a thematic readout of twenty earnings calls, or a draft market-entry framework in minutes. If your strategy team's value proposition is "we read a lot and summarize well," you are competing with a commodity.

What gets automated, and what gets amplified

Having built and led research teams across GlobalData, Cyient and HCLTech, my view is that the work splits three ways. Collection and first-pass synthesis are being automated — fast. Verification, triangulation and judgment are being amplified: the analyst who can interrogate an AI draft, spot the hallucinated market size, and reconcile it against primary signals becomes more valuable, not less. And the third layer — stakeholder trust — was never about analysis at all. A CFO doesn't act on a document; they act on a person whose track record they can audit.

The intelligence function isn't disappearing. It's becoming a product team — with AI as the manufacturing line and humans as quality, design and distribution.

When we built the Intelligence Command Centre at Cyient, the hard part was never the research itself. It was standardizing definitions across regions, deciding which signals deserved escalation, and earning the credibility that made leaders act on a flag about an executive transition or a contract renewal. Those are governance and trust problems. AI makes the inputs cheaper; it makes the governance more important.

Three moves for strategy leaders

• Re-price your deliverables. If a deliverable can be 80% machine-generated, stop staffing it like a craft product. Redirect the saved hours toward primary research and exec engagement — the parts machines can't fake.
• Build a verification culture. Treat every AI output like a junior analyst's first draft: useful, fast, and wrong in ways you must find. Codify the checks.
• Move analysts up the stack. The career path I now coach toward: from gatherer, to verifier, to question-framer. The person who decides what to ask the model owns the value.

The uncomfortable truth is that strategy teams have spent years advising clients on digital disruption while assuming their own work was immune. It isn't. But teams that treat AI as a force multiplier on judgment — rather than a threat to throughput — will produce better strategy than either humans or machines alone.

Benchmarking fifteen-plus global engineering services peers teaches you something the quarterly decks don't: most providers are far more alike than their positioning claims. Similar delivery pyramids, similar vertical mixes, similar pricing logic anchored on headcount. For two decades that homogeneity was fine, because the demand engine — offshoring engineering work at labor arbitrage — kept growing.

AI breaks the arbitrage math. When code generation, test automation and design iteration compress effort by meaningful double digits, a pricing model built on billable hours starts working against the provider: efficiency becomes revenue leakage. The firms I study are all converging on the same answer in their messaging — "outcome-based engagement" — but very few have rebuilt the machinery underneath it.

What the forecast models actually say

In my forecasting work, integrating macroeconomic indicators with firm-level signals, the divergence between scenarios is wider now than at any point I've modeled. The upside case isn't simply "AI adds a new service line." It's that ER&D providers become orchestrators of engineering throughput — owning toolchains, accelerators and domain IP — and capture value tied to product velocity rather than effort. The downside case is brutal: clients internalize AI productivity, renegotiate rates, and the industry experiences margin compression that consolidates the mid-tier.

The strategic question for every services CEO is no longer "how do we use AI to deliver faster?" It's "what do we charge for when speed is free?"

Where differentiation will actually come from

• Domain depth over delivery scale. Vertical expertise — regulatory pathways in medical devices, safety cases in auto, certification in aero — is the part AI commoditizes slowest. Portfolio choices like moving into Pharma & MedTech or energy-transition niches such as CCUS are bets on exactly this.
• Proprietary accelerators with measurable claims. Not "AI-enabled delivery," but auditable productivity data a procurement team can verify.
• Commercial model courage. Outcome pricing requires risk appetite, baselining discipline and contract sophistication. The firms that build that muscle early will set the reference points everyone else negotiates against.

Engineering services has reinvented itself before — Y2K to ADM, ADM to digital, digital to ER&D. The difference this time is that the disruption targets the unit of sale itself. Strategy teams in this industry should be running scenario models, not extrapolations.

Every few months a quantum computing announcement sends a wave of "should we have a quantum strategy?" questions through corporate strategy teams. Having sized emerging-technology markets for a decade, my honest answer is: most companies don't need a quantum strategy yet — but every strategist needs a quantum reading discipline.

The mistake is treating quantum like a normal technology adoption curve. TAM models for quantum are mostly conditional forecasts wearing the costume of market sizing: they assume error-correction milestones, then assume application discovery, then assume enterprise integration. Multiply three uncertain probabilities and the confidence interval swallows the headline number. When I build market models, I'd rather present a small, honest range with explicit trigger conditions than a large number with invisible assumptions.

A trigger-based posture instead of a timeline bet

• Watch error correction, not qubit counts. Logical qubits and error rates are the economically meaningful milestones; raw qubit announcements are marketing.
• Map your exposure asymmetrically. For most firms the near-term material issue is cryptography migration — "harvest now, decrypt later" risk — not quantum advantage in their products. That's a security roadmap item with real budget implications today.
• Position via optionality. For services and engineering firms, the rational play is small: a few practitioners fluent in the stack, partnerships over capex, and pre-built points of view for the verticals (chemistry, logistics optimization, financial risk) where advantage may land first.

In frontier tech, the strategist's edge isn't predicting the breakthrough. It's knowing — in advance — exactly what you'll do in the first 90 days after it happens.

This is scenario planning in its purest form: define the states of the world, attach observable triggers, pre-commit responses, and spend almost nothing until a trigger fires. It's also a useful template for the next frontier wave, whatever it is — the discipline transfers even when the physics doesn't.

Most corporate forecasting rests on a quiet assumption: the structure of the economy changes slowly enough that historical relationships hold. My regression models — like everyone's — are trained on a world where productivity growth, labor cost curves and technology diffusion behaved within familiar bands. The debate around increasingly general AI is really a debate about whether that assumption survives the decade.

I don't claim to know when, or whether, AI crosses into something economists would call transformative. But as a practitioner who presents best-to-worst-case scenarios to leadership teams, I've had to confront an awkward fact: the "worst case" and "best case" in most corporate models differ by a few points of CAGR, while the genuine uncertainty in the world now includes structural breaks — step changes in the cost of cognition itself.

How I've adjusted my own practice

• Add a structural-break scenario. Alongside conventional best/base/worst cases, model at least one world where a key input cost (software engineering, research, analysis) falls 50%+ within the planning horizon. It feels extreme until you price recent trends.
• Shorten the commitment horizon, not the planning horizon. Think in ten-year directions but commit capital in stages with explicit review triggers — the corporate version of options thinking I leaned on in my applied finance training at IIM Kozhikode.
• Forecast adaptation capacity, not just markets. The most decision-relevant metric may be internal: how fast can your organization redeploy people and budgets when an assumption breaks? That's measurable, and most firms have never measured it.

Under structural uncertainty, the quality of a plan matters less than the speed of the planning loop.

Economic history offers a consolation: general-purpose technologies — electricity, computing — took decades to diffuse, and the winners weren't the best predictors but the fastest learners. Strategy teams should hold their forecasts a little more loosely, instrument the world a little more densely, and treat every planning cycle as a chance to update rather than defend last year's number.