At Lease Renewal, AI Has to Show Its Headcount

On July 14, JLL published a survey of more than 2,200 C-suite and corporate real estate leaders. Seventy-eight percent said AI would affect their real estate strategy within three to five years. Only 15% had reached the optimizing phase in their real estate operations. Just 31% were preparing workplaces for collaboration between people and AI.

Put those percentages on a lease-renewal agenda and the timing problem appears immediately.

A five-year extension signed this summer can still be binding in 2031. A ten-year lease can outlive several generations of models, a return-to-office policy, two leadership teams and most of the jobs in a current workforce plan. Yet the renewal package still arrives with fixed questions. How many people will work here? Which teams need seats? How often will they come in? What power, network and security upgrades will they require? How much flexibility is worth paying for?

The answers sit in different systems. HR owns the headcount plan, finance the cost envelope, IT the AI stack and business leaders the deployment targets. Corporate real estate owns the calendar, and that calendar keeps moving while the other forecasts are unfinished.

AI turns that separation into a capital risk. Productivity may reduce work in one function while a new product adds engineers, salespeople or implementation teams elsewhere. An agent can remove a sequence of tasks without eliminating the role around them. Fast-growing AI companies can take more office space while mature employers consolidate floors. Attendance can raise peak demand without changing headcount; a smaller workforce can still require more expensive rooms, power and privacy controls.

The lease cannot wait for one definitive AI employment forecast, because no credible forecast can supply one. It can, however, require a range that HR, finance, IT and real estate are prepared to sign together. At renewal, AI has to show its headcount, its role mix, its attendance pattern and the conditions that would make those numbers change.

July 14 exposed the lease-planning gap

JLL’s Future of Work Survey 2026 was fielded from January through April across 21 countries. It captures an unusual planning moment. Leaders largely accept that AI will alter work and the workplace, but most companies have not translated that belief into a portfolio decision.

The gap is wider than the 78% and 15% figures suggest. Forty-six percent of respondents were monitoring AI’s effect on real estate, 40% were analyzing it and 33% were modeling scenarios. Those activities can produce useful information, but none commits a company to a floor, a term or an exit option. They describe organizations watching an approaching decision from several distances.

JLL’s account of the bottleneck is direct: corporate real estate teams are waiting for workforce strategy from the CEO and CHRO. That dependency makes sense. A property team should not invent a hiring plan. It becomes dangerous when workforce strategy treats real estate as a downstream implementation detail. Lease expirations, break dates and construction windows create deadlines that a workforce planning process cannot move.

The survey also found that skills gaps had overtaken budget constraints as the main obstacle to real estate transformation for the first time in the study’s 15-year history. Respondents named AI, analytics and emerging technology skills most often, at 36%. Change management followed at 26%, organizational silos at 25% and measurement at 23%.

The gap is larger than a shortage of AI specialists on property teams. A company can hire a data scientist and still fail to connect an agent rollout to a lease decision. The scarce capability is translation: turning a product roadmap into assumptions about jobs, teams, attendance, locations, rooms, infrastructure and time. That work crosses reporting lines, which is why the organizational silos matter almost as much as the technical skills.

The investment priorities reveal another split. Forty-six percent of leaders prioritized advanced technology and AI support in the workplace, while 44% prioritized reliable technology. Only 31% prioritized adaptable spaces, and 24% prioritized wellbeing. Many companies appear ready to fund the digital layer before they know what physical work the layer will produce.

That sequence can create a costly mismatch. A company may buy collaboration software, sensors and an internal agent platform, then discover that its office has the wrong mix of open desks, secure rooms and team space. It may add power and cooling to a floor it plans to surrender. It may fit out a long lease around an attendance policy that managers cannot sustain. Technology spend does not rescue an inflexible property decision.

JLL also reported broad expectations of workforce growth and role redesign. In its release accompanying the survey, 60% of leaders expected their workforce to grow over the next five years, while 40% expected it to shrink. Sixty percent expected AI to reinvent roles, compared with 40% who expected replacement.

Those figures should not be read as a labor forecast. The respondents are leaders describing expectations, not a representative model of employment outcomes. The categories are broad, and an employer can experience growth, shrinkage, reinvention and replacement in different units at the same time. The value of the survey lies elsewhere. It shows that companies making property decisions do not share one AI workforce trajectory. A renewal process that assumes a single number conceals that disagreement.

Consider a company with 3,000 employees and a headquarters lease expiring in 18 months. The CHRO expects total headcount to remain near 3,000, with fewer transactional roles and more product, implementation and security jobs. The CFO’s productivity case assumes a 10% reduction in labor cost. A business-unit leader expects demand created by an AI product to add 250 customer-facing roles. IT needs controlled rooms for model testing. The chief operating officer wants three office days a week. Corporate real estate receives the combined message as “flat headcount.”

Flat headcount is not a space plan. It hides the jobs moving between cities, the teams that need to sit together, the daily attendance peaks and the technology load. It also hides the confidence interval. If the product succeeds, the company may need more capacity. If productivity arrives before revenue, it may need less. A five-year commitment should expose both outcomes before anyone compares rents.

Five years of rent meet a quarterly workforce forecast

Every function arrives at the renewal with a reasonable horizon, and none matches the lease.

Finance revises forecasts by quarter. HR often works from an annual operating plan, with monthly hiring controls. Product and technology teams can change AI vendors or deployment priorities within weeks. Real estate teams begin renewal work 18 to 36 months before a major expiration, then commit the company for years. Construction, permitting and relocation add their own lead times.

The horizons meet at the lease signature. By then, flexibility has a price.

A shorter term can carry higher rent or fewer landlord concessions. A contraction right can require a premium, notice period and a clearly defined portion of the premises. An expansion option is useful only if the adjacent space is available when the company needs it. Subleasing transfers some risk but introduces market, timing and credit uncertainty. A fitted floor can be cheaper today and expensive to reconfigure later. A new building may offer better infrastructure while imposing a longer commitment.

AI widens the distribution of outcomes inside those choices without revealing which one will occur.

Suppose a customer operations group has 500 employees. An internal agent may reduce handling time and administrative work. One forecast holds volume constant and reduces staffing. Another assumes better service increases demand, so staffing stays level while throughput rises. A third moves people into retention and account expansion. A fourth encounters weak adoption and changes little. All four can be honest interpretations of the same pilot.

If the company writes only the expected case into its property model, the expected case acquires false authority. A range is more useful:

The low case shows the space needed if productivity converts quickly into fewer roles.
The operating case shows the space needed if jobs are redesigned while headcount stays close to plan.
The growth case shows the space needed if lower costs or a new AI product increase demand.
The delay case shows what happens if adoption, data access or regulation slows deployment while the lease clock continues.

These cases should use the same dates as the lease. A workforce range for next year does not support a decision through 2031. The planning team needs annual ranges, trigger dates and a view of which assumptions can be reversed. Precision should decline with time. The first year can use approved requisitions and known attendance. Years four and five may be wider bands tied to product revenue, automation coverage and location strategy.

This approach changes the role of a productivity estimate. A claim such as “the agent saves 20% of task time” is an input, not a headcount output. The company still needs to decide whether saved time becomes lower staffing, higher volume, better quality, shorter hours, new services or unused capacity. It needs to identify the date when that conversion can affect seats. A task saving measured in October cannot reduce a lease signed in August.

Attendance adds a second conversion. Headcount does not equal peak occupancy. A company with 2,000 assigned employees and two coordinated office days can have a larger Tuesday problem than a company with 2,500 employees spread across the week. If managers pull teams together for AI training, review or client work, collaboration can become more synchronized even when individual work becomes more flexible.

That is why badge averages are weak renewal evidence. An average of 45% occupancy can conceal 80% on Tuesday and 25% on Friday. It can also conceal a shortage in one building and excess space in another. Renewal models need the distribution: peak day, peak hour, team overlap, meeting-room pressure and the number of people unable to find an appropriate setting.

The forecast also needs a cost boundary broader than rent. JLL describes a new capital layer that includes AI automation and technology infrastructure alongside energy, occupancy and operating expense. A smaller but more intensive office can cost more per employee. Secure computing areas, upgraded connectivity, acoustic privacy, sensors and reservation systems can consume capital that a rent-only model misses.

The CFO therefore needs two denominators. Cost per leased square foot reveals the property commitment. Cost per productive work setting reveals whether employees can do the intended work. Neither should be replaced by cost per employee, because the employee count is precisely what AI has made uncertain.

This is where lease optionality earns or loses its premium. A company should not buy every possible option. It should pay for the options that correspond to named workforce triggers. If the growth case depends on an AI product reaching a revenue threshold, an expansion right can be tied to that threshold and its expected timing. If the low case depends on automating a defined process, a contraction date should fall after the adoption and workforce transition period, not before it.

A generic desire for “flexibility” produces expensive contract language. A dated workforce scenario produces an option the company can evaluate.

Sixty percent expect growth across a different role mix

Two payrolls with the same headcount can require very different offices.

JLL’s finding that 60% of leaders expect growth sits beside its finding that 60% expect role reinvention. Together, the figures point to a planning problem that total headcount cannot solve. Even if the number of employees rises, the work can move toward different functions, levels and interaction patterns.

An AI-enabled service team may need fewer people processing routine cases and more people handling exceptions, improving knowledge sources and working with customers on complex problems. A software company may reduce manual quality checks while adding model evaluation, security and implementation roles. A marketing team may produce more variants with fewer production handoffs, then need more time for brand review and campaign analysis. These shifts change the rooms and relationships that matter.

Role level is one variable. Early-career employees often depend on observation, feedback and access to experienced colleagues. If AI handles the smaller tasks through which new employees once learned, the office may need to support more deliberate coaching. A bank of identical desks will not create that contact. Team rooms, review sessions and predictable overlap can become part of the talent system.

Location is another. A company may keep global headcount flat while concentrating scarce AI, product or security talent in a few markets. It may distribute customer operations while centralizing model governance. It may hire implementation teams close to clients. Each pattern produces a different portfolio, even before attendance changes.

The work mode matters too. Gensler’s 2026 Global Workplace Survey, based on 16,400 office workers in 16 countries, identified 30% of respondents as AI power users. Those workers reported spending 37% of their time working alone, compared with 42% among late adopters. They spent 12% of their time learning, compared with 8%, and 11% socializing, compared with 9%.

The differences are modest, and the survey does not establish that AI caused them. They still challenge a common property assumption: more AI automatically means more isolated work and fewer shared settings. Heavy users in this sample reported a somewhat more varied workday. If that pattern holds inside a company, it argues for a mix of focus, learning and social spaces rather than a simple reduction in desks.

Gensler also found that the office accounted for 55% of work time, home for 18% and other locations for 26%. Two-thirds of workers said they had modified or improvised their workspace; one in four had created a fix for ergonomics, temperature or privacy. The figures describe employees compensating for settings that do not fit their work.

At 8:55 on a busy Tuesday, those improvised fixes become a capacity problem. A model evaluator searches for a quiet room with a secure screen. A sales team needs client-ready space. A junior analyst needs ten minutes beside a manager. Someone using a speech interface needs acoustic privacy. Facilities sees each request as a room problem; the employees experience lost time and an office that cannot support the work they were asked to do.

The employee perspective also supplies a needed counterweight to executive forecasts. Leaders can describe a role as redesigned while the employee experiences two jobs at once: the old process remains, and the AI process adds checking, correction and escalation. During that period, space demand may not fall. Training sessions, support desks and cross-functional reviews can increase it.

Managers face a related burden. They must decide when AI output is good enough, who can approve exceptions and how performance should be measured when output rises. Those decisions often happen through recurring team contact. If an employer removes space before the management system catches up, it can turn a technology transition into a coordination problem.

The corporate real estate model needs a role-mix bridge between HR’s job architecture and facilities data. That bridge does not require a perfect taxonomy. A practical version can group roles by the work settings they need:

Concentrated individual work with strong privacy or security requirements.
High-frequency team coordination and project work.
Client or candidate interaction.
Learning, apprenticeship and review.
Laboratory, studio, hardware or other specialized work.
Mobile or field work with occasional office use.

HR can map planned role changes into those groups. Real estate can translate the groups into work settings and peak capacity. IT can attach device, network, access and data requirements. Managers can validate whether the map resembles actual work.

This step prevents a familiar error. When 50 administrative positions disappear and 50 engineers arrive, a spreadsheet can label the change neutral. The office may experience it as a material increase in project rooms, secure environments, power and hiring competition in a different city. Neutral headcount can carry a non-neutral property bill.

It can also carry a different talent bill. A company that wants scarce employees in three days a week has to offer a place worth the commute and locate it where those employees can be hired. The lease is then part of workforce strategy, not a passive container for it.

San Francisco proves how concentrated demand can get

The San Francisco Bay Area is a useful warning against national averages.

CBRE reported in May that technology and AI companies leased more than 14 million square feet across San Francisco and Silicon Valley in 2025, representing 55% of leasing activity in those markets. AI companies had leased a cumulative 21 million square feet there since 2019. San Francisco and Silicon Valley each had more than 5 million square feet of active AI-related tenant requirements in the market.

The concentration becomes clearer in CBRE’s 2026 Tech Gateway Office Markets report. San Francisco accounted for 10.6 million square feet of AI leasing and Silicon Valley for 10.4 million from 2019 through the first quarter of 2026. Together, they represented 66% of AI leasing across the six major markets CBRE tracked.

CBRE placed another growth series beside the leasing data. It estimated that the global workforce of the 15 largest venture-backed AI companies rose from roughly 7,500 people in 2020 to 48,000 in 2025. The report does not say how many of those employees worked in the Bay Area or establish that hiring caused the local leases. Read together, the two series describe selected AI firms growing at the same time that their property demand concentrated in two markets.

The Bay Area surge cannot be extended to every central business district. It does show how an aggregate automation story can miss local expansion. A company deciding whether to surrender space in a talent hub needs to model the value of future access alongside current utilization.

The broader U.S. technology sector also increased its share of leasing. CBRE’s release on the Bay Area surge said tech companies leased 36.7 million square feet nationally in 2025, or 16.8% of total leasing. In the first quarter of 2026, the sector leased 11.5 million square feet, or 22.7%, up from 7.9 million and 15.3% a year earlier.

Demand did not erase excess supply. San Francisco’s office vacancy rate was still 32.8% in the fourth quarter of 2025, according to CBRE. A market can have severe vacancy and intense competition for a particular building, neighborhood or floor type at the same time. The relevant question for a tenant is not whether the city has empty offices. It is whether suitable space will be available, on workable terms, when a named team needs it.

That distinction changes a renewal negotiation. If a company’s growth case depends on hiring AI engineers in San Francisco, giving up a well-located building may destroy an option that is expensive to recover. If its workforce plan moves those jobs elsewhere, the same space may be unnecessary. Corporate real estate cannot choose between those cases using vacancy data alone.

Concentration also affects portfolio balance. A company can shrink total square footage while expanding in one gateway city. It can leave a large legacy office and take a smaller, higher-quality site near a specific talent pool. It can use a satellite hub to enter a market before committing to a headquarters-scale lease. These moves look contradictory only when the portfolio is reduced to one net number.

The Bay Area case carries an important source caveat. CBRE advises property owners and occupiers and benefits from real estate transactions. Its market data is valuable, but its framing naturally emphasizes leasing activity and opportunity. The 15-company employment estimate also covers a selected group of successful venture-backed firms. It should not be generalized to all AI startups or all employers.

The evidence is strongest when used as a scenario, not a prophecy. It shows that selected AI firms have generated large, concentrated office requirements. It does not establish the probability that another employer will grow, or that the pattern will persist through a full lease term.

For a renewal team, the practical question is whether its company has the same demand ingredients: product growth, funding or cash flow, a location-specific talent need, leadership preference for in-person work and a role mix that benefits from proximity. If those ingredients are absent, the Bay Area numbers are context. If several are present, they belong in the growth case.

The vacancy rate makes a national countercase

The national office market supplies the opposite warning. A company can overpay for growth optionality in a market where supply remains abundant and its workforce does not need a concentrated hub.

Moody’s Analytics reported a 21% U.S. office vacancy rate in the first quarter of 2026 across 79 markets, according to Axios’s account of the data. The comparable rate was 17% in 2020. Employees were spending about a quarter of workdays remotely, compared with 7% in January 2020. The article also noted a stubborn contractual fact: office leases often run for five or ten years.

High vacancy gives many tenants leverage. It can support shorter terms, better concessions or a move to higher-quality space. It can also make the cost of holding excess floors visible. An AI growth narrative should not become an excuse to preserve a legacy footprint that current employees do not use.

Cushman & Wakefield reached a more expansionary conclusion in its May scenario analysis of AI and commercial real estate. Its baseline scenario, assigned a 50% probability, projected that AI could add 330 million square feet of total U.S. commercial real estate demand by 2035 relative to a pre-AI baseline. The office component was 24.4 million square feet, or 9.2% above the baseline.

The headline sounds like a directional forecast. The scenario structure is more informative. Cushman & Wakefield assigned 15% to a productivity expansion scenario, 25% to an AI bust and 5% to a displacement scenario. Those figures and the 50% baseline add to 95%, as published, so they should not be imported as a complete probability distribution. The firm also expected U.S. job gains of 400,000 to 700,000 a year through 2030 in the near term, about half the long-run average.

The spread is the point. A property adviser modeling an AI boost still included outcomes in which enthusiasm fades or displacement dominates. Tenants should be at least as explicit about their downside case.

The analysis also said U.S. office deliveries over the next few years were tracking near 5 million square feet annually, compared with a historical annual pace around 50 million. Limited deliveries can tighten selected markets later even while today’s vacancy remains high. Once again, national supply and local suitability can move differently.

Cushman & Wakefield, like CBRE and JLL, participates in commercial real estate markets. Its business gives it access to useful data and expertise. It also gives the firm an incentive to frame AI as a source of future property demand. A tenant should not import the 330 million square foot figure into a board deck as an independent fact about its portfolio. The number belongs to the adviser’s stated model and assumptions.

The same discipline applies to workplace technology vendors. CBRE describes a case in which its AI planning products helped a manufacturer fit all employees into an existing building, add a floor’s population without expanding the footprint and avoid capital and occupancy costs. Its April case study says planners could adjust a workshop scenario in under an hour.

That result demonstrates the value of faster scenario testing. It does not tell another company how many seats to remove. The case omits the manufacturer’s identity, detailed baseline, cost figure and longer-term employee outcomes. It also promotes CBRE products. A renewal team can learn from the method while demanding its own evidence.

Current vacancy and hybrid work make excess space a real cost. Vendor projections of AI-driven demand remain uncertain and commercially interested. Limited construction complicates the picture because a choice that looks easy to reverse nationally may be difficult to reverse in a specific submarket.

“AI will empty offices” and “AI will refill them” make the same mistake. Each turns a local, company-specific capital decision into a slogan.

A mature employer with slow growth, dispersed talent, low attendance and abundant local supply may rationally contract. A funded AI company hiring rapidly in San Francisco may rationally expand. A third company may shrink its total footprint while protecting an option in one market and investing in a better collaboration hub. The same technology wave can support all three decisions.

One forecast for HR, finance and real estate

The renewal team needs a common artifact before it needs a common conclusion. An AI-to-space scenario charter can supply one.

The charter is a one-page control document backed by the detailed models each function already maintains. It does not replace HR’s workforce plan, finance’s forecast, IT’s architecture or real estate’s portfolio data. It records the handful of assumptions that connect them and identifies who is accountable for updating each one.

Decision field	Low case	Operating case	Growth case	Evidence and trigger	Accountable owner
Total workforce	Annual range after approved automation and attrition	Approved plan plus funded roles	Product and revenue-linked hiring range	Requisitions, attrition, automation adoption, revenue	CHRO
Role mix	Functions and levels expected to decline	Roles redesigned within current teams	New product, implementation, security and sales roles	Job architecture, skills inventory, hiring pipeline	CHRO and business leaders
Location mix	Sites consolidated or hiring made location-flexible	Current hubs retained	Talent-critical hubs expanded	Candidate supply, compensation, client proximity	CHRO and corporate real estate
Peak attendance	Observed low-day and policy floor	Coordinated team schedule	Launch, training and client peaks	Badge distribution, room use, employee research	COO and workplace leader
AI deployment	Delayed adoption or limited process coverage	Funded use cases reach target teams	Product success or broad workflow adoption	Active use, task coverage, quality and cycle time	CIO and business leaders
Work settings	Fewer assigned seats, protected specialist settings	Mix aligned to redesigned roles	Added project, learning and client capacity	Role-to-setting map, unmet demand, booking data	Workplace leader
Technology load	Reuse current infrastructure	Planned network, security and device upgrades	Higher compute access, power, cooling and secure-room need	Architecture review, energy and network capacity	CIO
Lease action	Contract, sublease or exit	Renew with defined flexibility	Expand or protect adjacent capacity	Break dates, option prices, suitable local supply	CFO and real estate leader
Decision trigger	Named threshold and latest safe date	Named threshold and latest safe date	Named threshold and latest safe date	Monthly control review until signature	Executive sponsor

The table forces several useful conversations.

HR and business leaders build the workforce range together. They separate positions expected to disappear from roles expected to change or grow, map them by location and level, and name the workflow behind each productivity claim. When an AI product creates demand, the growth case names the revenue or customer milestone that releases hiring. A global headcount total or a general promise of “AI productivity” cannot do that work.

IT and workplace teams then translate the range into physical requirements. Some AI services run in external clouds and have little direct building impact. Others increase network traffic, secure access, device needs, audiovisual use or local support. Badge data, reservations, room demand and employee research reveal the occupancy side, but each has a blind spot: entry differs from productive use, intent differs from presence, and reported pain does not measure prevalence.

Finance prices rent, concessions, fit-out, technology, energy, operating expense, moving costs, sublease risk and the available options over the lease term. Corporate real estate supplies the irreversible dates. A board may hear “the lease expires next year” and assume there is time, even when an option notice, construction schedule or competing tenant closes the choice months earlier. If the growth and labor-cost cases cannot coexist, the executive sponsor owns that disagreement instead of letting the property model quietly average it.

Each scenario also needs a confidence label. Approved hires and observed attendance carry more weight than an estimate of model capability in 2030. The label belongs to the assumption, regardless of who presents it.

A monthly control meeting can keep the charter current during a renewal. The meeting should be short and tied to changes:

Did the workforce range move?
Did role or location mix move?
Did peak attendance move?
Did an AI deployment cross its adoption or quality threshold?
Did the property market change the price or availability of an option?
Is any decision date now inside the next 90 days?

If nothing moved, the scenario remains. If one assumption moved, every dependent field is checked. A product delay may preserve headcount longer. A hiring freeze may reduce near-term seats but make an expansion option cheaper than empty space. A return-to-office change may alter peak occupancy before it alters total workforce.

The charter should preserve employee and manager signals as evidence, not decoration. If teams report that AI adds review work, the expected productivity conversion date may need to move. If new employees cannot access coaching, the role-to-setting map may need more learning space. If facilities staff see repeated privacy workarounds, the fit-out plan may need enclosed rooms.

Procurement has a role as well. An AI contract can create a three-year software commitment inside a ten-year property commitment. Vendor exit terms, data portability and adoption targets affect how credible the workplace case is. Real estate counsel tests whether the physical options work as advertised: the exact expansion floor, notice date, landlord consent, restoration duty and sublease restriction. A workforce trigger has little value if the contract cannot execute it in time.

The artifact becomes especially useful when a board asks for one number. The team can give the recommended commitment and show the protected range. For example: renew 70% of the current area for five years, secure an option on 15% for two years, redesign the retained space for the operating role mix and set a contraction trigger if adoption and attrition cross named thresholds by a named date.

That statement is auditable. A promise that “AI will make us more efficient” is not.

At renewal, optionality has a price

Picture the renewal meeting 18 months before expiration.

The landlord has offered a five-year extension with a tenant improvement allowance. An adjacent floor is available now but may be leased within six months. The current office averages 47% attendance, reaches 79% on Tuesday and has a persistent shortage of small secure rooms. HR’s approved plan is flat for the next year. The AI product team believes a successful launch could add 180 implementation and sales roles in two cities. Finance expects automation to remove 120 positions over three years, but the first workflow has reached only 35% active use. IT needs network and privacy upgrades in either footprint.

Without a charter, each fact supports a different argument. The attendance average supports contraction. Tuesday supports more capacity. Flat headcount supports renewal. The product plan supports expansion. The productivity case supports waiting. The lease calendar supports deciding now.

With a charter, the team can distinguish commitments from options.

It can retain enough space for the operating case and redesign the settings that are demonstrably scarce. It can price the adjacent floor against the probability and timing of product-led hiring. It can negotiate a contraction right that falls after the automation transition, when the company will have evidence. It can compare that premium with the expected cost of carrying empty area. It can schedule the IT work only in the space that survives each scenario.

The decision still involves judgment. The charter makes the judgment visible.

That visibility matters because optionality is not free. A company that protects every upside case will carry too much space. A company that plans only for the lowest headcount case can lose access to talent, disrupt teams and pay more to re-enter a tight submarket. A company that delays until its AI forecast feels certain may discover that its strongest real estate options have expired.

The right amount of flexibility depends on the cost of being wrong in each direction. In a market with 21% vacancy and many suitable alternatives, the cost of taking too little space may be manageable. In a talent cluster where suitable buildings are being absorbed, the same error can be severe. In a specialized facility, replacement lead time can dominate rent. In a distributed service organization, location flexibility may reduce the need for a property hedge.

Leaders should ask for four numbers before approving the lease:

The workforce range for each lease year, with role and location mix.
The peak attendance range, with the team overlap that produces it.
The cost of the recommended commitment under the low, operating and growth cases.
The price and expiration date of each option that protects a material scenario.

They should also ask what evidence would change the recommendation. If the answer is “better AI,” the plan is unfinished. A useful answer names an adoption rate, a product revenue level, a hiring threshold, a peak-occupancy measure or a market event.

JLL’s 78% figure shows that leaders expect AI to reach the portfolio. Its 15% figure shows how few have turned that expectation into optimization. The distance between them will be closed one lease, fit-out and consolidation at a time.

No company will possess a perfect forecast of AI employment. The useful discipline is to put the uncertain forecasts in the same room before a long-term commitment is signed.

Then the lease can record what the company actually knows: the space it needs now, the outcomes it is willing to protect and the date when evidence must replace assumption.

At the next renewal meeting, the most useful AI slide may contain no model name at all. It may be the floor plan with three headcount ranges written beside the signature line.