Work Transformers Labs · Research

Project Horizon

Applied research into how built-environment firms move from AI pilots to AI embedded in everyday operational work: model flexibility, system integration, governance and safe deployment.

Applied researchOngoing programme
The state of play
Everyone is piloting. Almost no one is operating.

Nearly nine in ten real estate investors and owners are running AI pilots, and the figure is similar for occupiers. Yet only around one in twenty report hitting all of their goals, and more than six in ten say they are not ready to scale beyond the pilot. The gap is rarely the model. It is the absence of the operational setup around it.

88%
of investors and owners are piloting AI
JLL · 2025
5%
have achieved all of their AI goals
JLL · 2025
60%+
are not ready to scale beyond pilots
JLL · 2025
The pilot trap
Why most AI stalls before it reaches the work.

A pilot proves a model can do something interesting. Production demands something harder: the same thing, the same way, every time, inside real systems, under real governance, with people who trust the output. Three patterns stall the move.

Tool sprawl

A dozen disconnected tools, each impressive alone, none of them part of how the work actually flows.

Single-model dependence

A system wired to one provider turns fragile the moment that model changes, reprices or fails.

No path to production

A demo with no architecture for evidence, review or governance has nowhere to go once the applause stops.

How it fits together
How AI actually reaches the work.

Connect what you already run, orchestrate it through a operational core, and deliver outcomes your team can trust.

Your documentsOccupancy dataFinance systemsEmail and filesExisting toolsFaster reviewsConsistent outputFull audit trailLess manual workModel-flexibleLower riskConnectNormaliseRoute modelApply rulesOrchestrateHuman reviewAuditMonitorDeliverYOUR SYSTEMSTHE OPERATIONAL COREOPERATIONAL OUTCOMES
The research question

Project Horizon studies the layer that turns AI from a set of tools into something embedded in how a firm actually works: how to stay flexible across models, integrate with the systems work already lives in, keep people in control, and deploy safely at scale. In short, what has to be true for AI to survive contact with real operations.

The lock-in tax
Betting the operating model on one provider is the expensive mistake.

Single-vendor stacks look fast until you try to move. Integration costs routinely run three to five times initial estimates, only about six percent of enterprises can switch providers without serious disruption, and the average switch costs more than three hundred thousand dollars per project. The mature pattern is the opposite of lock-in: staying model-agnostic, designing deployment so models can change underneath, and keeping the freedom to route each task to the provider best suited to it. The largest software firms now market multi-model optionality as a feature, not a compromise.

3 to 5x
the true cost of single-vendor integration versus estimates
Industry analysis · 2026
6%
of enterprises can switch model providers cleanly
Industry analysis · 2026
$315k+
average cost to switch provider, per project
Industry analysis · 2026
What we are exploring
Six questions that decide whether AI scales.
Model flexibility

How to route each task to the right model, and swap providers without rebuilding the work.

System integration

How AI reads from and writes back to the systems work already lives in, from data rooms to Excel to CAFM.

Safe deployment

What guardrails, permissions and review a system needs before it touches a real decision.

Operational orchestration

How supervised, multi-step systems replace one-shot prompts, with escalation and audit.

The economics of availability

How negotiated capacity changes reliability, latency and price-performance at scale.

Measurement

How to measure deployed AI by operational outcome, not by tools bought or prompts run.

Design principles
Six principles for AI that survives contact with operations.
01
Model-agnostic by default

No system is wired to a single provider. The right model is chosen per task, and can change without rebuilding the work.

02
Model-flexible by design

Deployment is designed so the model can change underneath without disturbing the system above.

03
Evidence and traceability as a requirement

Every figure and claim traces to a source. An output that cannot be checked does not ship.

04
People at the decisions that matter

Automation carries the manual load; a named person owns the judgement and the sign-off.

05
Governance designed in, not bolted on

Permissions, audit and escalation are part of the architecture from day one, not a later compliance project.

06
Measured by outcome, not activity

Success is a faster, more consistent system, not the count of tools deployed or prompts run.

What clients can learn

Firms working with us through Horizon leave with a clear view of their own AI estate: where pilots are stuck and why, where single-vendor risk is accumulating, and what embedding AI into their operational work would actually look like. The research is not abstract. It is the same architecture we deploy.

Where this is heading
The firms that embed AI into everyday operational work will move faster than the firms still running disconnected pilots.

Three forces are converging: models are becoming interchangeable commodities, governance and operational integration are becoming increasingly important, and the value is shifting from cost-cutting to competitive advantage. The firms that embed AI into everyday work now will move ahead. The rest will keep buying tools.

Related research and outputs
How we publish

What we share, and what we keep.

Project Horizon is open research into deployment and governance: what has to be true for AI to survive contact with real operations. We publish the questions we are wrestling with, the patterns that recur across real deployments, and the principles that hold up under pressure. We do not publish client data, or the parts of the method still being worked out. It is research, not a product: the questions are the interesting part, and the answers are earned in the work.

Work Transformers Labs

Following the deployment research?