The Short Answer
HP and Legora show that enterprise AI value sits in the deployment and workflow layers, not the model. HP spreads AI across 80,000 devices and the cloud by workload; Legora passed $200M ARR on other companies' models. ibl.ai gives an enterprise both layers to keep: you own all the code and the data, stay model-agnostic, and deploy anywhere.
Two announcements a week apart make the same argument from opposite ends of the market. One is a hardware company describing its own workforce. The other is a legal software company describing its revenue.
Neither is a story about a better model. That is the part worth taking seriously, and the part worth checking before repeating.
What did HP actually report about its hybrid-edge AI deployment?
HP reported that a hybrid-edge AI strategy across 50,000 employees and 80,000 devices produced a 38% productivity increase and 4 hours saved per employee each week.
The claim was made in a post on HP's official X account on 29 September 2026. It describes "bringing AI closer to where work happens instead of relying solely on the cloud."
That wording matters: "solely". HP is not describing a cloud exit. It is describing workload placement, where some inference runs on the device and some stays in the data center.
HP's chief strategy and transformation officer, Prakash Arunkundrum, set out that split in an HP newsroom post on 24 September 2026. Local inference handles latency-sensitive and security-sensitive work; the cloud keeps the largest, most complex, most compute-intensive workloads.
He also writes that 63% of workers use AI in their workflows daily or weekly, citing HP's own research, and that his own laptop runs a 20B-parameter model locally.
One disclosure belongs next to every number above: HP sells the AI PCs this strategy runs on. Its internal deployment is a genuine data point and also a product demonstration.
Why do HP's AI productivity numbers range from 16% to 38%?
Because HP has published at least three internal productivity figures in 2026, each from a different cohort, and only the smallest with its testing basis footnoted.
| Date | Source | Cohort | Productivity gain |
|---|---|---|---|
| 24 Mar 2026 | HP press release | 4K+ employees, 4.5 months, AI PCs | 16% |
| 26 Mar 2026 | InformationWeek interview | "wherever we have deployed" AI tools | about 20% |
| 29 Sep 2026 | HP on X | 50,000 employees, 80,000 devices | 38% |
The 16% comes from HP's March 2026 "Reimagines the Future of Work" release, footnoted to 4.5 months of testing in HP's internal Insider Program, 4K+ employees on EliteBook X models.
The "about 20%" is Arunkundrum speaking to InformationWeek two days later, when he also said about half of HP's employees had some kind of agent tool deployed.
The 38% is the September figure, at full scale, with no methodology we could find published alongside it.
These are not contradictions. A small pilot, a partial rollout and a full rollout should produce different numbers. But they are a warning about lifting one figure into a slide: the 38% is HP's to defend, and it arrived in a social post.
What survives the caveat is the direction. Every figure HP has published points the same way, and the largest came from the widest deployment.
Why would an enterprise run AI on the device instead of only in the cloud?
Because cost, latency and data control all change when inference moves closer to the user, and the cloud is not the cheapest place for routine work.
Arunkundrum told Fortune in May 2026 that edge AI "drives down the total cost of ownership by slashing token use." Every cloud request is a metered token; a request served on hardware you already own is not.
The same interview tied local inference to data sovereignty, and reported that 78% of HP's APAC staff use AI in their daily workflow.
That is the economic logic behind every hybrid design: route the high-volume, low-difficulty work to the cheapest capable compute, and pay frontier prices only where the task needs a frontier model.
We laid out the device side of this in On-Device AI Agents Are Enterprise's Next Moat, and the deployment patterns in the hybrid cloud and on-premise AI platform guide.
The precondition is control over placement. A workload can only be moved to the edge, a private cloud or an air-gapped server if the platform running it can be deployed in those places. Hybrid is a deployment decision, and only the party that controls the stack gets to make it.
How did Legora reach $200M ARR without building its own legal model?
Legora built the legal workflow and ran it on other companies' frontier models, betting that general models would keep improving instead of fine-tuning a legal model of its own.
Legora announced on X on 22 September 2026 that it had passed $200M in ARR. It took 18 months to go from $1M to $100M, and less than six to double it, with 130,000 lawyers using it monthly across 2,100 firms and legal teams in more than 80 countries.
CFO David Eckstein followed on 1 October with Q3 figures: 2,300+ customers across 80 countries, 800 of them added during Q3, and 40%+ of new business from in-house legal teams.
On the model question, CEO Max Junestrand explained the choice at YC's Startup School 2026. The early consensus was that you had to fine-tune a model; Legora went the other way, partly for lack of money and partly from conviction that "the models will keep improving."
A 2025 Microsoft customer story describes the same posture from the infrastructure side: Legora relies on "the latest, and multiple AI models," including GPT-4o and "GPT-o1" via Azure OpenAI.
So "own the workflow, rent the model" describes Legora fairly. It competes on the workflow (review, research and drafting), and treats the model as a supply it can swap.
What do HP's and Legora's results have in common?
Both put the value in a layer above the model, and both treat the model as the interchangeable part.
| Layer | HP | Legora |
|---|---|---|
| Model | Local models on device plus cloud models | Multiple third-party frontier models |
| Where the value came from | Deployment: placing each workload on device or cloud | Workflow: legal review, research and drafting |
| Headline number | 38% productivity, 80,000 devices | $200M ARR, 2,300+ customers |
| Who owns that layer | HP, for its own workforce | Legora, rented to its customers |
The last row is where the two stories diverge, and it is the row most commentary skips.
HP owns its deployment layer because it builds the hardware and runs the fleet. Legora owns its workflow layer because it is the product. The 2,300+ customers using it rent that layer; the vendor runs the models and sets the hosting options their matters can use.
If the workflow layer is the moat, should an enterprise rent it or own it?
For a firm whose workflows are its competitive knowledge, renting the layer where value accumulates moves the lock-in from the model to the workflow, which is harder to leave.
"Own the workflow, rent the model" is good advice, and Legora's growth shows the market agrees. The catch is who does the owning: for Legora's customers, the workflow is rented too.
That is a reasonable trade for many teams. It is a harder one for a firm whose playbooks, clause libraries and review standards are the reason clients hire it, or for a regulated organization that has to answer where privileged data is processed.
We covered the legal version in the Harvey AI alternative and on-premise legal AI posts.
HP's lesson points the same way. Workload placement saved HP money because HP could choose where each workload runs. An enterprise that rents a managed AI product inherits the vendor's placement, and its cost curve.
The model is the part both companies proved you can rent. The workflow and the deployment are the parts they proved are worth keeping.
How does ibl.ai let an enterprise own the workflow and deployment layers?
ibl.ai is an agentic AI platform where you own all the code and the data, delivered as source code you run on your own infrastructure.
It is model-agnostic: the same agents and workflows run on Claude, GPT, Gemini, Llama or a local open-weight model, and you can switch providers without rewriting them. That is Legora's model posture, held by the customer instead of the vendor.
You can deploy anywhere: your cloud, on-premise, GovCloud or fully air-gapped, so routine inference can sit next to the work while heavier calls go to a frontier API. That is HP's placement decision, made by the organization that pays the bill.
The placement reaches the device too: the ibl.ai Mac app installs a local model runtime, so inference happens on the laptop and keeps working with the network off.
Billing is usage-based, with no per-seat pricing, so cost follows the tokens actually consumed rather than headcount. 1.6M+ users across 400+ organizations run the platform this way, including NVIDIA, MIT, and Syracuse University.
ibl.ai is family-owned and operated from New York, NY, a long-term partner rather than a vendor that sells licenses and moves on. Agentic OS is where to start, and air-gapped AI covers the most restrictive deployments.
Related: On-Device AI Agents Are Enterprise's Next Moat · Harvey AI Alternative · Hybrid Cloud and On-Prem AI Platform