Bring AI into daily workflows and keep it running every day
Horizon AI is KONST's enterprise AI adoption business line. One team sees the work through from start to finish, covering process inventory, custom application development, enterprise system integration, and usage governance after launch. The delivery standard is production go-live (a real data pipeline, real permission boundaries, and real cost records), not a proposal or a demo environment. Model access and usage governance run on the group's own ATP Token platform, which is in place from the first day of the proof of concept.
Four hurdles that keep AI projects stuck at proof of concept
The bottleneck rarely lies in model capability. It lies in the unengineered distance between the proof of concept and the production environment. These four hurdles have nothing to do with how well the model was chosen, but any one of them can keep a project at the demo stage.
Assessment, development, integration, and governance: four phases handled by one team
We do not sell packaged software, and we do not stop at a proposal. The only acceptance standard is going live. One team takes all four phases from start to finish, with no handover gaps and no "that part is out of our scope."
- Problem
Wrong scenario choice: starting from the model and looking for a use
When the model is chosen first and a use is found afterward, the resulting feature is not connected to any process, and at acceptance nobody can say whether it succeeded or failed.
Solution01 · ASSESSMENTAI adoption assessment and consulting
Take inventory of scenarios, data, and existing systems, and produce an actionable adoption blueprint and ROI assessment. First confirm that inputs and outputs are clear, data is accessible, and results can be verified, then decide how to implement.
Solution02 · DEVELOPMENTCustom AI application development
Knowledge retrieval, document processing, multilingual communication, and workflow automation are custom built around your actual process, not fitted to templates. Existing applications can stay in use, and adoption can be a hybrid.
- Problem
Go-live conditions are not defined up front
If the data pipeline, permission boundaries, cost caps, and service levels are not built to production standards at the proof-of-concept stage, almost everything has to be redone after the demo.
ProblemUsers have no reason to change how they work
The feature goes live, but it is not in the interface users open every day, and nobody guides the first group to use it. Adoption does not grow on its own.
Solution03 · INTEGRATIONEnterprise system integration
Connect AI to ERP, CRM, and existing workflows, with the data pipeline and permission design done right in one pass, so AI sits inside the interface users already use.
- Problem
Nobody owns it end to end
One group builds the demo, another takes over for go-live, and yet another runs operations. Data governance, permission boundaries, and change management in between have no owner, and the project stalls at the handover.
Solution04 · GOVERNANCELive operations and
usage governance Long-term operations for model access, cost control, and security and compliance. Usage, cost, quality, and permissions are visible in real time and auditable on the ATP Token platform.
Four types of applications ready to adopt, the rest customized by industry
The applications below already have production deployments. After assessment they can be adopted directly or adjusted to your process, and they can also be combined with custom development.
| Application | What it does | Suited for |
|---|---|---|
| Enterprise Knowledge Base | Product specifications, legal terms, and pricing methods are stored in one place and synced automatically. Ask in the group chat and get an answer right away, with a source cited for every answer. | Organizations where repetitive Q&A takes up specialist staff time |
| Real-time multilingual AI meeting translation | Cross-border meetings get real-time multilingual translation. After the meeting, a structured summary and action items are produced automatically, and decisions are stored and traceable. | Teams with sites in multiple countries and meeting notes that depend on manual work |
| Document AI | Extraction, comparison, and generation for contracts, specifications, and reports, connected to your existing document systems and review workflows. | Processes with high document review volume and highly repetitive formats |
| AI Factory Digital Twin | Builds a digital twin from production line and facility data for monitoring, simulation, and anomaly diagnosis. | Enterprises with a data collection foundation on the manufacturing floor |
The path from the first meeting to go-live
Every step has a clear output and acceptance point, so you always know where things stand and what comes next. Below is the pace of a typical project. The actual schedule depends on scope and your existing IT architecture.
- 01Adoption diagnosticWeeks 0 to 2
Scenario inventory, data health check, and feasibility and ROI assessment, producing an actionable adoption blueprint.
- 02PoC validationWeeks 2 to 8
Build the proof of concept to production standards, with a real data pipeline, real permissions, and real cost records, and deliver verifiable results in six weeks.
- 03System integrationWeeks 8 to 16
Connect to ERP, CRM, and existing workflows, and place AI inside the interface users already use.
- 04Production go-liveGo-live
Change management and rollout to seed users, with usage quotas and permission boundaries set. Adoption rate comes from design.
- 05Operations and governanceOngoing
Long-term governance of usage, cost, quality, and security, visible in real time and auditable on the ATP Token platform.
After go-live, four roles each get what they need
The benefit of adoption goes beyond "one department saved time." When the same architecture goes live, business, IT, finance, and audit each get the part that belongs to them.
Workflow steps really do get taken over
Meeting notes automatically become structured summaries and action items, knowledge questions are answered by the knowledge base with sources cited, and document and content production drops from "half a day" to a single command in a conversation. AI appears in the interfaces people already use, so there is no new tool to learn.
One access framework, no code changes to switch models
Model access converges on a single integration point, with an interface compatible with the SDKs of major providers, and in most cases only base_url and the key change. New models fall under the existing permission framework as soon as they are listed, and when a single upstream has a problem, the routing layer hands over to another source.
AI spend can be attributed and allocated
The token count and charge of every call map to an organization, workspace, and project. Quota is allocated from the top down and usage is attributed from the bottom up, so each department's cost comes straight from the platform with no separate estimate and no waiting for the end-of-month bill.
Logs are the audit evidence
Every request keeps its model, key, token usage, and status. Permissions are scoped to the key, the project level decides which models can be called, and logs remain in the audit record after a key is revoked.
Stuck at proof of concept or ready to go live: the difference is these six things
For the same scenario, how you work decides whether it stays at the demo stage or actually goes live. These six items are the conditions we check one by one in the adoption diagnostic.
| Item | Projects stuck at proof of concept | Projects ready to go live |
|---|---|---|
| Scenario selection | Pick a model first, then look for a use | Take inventory of processes first, then pick steps with clear inputs and outputs and verifiable results |
| Data pipeline | Works on exported sample data | Built on production data sources and update frequency |
| Permission boundaries | Developers share one key | Permissions are scoped to project keys, with callable models set key by key |
| Cost control | You learn what you spent only when the bill arrives at month end | Usage is attributed to projects request by request, with an alert before quota is exceeded |
| Change management | Counted as done once the feature is live | Seed-user rollout and process adjustment happen together |
| Operations ownership | No clear owner after go-live | One team from assessment to operations, including tracing of abnormal usage |
Run it on ourselves first, then deliver it to customers
Horizon AI treats itself as its first customer: every approach we recommend to customers has first run on our own projects. The figures below come from projects already live, and customer names are anonymized under confidentiality agreements.
Internal enterprise AI Agent: bringing agents into daily workflows
An internal AI agent deployed on a collaboration platform: meeting notes automatically become action items, the knowledge base answers questions on the spot, and a single command produces brand-grade presentations. All model usage is governed centrally by ATP.
- 100%Meeting notes and action items generated automatically
- 10×Faster production of presentations and marketing materials
- 24HUsage anomalies located and fixed within one day
Token governance for an enterprise AI Agent: turning wasted usage into predictable cost
An internal AI agent running on multiple models uses project-level governance to bring scattered token usage into a single billing point, so cost can be attributed and waste is visible.
- 100%of model calls consolidated into a single governance gateway
- 1consolidated bill replacing per-provider reconciliation
- 0token spend with no budget owner
Adoption services are quoted by scope, and usage after go-live is quota-based
Adoption is priced in two parts. Consulting, development, and integration during the project are quoted by scope. After go-live, model usage runs on ATP Token's quota system: top up first, charged as you use, no monthly fee. Both parts are explained together in a single proposal.
| Item | Pricing basis | Description |
|---|---|---|
| Adoption assessment and consulting | Per project | Quoted by inventory scope and number of scenarios, producing an adoption blueprint and ROI assessment. The output of this phase stands on its own and is not tied to later development. |
| Custom application development | By feature scope | Quoted by feature scope and acceptance criteria. You can choose existing applications, custom development, or a mix, which avoids paying twice for features that already work. |
| Enterprise system integration | By integration depth | Quoted by existing IT architecture, number of connected systems, and data pipeline complexity. |
| Model usage (after go-live) | Quota-based | Quota is deducted by input + output tokens multiplied by each model's rate. Top up first, charged as you use, no monthly fee, and quota does not expire. |
| Operations and usage governance | Annual | Long-term operations, quota control, and security and compliance support are agreed by service scope. Customers with larger usage can ask about an enterprise contract. |
Common questions about adoption assessment, schedule, security, and pricing
Where should an enterprise start when adopting AI applications?
Take inventory of processes, not models. First find the steps where inputs and outputs are clear, data is accessible, and results can be verified. A step is ready to go first only when all three hold.
With a single access point in place, models can be swapped at any time, but changing a process design costs far more. So the first step is a process inventory and data health check, not a model comparison.
How soon can we see verifiable results?
The typical schedule is: weeks 0 to 2 for the adoption diagnostic, weeks 2 to 8 for the proof of concept, and weeks 8 to 16 for system integration, followed by go-live and a move to operations and governance. The proof of concept aims to deliver verifiable results in six weeks and is built to production standards: a real data pipeline, real permissions, and real cost records.
The actual schedule depends on the number of scenarios, data maturity, and the depth of integration with existing systems. The adoption blueprint states the outputs and acceptance points of each phase.
Does adoption always require a full custom build?
Not necessarily. After the assessment you can choose an existing application, a custom build, or a mix of the two. Four types of applications already have production deployments: enterprise knowledge base, real-time multilingual AI meeting translation, document AI, and AI factory digital twin. They can be adopted directly or adjusted to your process.
The scope of system integration depends on your existing IT architecture.
How much code in an existing application has to change to connect to the platform?
The platform sits on top of your existing call pattern and does not require the application to adopt a new protocol. The interface is compatible with the SDKs of major providers, and in most cases you only change base_url and the key.
During adoption we also take inventory of existing self-provisioned keys and consolidate them under project keys with unified governance.
How are data security and data location handled?
In the first phase of adoption, deployment location and access boundaries are decided by data classification: which workloads can use public models, and which must stay inside the enterprise network or in a specified region.
Permissions are scoped to the key, and the project level decides which models can be called. Every request is logged, and logs remain in the audit record after a key is revoked. Workloads that must be hosted in-country or in your own data center are assessed per project.
How is AI usage cost allocated to each department?
The token count and quota deducted for each call can be traced back to the organization, workspace, and project. Quota is allocated from the top down and usage is attributed from the bottom up, so each department's cost comes straight from the platform with no separate estimate.
An alert fires before quota reaches its limit, which turns AI spend from an end-of-month result into a variable you control in advance.
Who operates it after go-live?
The same team handles assessment, development, integration, and operations, with no handover at go-live. Operations cover model access, quota and permission control, tracing of abnormal usage, and security and compliance support.
Your own IT team can also get access to the platform console and share the full picture of usage and permissions.
How are Horizon AI and ATP Token related?
Horizon AI is KONST's enterprise AI adoption business line. ATP Token is an enterprise AI model management platform developed in-house by Horizon AI, and it handles model access, metering, and governance at the base layer of adoption projects.
ATP can also be adopted on its own, connecting existing applications as a Token as a Service offering. Learn about the ATP Token service →
Ready to connect AI to your business processes?
Start with a process inventory and plan your company's adoption path. In the first meeting we can tell which steps are best to do first.