Nexus brings quotes, orders, inventory, accounts and team workflows into one ERP. Its agents work on that data alongside your team, under approval and audit rules.
They notice the unanswered quote and the overdue invoice, and suggest what to do. On every step that reaches a customer, an authorised person has the final say.
Built to run Karea Bilgi Teknolojileri's own sales and operations. In production at two companies today.
The quote was sent on 12 Sep 2026 and has had no reply for 12 days. It expires on 30 Sep 2026.
of enterprises will demote or decommission autonomous AI agents by 2027 because of control gaps that surface in production.
Gartner, 26 May 2026
of Turkish firms with 10-49 employees use ERP software. For small firms in the EU the figure is 41%.
TÜİK 2025 · Eurostat 2025
of Turkish firms with 10 or more employees use AI. The EU average is 20%.
TÜİK 2025 · Eurostat 2025
The global platforms that offer governed agents either lack Turkish support or process data outside Türkiye. In local ERPs, AI is still limited to assistants that answer questions. Nexus is being built so you can give agents work and keep control.
About three in four Turkish firms with 10-49 employees do not use an ERP (TÜİK 2025). These follow-ups often depend on one person remembering.
The quote went out 12 days ago and nobody replied.
It is about to expire and no one noticed.
An overdue invoice reaches no one's desk until the month-end report.
The call-back promised in a meeting stays in someone's diary. Meanwhile the customer calls your competitor.
Nexus agents are designed for three results.
Quotes waiting for a reply and overdue follow-ups are found every morning. They reach the record's owner as a task or a notification.
Receivables 7 or more days overdue are found without waiting for month end. A payment reminder e-mail comes to you for approval.
Requests waiting for approval reach the right person with the reason attached.
Three rules make this safe
What the agent does
The check runs every day at 07:30. When a quote is accepted, the agent looks again within minutes. Depending on its level, each finding becomes a suggestion, a follow-up task or an action waiting for approval.
Autonomy
The agent starts at L1, in shadow mode. An administrator sets the level, and every change is written to the audit trail.
Finds and counts the records that need follow-up. No model call, no cost.
Nothing changes
Writes a suggestion for each finding. An admin rates it Right or Wrong.
Nothing changes
Also writes a follow-up task and a notification for the record's owner.
An internal task, logged under the agent's name
Puts the customer-facing step up for approval: a quote revision or a reminder e-mail.
A person approves; the action runs in their name
Act alone Not offered in Nexus.
This matches the four autonomy levels Gartner recommends (Observe, Advise, Act with Approval, Act Autonomously).
Nexus deliberately does not offer the fourth level (Act Autonomously).
Workflow
Every morning at 07:30, and when a quote is accepted, rule-based code finds the records that need follow-up. If there is nothing to find, the model is not called and the run costs nothing.
Company, project and quote names become random codes. Amounts are rounded and dates become day counts. Names, e-mails and phone numbers are never sent.
The summary goes to the model your company has approved, in a single call.
The call passes through Cloudflare AI Gateway, which keeps no copy of the data.
The suggestion reaches an admin, who rates it Right or Wrong.
At L2 a follow-up task goes to the owner; at L3 the action is put up for approval.
Each run records the model, the cost and the summary sent.
Every agent action is written to the audit trail under the agent's name.
Every suggestion shows what the model saw
Fields sent to the model
Fields withheld
Route: Amazon Bedrock · Cloudflare AI Gateway
Data
The summary the agent sends does not reveal who it is about. Codes are generated fresh for every run, and the key to them lives in Nexus's memory only while the run lasts, then it is dropped.
Example. In a breach, this is all anyone could see: "[C-7KQ]'s quote: 120,000 TRY, unanswered for 9 days."
Every column the agent reads is classed as coded, clear or never sent. An unclassified column stops the tests.
The call passes through Cloudflare AI Gateway with request and response logging off. Cost is metered in Nexus.
The suggestion card shows which fields the model saw. The summary sent in each run is stored exactly as sent.
Platform
Ten modules share one customer record, one permission model and one audit trail.
The quote sales enters and the invoice finance tracks belong to the same account.
| FTR-2026-000517 | Baltacıoğlu Metal San. ve Tic. A.Ş. | 1 Sep 2026 | TRY 312,400.00 | TRY 162,400.00 | Overdue | Elif A. |
|---|---|---|---|---|---|---|
| FTR-2026-000664 | Karaduman Tekstil A.Ş. | 4 Sep 2026 | TRY 94,250.00 | TRY 94,250.00 | Reminder awaiting approval | Sales Operations Agent |
| FTR-2026-000702 | Doğubey Rüzgar Elektrik Ltd. Şti. | 1 Oct 2026 | TRY 58,600.00 | TRY 58,600.00 | Open | Elif A. |
| ALS-2026-004411 | Yeşilova Gıda Sanayi A.Ş. | 23 Aug 2026 | TRY 133,700.00 | — | Paid | Murat T. |
Infrastructure
AI exit and front door
Every model call leaves through AI Gateway. The gateway holds the cloud keys, so no AWS key sits on the server. No copy of requests or responses is kept. Next: a firewall and access control in front of every installation.
Primary model route
Agents run on the model your company approves, through Amazon Bedrock. Bedrock does not store prompts and responses or use them to train models. Bedrock invocation logging is set off when the account is set up.
Second route and identity
Azure AI Foundry is ready as a second route to the model. Automatic failover is planned. Sign-in with Microsoft work accounts is planned.
Engineering measures running today in production installations.
The session cookie is encrypted and signed with iron-session. Page scripts cannot read it, and the session closes on its own after 12 hours.
Every installation is served with an automatic TLS certificate. HSTS forces the browser to use encrypted connections only.
CSP lets only Nexus's own scripts run on the page. The page cannot be embedded in another site.
The reverse proxy rewrites the client IP header on every request, so the sign-in limit cannot be bypassed with a fake header.
The app, background jobs and database run in separate Docker containers for each company.
Every change passes a 148-scenario browser test suite. Database permissions are checked by their own tests.
Cloudflare, Amazon Web Services, Amazon Bedrock, Anthropic, Microsoft Azure and Azure AI Foundry are trademarks of their owners.
Security
Protection in four layers, from sign-in to database.
Scenarios
Sign-in fails without the second-step code. Guessing locks after 5 failed tries. Resetting the password closes every open session.
They can touch only the records their role allows. Not even the app itself can change audit records; if an old entry changes, the chain breaks and it shows.
Encrypted backups sit in an account separate from production. The target is to be working again within 4 hours, with at most 6 hours of data lost.
The agent works only through controlled actions and cannot approve its own request. Every step that reaches a customer waits for approval; internal follow-up tasks are logged under the agent's name. An admin stops the agent with one switch, which also closes its account.
All that leaks is codes, rounded amounts and day counts: "[C-7KQ]'s quote: 120,000 TRY, unanswered for 9 days." The key to the codes stays in Nexus, and only while the run lasts.
Each call is priced at its worst case before it is made. If the daily cap or the weekly allowance is used up, the agent does not run and the run is recorded in red.
Usage
The subscription is priced by the modules you choose, with one of three AI tiers on top. Each tier grants a weekly allowance that your team sees as a percentage. The allowance renews every Monday at 00:00.
We share module and tier prices in the demo, based on the scope of your installation.
Demo
Share how your team is organised and what matters most. In the call we show the platform on your own processes and work out together where agents would save your team time.
Nexus brings quotes, orders, inventory, accounts and team workflows into one ERP. Its agents work on that data alongside your team, under approval and audit rules.
They notice the unanswered quote and the overdue invoice, and suggest what to do. On every step that reaches a customer, an authorised person has the final say.
Built to run Karea Bilgi Teknolojileri's own sales and operations. In production at two companies today.
The quote was sent on 12 Sep 2026 and has had no reply for 12 days. It expires on 30 Sep 2026.
of enterprises will demote or decommission autonomous AI agents by 2027 because of control gaps that surface in production.
Gartner, 26 May 2026
of Turkish firms with 10-49 employees use ERP software. For small firms in the EU the figure is 41%.
TÜİK 2025 · Eurostat 2025
of Turkish firms with 10 or more employees use AI. The EU average is 20%.
TÜİK 2025 · Eurostat 2025
The global platforms that offer governed agents either lack Turkish support or process data outside Türkiye. In local ERPs, AI is still limited to assistants that answer questions. Nexus is being built so you can give agents work and keep control.
About three in four Turkish firms with 10-49 employees do not use an ERP (TÜİK 2025). These follow-ups often depend on one person remembering.
The quote went out 12 days ago and nobody replied. It is about to expire and no one noticed.
An overdue invoice reaches no one's desk until the month-end report.
The call-back promised in a meeting stays in someone's diary. Meanwhile the customer calls your competitor.
Nexus agents are designed for three results.
Quotes waiting for a reply and overdue follow-ups are found every morning. They reach the record's owner as a task or a notification.
Receivables 7 or more days overdue are found without waiting for month end. A payment reminder e-mail comes to you for approval.
Requests waiting for approval reach the right person with the reason attached.
Three rules make this safe
What the agent does
The check runs every day at 07:30. When a quote is accepted, the agent looks again within minutes. Depending on its level, each finding becomes a suggestion, a follow-up task or an action waiting for approval.
Autonomy
The agent starts at L1, in shadow mode. An administrator sets the level, and every change is written to the audit trail.
Finds and counts the records that need follow-up. No model call, no cost.
Nothing changes
Writes a suggestion for each finding. An admin rates it Right or Wrong.
Nothing changes
Also writes a follow-up task and a notification for the record's owner.
An internal task, logged under the agent's name
Puts the customer-facing step up for approval: a quote revision or a reminder e-mail.
A person approves; the action runs in their name
Act alone Not offered in Nexus.
This matches the four autonomy levels Gartner recommends (Observe, Advise, Act with Approval, Act Autonomously). Nexus deliberately does not offer the fourth level (Act Autonomously).
Workflow
Every morning at 07:30, and when a quote is accepted, rule-based code finds the records that need follow-up. If there is nothing to find, the model is not called and the run costs nothing.
Company, project and quote names become random codes. Amounts are rounded and dates become day counts. Names, e-mails and phone numbers are never sent.
The summary goes to the model your company has approved, in a single call. The call passes through Cloudflare AI Gateway, which keeps no copy of the data.
The suggestion reaches an admin, who rates it Right or Wrong. At L2 a follow-up task goes to the owner; at L3 the action is put up for approval.
Each run records the model, the cost and the summary sent. Every agent action is written to the audit trail under the agent's name.
Every suggestion shows what the model saw
Fields sent to the model
Fields withheld
Route: Amazon Bedrock · Cloudflare AI Gateway
Data
The summary the agent sends does not reveal who it is about. Codes are generated fresh for every run, and the key to them lives in Nexus's memory only while the run lasts, then it is dropped.
Example. In a breach, this is all anyone could see: "[C-7KQ]'s quote: 120,000 TRY, unanswered for 9 days."
Every column the agent reads is classed as coded, clear or never sent. An unclassified column stops the tests.
The call passes through Cloudflare AI Gateway with request and response logging off. Cost is metered in Nexus.
The suggestion card shows which fields the model saw. The summary sent in each run is stored exactly as sent.
Platform
Ten modules share one customer record, one permission model and one audit trail. The quote sales enters and the invoice finance tracks belong to the same account.
| FTR-2026-000517 | Baltacıoğlu Metal San. ve Tic. A.Ş. | 1 Sep 2026 | TRY 312,400.00 | TRY 162,400.00 | Overdue | Elif A. |
|---|---|---|---|---|---|---|
| FTR-2026-000664 | Karaduman Tekstil A.Ş. | 4 Sep 2026 | TRY 94,250.00 | TRY 94,250.00 | Reminder awaiting approval | Sales Operations Agent |
| FTR-2026-000702 | Doğubey Rüzgar Elektrik Ltd. Şti. | 1 Oct 2026 | TRY 58,600.00 | TRY 58,600.00 | Open | Elif A. |
| ALS-2026-004411 | Yeşilova Gıda Sanayi A.Ş. | 23 Aug 2026 | TRY 133,700.00 | — | Paid | Murat T. |
Infrastructure
AI exit and front door
Every model call leaves through AI Gateway. The gateway holds the cloud keys, so no AWS key sits on the server. No copy of requests or responses is kept. Next: a firewall and access control in front of every installation.
Primary model route
Agents run on the model your company approves, through Amazon Bedrock. Bedrock does not store prompts and responses or use them to train models. Bedrock invocation logging is set off when the account is set up.
Second route and identity
Azure AI Foundry is ready as a second route to the model. Automatic failover is planned. Sign-in with Microsoft work accounts is planned.
Engineering measures running today in production installations.
The session cookie is encrypted and signed with iron-session. Page scripts cannot read it, and the session closes on its own after 12 hours.
Every installation is served with an automatic TLS certificate. HSTS forces the browser to use encrypted connections only.
CSP lets only Nexus's own scripts run on the page. The page cannot be embedded in another site.
The reverse proxy rewrites the client IP header on every request, so the sign-in limit cannot be bypassed with a fake header.
The app, background jobs and database run in separate Docker containers for each company.
Every change passes a 148-scenario browser test suite. Database permissions are checked by their own tests.
Cloudflare, Amazon Web Services, Amazon Bedrock, Anthropic, Microsoft Azure and Azure AI Foundry are trademarks of their owners.
Security
Protection in four layers, from sign-in to database.
Scenarios
Sign-in fails without the second-step code. Guessing locks after 5 failed tries. Resetting the password closes every open session.
They can touch only the records their role allows. Not even the app itself can change audit records; if an old entry changes, the chain breaks and it shows.
Encrypted backups sit in an account separate from production. The target is to be working again within 4 hours, with at most 6 hours of data lost.
The agent works only through controlled actions and cannot approve its own request. Every step that reaches a customer waits for approval; internal follow-up tasks are logged under the agent's name. An admin stops the agent with one switch, which also closes its account.
All that leaks is codes, rounded amounts and day counts: "[C-7KQ]'s quote: 120,000 TRY, unanswered for 9 days." The key to the codes stays in Nexus, and only while the run lasts.
Each call is priced at its worst case before it is made. If the daily cap or the weekly allowance is used up, the agent does not run and the run is recorded in red.
Usage
The subscription is priced by the modules you choose, with one of three AI tiers on top. Each tier grants a weekly allowance that your team sees as a percentage. The allowance renews every Monday at 00:00.
We share module and tier prices in the demo, based on the scope of your installation.
Demo
Share how your team is organised and what matters most. In the call we show the platform on your own processes and work out together where agents would save your team time.