Open AI to the whole company. Keep control, records, and cost with you.
ajan runs every request through your identity, access, guardrail, cost, and approval rules. Teams move faster; IT, security, and finance see what happens at any moment.
A request's journey
Selin Β· Sales team
βSummarize Ege Building's orders last quarter and their contact details.β
An answer for everyone at the buying table
No one person makes the enterprise AI decision. Each stakeholder asks a different question; ajan answers each in its own terms.
βHow does it connect to our systems, and how is it managed?β
- Ready integrations, MCP, and your own API behind one permission model
- SSO and SCIM connection paths with your identity provider
- Model-independent: switching providers doesn't change the agent
Documents and resources
What the agent knows is exactly what you manage.
Documents, web pages, API sources, and incoming email go through the same pipeline, with every stage tracked. You decide which document is published, which has expired, and who can access which knowledge area.
1
Intake
File, web, API, email
2
Scan
Malware check
3
Text
Extraction incl. OCR
4
Structure
Sections and chunks
5
Index
Semantic and keyword search
6
Publish
Version and validity
Version and validity
Publication state and validity date are tracked; old versions leave search.
Connected sources
Websites, HTTP APIs, and connections are synced regularly.
Access per knowledge area
Each knowledge area is opened to specific agents and people.
Trash and retention
Deleted documents can be restored during retention, then cleaned up.
Guardrails
Limits on what an agent says and does, defined per agent.
Guardrails are modules added to an agent; they work together and their effective state is visible in one view. When an answer is stopped by a guardrail, that goes into the disclosure record too.
Topic boundary
The agent doesn't answer questions outside its remit.
"I can't help with that; ask me about HR policies."
Output PII masking
Personal data in the answer is masked before it reaches the user.
0532 β’β’β’ β’β’ 17
Tool risk ceiling
Limits the highest-risk tier of action the agent may call.
Data-deleting action: closed to this agent
Source-limited answers
Answers only from company sources; says so when there's no basis.
"I couldn't find an approved source on this."
Extra technical protections apply to connected web addresses and database tools (blocking internal network addresses, mandatory tenant filter).
Personal data (PII)
Personal data stays where it's needed and stays out of sight where it isn't.
Personal data in answers is hidden before it reaches the user by the masking module you enable for the agent. Documents are flagged for personal data at intake; tool-call records keep requests and responses masked.
Customer: AyΕe YΔ±lmaz, ID 12345678901, phone 0532 555 44 17, email ayse.yilmaz@example.com
Customer: AyΕe YΔ±lmaz, ID β’β’β’β’β’β’β’β’β’01, phone 0532 β’β’β’ β’β’ 17, email aβ’β’β’@example.com
Output masking
A guardrail module enabled per agent.
Flagging at intake
A personal-data flag in document preprocessing.
Masked action records
Tool requests and responses are stored masked.
Per-person boundary
Answers based on personal data go only to authorized people.
Access rules
Who uses what and who manages what, from one place.
Members are managed with roles and fine-grained permissions. Agents and knowledge areas are shared individually; a resource without access doesn't even appear in lists. Tenant isolation is enforced at the database level.
| Access rules | Use agents | Build agents | Manage knowledge | Channels | Analytics | Billing |
|---|---|---|---|---|---|---|
| Employee | ||||||
| Team lead | ||||||
| IT admin | ||||||
| Finance |
SSO and SCIM
Sign-in and user sync paths with your identity provider.
Per-resource permissions
People and groups per agent and knowledge area.
Model permissions
Which model is open to which user.
IP and rate limits
Restrict API access to IP addresses; rate limits per tenant and IP.
Monitoring and audit
Every management step and every action, recorded with before and after.
Management records keep the actor, action, resource, previous and new value, IP, and client, and can be exported. Whether an answer came from sources or general model knowledge, and whether a guardrail stopped it, is recorded separately.
Management recordSample scenario
| Time | Who | What | Change |
|---|---|---|---|
| 09:14 | e.demir | Agent prompt updated | Version 7 β 8 |
| 09:31 | a.kaya | Knowledge area shared | HR Policies β Sales team |
| 10:02 | m.sahin | Action setting changed | Create refund: automatic β approval |
| 10:47 | e.demir | Model permission granted | Reasoning model β Finance |
Disclosure record
Sourced answer, general knowledge, or guardrail block.
Tool execution record
Every action attempt, its result, and its approver.
Turn metrics
Tokens, time to first answer, and latency per turn.
Personal data requests
Export, delete, rectify, and restrict requests are tracked.
Cost analysis
AI spend is a report line, not a surprise.
Usage is metered in AI Credits, visible and exportable by agent, user, and model. Daily token and monthly credit caps are set per user; on eligible models you keep spend on your own provider account with your own key.
- Support agent42,300
- Quote builder27,100
- HR knowledge agent18,400
- Invoice pre-check12,900
Cache savings
Its own panel card
User cap
Daily tokens Β· monthly credits
Usage breakdown
By agent, user, and model.
Budget caps
Daily and monthly limits per user.
BYOK
Your own provider key on eligible models.
Export
Usage data into your own reports via API.
Caching and background jobs
Don't pay twice for repeated context; leave heavy work to the background.
Repeated parts such as agent instructions, knowledge context, and conversation history are read from the model provider's prompt cache. Cached input is billed at a discount, and the credits it saves show separately in the panel. Document processing, indexing, scheduled processes, and heavy model calls run in the background through queues.
- Turn 1
- Turn 2
- Turn 5
- Turn 10
From cacheNew input
Queued document processing
Intake, scanning, extraction, and indexing run as step-tracked jobs.
Scheduled processes
Workflows are scheduled; as scale grows, the dispatcher spreads work in batches.
Retries
Failed jobs are retried; persistent failures move to a separate queue and stay visible.
Human approval and improvement
Critical decisions stay with people; every correction improves the next answer.
Actions put behind approval wait until approved, then resume where they left off. Low-confidence, unsourced, or negatively rated answers go to a review queue; approved corrections are published to the knowledge source.
Catch
Unsourced, low-confidence, or negatively rated answers enter the queue.
Review
The assignee reviews the answer and the content behind it.
Approve
The correction is approved or rejected; deadlines are tracked.
Publish
Approved knowledge joins the source; the agent uses it next time.
Test
Prompt changes are tried against a test set before publishing; prompt versions are kept.
The model doesn't train itself; improvement comes from the knowledge and instruction changes you approve.
Deployment and rollout
From pilot to company-wide, by measuring.
The path we recommend starts with one job and widens by measuring what works at every stage.
- 1
Pilot
One team, one job, a source-limited agent.
Answer accuracy and usage become visible
- 2
Expand
Actions open behind approval; new teams join.
The first jobs handed off
- 3
Govern
Roles, budgets, guardrails, and records are standardized.
One setup for IT and security
- 4
Scale
Processes, channels, and the API company-wide.
AI as part of daily work
Deployment options
The default is multi-tenant cloud. A dedicated environment, your own cloud, or on-premises deployment is scoped in a sales conversation; the hosting region is decided at onboarding.
Ready for security and procurement review
We can walk through these topics with your team during review. What's live and what isn't published yet is written on the open status board of the security page.
See the open status board- Architecture and tenant isolation
- Identity, roles, and access model
- Guardrails and personal data handling
- Model providers and data flow
- Records, retention, and deletion
- Cost model and budget controls
Enterprise questions
What people ask most in an enterprise evaluation.
All FAQs
No. Prompts, uploaded documents, and outputs are not used to train or fine-tune foundation models.
Let's map your enterprise roadmap together.
Pin down the first job, the control points, and the rollout plan in one conversation.