SAIE development roadmap

Creating effective system instructions and workspace settings requires experience that not every process owner possesses

A first encounter with a complex application can be overwhelming, delaying productive use and increasing support requests

Meeting decisions must be copied manually into the task system, resulting in delays and missed commitments

Token consumption by tools running on employee computers remains outside centralized cost reporting and limit management

Without named checkpoints, it is difficult to return to an approved version, compare changes, and identify the currently binding document

Meeting recordings are time-consuming to review, and without separating speakers it is difficult to assign decisions and commitments

Multilingual teams spend time on manual translation and struggle to maintain consistent industry terminology

Long and complex documents require navigation through structure and dependencies that simple text chunking does not preserve

Creating cross-project summaries from multiple sources requires manual data consolidation and format standardization

Calculations performed solely by a language model can be inaccurate and difficult to reproduce, especially in reports subject to audit

Before launching a solution, it is difficult to forecast model costs and identify a configuration that delivers the required quality within budget

Executives need a concise view of AI costs, risks, quality, and usage without analyzing operational reports

Disconnected environments need a controlled local catalog of model packages, versions, and compliance information

In large organizations, documents vary in importance, freshness, and scope, so AI responses must select the right sources based on the user's role and the context of the matter

Traditional context search loses the relationships between documents, concepts, and agent outputs, resulting in fragmented answers that are difficult to justify

Building integrations and interface components usually requires development work, which delays the rollout of new capabilities for business users

Shared infrastructure should, when needed, allow individual components to be separated and used by other teams or in other situations — including emergencies — and easily reintegrated afterwards

Agents and applications built outside the core product need secure, consistent access to search across corporate knowledge

High volumes of scanned documents create queues where urgent files compete with lower-priority jobs and processing status is difficult to track

Without recognizing the context of an individual session, it is difficult to apply the right limits, policies, and behavioral analysis to a specific conversation

In a multi-node environment, requests must reach available resources while taking workload, location, and model requirements into account

The organization needs to enforce regulatory requirements automatically during model use instead of relying solely on manual compliance procedures

Different teams and use cases require consistent rules for access, permitted models, data, and response generation

Distributed AI infrastructure requires a single place to control nodes, capacity, availability, and workload allocation

Manually collecting configuration information creates gaps, inconsistencies, and lengthy coordination at the start of an implementation

Multiple project configurations are difficult to compare, find, and reuse without an organized registry

Changes to standards and procedures require parallel maintenance of different methodology releases and clear assignment of each release to the relevant projects

Implementation knowledge is scattered across files and conversations, leaving the assistant without complete context for a given project

Project initiation is delayed by the manual collection of basic materials needed by the team and the assistant

Experts need a simple way to add missing information during ongoing work without involving a system administrator

New decisions and artifacts created during a project should enter the shared context without recurring manual imports

Copying information into forms consumes consultants' time and increases the risk of omissions or inconsistent answers

Completing extensive tables from multiple source materials is repetitive, time-consuming, and prone to errors

Project managers need a current, data-driven view of execution rather than declarations collected only before reporting

Static screens do not suit different roles and project stages, causing users to waste time looking for the right information

New users need contextual guidance through the process so they can become productive and independent more quickly

Memory, documents, activities, and agent outputs remain in separate silos, limiting the understanding of relationships and the reuse of knowledge

Text alone is not sufficient for convenient handling of forms, tables, and interactive results generated by an assistant within a conversation

Building multi-step agent processes requires specialized knowledge of step sequencing, tools, and context transfer

Ready-made agents, automations, interface components, and knowledge packages are difficult to discover and deploy again across teams

Manual agent validation covers too few scenarios and fails to detect regressions before release to users

Public institutions need to use national services and registries without building every integration from scratch

Autonomous operations must be reproducible, auditable, and reversible when a sequence of actions produces an unwanted outcome

Training content, assistant conversations, and progress measurement operate separately, making it difficult to personalize learning and identify competency gaps

Long conversations exceed model limits and increase costs even though only part of the earlier context is needed for the next response

Administrators lack a single view of which agent instances are running, what profiles they use, and whether they behave as expected

Raw logs are scattered and lack business context, making event analysis and the preparation of audit evidence unnecessarily slow

Without unified oversight, it is difficult to assess source quality, retrieval effectiveness, latency, and the causes of weak knowledge-grounded responses

Unusual user behavior and sudden changes in AI usage patterns may remain unnoticed until damage occurs

Requirements from standards and European regulations are fragmented, while teams need a consistent set of controls and audit-ready evidence

Before deploying an AI use case, its impact on people, data, processes, and the organization must be assessed consistently, with decision rationale retained

Inconsistent, incomplete, or outdated data reduces model reliability, yet data quality is often neither measured nor assigned to accountable owners

Auditors and process owners need a tamper-resistant history of decisions, configuration changes, and model usage

An ARS assessment requires extensive evidence gathering and interpretation of criteria, which prolongs the team's work without expert support

Completing the F0-AC-01 control requires consistent links between answers, documents, responsibilities, and execution status

The rationale for applying or excluding safeguards must be complete, current, and linked to evidence in order to pass a compliance review

Assessing the impact and risk of an AI system requires structured guidance through criteria and related artifacts

Managing an AI system according to the adopted methodology requires continuous access to the correct procedures, records, and implementation guidance

Traditional technical indicators do not show whether an assistant actually saves time, improves decisions, and delivers responses of the expected value

Preparing F6 reports from multiple projects and sources consumes time and makes it difficult to maintain a consistent data scope

Compliance deadlines can be missed when they depend on multiple dates, owners, and stages of a case

Users need timely reminders about planned tasks to reduce delays and the manual tracking of commitments

Access to projects and materials must reflect each user's role and responsibilities without granting excessive permissions

Rigid document fields restrict filtering, access rules, and search based on organization-specific attributes

Tools created within the product should also be available to external automation and development environments through a controlled interface

The same assistant should respond differently according to a user's permissions, responsibilities, and type without maintaining multiple separate configurations

Before an extended absence, knowledge about open matters and decision context must be transferred securely to a substitute

Tasks originate in many conversations and processes, so without a shared registry it is difficult to determine ownership, deadlines, and completion status

Automations maintained outside the product need a simple way to invoke agents, pass data, and receive results

Organizations want to make an assistant available on existing websites without rebuilding the site or creating a separate interface

Model and infrastructure consumption must be controlled by user, team, or project to prevent overload and unplanned costs

Diagnosing workstation failures requires manual review of many technical files and specialist administrator knowledge

Users waste time manually locating local files and attaching them to conversations, especially when they do not remember the exact location

A low cache hit rate unnecessarily increases cost and latency, while the cause of the decline often goes unnoticed

Unlimited or malicious requests can rapidly exhaust an organization's budget before an administrator notices the increase in usage

Repeated questions continue to consume expensive model capacity when semantic similarity is not used effectively to reuse responses

Users and auditors must understand why a request was blocked or modified instead of receiving only a refusal message

Model responses often require data masking, format validation, filtering, or enrichment before they can be passed to an application

Risks related to AI systems are scattered across documents and spreadsheets, making it difficult to track owners, actions, and exposure levels

Models move through assessment, testing, publication, and retirement, while the absence of a consistent version history makes change management unsafe

High-impact decisions require clearly defined human control points, accountability, and the ability to stop automation

Compliance evidence quickly becomes outdated and is stored in many places, increasing the cost of preparing for an inspection

An organization cannot manage risk effectively if it does not know all deployed systems, owners, models, and intended uses

A new model should be released only after the appropriate people document and approve its quality, security, cost, and compliance assessment

Declines in response accuracy, completeness, or safety may remain invisible without continuous measurement against the expected level

Users need consistent sign-in and permission recognition across components without creating multiple accounts

Protective controls and security events are difficult to oversee when configuration and safeguard status are distributed across systems

Implementation information must be recorded in a tamper-resistant way so that the process and accountability can be demonstrated later

Records with high evidentiary value require protection against deletion or modification throughout the required retention period

Field employees and executives need access to the company assistant from a phone without having to start a workstation

A lack of connectivity should not block access to previously shared knowledge and essential assistant capabilities on a mobile device

Commitments regarding assistant quality and effectiveness should be measured and accounted for in one place using consistent definitions

Messages from different processes get lost across multiple channels, so users need one view of events requiring attention

Technical support lacks current information about software installed on a user's computer, extending diagnosis and ticket handling time

Preparing documents, presentations, and spreadsheets requires switching applications and manually transferring content from a conversation

Local tools are difficult to discover and control without an inventory of their configuration, permissions, and availability

When working with long documents, users need to modify a selected section without risking unintended changes to the remaining content

Administrators need a single catalog of available models, versions, parameters, owners, and approved uses

AI-related events require coordinated registration, impact assessment, accountability, and corrective action

Creating and updating required control materials is labor-intensive when evidence and templates remain scattered

After services go live, sources of excessive spending must be identified continuously and changes recommended without reducing response quality

Employees need a simple and confidential channel for raising concerns about AI behavior, together with visibility into subsequent actions

Reports submitted by email should be recognized, organized, and assigned automatically without manual re-entry

Cases arriving through different channels require a shared record of status, communication, evidence, and corrective actions

Each report must quickly reach the correct category and accountable people to avoid inactivity or duplicated effort

Critical cases require measurable response deadlines and automatic escalation when thresholds are exceeded

The history of sensitive case handling must be resistant to modification in order to preserve the evidentiary value of decisions and actions

Required information about concerns and actions taken must be complete and available in a structure compliant with the F6-AFC-01 form

Information from reports should feed management and impact-assessment processes without manually copying data between registers

Planning, execution, and monitoring require a shared model of concepts and relationships so that artifacts are not interpreted differently by each module