We go beyond an assistant that answers questions and build workflows that complete work in multiple steps: classifying the incoming request, pulling data from your systems, checking the rules, opening the record and dropping to human approval where needed. Every step has defined permissions; every step leaves a recorded trail.
Prepared by Pan Innovation House, a software company based in Gaziantep, Türkiye · Last reviewed:
The first wave of enterprise AI was built on question answering: assistants connected to a document pool, replying to questions in text. They are useful, but they have a limit. They answer; they do not do the work. The user still logs into the system, opens the record themselves, requests the approval themselves. The time saved stays at information access; it never reaches the work itself.
The second wave comes with language models able to use tools. The model no longer just produces text; it can call defined functions. That means steps such as classifying a request, pulling data from the relevant systems, checking the rules, opening the record and reporting the result can run within a single flow. The critical point is this: that power is limited to the tools defined and the permissions granted. An agent can do no more than what you have defined.
That is why four things sit at the centre of the flows we build. Tool definitions: which functions the agent may call and what each one does is written down explicitly. Permission boundaries: which operations may run directly, which require human approval, and above which amount no automatic operation happens at all. Rollback: how to reverse a step that went wrong is built in advance. Audit trail: which decision was made with which data and through which steps is retained.
The level of autonomy is a business decision, not a technology decision, and it is set together with the customer. In some flows the agent only prepares and presents everything to a human; in others it completes low-risk operations itself and drops anything above the line to approval. The right start is almost always low autonomy: the system suggests, a human approves, and the boundary widens as trust accumulates. Set-ups that start with high autonomy get switched off entirely at the first serious mistake, and the project loses trust irreversibly.
Processes where an incoming request is classified and routed, data is gathered and checked, and a record is opened at the end. This work is too variable to run on pure rules, and too repetitive to leave entirely to people.
Companies where order, quote, stock and lead-time questions keep landing on the same people. Some of these requests can be answered directly by pulling data from the system; the rest reach a human in far better-prepared form.
Purchasing, foreign trade and finance teams. Where document reading meets rule checking, an agent set-up does more work than document reading alone.
Structures where the team is small and each person carries several jobs. The gain here is not headcount reduction; it is the same team being able to carry a larger volume.
The functions the agent may call are defined one by one: what may be read from which system, what may be written, which operation may never happen. Permissions are separated per user and per flow. The agent cannot call an undefined function; the boundary is enforced technically, not by instruction.
The steps are defined from classifying a request through to opening the record. Each step's input, output and failure behaviour is known. The flow is not a black box; which step it is on can be tracked.
Operations above the defined thresholds drop to approval; the approval screen shows what the agent wants to do and what it bases it on. The user can take the flow over and complete it by hand. A set-up that proceeds without approval puts the whole system in question at the first mistake.
How a wrongly opened record is cancelled and how a half-finished flow is closed is built in advance. Errors are not swallowed silently; they land in a queue and their owner is notified.
For every flow, the data it started with, the tools it called, what it decided and who approved it are retained. When a result is questioned, it can be traced back. Without this trail, automation turns into a box no one can defend.
ERP, CRM, email and document systems are connected as tools. The agent is not a new system but a layer running on top of your existing ones; your systems remain the owners of the data.
We choose together which job the agent will run. A good first flow has these traits: it repeats often, its rules can largely be written down, its mistakes are reversible and its outcome is measurable. Starting with critical, irreversible work is not the right approach.
How the work is done today is mapped step by step; which decision rests on which information is written down. The autonomy level is set here: what runs by itself, what drops to approval, what never happens.
The systems are given read access and, for a period, the agent only produces suggestions; it opens no records. Its suggestions are compared with the work a human actually did. This period is where trust is built by measurement.
Write access is opened first on low-risk operations, with the approval flow active. Results are monitored, error cases are examined and the boundary is widened gradually. Every widening is a separate decision.
How much work the flow completes, how much drops to approval and the error rate are measured. Evaluation tests are set up; when the model or the rules change, behaviour is checked for regressions. Maintenance responsibility is handed over in writing.
An assistant makes information easier to reach; an agent completes the work. An assistant answers a question, while an agent classifies the request, pulls the data from your systems, checks the rules, opens the record and drops it to human approval where required. The difference is that the output is not text but a completed piece of work. The two do not exclude each other; in most organisations the assistant stays and the agent is added alongside it.
We limit that risk in three places: the agent can only call defined tools, operations above the set threshold drop to human approval, and every operation's rollback path is built in advance. On top of that, no write access is granted at the start; for a period the system only produces suggestions. The risk cannot be reduced to zero, but it can be bounded and made traceable.
We decide that together, and it is a business decision. Our recommendation is to start with low autonomy: the system suggests, a human approves. As trust accumulates measurably, the boundary is widened. Set-ups that start with high autonomy get switched off entirely at the first serious mistake, and the way back is hard.
That is an architectural decision, and it is made together with you. It can be built with models running inside the organisation, or, if an external service is to be used, what data goes out, how long it is retained and what the contract says are settled from the start. Masking sensitive fields is also an option. This topic is covered in detail on a separate page.
It can, and that is why we set up evaluation tests. A test set is built where known inputs must produce expected outputs; when the model or a rule is updated, the set is run. An agent set-up without tests changes behaviour silently, and no one notices.
One or more workflows running with defined tools and permission boundaries; an approval screen and the ability to take over; rollback paths and an error queue; an audit trail recording every step; an evaluation test set; and documentation explaining how to extend the flow. The source code we write and your data are handed over to you as defined in the contract; third-party and open-source components remain subject to their own licences.
An AI agent is software in which a language model carries out a multi-step task itself by calling defined tools: it classifies the request, pulls data from systems, checks rules, opens a record and reports the result. What the agent can do is limited by the tools defined for it and the permissions it is given. In the flows we build, tool definitions, permission limits, a way to roll back and an audit trail are written down from the start.
We listen to what you need and tell you whether custom development or an off-the-shelf product fits better. The first 30-minute call is free; the scope of an on-site discovery, and its fee if any, is agreed in writing during that call.
We build a layer that reads incoming invoices, delivery notes, specifications and customs documents field by field and posts them into your ERP. It validates every field it reads against your own records, routes anything it is unsure of for approval, and keeps a record of which value came from which part of which document.
Details →We develop solutions built on Turkish speech recognition and synthesis: call recordings transcribed and made searchable, topic and sentiment analysis, a voice assistant on the order line, and voice-driven data entry for field staff whose hands are full.
Details →For organisations that will not let their data leave the building, we build model infrastructure that runs on your own servers or in your own cloud tenancy: hardware planning, model selection, internal document search, routing that decides which job goes to which model, and cost measurement.
Details →We install workflow engines and internal tool builders that run on your own servers, connect them to your systems and — most importantly — set up their governance: who can build flows, where secrets live, who owns each flow and who fixes it when it breaks.
Details →We build a system that gathers occupational health and safety (OHS) data in one place — from risk assessments and periodic equipment inspections to per-employee training and medical examination validity and near-miss records. When an audit arrives, the requested file is not compiled; it is already there.
Details →We build a system that runs your personal data inventory, retention and disposal schedule, data subject requests and breach response records. The software manages the process and the evidence; legal interpretation and the content of your notices belong to your lawyer.
Details →
