OSNO-VA: AI accountant
- Built an AI platform
Learn critical questions to ask AI implementation contractors about data security, response quality, contract terms, and how to structure a successful pilot.
In custom development, results are deterministic: an agreed specification describes system behavior, acceptance verifies the implementation. An AI agent responds probabilistically—the same request can yield different answers. Quality is maintained by a control loop: acceptance scenarios, accuracy metrics, audit log, and confirmation routes for risky actions. You need to ask the contractor about this loop: without metrics and acceptance scenarios, the word 'works' has no definition.
The second difference is data. An agent performs as well as its context is prepared: master data, reference guides, regulations, knowledge base. A significant portion of an implementation project is cleaning up this data, and before signing, it must be clear who owns this work: duplicates, empty fields, and conflicting reference guides will be inherited by the agent if not cleaned.
The third difference is operations. After deployment, traditional systems rarely change; an agent lives with the process—regulations and reference data change, models update, rules become outdated. So a contract with an AI contractor is primarily an operations contract: who manages the rules, how quality is measured after changes, what's included in support. How these terms sum to total cost is discussed in AI agent pricing and TCO article.
This guide is structured as: three groups of contractor questions—data, security, quality—a red-flag table, contract framework, and a healthy selection process. These questions apply to contractors of any size, including us
Phrasings from the left column appear verbatim in commercial proposals. Each has a verifiable answer you can rightfully demand: it translates the promise into calculation, boundaries, and architecture.
| Red flag | A healthy contractor response |
|---|---|
| «Guaranteed 30% efficiency gains» | Baseline measurement of your process, target metric, and effect calculation on your data; percentage before measurement is a marketing figure |
| «We'll implement any process in a week» | A pilot on one process with fixed boundaries, deadline and price; scaling comes after pilot metrics |
| «Move data to the cloud, it's all secure» | Specific loop architecture: which data, into which model, through which anonymization gateway, what gets logged |
| «Price is fixed, we quote immediately» | TCO breakdown: development and integration, tokens, security gateway, support; price fixed after process analysis |
| «We'll discuss metrics after launch» | Acceptance scenarios and target metrics are agreed before start—without them, the project lacks a completion criterion |
TCO components and why honest price fixing is possible only after process analysis—in AI agent cost article. What a control circuit with a gateway, narrow tools, and audit log looks like in practice—in our breakdown AI agent for 1C.
This section is a negotiation framework; your lawyer prepares legal language. First: data—a personal data processing agreement, list of categories, data transfer circuit, and the contractor's obligation to work through an agreed-upon security architecture.
Second is intellectual property. Prompts, rules, tool configuration, and accumulated fine-tuning—these are part of your process; the contract must establish that they belong to you and are transferred in readable form. Without this clause, switching contractors means rebuilding the control loop from scratch.
Third: acceptance criteria—acceptance scenarios, target metrics and data they're measured against—are transferred to the contract from the pilot. Fourth: exit strategy—what you receive upon termination: prompts, logs, documentation, tools—in what format and timeframe, and how support is transitioned.
The contractor analyzes your process, data, and constraints. A good sign is more questions than promises, and 'no agent needed here' is among acceptable answers
One process, fixed boundaries, deadline and price. The deliverable is a working result on your data that can be measured
Quality and savings measurement using pilot acceptance scenarios. Scaling decisions are based on measurement numbers.
Scale to adjacent processes. The contract establishes operations: support SLA, system ownership, exit strategy.
FAQ
Company size predicts little on its own—specific project team composition matters. Team-size research (QSM) shows 3–7 people optimal for mid-scale systems; an AI agent is a typical such project. Compare the composition of the team building your loop, direct access to its engineers, and answers to this guide's questions: the checklist applies to contractors of any size
Request a walkthrough of a deployed agent's architecture: data sources for context, how the security gateway works, what gets logged, how quality is measured. Example breakdown— AI agent for 1C. The contractor's public materials—articles, architecture breakdowns, calculators—are verifiable before meeting. A quick way to see the team in action— workshop or pilot sprint.
What agents operate within the contractor's company, which processes they cover, how their quality is measured, and what had to be disabled from deployed implementations. A contractor who daily operates their own agents knows operational issues before they hit your systems. The answer 'we deploy for clients, but haven't used them internally yet' is a red flag.
A free pilot usually means a templated demo on prepared data: a contractor can't afford real work with your systems and data for free. A paid pilot with fixed price, timeline, and metric-based acceptance disciplines both parties and provides honest material to decide on scale. The format for this approach— agent sprint.
Bring them to a common baseline: total cost of ownership over two to three years—development, tokens, gateway, support (calculation methodology); identical acceptance metrics; what stays with you when parting with the contractor. Cost breakdown for your configuration is calculated by AI automation calculator.
Testing by doing
Agent Sprint—an AI solution for one real business challenge within a fixed timeframe: a report, data export, regular summary, or analytics. You get a working result and metrics to decide on scale. We consider the questions in this guide sound and take them ourselves—ask us at the brief.
Verification date: July 24, 2026