AI-discovery
Ten business days to answer three questions that stall AI projects: which scenario will deliver measurable impact, whether the data is ready for it, and what the security team will say. Ends with a decision—including an honest “we’re not taking this scenario.”
Login Format
This is a short phase prior to the implementation decision: fixed scope, ten working days, with no obligation to launch a large project afterwards. It includes a separate contract and delivers a distinct outcome—the company receives materials it can use for procurement or hand over to another contractor.
We deliberately separate this phase from the pilot. The pilot is more expensive and answers the question “does it work?”, but doesn't protect against the main risk—spending a quarter on a scenario doomed from the start due to data quality issues or security restrictions. Discovery addresses this in ten business days and ends with a decision, including an honest “we won’t pursue this scenario.”
- 10 working days — fixed first stage.
- The result is applicable immediately —for procurement, internal decision-making, and the roadmap, regardless of whether we continue working together or not.
- Can be done remotely, if interviews and system descriptions are available.
When discovery is needed—and when it isn't
Required if:
- There are many ideas, no priorities, and each department pulls in its own direction;
- The pilot has already shown results on the export, but it's unclear what will happen with real data;
- Information security and legal teams are blocking the initiative, and there has been no substantive discussion about risks;
- a justified budget and roadmap are needed for the board approval;
- The company already uses public AI services, and it is necessary to understand the scale of shadow usage.
Not needed if: the scenario is one and obvious, the data for it has already been collected, and the information security mode has been agreed upon. In this case, proceed immediately to RAG pilot — There's no need to pay for confirmation of something you already know. If the main concern is the safety of an already operational environment, "tochnее popadёт" Secure AI audit.
What we do in ten working days
01. Interview. CIO, CISO, process owners, data owners, IT, and business stakeholders. The goal is not to collect wishes, but to build a picture: where time is currently being lost, which decisions are made based on incomplete data, and what has already been attempted.
02. AI Capabilities Map. A complete list of scenarios without filtering, with a rough estimate of the effect. Usually 15–30 items are collected; the value lies in the fact that they are subsequently compared against the same criteria, rather than by how loudly the proposer advocates for them.
03. Data verification. For candidate scenarios: what sources are required, is there programmatic access to them, who owns them, what is the state of reference data, and whether they contain personal data or trade secrets. This usually eliminates half the list right away.
04. information security restrictions. Talk to the CISO before choosing a scenario, not after: what cannot be exposed externally, which perimeter is acceptable, whether you need your own boundary, and what specific regulatory requirements apply to your systems.
05. Shadow AI. What employees are already using and what data is being sent there. It's almost always discovered; it's more useful to provide a legal alternative than to ban it.
06. Pilot architecture and plan. For one or two selected scenarios: diagram, team composition, timeline, budget, success criteria, and stop criteria.
By what criteria do we select a script?
Priority is determined not by "importance" but by four measurable criteria. This way, the decision can be defended before the board and independently replicated in the next cycle.
| Criterion | What we check | Bad sign |
|---|---|---|
| Value | How much time or money does the script save per month, and who will confirm this with a number? | The effect is formulated only qualitatively. |
| Data readiness | Is there software access, an owner, acceptable quality, and permissible lag | The only way to access the data is by request |
| Access mode | Is it possible to work with this data within the selected perimeter without violations? | The answer comes down to “you need to ask lawyers,” and there’s no one to ask. |
| Reversibility | What happens if the model gives an incorrect response, and can the consequences be undone? | The error is immediately sent to the client or to reporting |
The first step in the workflow is almost always a highly reversible scenario: where a person checks the result before applying it. This delivers quick results and avoids reputational risk at the outset.
How to later validate the value of the scenario in the pilot — Methodology for measuring the benefits of AI for employees.
Volume limits
The package is designed for one business unit. The boundaries are fixed in the contract and define what is included in the fixed price.
- one business block, up to three processes, up to eight interviews;
- The assessment includes up to 20 AI scenarios; how many of them turn out to be substantive depends on the identified tasks—the package does not include an obligation to “come up with 20 useful ideas”;
- for two scenarios, we are verifying data readiness and security constraints; for one, we are preparing a pilot concept;
- up to five source systems and up to 15 documents with a total volume of up to 300 pages;
- if necessary, we analyze up to two anonymized samples totaling up to 100 MB—without full database profiling.
What is needed from the client
- Five to eight interviews of 40–60 minutes each: CIO, CISO, process and data owners.
- Description of key systems and integrations—diagrams, if available; if not, we'll discuss during the interview.
- Access to the data structure description (not to the data themselves) for candidate scenarios.
- One responsible person from the client side who can gather people and make the final decision.
To assess the economics, the customer confirms baseline metrics for process volume, labor input, and cost, or agrees to use clearly stated assumptions. Ready-made data is not required—discovery precisely reveals what is ready and where a preparatory phase is needed. Why this is the key issue for the entire project is explained in the article. “Why an AI project starts with data”.
What is not included
- development and demonstration of a working AI solution;
- data cleansing and technical connection of sources;
- legal opinion;
- pentest.
Shadow AI is assessed based on interviews and provided information—without deploying agents, analyzing all network events, or guaranteeing full detection.
When the work is completed
- the reviewed scenarios have been compared using an agreed methodology;
- data, constraints, and risks are described for two priorities;
- According to the base scenario, a decision has been made: launch the pilot, defer until conditions are met, or reject.
- Success and stop criteria have been prepared, and the results have been defended.
A negative decision is a valid discovery outcome. A justified "not launching now" decision saves a pilot quarter and is compensated the same as a positive one.
Output documents
| Artifact | What's inside | Who needs it |
|---|---|---|
| AI Capabilities Map | All collected scenarios with ratings across four criteria and justification of priority | CIO, business customer, board of directors |
| Data Readiness Assessment | Sources, owners, quality, gaps, and what needs to be closed before the pilot | To the data owner, architect |
| List of IS security constraints | What is prohibited, what is allowed with reservations, and what measures are required for the selected environment | CISO |
| Pilot architecture | Solution architecture, integration, deployment, access control | To the architect and implementation team |
| Roadmap and budget | Stages, timelines, team composition, budget, success criteria, and stop criteria | For procurement and budget protection |
The listed items are supplemented with a survey passport specifying the fixed perimeter and a management presentation to support the decision. The documents are provided in editable format: they will be further used within the company, including without our involvement.
Package budget
Package budget: 600K – 1.2M RUB excluding VAT.
The exact fixed price is determined after the initial meeting and scope verification; included work, deliverables, and exclusions are confirmed prior to commencement; any scope expansion is possible only upon agreement.
Team composition
- AI Architect — scenarios, approach selection, pilot architecture.
- Data Architect — sources, quality, delineation, ruptures.
- Information Security Architect — processing mode, environment, regulatory requirements, shadow AI.
- Business Analyst and Project Manager — interview, effect evaluation, timelines, solution support.
The work occurs at the intersection of four practices, and this is not a formality: a scenario selected without cybersecurity is rejected during approval, while one selected without data is rejected during the pilot phase.
Frequently asked questions
What if none of the scripts passes the selection? This is a valid and useful outcome: you receive justification for why now isn't the right time, along with a list of what needs to be addressed in data or processes. It's cheaper to learn this in ten business days than during a quarter-long pilot.
We've already tried AI, and it didn't work. Is it worth it? Usually, yes—the unsuccessful experience typically ran into issues with data or security compliance, not the model. Discovery starts precisely with these two checks.
Is it mandatory to continue with you after discovery? No. The artifacts are self-contained and can be handed over to another contractor or implemented independently.
Will the data leave our perimeter? At this stage, we are working with the data structure description, not the data itself. If the scenario requires its own perimeter, this will be reflected in the pilot architecture—the principles are described in the material. about safe corporate AI.
What's next? Usually RAG pilot according to the selected scenario, the production environment on Restart AI Enterprise Platform either the preparatory stage based on data — Data, BI, DWH.
How we fix the volume and deadline
The terms below are identical for all packages and are incorporated into the contract without changes.
- Perimeter — prior to the agreement. We hold an initial meeting of up to 60 minutes and agree on the perimeter passport: systems, processes, sources, participants, and outcomes.
- Countdown of the term. The timeline starts after the handover of the approved set of source materials, provision of access credentials, and confirmation of the interview schedule. Delays on the client's side will postpone the timeline by a corresponding amount of time.
- Data. We operate within a coordinated secure channel. Production data is not transmitted to external AI services without the customer's written permission.
- Result transmission. We provide materials in PDF and editable formats (DOCX, XLSX, PPTX, or agreed equivalents) and conduct one results defense session of up to 90 minutes.
- Acceptance. Five business days are allocated for reviewing the results. One consolidated round of clarifications within the agreed scope is included; however, correction of errors and non-compliances with the contract is not limited to this round.
- Volume change. New systems, sources, and tasks are included only upon agreed changes to scope and cost.
- Work format. The default format is remote, unless an in-person component is explicitly planned for the package.
The statuses and licenses on the basis of which the works are performed are collected on the page "Licenses and statuses".
AI-discovery - the first stage implementation of AI in business: Followed by Pilot, Industrial Loop and Operation.
Email us
Write to us at info@restart.re or click the button below. For the first letter, it is sufficient to briefly describe the outline, the role of the organization, current systems, limitations and the desired outcome.
Email usRequest format
We agree on the volume, timing and output artifact - and launch it in a short time.
