How to Implement AI in Business: 7 Steps to Launch | Tantal

Eight out of ten AI projects at large and mid-sized companies never reach production. Among those working with a specialized contractor, the failure rate drops to three out of ten. That two-to-one gap does not come from different technology — the underlying tech stack is roughly the same across the board.

The gap comes from the implementation process itself. From seven sequential steps that most companies execute carelessly, skip entirely, or hand off to the wrong people. Below is the plan we use at Tantal on every project — from AI assistants in B2B sales to computer vision on the production floor. This is not theory; it is the exact sequence every successful project follows.

Step 1. Process Audit and Readiness Assessment

Every implementation starts not with model selection, but with an inventory. By the end of your discovery session, you should have a list of 5–15 candidate operations, each scored on three dimensions.

Volume. How many times per week the operation is performed. Fewer than 100 repetitions — defer it. AI pays off at scale.

Cost of manual execution. Employee hours × hourly cost, plus losses from errors. A high figure makes an operation a strong pilot candidate.

Input data quality. Free-form text with no consistent structure, unlogged calls, manual spreadsheets — these all signal that you will need 4–8 weeks of data normalization before the pilot can start. Build that into the plan.

By the end of this step you should have a table with five candidates and a preliminary ROI estimate for each. According to RAND 2025, 33.8% of AI projects never reach production — and nearly half of those fail at this exact stage because the audit was treated as a formality.

Step 2. Selecting the First Task Using the ROI Matrix

From the table produced in Step 1, choose one operation. Just one. Not a "comprehensive five-module solution." Every additional module in a first project multiplies timelines and costs by a factor of two to three.

Selection criteria:

• High volume and repeatability
• Structured input data (or straightforward normalization)
• A clear, measurable KPI before automation (average processing time, conversion rate, error count)
• Low cost of failure during the pilot — do not start with an operation where a single AI error costs millions
• A designated internal owner — one person on the client side who is accountable for the outcome

At Tantal, 281 projects have gone through this matrix. The most consistently correct choice for SMB companies is handling inbound requests, document processing, and quality control. The most consistently poor choice is a "creative" content generator with no defined business metric.

Step 3. Data Preparation and Infrastructure Setup

This step is routinely underestimated: data preparation consumes 30–40% of total project time. Mark Watson of RAND identifies infrastructure costs as the primary barrier — data and infrastructure expenses at production scale turn out to be three to five times higher than pilot-phase estimates.

What to do:

• Collect 3–6 months of historical data for the selected operation (calls, tickets, requests, camera footage — whatever the task requires)
• Anonymize personal data in accordance with applicable data protection regulations
• Manually label 100–500 examples for training and validation
• Connect target systems (Bitrix24, amoCRM, 1C, Telegram) via API
• Set up a sandbox environment isolated from production

If the selected operation does not have six months of data in your systems, add a data collection phase to the plan before the pilot begins. There is no shortcut around this step.

Step 4. Pilot — 6 to 12 Weeks, No More

The pilot must have a hard time limit. Best practice is two iterations of four to six weeks each.

First iteration (Weeks 1–6): a working baseline version that handles 60–70% of cases. No polish. The goal is to confirm that the technology can solve the problem at all.

Second iteration (Weeks 7–12): bring coverage to 85–95%. Address edge cases, integrate with production systems in read-only mode, and build monitoring dashboards.

If the pilot has not hit its target metrics by week twelve, that is a signal to reassess either the task or the approach. MIT Sloan 2025 data shows that 95% of GenAI pilots never reach production precisely because of "pilot purgatory" — an open-ended pilot with no defined exit point.

In our AI assistant case study, the pilot took nine weeks. After it concluded, the assistant handled 95% of initial inquiries — that is the production metric, not a pilot hypothesis.

Step 5. Metrics and Baseline Comparison

The most common post-pilot objection is: "things were working fine without AI." The way to address it is to record a baseline before the project starts.

Minimum required metrics:

• Volume metric — number of operations processed per hour or per day
• Quality metric — accuracy, conversion rate, error count
• Financial metric — cost per operation (payroll + systems) before and after
• Time metric — average response or processing time

The comparison should cover identical periods (four weeks before vs. four weeks after) at identical input volumes. If your business has seasonal patterns, use the equivalent period from the prior year as a control.

In a Deloitte survey of 3,235 executives conducted in 2026, 74% of companies expected revenue growth from AI — but only 20% actually achieved it. Every company in that 20% had a documented baseline before the project. Almost none of those in the remaining 80% did.

Step 6. Production Launch and Team Training

On the technical side, this means switching the system from read-only to write mode and connecting it to production. On the organizational side, it means working with people — and that takes just as long as the technical work.

What needs to happen:

• Train the operations team (2–4 hours of hands-on practice, not a slide presentation)
• Provide an honest list of what the AI does and what it does not do — without marketing language
• Run a parallel mode (AI alongside a human) for 2–4 weeks for safe validation
• Assign an escalation point of contact for the first month
• Establish a feedback loop so staff can report AI errors for retraining

Quiet resistance is a real risk. Based on our observations, it happens when a team is told "you are now working with AI" without training or a channel for feedback. The solution is early involvement of the people who will work alongside the system every day.

Step 7. Scaling and the Next Task

One to two months after the production launch, move to scaling. This means either increasing the load on the same operation or applying the same approach to an adjacent process.

Scaling readiness checklist:

• Metrics have been stable for four or more consecutive weeks
• The system handles two to three times the current load with headroom to spare
• Infrastructure costs have been confirmed at production volume (remember the three-to-five times pilot estimate)
• The team independently resolves 90%+ of routine situations without contractor support

Once the checklist is complete, return to the table from Step 1 and select the next operation. In our experience, after three successful implementations a company builds enough internal capability to run some projects independently. Until that point, working with a contractor is the better path: BCG puts specialized vendor success rates at 67%, compared to 33% for internal builds.

What to Do This Week

If you have read this far and are thinking about your first AI project, here are three concrete actions.

First: spend two to three hours building a table of five to ten candidate operations using the criteria from Step 1. Most companies are surprised by how quickly it becomes clear which operation to prioritize.

Second: assess data readiness for each candidate. If none of them has six months of historical data in your systems, revise the plan: start with data collection, then move to AI.

Third: get an external assessment of realistic timelines and ROI before committing to a pilot. AI implementation is specialist work, and 80% of projects fail precisely because the goal was defined incorrectly at the outset.

Tantal is an IT agency and Skolkovo resident with 10 years in the market and 281 full-cycle projects delivered — from AI assistants to industrial computer vision. We have 253 in-house specialists, including a dedicated data engineering team that handles Step 3 — the most underestimated part of any implementation. Request a project audit through the form on our website, or browse our case studies to see how this approach applies to projects at your scale.