Work we can talk about

Employees were using public AI tools with company data, out of sight. Those tools knew nothing about the company. Nobody knew how much AI was used, what worked, or kept the know-how when a person left. Solutions that did work did not spread beyond one team.

• A model gateway — one entry point to several providers: proprietary, open-source and regional models, cloud and on-premise, switchable per use case.
• A web chat for every employee — dictation, speech output, attached documents; deep research over internet sources where policy allows.
• A builder for assistants and agents — any employee can create an assistant with its own knowledge base. First ones: incident investigation, goal setting.
• API keys for internal systems and for AI extensions in developers' IDEs, with usage accounted per key, consumer and department.
• Analytics — product and technical metrics of the platform; cost per consumer.
• Sensitive data detection on the way out: access keys and personal data.

One observed environment instead of shadow AI. Cost per department. Knowledge and assistants that stay in the company.

Several thousand internal regulations describe norms, processes and procedures. Finding the right clause in the right document is slow, so people ask the expert instead — and the expert stops doing their own work.

• A retrieval-based assistant over the regulations, built on the corporate AI platform.
• Embedded in the portal's search, where people already look.
• Documents and the search index synchronised automatically.

Answers with the source document, in the place people already search. Experts answer fewer repeat questions.

Give the largest number of employees the cheapest, safest and simplest way to work with AI — without data leaking to outside services.

• A plain chat agent in the corporate messenger with access to language models through the gateway.
• Cloud models connected where policy allows; local where it does not.

AI for everyone at the cost of an agent. No company data in outside services.

Hundreds of meetings a day went unrecorded. Writing up the course and the outcomes took the participants' time, and people turned to outside services that leaked data.

• An agent in the corporate messenger: send a link to the recording, get the result in the agreed format.
• Integrated with the video conferencing service and with cloud models: transcription, speaker separation, summary.

Meetings documented without anyone taking notes. Decisions and actions in one place.

No time or hands to process the volume of information sitting in documents.

• An agent in the corporate messenger: upload a document in the chat, get a short summary, ask follow-up questions.

Long documents read in minutes; the questions that matter asked right there.

Some departments have their own documentation that matters only to them. Onboarding a newcomer takes colleagues' time; reading everything is not always justified.

• An agent with knowledge bases for group chats, consulting on a specific subject area or a department's context.
• Connected to the group chat; answers colleagues when addressed.

Newcomers get answers from the department's own documentation without pulling colleagues away.

Support requests arrived as messages to whoever was online. Routine requests — access, accounts, a DNS name — waited for a person, and nobody knew which team a request belonged to.

• An agent in the corporate messenger. Any employee describes the problem in plain words.
• A catalogue of actions it can perform itself — grant an access, create a DNS record, provision an account or an environment — executed on the spot and logged.
• When it cannot resolve the request, it creates a ticket in the service desk, correctly classified and attributed to the right support team, and notifies that team.
• Runs in our own company.

Routine resolved in minutes without a person. Every other request becomes a ticket with the right team on it. Nothing lost in chats.

The CRM is built for a desk. A salesperson on the road needs the state of a deal now, often on a poor connection.

• A sales agent in the corporate messenger integrated with the CRM.
• Deal and account data on the go; interactions created and summarised from the chat.
• Functionality keeps growing.

The state of every deal in one message, wherever the salesperson is.

A manager runs many deals at once and has to know where each customer stands. A newcomer inherits hundreds of relationships. Every call is another hour of typing.

• A relationship summary service: calls transcribed through the telephony provider, speakers separated, summarised.
• Integrated with the CRM: a summary of the last interactions shown inside the customer card.

Onboarding on a customer base of hundreds without reading it all. Calls become notes without typing.

Dashboards multiplied until nobody could find the right one. Adding an indicator took months. Management wanted a simple answer to one question, now.

• A web chat over the data warehouse: the assistant knows the business context and the database structure, writes complex SQL itself, and continues the investigation as a dialogue.
• Integrated with the DWH; models on a local LLM server.

Ad-hoc questions answered in minutes. Dashboards for monitoring, the assistant for questions.

CVs arrive as DOCX, PDF and scans in every layout; standard parsers fail.

• An AI service: any file in, an approved data structure out.
• Cloud LLM providers connected; API for the recruiting systems.

Every CV becomes structured data the recruiting systems can use.

A popular vacancy gets hundreds of responses a day, from people and from bots; many are unrelated to the role. Processing them is heavy and unavoidable.

• An AI service for automatic processing of responses: a coarse fit check first, an extended check for the rest.

Hundreds of responses a day handled; recruiters see only the relevant ones.

First screening for mass positions is routine work that eats the recruiter's week.

• An AI recruiting system: first screening of candidates through agents in messengers.
• Reports on the results.

Recruiters on the roles that need a recruiter.

Internal surveys ask open questions. The free-text answers are the most valuable part for HR and the hardest to process.

• A service for automated processing of survey results: upload free-text answers, set the clustering criteria, get themes.

Open answers become themes and quotes instead of a spreadsheet nobody reads.

Open answers bring more value if you can ask a follow-up question — which a form cannot do, and a sensitive topic risks the respondent walking away.

• A system of employee surveys in the corporate messenger with AI agents that go deeper on open questions.
• AI reports on the results.

Follow-ups without an interviewer; the "why" behind every answer.

There are moments when text is about to leave the company perimeter — a prompt, a message, an upload — and nobody checks what is in it.

• An AI service that detects access keys and personal data in text.
• Cloud LLM providers connected; API for other systems; used by the AI platform.

A check at the exit instead of a policy nobody reads.

Corporate mailboxes receive a large flow of letters processed by hand, because the internal recipient can be determined only from the content.

• An AI service for automatic processing of inbound e-mail.
• Recipient rules configurable by developers from the content of the letter.
• Integrated with corporate mail and the document management system.

Letters reach the right desk without a person reading them first.

Several thousand internal regulations; links between them exist only in their content; nobody knows which documents to update when a process or the org structure changes, so the documentation is never current.

• A service that indexes a document store and builds the graph of links.
• AI for links, tagging, summaries and vectors; retrieval-based search on top.

When a process changes, the list of documents to update is known before anyone asks.

Market analytics has to be collected regularly from heterogeneous sources; the work is hard to automate and needs domain understanding to interpret.

• An AI service: collection from company reports and news portals, AI processing, a report in the agreed format.

A regular market report without a week of manual collection.

• Promotions planned in Excel, with data collected by hand from sales and trade marketing
• No cross-functional process and no common method to measure promo effectiveness
• No control over promo closing and allocation of actual costs
• A growing share of promo volume was eroding forecast quality every year

• A TPM system for the full promo cycle: plan, approval, execution, closing
• A new KPI model and a P&L for every promotion
• Integrations with adjacent systems and a data model in the client's data platform
• Moved onto our TPM product: distributors, complex approvals, real-time recalculation

• Routine person-hours taken out of the promo cycle

• A legacy system built by the parent company abroad, with no source code
• Calculations took four hours or more; part of the data was calculated by hand
• Incomplete integration with SAP and no contract balancing

• A new application on our platform fully replaced the old one
• Master contracts, contracts and terms at any level of customer and product hierarchies
• Terms calculated on primary and secondary sales from BI
• Accruals and settlements posted to SAP, reversals, balancing, year-end closing

• The legacy functionality fully replaced
• Embedded in the existing IT landscape and SAP
• The old system was supported until the new one went live
• In production and under our support

• Finance struggled to manage trade investments, especially at period close
• Data was collected from many Excel files in different formats
• Preparing data for financial systems took a lot of time

• An Anaplan model with automated data loads and manual adjustment
• Scenarios, allocations and splits, metric calculation including RGM
• Input from the sales team and integration into S&OP
• Financial documents and export to financial systems

• Data collection and preparation for finance automated
• Trade investments built into the S&OP process
• The project became the basis of our trade terms product

• Distributors without EDI; orders sent by e-mail were distorted and lost
• Orders keyed into the ERP by hand, including orders for unavailable items
• No way to manage the order flow or set limits
• No link between logistics and finance

• The full order-to-cash cycle in the indirect channel
• Cart, delivery requests, recommended order
• Cost and availability of shipment from a specific warehouse, statuses, receiving
• Settlements, consignment, quotas and permissions

• Several business units and their integrations on one platform

• Allocations at distributors led to out-of-stocks
• Distributors planned demand poorly and bought unevenly
• Stock with the right shelf life had to be sold quickly
• Complex sourcing from several warehouses

• A demand model on sales history, with the model chosen by the pattern of each series
• Recommended volume that accounts for stock, transit and open orders
• Confirmation with posting to the ERP, or automatic ordering in VMI mode
• Sell-through forecast by shelf life and an automatic discount proposal

• Multi-model demand forecasting
• Manual mode and VMI mode
• A storefront for discounted stock, with reservation
• Works together with the distributor order system

• Regulation does not allow setting tasks directly for the distributor's staff
• Distributors learned late about launches, promotions and process changes
• Performance assessment was not transparent
• Slow document signing and communication failures

• Tasks with completion control, publications with acknowledgement
• Distributor hierarchies, contacts, change requests
• Capability building, shelf-life promotions, reservations
• Reporting: KPIs, relationship balance, sell-in and sell-out

• Fewer errors and adjustments in settlements

• Merchandising ROI below target
• Manual paperwork with agencies and opaque KPI-based payments
• An outdated outlet base and frequent changes to tasks

• Outlet program, visit targets and a coverage model
• KPI plan versus actual, budgets and their approval
• Service orders, e-documents, payment statements
• Outlet base validation and integration with SFA

• Higher merchandising ROI
• Less effort on documents and target setting

• No single space for everyone in the supply chain
• Contract obligations and transport quality had to be controlled
• High costs of transport contracts and returnable packaging
• Complex integration with different ERPs

• One environment for carriers, warehouses and customers
• Optimisation of contract costs and trip allocation, auctions
• Dock slots, a live queue, returnable packaging, electronic waybills
• Two-way integration with SAP and SAP EWM

• Carriers, delivery points and warehouses in one system
• No downtime longer than four hours

• ERP transactions did not match the system's requirements for content and detail
• Recording thousands of operations a day by hand was impossible
• Several different systems served the supply chain

• An integration hub between ERP, MES, TMS, WMS and the government API
• Existing messages intercepted, so business processes did not change
• Traceability down to material, batch and weight
• Discrepancy control, workplaces for staff and veterinarians

• A small team instead of the large one first estimated, with high service quality
• Not a single shipment stopped because of the integration

• A high volume of shipments every day at every plant
• Paper sheets and manual re-entry of data into the system
• Slow paperwork with errors in documents

• An app for handheld terminals, integrated with the TMS
• Vehicle check-in and dock assignment, vehicle quality inspection
• Start and end of loading recorded, unloading supported
• Documents issued automatically

• No single point of contact between employees and the employer
• Services of different functions scattered across systems
• Many manual operations and a low level of service

• One window for employee services, like a public services portal
• HR, transport, office, business travel, IT and legal services
• Integrations with HR systems, including Workday and SAP HR
• A messenger agent for part of the services

• Voted best IT project by employees
• Cheaper outsourced HR administration and a move to self-service

• Employee data of different legal entities lived in different systems and was out of date
• Several versions of the organisation structure
• Data passed to internal consumers by hand

• One master source of employee data across legal entities
• A flexible organisation structure and virtual teams
• Enrichment with data from adjacent systems
• Role-based access to personal data, an integration portal for the company's systems

• Lower risk of personal data leaks, data governance in place

• The global SAP HR had no attendance and absence tracking
• Errors in timesheets and no integration with access control
• A move to flexible working with no mechanism for substitutions

• A portal: employee profile, attendance, absences, leave
• Timesheet, working modes and schedules, substitutions
• Integration with HR systems and access control

• Time released in the HR department
• Automatic tracking instead of manual operations
• The accumulated data enabled a move to a full time-management system
• Replaced later by the employee service portal

• KPIs and analytics calculated by hand, some impossible to calculate
• No end-to-end work on an order from lead to payment
• Lots of routine, with no AI in the process
• No integration with e-mail, messengers and internal systems

• The full hiring cycle and a flexible recruitment funnel
• Publishing to job boards, ranking of responses
• AI: vacancy texts, candidate matching from the database, de-duplication
• Consent to personal data processing, KPIs, dashboards, NPS

• Feedback from the real consumers — cats and dogs — is hard to obtain and interpret
• Many raw material suppliers, each needing a full cycle of tests
• Strict corporate rules for running tests

• Three hardware and software systems of our own design
• Smart feeding stations with video monitoring
• A portable station for observation at home
• Test management software and a mobile app

• Losses of several million per testing cycle avoided
• No dependence on equipment vendors

• Product samples are sent to households across the panel
• Logistics, data collection and surveys were run by hand through an outsourcer
• Research was expensive and the data inaccurate

• A smart feeder with precise scales and a 4G module
• A test management system
• A courier app: logistics, calibration, data readout
• An app for pet owners with surveys

• In production; the fleet has since been expanded

• Instrument readings were entered into the SPC system by hand
• Reports were assembled in Excel
• Data arrived late, incomplete and with errors
• Users stopped trusting SPC metrics

• Direct reading from instruments by reverse-engineering their protocols
• Data from LIMS, SAP, MES and process control systems in one model
• Written directly into the SPC system's database
• Reports that surface deviations

• Low forecast quality and scattered master data
• No detailed plan by customer, warehouse and unit of measure
• No scenario planning; new factors could not be added quickly
• Lots of manual data cleansing

• An Anaplan application for the whole demand planning cycle
• A short-term forecast with ML and a long-term top-down forecast
• Sell-in and sell-out, seasonality, promotions, launches, prices
• Integration with the ERP, TPM and data lake

• Better CFR, stock levels and bias
• Several times more DFUs per FTE without growing the team

• The forecast's level of detail did not suit the supply team
• The short-term plan had to be corrected by actual sales
• A regional forecast was hard to split by warehouse

• A product-to-warehouse allocation matrix
• Online adjustment of volumes by the sales flow
• A controlled model of forecast disaggregation

• Fewer errors in warehouse allocation, higher CFR
• Less effort on master data

• Promo shipments could not be adjusted quickly without changing the promotion itself
• Incorrect warehouse stock data
• Target stock levels could not be set without promo data

• A model that adjusts volumes by actual shipments and orders
• Manual adjustment of the model's result
• Approval and plan versions
• Shipment intensity by customer and region

• More accurate stock in promo periods and short-term forecast
• Lower storage costs

• No assessment of forecast quality across dimensions
• Slow reaction to changes in the sales mix at retail chains
• Hard to assess the impact of sell-out on sell-in

• BI on the corporate data lake
• All demand planning KPIs across horizons
• Forecast quality models, visualisation of deviations and drivers
• Metrics of the planning team's work

• One version of the truth for the sales forecast
• Weak points of models and processes found and fixed
• Higher SFA and CFR

• Sell-out could not be matched with sell-in
• No sell-out detail for national retail chains

• Data products to analyse the impact of sell-out on the sell-in flow
• Changing granularity with correct aggregation
• Points of sale mapped to shipments
• Streaming search for relationships

• A clear view of how retail sales drive shipments
• A more accurate short-term forecast
• Sell-out factors added to sell-in models

• Systems built by us and by other vendors had to keep running and changing
• Legacy systems with no one left to support them
• A high volume of integrations with ERPs and partners

• Application support with continuous refactoring for a long life cycle
• A SQL Server cluster and deployment systems
• Legacy systems from other vendors taken over and rewritten without impact on users
• Integrations at high volume across many message types — SAP (IDoc, RFC), e-document and 3PL providers

• A large estate of integration instances under support

• The forecast had to follow off-the-shelf sales, not only shipments
• Planners needed one forecast across days, weeks and months
• Many inputs — promo, routes, customer forecasts, prices, launches, stock — lived in different places

• A sell-out demand forecast by customer warehouse, shipping warehouse, SKU, week and unit of measure
• Short-term daily, mid-term weekly and long-term monthly forecasts in one model
• Many data sources integrated automatically; inputs at any level, as absolute or %, calculated to all levels
• The forecast broken down into its building blocks, showing what each factor contributes
• Adaptable for customers without sell-out data

• Better forecast accuracy and speed
• The drivers of change understood
• Higher customer service level
• Optimised inventory and production

• Procurement decisions and the cost of recipes were planned apart
• The financial effect of a change in suppliers, currencies or volumes was hard to see in advance

• Raw material and recipe costs forecast from bills of materials, raw material prices and production volumes
• One model connecting the commercial function with finance and production
• What-if analysis for procurement: supplier share, procurement currency, volumes, exchange rates
• Price, spend and inflation reporting by finished good and raw material, with the drivers of cost change

• Lower raw material cost in the finished product
• Better margin
• Quick switches to alternative suppliers and currencies
• Less manual work

• Recipe formulations were managed with slow, opaque methods
• The financial effect of a recipe change was hard to forecast

• Scenario planning for recipe formulations by SKU, brand or group of SKUs
• Ingredients adjusted or substituted across SKUs, brands and product groups
• Built-in validation against the company's quality thresholds
• R&D, commercial and production teams planning in one place, with the result sent to the ERP

• Lower cost of goods
• Faster, better decisions on raw material substitutions
• Shorter new product launch cycles
• Less manual work

• Logistics and storage costs were forecast by hand
• Finance, commercial and logistics teams worked from different numbers

• Logistics and storage costs forecast from sales and production volumes
• Detail down to customer, SKU and warehouse
• Collaborative planning and scenario analysis across finance, commercial and logistics
• One source of truth fed by the model's data integrations

• Lower logistics costs
• Faster, better logistics decisions
• Better planning accuracy
• Less manual work

• Logistics and storage costs depend on a wide range of variables
• Questions about launches, packaging or supply chain changes had no quick answer

• Scenario modelling of logistics and storage costs, factor by factor
• Immediate answers: the cost of a launch, a new pack size, a palletisation change, a different supply chain for SKUs or customer segments
• Scenario comparison with the drivers behind each change

• Lower logistics costs
• Better margin
• Faster, better decisions in logistics and production
• Faster P&L calculations for new launches

• Production costs had to be tracked and explained in detail
• Budget allocation needed the specific cost drivers

• Production costs forecast from sales and production volumes by plant, line and cost item — electricity, personnel, depreciation and others
• Bridge reporting that shows the drivers of change by cost item

• Optimised production costs
• Better margin
• Faster, better decisions

• Promo calendars were built by hand
• Scenarios were hard to compare before sending a plan to the TPM

• Users set parameters and constraints — discount depth, maximum discount and others — and a goal: volume or margin
• An ML calculation generates a promo calendar by SKU
• A report with total volume, ROI and incremental volume and value by customer
• Recalculation after changed inputs or manual edits; the approved scenario sent to the TPM

• Better promo ROI
• Better margin
• Higher sales volume

• The volume effect of a price change could not be seen before it reached the market

• Pricing effects calculated from market price elasticity data: elasticities, price ladders, thresholds
• Calculation by SKU, customer and sales channel, connected to the demand forecast
• What-if simulation of pricing strategies, with reporting in BI

• Granular pricing management
• Faster what-if modelling and decisions
• Better margin

• Savings and overspending on trade investments were seen only in actuals
• Sales, finance and demand planning planned trade spend separately

• Integrated planning of trade investments across sales, finance and demand planning
• Scenario planning for risk assessment and opportunities
• Foresight into savings and overspending by brand and customer, before the actuals
• Trade spend efficiency tracked and aligned with the sales forecast and financial goals

• More efficient trade investment
• Higher sales volume
• Better margin

• A manual planning process
• Risks and opportunities were collected separately by each function

• An FP&A forecast with what-if modelling
• One process for collecting risks and opportunities across functions
• Scenarios built on risks and opportunities, with their impact on the P&L at once
• A move to an automated rolling forecast, with reporting in BI

• Faster, better decisions
• Better resource allocation
• Less manual work

• Margin by customer and product segment was hard to plan

• Margin scenarios from shelf prices, shipping costs to customers, sales forecasts and production costs
• The company's own P&L and its customers' P&L modelled down to product line and month
• Levers: shelf pricing, trade investments, COGS and others

• More efficient trade investment and promotion
• Better relations with customers
• Higher revenue and margin

• Customer assortment matrices had to be planned and their execution followed

• Analysis, planning and management of customer assortment matrices
• Objectives set and deviations from the assortment strategy monitored
• Integrated with merchandising tools, master data management and other planning models
• Analytics of shelf space and product portfolio by customer, channel and point-of-sale type

• Better distribution
• Better use of shelf space
• Higher sales volume
• Better margin

• The product portfolio had to be planned over its full life cycle

• Oversight by project, SKU, customer and week
• Project parameters set up from benchmarks, adjustable by hand
• Seasonality, logistics costs, sales per point of sale and distribution taken from the chosen benchmark
• Projects viewed by life-cycle stage, from concept to distribution

• Faster product launches
• Better return on new launches
• Higher sales volume

• Interest rates and currencies fluctuate; the cash position had to be clear in advance

• Cash movements forecast by bank account, legal entity and currency
• Actual transactions combined with the financial plan
• Changes in payment terms, capital investments, customer bonuses and more assessed from the business functions' inputs

• Faster, better cash flow decisions
• Efficient use of available cash
• Lower borrowing costs
• Less manual work