Frequently asked questions

Your historic ad units across whichever formats you've run, the campaign performance data that goes with them, and your brand guidelines. Ad server access is only needed if you want real-time optimization.

The model is built on your historical ad and performance data, so depth of insight depends on how much you have. The initial model is typically delivered within a few weeks of data onboarding.

Ad server access. With it, the system rebuilds the optimized variant and serves it during the campaign flight, swapping out underperforming versions automatically. Without it, you still get the full diagnostic output to brief against.

Yes. Results are produced at segment level, so you can see which attributes resonate with a geographic, demographic or behavioral segment rather than averaged across everyone. Paired with the segmentation model, that connects directly to your customer groups.

Over a thousand attributes. On the visual side: colors, contrast ratios, objects and people in frame, logo placement and size, CTA position, image quality, aspect ratio, black and white versus color treatment. On the audio and copy side: bit rate, codec, sentiment, video duration, copy character count, CTA language, and caption sentiment via NLP.

Any format. The model is trained on static images, video and audio, so it analyzes TV spots the same way it analyzes a display banner. Visual and audio attributes are both processed regardless of where the ad ran.

Traditional testing tells you which creative wins. Creative diagnostics tells you why it wins, down to the specific visual, audio or copy attributes driving the difference. That means the learning applies to future briefs, not just the campaign you tested.

Revenue, commercial and pricing teams who set prices and need to understand the demand effect before they do. Finance uses it to see where margin is being left on the table.

Sensitivity is related to product availability, so pricing reflects what you have left to sell rather than treating demand as unconstrained. Paired with inventory management, markdown goes where it clears stock instead of where it gives away margin.

Yes. Elasticity in subscription businesses varies sharply by tenure: long-standing customers tend to absorb an increase with little effect on retention, while recently acquired customers churn at a rate that can cancel out the gain. The model shows you where that line sits.

Historical pricing, sales volumes and promotional history for each product, plus competitor prices where you have them. Customer and segment data sharpens the curve considerably, so if you're already running segmentation, the two connect directly.

No, that's the aggregate approach and it's where most models stop. Ours resolves to segment and individual level. Averaging across your whole base hides the customers who'd pay more and the ones who'd leave at the first increase.

Benchmarking tells you what competitors charge. It doesn't tell you how your customers respond to your price. Most pricing decisions come down to competitor prices, last year's price and the judgment of whoever knows the category best. This model replaces the judgment call with a measured curve.

It uses GA4 as one input but models against your full transactional data, including customer records, order history and margins that GA4 can't see. That's what makes the output specific to your customer base rather than an aggregate view built from what Google can observe.

Yes. Interaction variables in the model quantify both halo effects, where brand advertising lifts search conversion, and cannibalization, where branded search captures people who would have arrived organically. You see where spend is compounding and where it's buying traffic you already had.

Media doesn't work instantly or stop when the campaign ends. A brand TV spot has a long lag and a long tail; a shopping ad converts in the same session and fades fast. The model estimates lag and half-life per channel so contribution lines up with how each one actually behaves.

It typically needs several sales cycles before a change in the mix shows up reliably. MMM is built for planning and budget allocation, not week-to-week optimization. Attribution is the tool for that.

Every platform reports inside its own silo, credits view-throughs to people who were buying regardless, and can't see offline. Add up their claimed conversions and the total usually exceeds actual sales. MMM reconciles to what you sold, which is why the numbers differ.

Attribution follows clicks and resolves the digital journey to individual touchpoints. MMM works in aggregate and includes everything attribution can't see: TV, offline, impressions, view-throughs. Most clients run both and fuse them, so channel-level incrementality and touchpoint-level detail come from one connected view.

Two or more years of media data is the working minimum, since the model needs enough variation across time and channels to isolate effects. It works best for businesses spending over £750,000 to £1 million a year across all channels.

Segment codes are an input variable across attribution, media mix, creative diagnostics and forecasting. That's what lets you see not just which channel or message is working, but which customers it's working for.

Yes. The clustering approach is the same. The inputs shift toward account-level behaviour, contract value and engagement patterns rather than individual purchase history.

Typically 20 to 30 for a B2C business. Enough to differentiate messaging meaningfully, not so many that the segments become impractical to act on.

As new transaction data comes in. Individuals move between segments over time, and the model tracks and quantifies the value change at each move.

The model performs best on large, structured datasets with a meaningful transaction history. We scope it to what you have and enrich from external sources where there are gaps.

Yes. Kleene's segments can be run alongside or benchmarked against your existing classification. For clients already using national geodemographic data, the bespoke model typically adds significant value because it reflects actual customer behaviour rather than postcode-level proxies.

Most existing segmentation is a fixed snapshot: postcode, age band, spend tier. This model is built from your own transaction data and updates as customers behave differently, so it shows you who is becoming more valuable and who is drifting away, not just which box they sat in last year.

It updates as new sales data comes in, so the projection and the driver rankings both move with recent performance rather than being rebuilt once a quarter.

Not strictly, the two are built independently. If you carry stock they're usually delivered together, since that's the pairing that gets you an 18 month planning horizon.

Historical sales and demand data, your product, market and channel hierarchy, and explanatory variables like seasonality. Customer segment data is optional and improves accuracy where you have it.

No. It gives your planners a modeled baseline and a ranked set of drivers to work from, rather than a spreadsheet built on last year plus a percentage. The judgement calls stay with your team.

No. Businesses carrying stock get the most granular version. For service and experience businesses, the value sits in smoothing demand and improving capacity utilization.

This model ranks the drivers behind demand, including external ones like weather and event calendars, and updates as new data arrives. You get the projection and the reasons for it, which is what makes it actionable.

Typically two to three years of daily or weekly data. Clients running demand forecasting alongside inventory management generally plan up to 18 months out.

A 15 to 30% reduction in excess stock, fewer stockouts, and a planning horizon of up to 18 months when it runs alongside demand forecasting.

Manual reorder points are set from past consumption and rarely revisited. The model relates each SKU's demand profile to supplier constraints and the cost of holding stock, and updates as new data comes in. It also handles variation by location and by customer segment, which is impractical to do by hand across thousands of SKUs.

No. It runs alongside it. Your ERP or WMS stays the system of record for stock and purchase orders, and the model reads from it. What you get back is the forecast-driven layer on top: reorder points, order quantities and risk flags that your existing system doesn't produce.

Historical sales and demand at SKU level, your product hierarchy down to variant, supplier lead times and availability, purchase order status, and holding costs with stock by location. Most of it comes out of your ERP and ecommerce platform, and we scope the model around what you actually have.

It works from forecast demand at variant level, with supplier constraints and segment level demand variation built in. It runs alongside your ERP rather than replacing it, and takes stock and purchase order data from it.

That's the reason this needs machine learning rather than conventional statistics. Demand varies by geography and by segment, and stock requirements vary with it. The factors are considered in unison rather than market by market, and stock is modeled per location.

Not strictly, the two models are built independently. In practice they tend to be delivered together, and that pairing is what gets you a planning horizon of up to 18 months.

The full walkthrough for Claude, ChatGPT, Cursor, and other MCP-compatible clients is in the Kleene documentation at docs.kleene.ai. You will need an active Kleene.ai account with MCP enabled in App Settings under the AI tab before you start.

Phase 2 adds full ELT pipeline creation through conversation. Phase 3 brings analytics model context, so KAI Analytics models such as segmentation, demand forecasting, and price elasticity can be queried directly from your AI client without opening a separate visualization layer.

Today the MCP covers what KAI Assistant Phase 1 does: finding and inspecting transforms, generating SQL, searching logs, browsing schemas, and answering questions from the Kleene.ai docs. It cannot build full ELT pipelines through conversation yet, and it cannot query analytics model outputs such as segmentation or forecasting. Both are on the roadmap.

It is the same capability reached from a different place. KAI Assistant has worked inside the Kleene.ai platform for a while, and the MCP integration makes that available through whatever AI client your team already uses. If you prefer working in the Kleene app, nothing changes. If you would rather stay in Claude or Cursor, now you can.

Yes. Answers are grounded in the Kleene.ai documentation, so you can ask how pipeline scheduling works, what the recommended pattern is for incremental transforms, or how to set up a specific source connector. It is useful when you are onboarding someone new, or when you cannot remember a setting for the fourth time.

Yes, within the tables and schemas your account can access. You can ask for revenue by customer segment for the last quarter, weekly order trends from the ecommerce schema, or a report covering session volume, conversion rate, and revenue by channel. The answers come from your warehouse data rather than generic assumptions.

Yes. You can see which tables sit in a schema, inspect column names and data types, and preview small samples without leaving your AI client. Previews are protected by default, so personally identifiable information is obfuscated before rows are returned, unless your organization has enabled raw preview access.

Yes. Ask something like "why did the orders pipeline fail last night" and the MCP searches the logs for the relevant transform, surfaces recent errors, and explains what went wrong. It replaces opening logs, finding the right transform, and reading through raw run output by hand.

Yes. Describe what you need in plain English and the MCP drafts the SQL, then shows you a diff. You decide whether to insert it into the editor or send it to sandbox for testing. Prompts like "write a transform that calculates 30-day rolling revenue by customer" or "rewrite this join to reduce query cost" both work.

Yes. You can search transforms and transform groups by name, pull the SQL for a specific transform, check when it last ran, and retrieve the latest version. Prompts like "show me the SQL for the customer LTV transform" or "find all transforms in the marketing pipeline group" work without opening the Kleene UI.

It depends what you are doing. Asking questions about your data, browsing schemas, pulling the SQL behind a transform, and checking why a pipeline failed all work in plain English and require no code. Generating or editing transforms is more useful if you can read SQL, since you review a diff before anything is saved, but you do not have to write the SQL yourself.

The proposed SQL lands in your sandbox environment, where you can test it before it goes anywhere near production. Nothing is committed to production automatically, write access follows your existing Kleene role permissions, and changes are reviewed before they are applied.

Yes. Write access through the MCP follows the same role permissions you have already set in Kleene. If someone cannot edit a given transform inside the platform, connecting through Claude or ChatGPT does not change that. The AI client is another way into your account, not a way around its controls.

No. When you ask the MCP to write or rewrite SQL, it shows you a diff first, and nothing happens until you choose to insert it into the editor or send it to sandbox. You are always the one who decides whether a change goes anywhere. There is no path where the AI quietly edits a live transform on its own.

By default, table previews use synthetic samples, with personally identifiable information, passwords, and payment data stripped out before anything reaches the AI client. When you preview the orders table, the assistant sees the shape and structure of your data, not your customers' actual details. Raw previews can be enabled in App Settings, but they stay off until you turn them on.

A Claude admin with permission to add connectors creates the custom connector once for the workspace. After that, individual users connect it themselves under Customize, then Connectors, and sign in with their own Kleene credentials. Each person's access reflects the Kleene role they already have.

In ChatGPT, open Settings, then Apps, then Advanced settings, turn on Developer Mode and click Create apps. Enter Kleene as the name, https://mcp.kleene.ai/ as the MCP server URL, and select OAuth for authentication. Click Create, sign in to Kleene to complete the authentication flow, and enable Kleene in conversations when you need it.

A Claude admin adds a custom connector under Customize, then Connectors, using the name Kleene, the server URL https://mcp.kleene.ai, and individual sign-in enabled. Each user then goes to Customize, then Connectors, searches for KleeneAI, clicks Connect, signs in to Kleene, and approves access. After that you can start asking questions about your workflows.

Go to App Settings, open the AI tab, and switch on External AI Assistants (MCP). Use a Kleene account with permission to access the workflows, transforms, schemas, and pipeline runs you want to query. Once MCP is on, follow the connection steps for your specific AI client.

Yes. Kleene MCP has been available since August 2026 for Claude, ChatGPT, Cursor, and other MCP-compatible clients. You need an active Kleene.ai account with MCP switched on in App Settings under the AI tab.

Claude, ChatGPT, and Cursor are supported, along with any client that implements the Model Context Protocol. Because MCP is an open standard, new tools that adopt it will work with Kleene without anything changing on your end. You connect once in App Settings, then use whichever client your team already prefers.

You can search and inspect transforms, generate or rewrite SQL, debug pipeline failures, browse warehouse schemas, preview table data, analyze data in your warehouse, and ask questions grounded in Kleene documentation. Where write access is enabled, proposed changes can be sent to sandbox for review and testing. Everything runs against your own Kleene account and your live data.

MCP is an open standard that lets AI assistants connect securely to external tools and data sources. Instead of copying results between tabs, your AI client queries the system directly inside the conversation. Kleene supports MCP, so assistants like Claude, ChatGPT, and Cursor can work with your real transforms, schemas, and pipeline logs rather than hypothetical examples.

Kleene works from the data already in each platform's reporting and your own first-party systems (orders, CRM), joined in your warehouse using shared keys, timing, and signals such as coupon codes and post-purchase surveys -- plus impression-level data for upper-funnel platforms. It does not require you to hand over user-level first-party tracking or rely on third-party cookies.

Beyond the initial calibration and daily QA checks, Kleene validates models by backtesting against your historical outcomes and comparing predicted versus actual results. Because the pipelines, transformations, and assumptions are fully transparent, you can review and challenge the logic rather than trusting a black box.

Yes. Kleene can stand up a scoped proof of concept on a subset of your data so you can validate the value before committing to a full rollout.

Kleene.ai is ISO 27001 certified. Kleene also operates inside your own cloud data warehouse, so your data never leaves your environment, with role-based access, minimum-privilege access, and optional SSO. Kleene is GDPR-aligned and privacy-first, with no reliance on third-party cookies or user-level tracking.

Implementation is included. Pricing is a flat recurring fee based on your package and the number of connectors you need, with hands-on onboarding included as standard. New connectors are built at no extra charge. Larger bespoke custom development beyond your package may be scoped separately, but standard implementation and ongoing maintenance are part of your subscription.

Kleene.ai is typically provided as an annual subscription, billed monthly or annually.

Kleene.ai unifies data from your CRM, ecommerce, marketing, and finance systems into your warehouse and resolves it into shared entities -- customers, orders, products, and revenue -- to give you a single customer view. It complements rather than replaces your CRM: your CRM stays the system of record, while Kleene provides the unified analytical view and can push enriched segments back out. It can also remove the need for a separate CDP by handling unification and activation directly in your own warehouse.

Kai is Kleene.ai's built-in AI assistant. It lets anyone ask questions of your unified data in plain language -- for example "how much has revenue grown for this franchisee?" -- and get answers, charts, and tables without writing SQL. Because Kai sits on top of your modelled warehouse data, it can join across sources, surface anomalies and inconsistencies, and explain trends. Kai will soon be available to view-only users as well as licensed users.

Kleene.ai combines data sources across the funnel, allowing top-funnel activity to be evaluated alongside downstream outcomes. This avoids relying on last-click logic and ensures early-stage channels are included in performance analysis.

Kleene.ai’s Data Consultancy handles architecture design, implementation, modeling, and validation for you. This removes the need to hire or coordinate multiple data roles and typically shortens time to value from months to weeks.

Kleene.ai prepares and maintains customer segments in your data warehouse. These segments can then be activated through downstream tools or existing reverse-ETL workflows, depending on your setup and connectors.

Yes. Customer Acquisition AI uses data ingested through Kleene.ai’s connectors, which include common marketing platforms such as ad networks, web analytics, and email tools. These integrations feed the underlying models through your central warehouse.

Kleene.ai operates directly on your own cloud data warehouse (Snowflake, BigQuery, or Redshift). Your data never leaves your environment, and Kleene only accesses it using credentials you control. Access is governed through role-based permissions, user-level credentials, and optional SSO.

Yes. Dashboards and reports refresh automatically and can be shared with stakeholders without manual effort.

Yes. Kleene.ai supports attribution modelling that combines top-down and bottom-up approaches.

Yes. Offline data can be ingested via files, databases, or custom connectors and modelled alongside digital channels.

Yes. Multi-tenant setups (coming in Q2 2026 to Enterprise package customers) allow agencies to manage multiple clients securely within one platform.

Using unified historical data and AI models, Kleene.ai helps forecast performance and identify which channels and campaigns drive the best returns.

Kleene.ai connects to platforms such as Google Ads, Meta, Klaviyo, CRMs, ecommerce platforms, and web analytics tools. We can also build custom connectors for any other data sources.

Yes. Kleene.ai works with your existing BI tools, including Power BI, Tableau, and Looker. Kleene.ai prepares the data in your warehouse, so you can keep using your current dashboards or add a BI tool through Kleene if needed.

Yes. Marketing spend, revenue, and outcomes are modelled together so ROI is consistent and comparable across channels.

Most marketing tools focus only on reporting within a single channel. Kleene.ai unifies marketing, finance, and ecommerce data into a single source of truth and adds predictive intelligence on top. Kleene also attributes both direct clicks and the impact of upper-funnel impression activity -- from platforms like Meta, TikTok, Pinterest, and YouTube -- giving a more accurate picture of true channel contribution than last-click models. Kleene does not rely on invasive user-level tracking or third-party cookies, making it future-proof as privacy regulations tighten.

Yes. Role-based access controls allow you to restrict access by user, role, or dataset.

Authentication methods depend on the source and include OAuth login, API keys, and access tokens.

Yes. Data can be corrected and reprocessed safely using controlled backfills and reruns.

You can extract full historical data from supported sources, subject only to source system limits.

Traditional tools focus only on moving data. Kleene.ai delivers ingestion, transformation, analytics, and AI-driven insight in one platform.

Most implementations take six to eight weeks. Every customer receives hands-on onboarding, a dedicated Customer Success Manager, and ongoing support.

All data flows into a single warehouse with shared definitions and models. This ensures finance, marketing, and ecommerce teams work from the same metrics.

Mapping tables can be managed directly inside Kleene.ai, or synced from external sources like spreadsheets if preferred.

Yes. Kleene.ai provides pre-built schemas and models that surface shared entities like customers, orders, products, and revenue across systems.

Data is normalised and joined at the transformation layer using shared keys and mapping tables. This ensures consistent reporting even when source schemas differ.

Yes. Kleene.ai supports flexible mapping tables and dimensional modelling to replicate and extend class code logic across systems.

Connectors and models are fully customisable. Kleene.ai can build new connectors or adapt existing ones to support bespoke schemas and business logic.

Yes. Once data is centralized, teams can run advanced analytics directly or use Kleene.ai’s built-in AI applications for forecasting, segmentation, and optimization.

Your data lives in your own cloud data warehouse -- Google BigQuery, Snowflake, or Amazon Redshift. You retain full ownership at all times and Kleene does not lock you in. Kleene accesses your data securely to operate pipelines and models, with access governed by role-based permissions and minimum-privilege principles. User-level credentials and optional SSO are also supported.