We build AI into the work that already makes you money: quotes, renewals, the monthly report, the questions that queue behind one person. We map the workflow, connect it to the systems you already run, run the repeatable steps as governed agents, and put every run, cost and approval on one screen your finance team reads. Four things: understand, think, execute, govern.

No. It's a system doing real work across the systems you already rely on, including the fifteen-year-old one without a clean API, with a person approving anything above the limit you set. If what you want is a chatbot or a strategy deck, we're probably not a fit, and we'll say so.

Where the baseline and the outcome can be measured together, part of our fee sits on the outcome. Agreed up front, capped both ways, trued up every ninety days. Where a baseline can't be measured from your systems, that workflow runs on a flat fee instead. No blended numbers.

The outcome share moves with the number, in both directions, and the cap is written into the agreement before anything ships. Ask any AI vendor what happens to their invoice when the outcome misses. Then ask us.

Your finance team. They own the definitions of a saved hour or a captured dollar, sign the measurement method before anything ships, and can rerun it without us. The baseline is trailing twelve months, agreed in writing in week one. A baseline set after we start means nothing.

Cost down, revenue up, hours back, capacity up, quality up, risk down. Every workflow we build has to move at least one of them, and we agree on how to measure it before we start. Every ninety days you get the same review, for every workflow, with your finance team's signature on it.

The ROI Dashboard is the XSparks screen that shows what every workflow costs, what it returns, and what still needs a human's approval. Per-run cost and return, attributed to workflow, team and user, with the approval queue beside it. Your finance team reads it. So does the board.

Week one, the baseline. Weeks two to six, the first system live. Every ninety days after that, the review. Bring us one stuck workflow and we'll tell you plainly whether the math works before any of that starts.

Numbers, not testimonials. Every published XSparks result comes with five things or it doesn't get published: the company or approved context, the baseline, the date range, the measurement method, and client finance sign-off.

Mid-market companies with real operations, legacy systems, and no appetite for another science project: manufacturing and industrial, field service and trades, PE portfolios, MSPs and IT service providers, architecture and engineering, professional services. Probably not a fit if the work doesn't repeat, if nobody is willing to measure, or if you want a chatbot.

12 questions, under 3 minutes, scored on screen. It covers what actually decides whether AI works here: documentation, process, knowledge, adoption, measurement, oversight, approvals, carry-over. Your weakest dimension is your real level. No deck, no sales call unless you ask for one.

In your own cloud infrastructure, in your own account, in the region you choose, billed to you. AWS, GCP, DigitalOcean, whichever you run. If you would rather we host it, we can, billed separately. Either way tenants are isolated and no tenant's data touches another's. If you operate partner or child companies, you can provision them their own tenants underneath yours.

Not your documents. Documents are processed first: the information is extracted, entities and relationships are built from it, and only that extracted context reaches a model. PII redaction is available on the way out. On the public models, enterprise retention is switched off, so the provider keeps nothing. If you would rather nothing left at all, run a private open-source model inside your own infrastructure.

No. The default is embeddings for retrieval, which is not training. Fine-tuning happens only when you ask for it and your use case needs it, it is triggered by a person in your team through the interface, and what it does is disclosed before you sign anything.

Any of them, public or private. We are not tied to one provider. You bring your own accounts and keys, which means your provider bills you directly and we never touch your token spend. That is what at cost, no markup actually means. If you would rather we provision the models, we can, and any key we hold sits in an encrypted, pen-tested vault. The model is chosen per task rather than per workflow, from a task table we agree during discovery. You can pin your own model instead, and that choice is yours to make against the cost.

One kill switch stops any workflow or agent, mid-run, across the parent company and every child company underneath it. You can also roll back. If a context change trickled into several workflows, revert the context to the previous version and replay them. Separately, you set dollar limits by company, by department and by model provider. You get alerts before a limit, and at the limit it falls back to another provider rather than stopping the business dead.

Every run. Who triggered it, whose account it ran from, the department and company, which workflow and agent, which task, which model and which provider, the inputs and the outputs. Cost is attributed automatically from the run up to the user, the team, the company and the workflow, with nobody tagging anything by hand. It exports to CSV, and your finance team can pull it and rerun the consumption numbers without us.

SSO and SAML, yes. Role-based access runs down to the individual workflow, agent and context. Admins and heads build workflows and set context. Everyone else runs what has been provisioned to them, by department or by team. The whole structure is configurable to how your organisation is actually shaped.

Your data is yours. Your logs are yours. Your prompts, and any agents or workflows your team built and configured, are yours, and you take the personas with them. The platform itself is licensed annually, so our code and the workflows we ship on day one stay ours. Nothing of yours is locked to our platform.