How we work.

The mechanism is the point. This page sets out how a piece of work actually runs here: what gets measured before anything is built, where a person stays in the loop, and what happens in the twelve weeks after go-live.

Talk to us

Three principles, and what each one costs.

Every principle here has a price. It is easier to trust the ones that name it.

Build it into the work

AI that sits beside the job gets opened twice and then forgotten. The test we apply is whether the thing appears inside a system somebody already has open: the ticket queue, the invoice run, the approval path. If a person has to remember to go somewhere else, adoption is already lost.

What it costs youIntegration is the majority of the effort and it is not the interesting part. We will spend more time on the wiring than on the model, and the estimate will say so.

Deploy it where the data lives

When the data is contracts, HR records or customer information, where it physically sits stops being an architecture preference and becomes a question your DPO has to answer. We deploy inside your network when the data demands it, and on frontier APIs when it does not. That decision is made per workload, with your security and legal people in the room rather than after them.

What it costs youPrivate deployment is slower to stand up and more expensive to run than an API call. We will model both and show you the gap before you choose.

Stay after go-live

The failure mode is rarely a bad launch. It is week twelve, when nobody owns the thing and the usage graph is flat. We agree what gets measured before we build, take a baseline while the old process is still running, and report against it after go-live.

What it costs youIt means agreeing a number in advance that we might miss, and telling you when we have.

How an engagement runs.

Five stages. The second one is the one most firms skip.

A conversation

One call to work out whether the work fits. If it does not, or if it belongs somewhere else in the group, we say so on that call. Nothing to sign and no discovery invoice.

A baseline, before anything is built

We measure what happens today: volume, handling time, error rate, and who touches the work. This is the least glamorous stage and the one that makes every later claim checkable. Without a baseline, an improvement figure is decoration.

One narrow build

A single workflow, wired to the systems that actually hold the records, running in production rather than in a sandbox. Narrow enough to finish inside a quarter, real enough to prove.

Go-live on your thresholds

You set the confidence threshold, the action set and the points where a person stays in the loop. We implement them, log every step, and hand you the trail rather than a summary of it.

Twelve weeks of measurement

The four numbers we agreed, reported against the baseline. Then a decision to extend, adjust or stop, made on evidence rather than on how the launch felt.

What happens inside.

The same shape whichever practice the work sits in.

Your systemsThe ERP, the ticket queue, the file store, the contract archive. We connect to what is there rather than asking you to move it first.
RetrievalThe system finds the records that answer the question, filtered by what the person asking is entitled to see, and keeps a reference to every source it used.
ModelA frontier model, or an open-weight model on your own hardware where the data demands it. The choice is per workload and it is written down.
GuardrailA defined action set, a confidence threshold you own, and a logged trail of both. Anything outside the set stops rather than improvising.
ActionThe invoice posts, the ticket routes, the answer lands in the system where the work already happens, carrying a link back to what it was based on.

A person stays in the loop wherever being wrong is expensive. Anything under the confidence threshold goes to a human, and the threshold is yours to set rather than ours to recommend.

What we measure.

Agreed before the build starts, baselined while the old process still runs, reported after go-live.

ContainmentHow much the system finishes end to end, and how much it hands back. Tracked weekly, not claimed once.

AccuracyPer field and per document type. A vendor name and a tax amount fail differently and are counted separately.

Latencyp95 response on your workload, measured against whatever it replaces.

Use at week 12Active users three months in, against the week it launched. The number that tells you whether it stuck.

What we build on.

Models: Frontier models from the major labs, and open-weight models where the data has to stay inside your network. We do not train foundation models and we will not imply otherwise.
Engineering: Infrastructure, security and data engineering from inside the Dhanesh Indore group, which has been running enterprise systems for other people’s businesses for considerably longer than DimenAI has existed.
Where our advantage is: In the retrieval, the guardrails and the measurement rather than in the model itself. If your problem needs a novel model architecture, we are the wrong firm and we will tell you on the first call.

What we do not do.

We do not sell a pilot that lives in a sandbox. If it cannot reach a real system, it is a demo, and a demo proves nothing about your data.
We do not train foundation models. Our work is the wiring, the guardrails and the measurement around models other people built.
We do not report a number we did not measure. Where a figure does not exist yet, we will describe what was done instead.
We do not let an agent act outside the action set you defined, and we do not let one approve spend, sign, or contact a customer without a person in between.

Where the boundaries are.

Anything touching SAP is TechOrbit’s

If a project needs SAP, Joule, BTP or Clean Core work, it belongs to TechOrbit, the group’s enterprise technology firm. Blurring that line confuses two account teams and helps nobody. We make the introduction and stay out of it.

Computer vision is not sold yet

It sits in a planned Operational Intelligence practice, and the boundary with DVelta’s inspection robotics has not been agreed. Until it is, we do not take the work.

Tell us where the work is.

One conversation, no pressure. We will tell you what we would build, and what we would not.

Talk to us