AI for Business
AI for business, applied to one real job at a time.
Sahla works out where AI is worth using in your operation, builds the systems that do it, and helps with the AI you run inside your own business. Each one starts from a single real job: reading documents, forecasting demand, answering from your own knowledge. Each gets tested for the ways it fails before it goes near production, and you own what we build.
Most AI is sold as magic. We start by telling you where it is not worth using, build only what is, and hand you a version you own.
What we build
Five jobs AI for business does, and what each one looks like running.
Pick a job to see the real input on one side and what the AI returns on the other.
Every one of these ships with the evaluation below. That part is not optional.
How we test it
We test the AI we build. Every time.
Every build ships with a set of checks, or it does not ship. We assume it will be attacked, and we assume it will sometimes be wrong. The checks exist to decide what happens then.
| # | Check | What it means | If it fails |
|---|---|---|---|
| 01 | Grounding & factualityPassing | Answers trace to a real source in your material, and the source is shown. | Block the answer |
| 02 | Permission leakPassing | It will not surface data the person asking is not allowed to see. | Block and log |
| 03 | Prompt injectionPassing | A crafted message cannot override the instructions it was given. | Flag for review |
| 04 | Task accuracyMeasured | Measured against an acceptance set we agree with you before the build. | Send to a person |
For anything with weight, a person signs off before it acts.
This is the shape of the checks, not a report on one build. The thresholds are set with you before work starts, and what each check returns is specific to the system it is testing.
Own your AI
Vendor-owned or your own. We build both.
An AI assistant runs on a model, and that model is either the vendor’s or yours. Send your work out to a top-tier third-party model and someone else keeps it running. Run a top-tier local model inside your business and nothing leaves. Both are real options, both are good ones, and the answer depends on whether your data is allowed to leave. We build either, and the parts that matter are the same in both.
- Where your work goes
- Out to the vendor, and back with the answer.
- What you run
- Nothing. The vendor keeps it running.
- Choose it when
- That is the right call for the job.
- Where your work goes
- Nowhere. It stays inside your own walls.
- What you run
- The model itself, on your side.
- Choose it when
- The data cannot leave.
- You control where anything is sent.
- Every question asked and every answer given is recorded inside your business.
- Switching from one to the other is a change to how it is set up, not a rebuild.
Where we would tell you not to use it.
Some decisions should not be handed to a model. We would rather say so in the first conversation than after you have paid for a build.
- Decisions with legal, financial or safety weight — a person stays in the approval step.
- Thin or inconsistent data — we fix the data question first, or we pass.
- Work where a wrong answer costs more than the manual step it replaces.
Proof
Two builds, and what they actually changed.
Years of sales history in a database and no way to use it. We built a forecasting model on that history for inventory, pricing and campaign planning.
Customer-facing assistants with one job: talk to social and website visitors, work out who is a real buyer, and collect their details so a salesperson can follow up.
Our approach
We do not start with software.
We start with how the business works, identify the real cause of the problem, and decide what needs to change before choosing the technology. Then we deliver the solution and stay involved until it truly works.
Diagnose
Understand what is actually happening.
Define
Decide what should change, and in what order.
Deliver
Deliver the agreed solution alongside your team.
Improve
Support adoption, measure results, and refine what we've delivered.
A continuous cycle from clarity to results and ongoing improvement.
Evaluation runs inside Deliver and Improve, not as a phase at the end. Not ready to build yet? We run readiness sessions and training for teams getting started, and a full assessment for corporates rolling AI out across the business.
Take the AI readiness assessmentQuestions we get
Before you talk to us.
Both, and the choice is yours to make with us. A top-tier third-party model where that is the right call; a local model when the data cannot leave your infrastructure.
Glossary: Large language modelWherever you decide it can go. If the answer is nowhere, we build on a local model inside your perimeter and log every prompt and response there.
Glossary: Data sovereigntyWe check it. Answers have to trace to a real source in your material, and we measure that on an acceptance set before anything ships. Where it cannot ground an answer, it says so instead of inventing one.
Glossary: HallucinationThat is designed in advance: block, flag, or send to a person, depending on the check that failed. Anything with weight waits for human approval either way.
Glossary: Human in the loopYes. That is a routing decision, not a rebuild.
A first working build is typically six to ten weeks, priced per scope with the evaluation set included. We do not price the checks separately, because they are not optional.
When the data is thin, when a wrong answer is expensive, or when a rule-based automation would do the same job more reliably and cheaper. We would rather scope it smaller or pass.
Bring us one job you want AI to do.
We will tell you whether it is worth building, how we would test it, and where it should run.