AI that survives its second quarter in production.

Two-week AI readiness sprint

The problem: You have been told to do AI. Nobody can tell you whether your data supports the use case, what it would cost to run, or what would have to be true for it to be worth doing.

One use case tested against your real data, ending in a decision rather than a deck:

• Current-state cost baseline for the process you want to change
• Ranked use-case list with effort, data readiness and payback per case
• A working thin slice on your data, not a vendor demo
• A written scope for step one, or a written recommendation to stop

What we commit to: One of three written answers: build, fix the foundation first, or stop. Two of those three save you money, which is why you get them.

When this is the wrong purchase: If you have no owner for the process today, the sprint will produce a good answer to the wrong question. Fix the ownership first.

EU AI Act & AI governance readiness

The problem: Enterprise buyers ask AI Act questions in security review, and your board wants to know which systems are high risk.

The paperwork and the controls, written to survive an audit:

• AI system inventory with risk classification per system
• Gap assessment against the obligations that apply to you, dated
• Technical file skeleton: intended purpose, data, oversight, monitoring
• Governance structure: roles, review cadence, incident process

What we commit to: We classify honestly, including telling you which of your tools are out of scope. Inflating scope is how this work gets expensive.

Governed natural-language BI

The problem: Copilot confidently returns wrong DAX because your semantic model has three names for the same customer and no descriptions.

The model work that decides whether natural-language BI is trustworthy:

• Semantic model hardening: naming, descriptions, relationships, hierarchies
• Verified-answer set: 40 questions with expected values, run as a regression test
• Row-level security that holds through the conversational layer
• Prompt and question patterns published to users, plus training

What we commit to: On a well-described model you should see 80 to 90 per cent of natural-language questions answered correctly. We measure that on your data before rollout and publish the number internally.

When this is the wrong purchase: If the underlying metric definitions are disputed, fix those first. Copilot makes a bad model faster, not better.

RAG over your documents and databases

The problem: Contracts, specs, regulatory guidance and account history are scattered across SharePoint, file shares and an ERP. Keyword search returns noise, and ungoverned chatbots invent compliance answers.

Retrieval that cites its sources and refuses to guess:

• Ingestion and hybrid retrieval over the corpora that matter, with access control inherited
• Answer contract: citation, confidence, and an explicit refusal when the source is absent
• Evaluation set with graded answers, run on every change
• Cost and latency card per 1,000 questions

What we commit to: Every answer carries its source. If the retrieval cannot find support, the system says so instead of writing something plausible.

When this is the wrong purchase: If the documents are not worth reading, they are not worth retrieving. We will say so in week one.

Agentic workflows for operations

The problem: Order intake, claims triage and supplier follow-up run on copy-paste between four systems, and the handoffs have no owner.

One workflow in production with human approval gates, not five in a sandbox:

• Workflow map with the exact steps that are safe to automate
• Tool-using agent with write access restricted to approved actions
• Approval gate design: what a human must sign before the action lands
• Full audit trail: input, retrieved context, tool calls, output, approver

What we commit to: We run the pilot in shadow mode first and report the error rate before anything writes to a live system.

When this is the wrong purchase: Autonomous agents with write access and no approval gate. Not with your ERP, and not on our insurance.

Forecasting & anomaly detection

The problem: Demand planning is a spreadsheet, stockouts and overstock happen in the same month, and nobody trusts the forecast enough to act on it.

A model you can defend in a planning meeting, with its error bounds shown:

• Baseline comparison against your current process, on your history
• Accuracy report: MAPE or MASE by product family, with the failures listed
• Anomaly rules for the two or three exceptions worth waking up for
• Monthly retraining and drift monitoring

What we commit to: We score against your existing process first. If the model does not beat it on your data, you should not buy it and we will tell you.

Document intelligence

The problem: Invoices, delivery notes and contracts are retyped by hand. The legacy scanner reads about two thirds of them and the rest land on someone’s desk.

Extraction with a validation layer and a human queue for the exceptions:

• Extraction pipeline for invoices, contracts or forms at your volume
• Validation rules: PO match, vendor list, arithmetic, duplicate detection
• Exception queue with the reason code for every rejection
• Throughput, cost per document and accuracy measured on a held-out set

What we commit to: Accuracy is reported on documents the model has not seen, with the sample size printed next to the number.

MLOps & model monitoring

The problem: A model launched well, then degraded for a year while everyone kept reading its output.

Monitoring that treats drift as a paging event:

• Deployment pipeline with versioning and rollback
• Drift, input-quality and performance monitors with thresholds
• Retraining schedule and the owner for each model
• Model register: purpose, data, owner, review date, current status

What we commit to: Every model in the register has a named owner and a review date. Unowned models get switched off.