Service 02
Applied AI
Most of what is sold as AI is a demonstration. What makes a difference in a business is narrower and duller: a model reading the documents nobody wants to read, and a person approving the result.
Where models earn their cost
Document processing
Invoices, delivery notes, contracts, technical datasheets. Extracting the fields you need, flagging what it is unsure about, and handing the uncertain cases to a person instead of guessing.
Answering from your own material
Retrieval over your documentation, past projects, product data or support history, with the source of each answer shown so somebody can check it.
Classification and routing
Sorting incoming enquiries, tickets or applications into the right queue with the right priority, which is a task models do well and people find tedious.
Drafting
A first version of a quotation, a reply, a product description or a report, produced from your own data and your own past output, for a person to correct rather than write.
Agents with an approval step
A sequence of actions the model completes on its own, with a defined stopping point where a person signs off before anything irreversible happens.
An assessment before you spend
A short engagement that tests the idea against your real data and reports the accuracy you would actually get. Sometimes that report says do not build this.
How an AI engagement runs
01
Find the task
Week 1
We look for a task with a clear input, a clear output and a measurable error. Tasks without those three are where AI projects run for a year and produce a slide deck.
You getA named task, a success threshold, and a sample of real data to test against.
02
Measure before building
Weeks 2 to 3
We run the task against your sample and report the accuracy, the cost per item and the failure modes. You see the numbers before committing to a build.
You getA written evaluation with accuracy figures and the cases it gets wrong.
03
Build with a human in the path
Weeks 4 to 8
The system goes into production with a review step, a confidence threshold, and a record of every decision it made. Nothing consequential happens without a person able to see it.
You getA running system, an audit log, and the review interface for your team.
04
Watch the error rate
Ongoing
Model behaviour drifts and so does your data. Accuracy is monitored against fresh samples, and the review threshold moves when the evidence says it should.
You getA monitoring dashboard and a monthly accuracy report.
What we will tell you not to do
- A chatbot on the website because competitors have one. It will annoy your customers.
- Replacing a rule that already works. If an if-statement is correct every time, a model is a downgrade.
- Any workflow where a wrong answer is expensive and no person reviews the output.
- Training your own model when a general one plus your documents does the job for a fraction of the cost.