A measured accuracy baseline
On a real sample of your own data, before anything goes live.
Demo to production
How AI actually reachesproduction.
A convincing answer takes an afternoon. A system a business can depend on takes evaluation, cost control, an approval gate and somebody who notices drift. That gap is where most AI projects stop.
Every mark here links to the engagement it belongs to. See the work →
AI implementation is the work between a model that produces a good answer in a demo and one a business can depend on daily: choosing the model on measured evidence, wiring it into a workflow people already use, capping what it can cost, and deciding what a person still approves. The hard part is almost never the model. This page is for companies deciding whether AI is the right answer at all, and what has to be true before it reaches customers.
Deliverables
What weactually deliver
The AI step, built into the process
Extraction, classification, retrieval or drafting, producing structured output that ordinary code can validate.
A human review path
Everything under the threshold reaches a person with the source document and the reason it stopped.
Decision logging
Every AI decision recorded with its inputs and confidence, so a wrong one can be found and the rest checked.
Where it starts
The four places thisusually shows up
Operations
Work moves between people by hand, and every handoff can stall.
- The same record is typed into two systems
- A spreadsheet sits between two applications
- Volume grows, headcount is the only lever
Finance operations
Invoices and approvals depend on somebody remembering to check a folder.
- Month end is a week of copying
- Approvals live in an inbox rather than a system
- Nobody knows the cost per document
Customer support
The answers exist, in documents and past tickets, and finding them is the job.
- The same answer is rewritten daily
- Response time depends who picks it up
- The demo worked, production did not
Reporting and data
The number the business runs on is assembled by hand, twice, differently.
- A report is rebuilt weekly from exports
- Two dashboards disagree
- The data cannot be queried safely
Proof
Engagements thatlooked like this
COCOLI How COCOLI automated furniture supply across three European markets
No baseline was captured on this engagement.
Fastyr AI How Fastyr turned an AI-agent vision into a coherent operating workspace
No baseline was captured on this engagement.
Right-hand.ai How Right-hand.ai automated personalized security training and simulations
No baseline was captured on this engagement.
Is this the right engagement for you?
Describe what is actually going wrong. We come back with the outcome, the engagement that fits and the price basis - after a free initial audit.
Recognise this?
Where the gap between an AI demo and productionusually shows up
The proof of concept never became productionA demo that impressed everyone and could not be connected to the systems where the work actually happens.01
Documents are read by peopleInvoices, orders, forms and specs arrive as PDFs and email, and someone keys the contents into a system.02
Knowledge is spread across four placesStaff search policies, shared drives and databases to answer questions that have a definite answer.03
Judgement calls at volumeClassifying, prioritising and routing work that is too varied for rules and too repetitive to be a good use of people.04
Nobody owns accuracySomething AI-shaped is running and no one is measuring whether it is right, so no one will notice when it stops being right.05
How we decide
AI at the edges, software in the middle.
The model does the part that resists structure. Everything downstream of it is ordinary, testable code - which is why we can tell you what the system will do.
- Unstructured input, fuzzy classification and retrieval: a good use of AI.
- Arithmetic, lookups, thresholds and validation: not a good use of AI, and cheaper as code.
- A confidence threshold decides what a human sees. It is set from measurement, not from a default.
- We will tell you which parts of the system we propose contain no AI at all, and why.
In practice
What fixing the gap between an AI demo and productionlooks like
PDF or email → structured record → validated → system of record
Document extraction
Varied layouts read into consistent fields, with confidence per field and low-confidence ones routed to a person.
Request → category → priority → the right person
Classification and routing
Incoming work sorted the way your best operator would sort it, at a volume they could not sustain.
Policies + documents + databases → answer with a source
Internal knowledge assistant
Staff ask in plain language and get an answer they can check, instead of asking a colleague who is busy.
Transaction stream → anomaly → review queue
Exception detection
The things that do not look right surfaced early, while they are still cheap to fix.
Go deeper
The decisions behindthis kind of work
- ArticleAI vs. traditional automation: when should you use each?A practical comparison of AI and deterministic software, including the inputs, decisions, costs and controls that decide which belongs in a workflow.
- ArticleWorkflow automation vs. AI agents: which should run your process?A practical comparison of deterministic workflow automation and AI agents, including where each fits and why most production systems need both.
- ArticleWhat happens when an AI automation fails at 2am?The question that decides whether an automation belongs in your operations.
Questions
About the gap between an AI demo and production
That depends entirely on your data, which is why we measure it on a real sample before committing to anything. What we can promise is that you will see the number before launch, that the confidence threshold will be set from it, and that we keep measuring it in production.
Whichever performs best on your task at an acceptable cost, and we tell you which one and what data reaches it. That choice is a maintenance decision as much as a technical one - models change, so the system is built so the model can be swapped and re-evaluated against the same baseline.
No. We use enterprise or API tiers configured so your content is not used for training, and the provider, the data sent and the retention setting are written into the project documentation. Where data cannot leave your environment at all, we say so up front and design around it.
Eventually, in a controlled way, for the systems the workflow touches. Access is named, scoped to what the workflow needs, documented and revocable. We prefer a dedicated service account in your own identity provider so you can turn us off independently of anything else.
The design decides that before launch. Where a wrong call would be expensive, the workflow stops and routes to a person instead of guessing. Failures are logged and alerted, failed items are queued rather than dropped, and there is a rollback path. We operate what we build, so the alert comes to us.
Not this one?
The otherengagements
Our roadmap needs a senior team behind it.AI-Native Engineering Pod3 months minimumPriced OutcomeThe AI demo has to work in production.AI Integration Sprint4-8 weeksPriced OutcomeManual work is eating our capacity.Agentic Workflow Automation8-12 weeksPriced OutcomeThe prototype cannot survive real users.Vibe Code Rescue Sprint4-6 weeksPriced OutcomeThe old system is too risky to change.Legacy Modernization2+ quartersPriced OutcomeNobody owns the product after launch.Product CareOngoingPriced Outcome
Before you commit
What you wouldbe signing up for
One bounded outcome
A named result, a fixed window and a scope agreed in writing before delivery starts.
- Commitment
- 4 to 12 weeks
- Price basis
- Priced Outcome
- The problem has an edge you can describe
- You want the cost known first
- Someone inside owns the result
A team against a roadmap
Senior capacity inside your codebase, moving a roadmap week after week.
- Commitment
- 3 months minimum
- Price basis
- Priced Outcome
- The work is continuous, not bounded
- Recruiting would cost a quarter
- You want the same people throughout
Ownership after launch
Monitoring, patching, upgrades and a named engineer for live software.
- Commitment
- Ongoing, cancellable
- Price basis
- Priced Outcome
- It is live and business-critical
- Nobody owns it out of hours
- The risk needs an owner
One agreed price for an agreed outcome, quoted after a free initial audit and approved before delivery begins.
Named quotes go here as clients approve them. See what the engagements produced →
Next step
Tell us what you were hoping AI would fix.
We will tell you honestly which parts of it AI is actually good at, which parts are a rules problem wearing an AI costume, and whether the whole thing is worth building.