How we work
A practical route from opportunity to production.
Every step of a process gets one question: does this need judgement, or does it need to be right? Code handles what must be right. AI handles what needs judgement. The system will not publish a figure it did not calculate, and because we cost the process before building, you can check afterwards what the work was worth.
The delivery sequence
Nine stages, one accountable path- 01
Discover
Understand the process, the people, the systems and the economics.
- 02
Baseline
Establish how the process performs today, in numbers that can be re-read later the same way.
- 03
Design
Redesign the workflow and define the role of software, models and people, including exceptions and escalation.
- 04
Build
Build the system and connect it to the tools you already run. Anything that has to be right is calculated by tested code.
- 05
Validate
Test accuracy, exceptions, permissions, failure cases and operating cost before anyone depends on it.
- 06
Deploy
Move into production with monitoring, logs and clear internal ownership.
- 07
Adopt
Train the people whose work changes, and adjust the workflow around actual use.
- 08
Measure
Compare the operating result with the baseline from stage two.
- 09
Improve
Use real production behaviour to improve the system over time.
The rule, as an architecture
Four stages, and the check that holds the lineInputs that can be checked
Your documents, your records, live reads from the systems you already run. Everything the system may answer from, and nothing it may not.
Code works out the numbers
Prices, scores, totals, checks, business rules. Worked out the same way every time, and the only place a figure can come from.
AI writes the words
Reading documents, sorting, summarising, drafting the reply. It may explain a figure it was handed. It may never produce one.
A person keeps the decision
Anything expensive has a named owner, an approval step, and a route out when the information is not there.
The line is held by a check, not by a promise. If a model writes a figure the code did not calculate, the change does not ship. That check runs across all four stages, which is the only reason the rule above survives contact with a deadline.
That is the rule inside one step. For the whole system around it, two diagrams show where an AI layer sits in the technology you already run, and how one request moves through it: how AI works, in two diagrams.
Code works out the numbers. AI writes the words. Nothing goes out with a figure the system did not calculate.
That is the standard we hold every system to, ours and yours. A model may write the words around a figure, or pull a fact out of a document. It never decides the figure, and the build stops if it tries. Where plain code does the job, we use plain code: it costs less to run and cannot invent anything. We have made that swap in our own systems 5 times.
Every price a customer sees is one the system worked out
A figure can only appear in a sentence if the code calculated it first. If it did not, the sentence never leaves the building, and a plain template goes out instead.
A child's homework is marked by arithmetic, not by opinion
In our learning product, maths is settled by a solver and grammar by a grammar engine. The model sets the question and explains the answer. It never decides who is right.
A passport number the system has no way to keep
When a model reads a guest's travel document, there is nowhere in the record to write an ID number. It cannot store what it has no place to store, so the exposure does not exist.
The capability matrix: proven, versus where it transfers
Read the engagements| Capability | Proven in: real evidence | Could apply to: not proven yet |
|---|---|---|
One system instead of scattered paperwork and memory Documents, spreadsheets and years of institutional memory, turned into one record that is searchable, correct, and does not live in one person's head. | 719 order documents from 30 years of trading became one priced, queryable record. | Insurance claims historyLegal matter filesBanking back-office recordsManufacturing specs and BOMs Old paperwork becomes an asset your whole team can use, without a year-long IT project. The mechanism is the same whether the source is function sheets, claim files or contracts. |
A number everyone in the business already trusts One fact store, refreshed on a schedule, with every report built from those facts in code. A model may write the sentence around a figure; the figure itself always comes from the code. | Nine operating companies now read from one dashboard, refreshed nightly instead of assembled by hand each quarter. | FP&A and board reportingMulti-brand or multi-subsidiary groupsPE portfolio company roll-ups Nobody has to fact-check the board deck. The check that stops an invented figure is already running in production, and the same rule marks a child's maths homework in one of our own products. |
An agent that completes the task, or hands it to a person An agent scoped to one job, wired into the systems it needs, with a written definition of what it may do, what it hands to a person, and a log of every action it takes. | A concierge answered 40 out of 40 test questions correctly, partly by refusing the ones it could not verify. A second build cannot exceed its ad budget, because the limit is written into the code rather than into a prompt. | Contact centres and support queuesSales development and qualificationClaims intakeInternal helpdesks Real autonomy, with proof it still works when someone tries to trip it up. Most AI vendors offer one of those two things, not both. |
Knowing exactly where AI should not be used We decide where AI is genuinely too risky, and remove it from that part of the system completely: no field to type a number into, no AI in that step at all, a person alerted before anything happens. | Zero client-facing AI in a clinical intake system; screening scored by arithmetic against published cutoffs. | Healthcare and clinical operationsHR and people decisionsLegal adviceRegulated financial advice The same team that builds the AI will tell you in writing where plain code is the cheaper buy, and has made that swap 5 times in its own systems. |
Built to run with nobody watching it Every engagement held to the bar of a public product: it monitors itself, tests itself, and arrives with a runbook and a proper handover. | Decifer Markets has run every day since March on a broker paper account, with every change checked automatically before it goes near a user. Hospitality and cateringEvents managementFinancial markets intelligence | Any implementation where the real question is whether it survives without supervision What you take over already watches and tests itself, because that is the only way we have ever shipped anything. Nothing is added at the end to make the handover look finished. |
The middle column links to a real engagement, named where the client agreed in writing. The right column is our own judgement about where the same pattern would apply next. Every engagement behind this matrix runs at the scale of a single business. The assessment tests the pattern at your scale first: two weeks, fixed fee, credited in full against the build.
What we build on, and what you keep
Why each one, and which model runs which jobFront-end experience designData and reporting architectureAI agents and automationChatbots across WhatsApp and TelegramSecurity and access controlEmail and Meta marketing integrations
Built on tools you will own at handover:
Every account is opened in your name. At handover the repository, the accounts and a runbook transfer to you. Listing a tool means we have shipped production systems on it, not that its maker endorses us.
What to check before an AI system touches your business
Ask us these questions| What to check | The market norm | Decifer |
|---|---|---|
| Where does AI actually touch the output | Wherever the demo looks impressive | Only at the point that needs judgement; named in writing before the build starts |
| Who owns the accounts and the code | Rarely stated | You do, from day one, with a runbook at handover |
| Can a number in the output be invented | Usually possible; rarely checked for | No. The system will not publish a figure it did not calculate, and the build stops if it tries |
| A written list of what not to automate | Not offered | In every assessment and every agent scope sheet |
| Evidence it holds up once nobody is watching | A demo, on request | Three systems of ours have run daily since March, unattended. They are public, so you can open them and judge for yourself |
| The first step | A free consultation that becomes a pitch | A two-week assessment at a fixed fee, credited against any build |
Ask every vendor the left column. Whoever you choose, you will make a better decision.
Tell us the process. We'll bring the method.
A two-week assessment at a fixed fee. It costs the process, ranks the opportunities and names the one to start with. The fee is credited in full against any build that follows.
Replies come from a named person in Dubai within one working day.