Industrial AI for the Gulf back office.
Everyone in the building already uses AI. The operation still waits on people to read, check and re-enter what a system could.
A reply within one working day, from the engineer who would do the work.
What there is to buy
Three routes. Two of them are work, one is a system you keep, and the first is where every engagement starts.
Every engagement starts with the first of these. The other two are what it leads to: a specified piece of work, or a system that stays behind when we leave.
Four stages, each ending in a decision rather than an assumption, and each with a stated duration. You can stop after any one of them.
How the work runsFourteen lines across five function groups, three of them with production systems running. Each states what it needs from you and what it changes.
See the servicesA product rather than a project: one private place your whole company works with AI, on your own infrastructure. We put it there and build the parts specific to you.
What is in itAnd how the first of them runs
Building is no longer the hard part - modern tooling made it cheap to build almost anything. The expensive decision is what is worth building, in what order, and whether the organisation can absorb it.
| Stage | The question it answers | Typical duration |
|---|---|---|
| Stage 1 · DiagnoseReadiness & Opportunity Audit | Is there a real, quantifiable problem here, which two or three opportunities are worth pursuing and in what order, and is the organisation ready? | 2-4 weeks |
| Stage 2 · SpecifyStrategy & Roadmap | How exactly do we do it, what does it cost, who does it, in what order, and what has to be true for it to work? | 6-10 weeks |
| Stage 3 · BuildImplementation | Does it work in production, and do your people actually use it? | 3-12 months, phased |
| Stage 4 · OverseeGovernance & Advisory | Are the systems still performing, what new risk has appeared, and is the roadmap still pointed at the right things? | Ongoing, reviewed annually |
An audit can conclude that no AI work is warranted, and sometimes does - a cheaper way to find that out than eighteen months of building.
The assets we work on
The system that reads a permit to work is not the one that reads an invoice, and neither is the one that reads a weld.
Twenty years of procedures across SharePoint and four shared drives. Your new engineer asks three colleagues, then uses an old revision because nobody can tell her which one is current.
Buying AI is easy now. Knowing which work is worth automating, and getting it to change a number on a report, is the part that did not get cheaper. The lane is the industrial back office - finance, procurement, operations, HSE and shared services - and the paper that moves between them is where we start, because it is where the cost is easiest to count and hardest to argue with.
SAPOracleMaximoPrimaveraCMMSWMSSharePoint
Computer vision, data science, voice
Computer vision, data science and voice, all in production. Every figure below is a before-and-after from delivered work.
- Computer vision Defect escape down 84%
- On a bottling line, defects reaching distribution down 84% and detection from hours to under two minutes - inside the same shift, so the line could still be corrected. Vision work from quality inspection to street-level survey.
- Data science 42% error to 19%
- Demand forecast error more than halved on the piloted categories, and a hospital network's no-show rate from 14.6% to 9.1%. Forecasting, anomaly detection, optimisation and process mining.
- Voice agents 2.9 minutes to 31 seconds
- On a petrochemicals switchboard, time to an answer down from 2.9 minutes to 31 seconds, with 64% of reception contacts handled without a person.
- Retrieval over your own text 34% to 89%
- Questions phrased the way people actually phrase them, not in the document's own terminology. In Arabic and English.
- Quality review at full coverage 2% sample to 100%
- Every call reviewed instead of one in fifty, with compliance-phrase detection at 94% recall.
Mission and vision
What this practice is for, and what the sector looks like if it works.
- What this practice is for
- Be the practice industrial operators in the Gulf call when AI has to actually work - measuring before we build, building only what the measurement justifies, and staying until it runs.
- What the sector looks like then
- Industrial operations across the Gulf running on AI that was built in the region, in its own language, by engineers who stayed.
Results, with numbers
One engagement, measured against the process that preceded it. Not an average.
A petrochemicals operator in the Gulf. Reception and the switchboard were the front door to the plant’s office - location questions, contact lookups, facility enquiries - and every one of them interrupted a person.
| Measure | Before | After |
|---|---|---|
| Reception and switchboard contacts handled without a person | no prior measure | 64% |
| Average time to a location, contact or facility answer | 2.9 minutes | 31 seconds |
| Time to reshape the existing voice platform around a Gulf petrochemicals operator | no prior measure | ~3 weeks |
Built by reshaping an existing voice platform rather than starting from scratch. Dialect spread - the gap between the best and worst-performing Gulf dialect - narrowed from 19 points to 5, because a system that works on average and badly on one region’s speech has failed for a whole governorate.
64%
Reception and switchboard contacts handled without a person
The ring is the whole of that measure. The grey arc is the other 36%.
AI Workspace, our own product
A system you keep, on your own infrastructure, rather than a project that ends.
One private place your whole company works with AI, running on your own infrastructure. We built it for ourselves and we run it: the numbers below are from our own environment.
The problem it solves is consolidation, not capability. Knowledge sat in separate repositories, each with its own search behaviour, and AI tooling kept arriving as more destinations rather than fewer. Another capable assistant added to that estate would have made it marginally worse.
- Procurement
- Month-end
- HSE
- Contracts
- 4 people
- 2 assistants
A shared space, not a private chat window. People and assistants in the same thread, with access decided once for the space rather than per tool.
| Measure | Before | After |
|---|---|---|
| Average knowledge-discovery time for indexed enterprise content | ~12 minutes | under 90 seconds |
| Common internal questions answered or routed to the correct source without a further repository search | no prior measure | more than 70% |
| Separate access points for knowledge, search, and AI tools replaced by a single workspace | no prior measure | at least 3 |
Consolidating repositories means consolidating permission models that disagree with each other. An index that ignores those differences will surface a document to someone entitled to see the repository but not the file. Getting that right is most of the work.
Audit before automate
Published estimates put enterprise AI failure at 70-95%. Usually from automating a process that was already broken.
On a bottling line we ran a defect-data audit before anyone proposed a camera. It found three failure modes accounting for roughly 90% of the cost, against twelve listed in the QA manual. We solved the three.
Real data, real systems, the workflow as it runs - not as documented.
A ranked map, including the explicit finding that a case is not worth doing.
Fix the broken process first. Automating it as-is only makes it faster.
In phases, into your environment, instrumented against a baseline.
Against the baseline we established, and documented for your team.
Eight countries, one team
What a client buys is the people who will do the work, and this is where those people have already done it - across government and private sector, in Arabic and in English, on both sides of the Atlantic.
Where the work was done
- 1United Arab Emirates
- 2Saudi Arabia
- 3Qatar
- 4Kuwait
- 5Egypt
- 6Syria
- 7United States
- 8Brazil
The Gulf first, then the rest. Most of these engagements are documented case by case in our library; the remainder is delivered work the same team carried out under arrangements that produced no case file. Every entry is work by people in this practice.
Start here
Tell us where you are and we’ll tell you where we’d start.
Or run the numbers yourself first. The document load calculator asks what we would ask in a discovery call - volumes, minutes, people, loaded cost - and the arithmetic runs in your browser.
You get a reply within one working day, from the engineer who would do the work.