Services
three lines in full, fourteen in the catalogue
What you actually get
Grouped by the function that buys it, not by the technology that builds it. Nobody has a computer-vision budget.
The four engagement stages answer how you buy: a diagnostic, then a roadmap, then a build, then oversight. Fourteen service lines answer the other question - what actually gets built or improved - in five groups named after the function that signs for them: finance and procurement, operations, HR and shared services, customer and revenue, and compliance and governance. Three of them are below in full, with the numbers behind them; all fourteen are in the catalogue.
AI Voice Agents
Each line is evidenced by a failure mode and a measured change.
Five voice systems in production - the phone answered and the transaction completed, not a chat widget with a microphone on it.
Voice pays off wherever a high volume of repetitive enquiries can be answered from a system of record and the caller is often not at a desk - on a plant, in a vehicle, on a site, or out of hours. Delivered in energy and petrochemicals, telecommunications, banking and healthcare. The same shape applies directly to utilities, ports and logistics, facilities management and O&M, and government service desks - applies to, not delivered for.
The industrial anchor: a petrochemicals operator, where 64% of reception and switchboard contacts were handled without a person and time to a location, contact or facility answer fell from 2.9 minutes to 31 seconds. It took about three weeks to stand up, by reshaping an existing voice platform rather than building from scratch.
- Numbers spoken aloud across an authentication boundary are where voice agents actually die.
- A misheard digit does not degrade the answer, it ends the call - and the customer starts again with a human, which is worse than never having offered the agent. First-pass capture of spoken numbers went from 84% to 98.6%, and that single metric is the difference between a deployable agent and a demonstration.
Put your own numbers through it → - Gulf dialect is not one accent.
- Word error rate went from 31% to 12.7%, while the spread between dialects narrowed from 19 points to 5. A system that works well on average and badly on one region's speech has failed for a whole governorate, and an average hides exactly that.
All five run in telecom and government, at national scale, in Arabic. The switchboard problem is the same one an industrial site has.
AI Document Intelligence
Eight delivered systems, across invoices, purchase orders, contracts and settlements, delivery notes, submittals, scanned paper and bilingual documents. One capability applied to many document types, not a finance-department tool.
Coverage is the prize, not speed. On a telecom operator's partner settlements, verification against contract terms went from around 15% sampled to 100% - and the first full-coverage cycle surfaced 2.3 times the discrepancy rate the sampled process had been reporting.
The multiple measures what sampling was hiding: discrepancies were never evenly spread across the partner book, and a sample assumes they are. The same shape repeats on accounts payable: sample to 100% three-way reconciliation, and validation time per invoice from 12 minutes to 45 seconds.
Extraction accuracy by input quality, and what happened to the remainder
- Accuracy splits sharply by input quality.
- On scanned paper in a public health ministry, extraction ran at 96% on clean digital forms, 84% on poor scans and 61% on handwritten entries. Those are not averaged into one figure: 14% of fields were routed to human review, each shown with the region of the page it came from.
- The system stopped re-catching the same error and started removing its source.
- Repeat mismatches from the same supplier fell 67% once correction notices went back automatically. That is a second-order effect and it is usually worth more than the extraction itself, because it reduces the volume rather than processing it faster.
Delivered on telecom and financial-services documents. An industrial invoice has the same structure, the same bilingual problem and the same three-way match.
Semantic Search
Six delivered systems. The value here is not speed - it is surfacing an answer that would never have been found, and stopping one that would have been confidently wrong.
The engineering problem is not relevance ranking. It is that the words the asker uses are not the words the document uses: someone asks about retiring early while the statute says commutation of entitlement prior to the qualifying period. Closing that gap is the whole discipline.
On a social-insurance authority's statutory corpus - Arabic legal language, which is where general-purpose embeddings degrade most - correct provision in the top three went from 62% to 93% on exact statutory terms. On queries phrased the way a member actually talks, the same measure went from 34% to 89%. That second figure is the vocabulary gap closing: questions that used to return nothing usable now return the right provision.
- A confident wrong answer is worse than no answer.
- A system that trusts embedding distance alone will return a provision that is topically similar and governs a different category of person entirely. Provisions that read as near-neighbours can carry materially different entitlements, so a relevance-refinement stage checks each candidate against the actual question: roughly 1 in 6 retrieved candidates were discarded before a member ever saw them. A companion system pushed downstream correction errors from 12% to 3% the same way - that last number is the cost of a confident wrong answer, measured directly.
Put your own numbers through it → - Permission models are the real work, and it runs on-premises where it must.
- An index that ignores repository-level differences will surface a document to someone entitled to see the folder but not the file. One of the six was built where nothing could leave the building at all, and that variant exists because a judicial body required it.
This line finds an answer nobody could find. The AI Workspace packages the same capability plus the consolidation of the access points around it, and measures something different: how fast a known answer is reached across too many places to look.
What an answer looks like
One question, its answer in the same language, and the sources behind it by name and revision.
One question, one answer, and where it came from
Asked What is the current revision of the hot-work permit procedure?
Answered Revision 7, approved 14 March. Revision 6 is still the one on the shared drive; it was superseded but never removed.
Drawn from
- 1 Method statement, rev. 7
- 2 HSE procedure register
- 3 Permit to work, template
Every answer carries the sources it came from, by name and revision, each in whatever language it was written in. An answer with no citation is a guess with better grammar.
Figures are from the practice's own delivery records for the engagement named, measured against the process that preceded it.
Start here
Ask for the two lines closest to your situation, not all 14.
Describe the documents and the volume and we will send the two that fit, with the delivered results behind each.
You get a reply within one working day, from the engineer who would do the work.