SYSTEMS

Engagements that
deepen.

Start with one useful system. Deepen only when the process, evidence and operational data justify it—from a focused product, to custom software, to a digital twin you can question.

  • 01 Start focused
  • 02 Prove the workflow
  • 03 Deepen with evidence
Depth of engagement
  1. 01Core productsBuilt once, shaped to fit.Three systems
  2. 02Custom softwareShaped to a process no product can express.Platforms · Dashboards · Portals
  3. 03Digital twinYour whole operation, modelled and questionable.By inquiry only
CORE PRODUCTS

Three systems, deliberately few.

Each one solves a problem common enough to be built once and shaped to fit, rather than rebuilt from nothing every time. Keep scrolling — they stack.

01

AI Sales Assistant

Inbound answered in the minute it arrives

Inbound rarely arrives when someone is free to answer it. The sales assistant sits on the channels enquiries actually come through — the web form, the shared inbox — and does the work a good SDR would have done in the first ten minutes, in the same minute the enquiry lands.

  • Deterministic scoring
  • CRM write-back
  • Drafted replies
  • Reason on every pass
What it does
Captures the enquiry, enriches it against the sources you already pay for, checks it against existing accounts and open opportunities, and answers the questions a prospect asks before a person is available. The qualification result, the reasoning behind it and a drafted reply land in your CRM against the right record, ready for a seller to send or edit. Where the enquiry does not clear your own bar, it is closed out with a reason attached rather than left to sit.
Who it's for
Teams where inbound volume has outgrown the number of people available to answer it, and where the first response decides whether a conversation happens at all. It is most useful where the qualification criteria are already understood but applied unevenly, because the person applying them is busy, new, or working from a different mental checklist than the person beside them.
Where automation stops and AI starts
Enrichment, deduplication, routing and the scoring rubric are ordinary code running server-side. The same lead scored twice gets the same score twice, and the rubric can be read, reviewed and changed without retraining anything. The model is used for the part that genuinely needs reading comprehension: interpreting free-text intent and drafting a reply in your voice. It does not set the score, it does not decide the routing, and nothing leaves the system unless the deterministic rules agree it should.
02

Document Intelligence

Extraction you can trace back to the page

Most document work is not hard, it is repetitive and unforgiving. Invoices, statements, purchase orders, delivery notes, claims — the same fields retyped into the same systems, where one transposed digit becomes a reconciliation problem three weeks later.

  • Template-first intake
  • Arithmetic validation
  • Exception queue
  • Field-level provenance
What it does
Takes documents from wherever they arrive — mailbox, upload, shared drive, scanner — classifies them, extracts the fields that matter, validates them against your own records, and writes the result into the system that needs it. Documents that pass validation clear without a person touching them. Documents that fail queue as exceptions, each one carrying the specific rule it failed and the value that failed it, so the person picking it up starts from a diagnosis rather than a blank page.
Who it's for
Finance and operations teams whose headcount scales with document volume, and any process where the audit trail matters as much as the throughput. It suits high-volume, stable document types first, then widens to the messy tail once the clean path is trustworthy.
Where automation stops and AI starts
Layout-stable documents are handled by templates and rules, with no model involved at all — a supplier who has sent the same invoice format for six years does not need inference. Validation is arithmetic and lookups against the ledger: a model is never the arbiter of whether a total adds up or an account exists. The model earns its place on the tail — unseen layouts, poor scans, free-text clauses, anything where the shape of the document is genuinely unknown in advance. Every extracted field keeps a reference back to where on the page it came from, so any number can be traced rather than trusted.
03

Internal Knowledge Assistant

An answer with the clause it came from

Organisational knowledge is rarely missing. It is spread across a wiki nobody has pruned, a drive with four versions of the same policy, a folder of PDFs, and a handful of people who are asked the same question every week.

  • Permission-scoped retrieval
  • Version-aware
  • Citation required
  • Says when it doesn't know
What it does
Gives people one place to ask, and returns an answer with the document, the clause and the version date it came from. Because the citation is part of the answer rather than a footnote, the answer can be checked. When the sources disagree — the common case with policy documents — it says so and shows both, instead of quietly picking one.
Who it's for
Organisations where the cost of not finding an answer is real: onboarding that takes months longer than it should, support teams escalating questions that were already documented, or compliance-adjacent work where using a superseded version of a policy is a genuine problem.
Where automation stops and AI starts
Permissions and retrieval scope are enforced in code before the model sees anything. The assistant cannot answer from a document the person asking is not allowed to open, and that is a property of the retrieval layer rather than an instruction the model is trusted to follow. Indexing, freshness and version tracking are ordinary pipeline work. The model's job is narrow: compose an answer from the passages retrieved, cite them, and say plainly when retrieval came back with nothing rather than filling the gap with something plausible.
CUSTOM SOFTWARE

When the process will not fit a product.

Same principle underneath: deterministic where the rules are known, models only where judgment is genuinely required. These are the shapes the work usually takes.

Internal platforms

A process that has outgrown the spreadsheet it lives in and needs to become a system — with roles, states, approvals and a record of who changed what. Usually the point at which a workaround has quietly become infrastructure nobody owns, running something the business cannot afford to have stop.

  • Roles and permissions
  • An explicit state model
  • Approval steps
  • Change history

Operational dashboards

Reporting that has to be trustworthy rather than reassembled by hand each month. One surface reading from the systems that hold the data, with the definition of each number written down and applied the same way every time.

Client and supplier portals

An outside party needs to submit, track or approve something without being given a seat in your internal tools. Scoped access, a clear state model, and a trail of what was submitted and when.

The common thread is that the process already exists and is already load-bearing. The work is making it explicit, observable and safe to change — not replacing it with something cleaner that nobody actually follows.

DIGITAL TWIN — TOP OF THE PROGRESSION

Your operation, as a model you can question.

A working model of how your operation actually runs - its processes, constraints, handoffs and decision points - kept in step with real data rather than with a diagram somebody drew last year. Questions that were arguments become experiments: what happens to lead time if this queue doubles, which approval step is actually the constraint, what breaks first if a supplier is a week late.

Why it sits at the top
A model is only as good as the data feeding it. The earlier work is what produces a clean, continuous record of how the operation behaves. Without that, a twin models an assumption rather than a business.
Why it is an inquiry, not a product
Scope is set by your operation, not by a feature list: which processes are worth modelling, what instrumentation exists, what decisions you want to test. That is a conversation before it is a proposal.
Start a digital twin inquiry

We will come back with questions before we come back with a proposal.